Uploaded sanitized BC250/ROCm Repository.

2026-08-20 00:45:43 +02:00
parent 7d2184f1e8
commit d7d22e93b3
678 changed files with 65963 additions and 1 deletions
@@ -0,0 +1,4 @@
+import torch
+print("Device props:", torch.cuda.get_device_properties(0))
+print("Arch list:", torch.cuda.get_arch_list())
+print("GCN arch:", torch.cuda.get_device_properties(0).gcnArchName if hasattr(torch.cuda.get_device_properties(0), 'gcnArchName') else 'N/A')
@@ -0,0 +1,58 @@
+#!/bin/bash
+set -e
+
+echo "============================================"
+echo " PyTorch 2.5.1 Build for gfx1010 (BC-250)"
+echo "============================================"
+
+cd ~/pytorch-build
+
+# --- Phase 1: Clean old partial build ---
+echo "[Phase 1] Cleaning old build artifacts..."
+rm -rf build dist
+python3.11 -m pip install --user numpy pyyaml typing-extensions cffi 2>/dev/null || true
+echo "[Phase 1] Done."
+
+# --- Phase 2: Environment ---
+echo "[Phase 2] Setting environment..."
+export HSA_OVERRIDE_GFX_VERSION=10.1.0
+export HSA_ENABLE_SDMA=0
+export HIP_VISIBLE_DEVICES=0
+export ROCM_PATH=/opt/rocm
+export PATH="/opt/rocm/bin:$PATH"
+export LD_LIBRARY_PATH="/opt/rocm/lib:$LD_LIBRARY_PATH"
+
+# PyTorch build config — minimal build for inference
+export USE_ROCM=1
+export USE_CUDA=0
+export USE_NCCL=0
+export USE_DISTRIBUTED=0
+export USE_MKLDNN=0
+export USE_FBGEMM=0
+export USE_NNPACK=0
+export USE_QNNPACK=0
+export USE_XNNPACK=0
+export USE_KINETO=0
+export BUILD_TEST=0
+export BUILD_CAFFE2=0
+export USE_NUMPY=1
+export PYTORCH_ROCM_ARCH="gfx1010"
+export HIP_PATH=/opt/rocm
+export ROCM_SOURCE_DIR=/opt/rocm
+export CMAKE_PREFIX_PATH=/opt/rocm
+export MAX_JOBS=8
+export CMAKE_POLICY_VERSION_MINIMUM=3.5
+
+echo "  PYTORCH_ROCM_ARCH=$PYTORCH_ROCM_ARCH"
+echo "  MAX_JOBS=$MAX_JOBS"
+echo "[Phase 2] Done."
+
+# --- Phase 3: Build ---
+echo "[Phase 3] Starting PyTorch build..."
+python3.11 setup.py bdist_wheel 2>&1 | tee ~/pytorch-build-progress.log
+
+echo ""
+echo "============================================"
+echo " BUILD COMPLETE"
+echo "============================================"
+ls -lh dist/*.whl 2>/dev/null || echo "ERROR: No wheel produced!"
@@ -0,0 +1,62 @@
+#!/bin/bash
+set -e
+
+echo "============================================"
+echo " PyTorch 2.5.1 Build for gfx1010 (BC-250)"
+echo " NO TRITON / NO AOTRITON"
+echo "============================================"
+
+cd ~/pytorch-build
+
+# --- Phase 1: Clean old build ---
+echo "[Phase 1] Cleaning old build..."
+rm -rf build dist
+
+echo "[Phase 1] Done."
+
+# --- Phase 2: Environment ---
+echo "[Phase 2] Setting environment..."
+export HSA_OVERRIDE_GFX_VERSION=10.1.0
+export HSA_ENABLE_SDMA=0
+export HIP_VISIBLE_DEVICES=0
+export ROCM_PATH=/opt/rocm
+export PATH="/opt/rocm/bin:/home/fabian/.local/bin:$PATH"
+export LD_LIBRARY_PATH="/opt/rocm/lib:$LD_LIBRARY_PATH"
+
+# PyTorch build config — minimal for inference, NO Triton
+export USE_ROCM=1
+export USE_CUDA=0
+export USE_NCCL=0
+export USE_DISTRIBUTED=0
+export USE_MKLDNN=0
+export USE_FBGEMM=0
+export USE_NNPACK=0
+export USE_QNNPACK=0
+export USE_XNNPACK=0
+export USE_KINETO=0
+export BUILD_TEST=0
+export BUILD_CAFFE2=0
+export USE_NUMPY=1
+export USE_AOTRITON=0
+export USE_TRITON=0
+export PYTORCH_ROCM_ARCH="gfx1010"
+export HIP_PATH=/opt/rocm
+export ROCM_SOURCE_DIR=/opt/rocm
+export CMAKE_PREFIX_PATH=/opt/rocm
+export MAX_JOBS=8
+export CMAKE_POLICY_VERSION_MINIMUM=3.5
+
+echo "  PYTORCH_ROCM_ARCH=$PYTORCH_ROCM_ARCH"
+echo "  USE_AOTRITON=0 USE_TRITON=0"
+echo "  MAX_JOBS=$MAX_JOBS"
+echo "[Phase 2] Done."
+
+# --- Phase 3: Build ---
+echo "[Phase 3] Starting PyTorch build at $(date)..."
+python3.11 setup.py bdist_wheel 2>&1 | tee ~/pytorch-build-progress.log
+
+echo ""
+echo "============================================"
+echo " BUILD FINISHED at $(date)"
+echo "============================================"
+ls -lh dist/*.whl 2>/dev/null || echo "ERROR: No wheel produced!"
@@ -0,0 +1,84 @@
+#!/bin/bash
+# =============================================================
+# PyTorch 2.5.1 Build for AMD BC-250 — native gfx1010 target
+# ROCm 7.2 on CachyOS
+# =============================================================
+set -euo pipefail
+
+echo "=========================================="
+echo " PyTorch Build — gfx1010 for BC-250"
+echo "=========================================="
+
+# ========== PHASE 1: Clone ==========
+if [ ! -d ~/pytorch-build ]; then
+    echo "[1/5] Cloning PyTorch v2.5.1 (full recursive)..."
+    cd ~
+    git clone --recursive --branch v2.5.1 https://github.com/pytorch/pytorch.git pytorch-build
+else
+    echo "[1/5] Source already exists, skipping clone"
+fi
+
+cd ~/pytorch-build
+
+# ========== PHASE 2: Venv + deps ==========
+echo "[2/5] Installing build dependencies..."
+source ~/ComfyUI/venv/bin/activate
+pip install -q cmake ninja pyyaml typing-extensions numpy setuptools wheel cffi
+
+# ========== PHASE 3: Hipify ==========
+echo "[3/5] Running hipify (CUDA → HIP conversion)..."
+python tools/amd_build/build_amd.py
+
+# ========== PHASE 4: Environment ==========
+echo "[4/5] Configuring build environment..."
+
+# ROCm paths
+export ROCM_PATH=/opt/rocm
+export HIP_PATH=/opt/rocm
+export ROCM_HOME=/opt/rocm
+export PATH="/opt/rocm/bin:/opt/rocm/llvm/bin:$PATH"
+export LD_LIBRARY_PATH="/opt/rocm/lib:${LD_LIBRARY_PATH:-}"
+
+# Target gfx1010 (RDNA1 — closest supported arch for BC-250 gfx1013)
+export PYTORCH_ROCM_ARCH="gfx1010"
+export AMDGPU_TARGETS="gfx1010"
+
+# Build config
+export USE_ROCM=1
+export USE_CUDA=0
+export USE_CUDNN=0
+export USE_NCCL=0
+export USE_DISTRIBUTED=0
+export USE_MKLDNN=1
+export USE_OPENMP=1
+export BUILD_TEST=0
+export USE_FLASH_ATTENTION=0
+export USE_MEM_EFF_ATTENTION=0
+export REL_WITH_DEB_INFO=0
+export CMAKE_BUILD_TYPE=Release
+export CMAKE_POLICY_VERSION_MINIMUM=3.5
+
+# Parallelism — 12 cores, 14GB RAM, be conservative
+export MAX_JOBS=4
+export NINJA_STATUS="[%f/%t %e] "
+
+# Compiler
+export CC=gcc
+export CXX=g++
+
+echo "  PYTORCH_ROCM_ARCH=$PYTORCH_ROCM_ARCH"
+echo "  MAX_JOBS=$MAX_JOBS"
+echo "  ROCM_PATH=$ROCM_PATH"
+echo "  Python: $(python --version)"
+
+# ========== PHASE 5: Build ==========
+echo "[5/5] Building PyTorch wheel..."
+echo "  Start: $(date)"
+
+python setup.py bdist_wheel
+
+echo ""
+echo "=========================================="
+echo " Build complete: $(date)"
+echo "=========================================="
+ls -lh dist/*.whl 2>/dev/null || echo "ERROR: No wheel found!"
@@ -0,0 +1,140 @@
+#!/bin/bash
+set -e
+exec > >(tee -a ~/pytorch-build.log) 2>&1
+
+echo "=========================================="
+echo " PyTorch 2.5.1 Build for BC-250 (gfx1010)"
+echo " ROCm 7.2.0 | HIP 7.2 | $(date)"
+echo "=========================================="
+
+BUILD_DIR="$HOME/pytorch-build"
+ROCM_ARCH="gfx1010"
+
+##############################################
+# Phase 1: Fix incomplete clone
+##############################################
+echo ""
+echo "[1/5] Fixing submodules..."
+cd "$BUILD_DIR"
+
+# Reset any partial checkouts
+git submodule sync --recursive
+git submodule update --init --recursive --force --jobs 4 2>&1 || {
+    echo "WARN: Some submodules failed, retrying one by one..."
+    git submodule foreach --recursive 'git checkout . 2>/dev/null; true'
+    git submodule update --init --recursive --force 2>&1 || true
+}
+
+echo "Clone/submodule size: $(du -sh "$BUILD_DIR" | cut -f1)"
+
+##############################################
+# Phase 2: Install build dependencies
+##############################################
+echo ""
+echo "[2/5] Checking build dependencies..."
+for pkg in cmake ninja gcc python3; do
+    if ! command -v $pkg &>/dev/null; then
+        echo "ERROR: $pkg not found!"
+        exit 1
+    fi
+    echo "  $pkg: $(command -v $pkg)"
+done
+
+# Ensure python build deps
+pip3 install --user cmake ninja pyyaml typing-extensions setuptools wheel 2>&1 | tail -3
+
+##############################################
+# Phase 3: Hipify (CUDA -> HIP conversion)
+##############################################
+echo ""
+echo "[3/5] Running hipify (CUDA -> HIP)..."
+cd "$BUILD_DIR"
+
+if [ ! -f "aten/src/ATen/hip" ] || [ ! -d "aten/src/ATen/hip" ]; then
+    python3 tools/amd_build/build_amd.py 2>&1 | tail -20
+    echo "Hipify complete."
+else
+    echo "Hipify already done, skipping."
+fi
+
+##############################################
+# Phase 4: Configure environment
+##############################################
+echo ""
+echo "[4/5] Configuring build environment..."
+
+# ROCm paths
+export ROCM_PATH=/opt/rocm
+export HIP_PATH=/opt/rocm
+export ROCM_HOME=/opt/rocm
+export HCC_HOME=/opt/rocm/hcc
+export HIP_PLATFORM=amd
+
+# Target architecture
+export PYTORCH_ROCM_ARCH="$ROCM_ARCH"
+export AMDGPU_TARGETS="$ROCM_ARCH"
+
+# Build configuration
+export USE_ROCM=1
+export USE_CUDA=0
+export USE_NCCL=0
+export USE_DISTRIBUTED=0
+export USE_MKLDNN=0
+export USE_FBGEMM=0
+export USE_KINETO=0
+export USE_QNNPACK=0
+export USE_PYTORCH_QNNPACK=0
+export USE_NNPACK=0
+export USE_XNNPACK=0
+export BUILD_TEST=0
+export BUILD_CAFFE2=0
+export USE_FLASH_ATTENTION=0
+export USE_MEM_EFF_ATTENTION=0
+
+# BC-250 has limited RAM — reduce parallel jobs
+export MAX_JOBS=3
+
+# Fix CMake 4.x compatibility
+export CMAKE_POLICY_VERSION_MINIMUM=3.5
+
+# Compiler
+export CC=gcc
+export CXX=g++
+export CMAKE_C_COMPILER=gcc
+export CMAKE_CXX_COMPILER=g++
+
+# HIP compiler
+export HIP_CLANG_PATH=/opt/rocm/llvm/bin
+export HIPCC_COMPILE_FLAGS_APPEND="--offload-arch=$ROCM_ARCH"
+
+# BC-250 specific: need managed memory
+export HSA_ENABLE_SDMA=0
+
+echo "  PYTORCH_ROCM_ARCH=$PYTORCH_ROCM_ARCH"
+echo "  MAX_JOBS=$MAX_JOBS"
+echo "  ROCM_PATH=$ROCM_PATH"
+echo "  Python: $(python3 --version)"
+
+##############################################
+# Phase 5: Build
+##############################################
+echo ""
+echo "[5/5] Building PyTorch wheel..."
+echo "  Start time: $(date)"
+
+cd "$BUILD_DIR"
+
+# Clean any previous build artifacts
+python3 setup.py clean 2>/dev/null || true
+
+# Build wheel
+python3 setup.py bdist_wheel 2>&1
+
+echo ""
+echo "=========================================="
+echo " BUILD COMPLETE: $(date)"
+echo "=========================================="
+ls -lh dist/*.whl 2>/dev/null || echo "ERROR: No wheel file found!"
+echo ""
+echo "Install with:"
+echo "  pip install dist/torch-*.whl"
@@ -0,0 +1,27 @@
+#!/bin/bash
+set -e
+
+source /home/fabian/ComfyUI/venv/bin/activate
+
+export HSA_OVERRIDE_GFX_VERSION=10.1.0
+export HSA_ENABLE_SDMA=0
+export HIP_VISIBLE_DEVICES=0
+export ROCM_PATH=/opt/rocm
+export PYTORCH_ROCM_ARCH=gfx1010
+export FORCE_CUDA=1
+export TORCH_CUDA_ARCH_LIST=""
+
+# Uninstall old torchvision
+pip uninstall torchvision -y 2>/dev/null || true
+
+# Clone torchvision matching PyTorch 2.5
+cd /tmp
+rm -rf torchvision_build
+git clone --depth 1 --branch v0.20.0 https://github.com/pytorch/vision.git torchvision_build
+cd torchvision_build
+
+# Build and install
+python setup.py install 2>&1 | tail -20
+
+echo "=== DONE ==="
+python -c "import torchvision; print('torchvision version:', torchvision.__version__)"
@@ -0,0 +1,63 @@
+#!/bin/bash
+set -e
+
+echo "============================================"
+echo " PyTorch 2.5.1 Build for gfx1010 (BC-250)"
+echo " NO FLASH/MEM_EFF ATTENTION = NO AOTRITON"
+echo "============================================"
+
+cd ~/pytorch-build
+
+# --- Phase 1: Clean ---
+echo "[Phase 1] Clean build dir..."
+rm -rf build dist
+echo "[Phase 1] Done."
+
+# --- Phase 2: Environment ---
+echo "[Phase 2] Setting environment..."
+export HSA_OVERRIDE_GFX_VERSION=10.1.0
+export HSA_ENABLE_SDMA=0
+export HIP_VISIBLE_DEVICES=0
+export ROCM_PATH=/opt/rocm
+export PATH="/opt/rocm/bin:/home/fabian/.local/bin:$PATH"
+export LD_LIBRARY_PATH="/opt/rocm/lib:$LD_LIBRARY_PATH"
+
+# PyTorch build config
+export USE_ROCM=1
+export USE_CUDA=0
+export USE_NCCL=0
+export USE_DISTRIBUTED=0
+export USE_MKLDNN=0
+export USE_FBGEMM=0
+export USE_NNPACK=0
+export USE_QNNPACK=0
+export USE_XNNPACK=0
+export USE_KINETO=0
+export BUILD_TEST=0
+export BUILD_CAFFE2=0
+export USE_NUMPY=1
+# Disable Flash/MemEff attention -> no aotriton dependency
+export USE_FLASH_ATTENTION=0
+export USE_MEM_EFF_ATTENTION=0
+export PYTORCH_ROCM_ARCH="gfx1010"
+export HIP_PATH=/opt/rocm
+export ROCM_SOURCE_DIR=/opt/rocm
+export CMAKE_PREFIX_PATH=/opt/rocm
+export MAX_JOBS=8
+export CMAKE_POLICY_VERSION_MINIMUM=3.5
+
+echo "  PYTORCH_ROCM_ARCH=$PYTORCH_ROCM_ARCH"
+echo "  USE_FLASH_ATTENTION=0"
+echo "  USE_MEM_EFF_ATTENTION=0"
+echo "  MAX_JOBS=$MAX_JOBS"
+echo "[Phase 2] Done."
+
+# --- Phase 3: Build ---
+echo "[Phase 3] Starting build at $(date)..."
+python3.11 setup.py bdist_wheel 2>&1 | tee ~/pytorch-build-progress.log
+
+echo ""
+echo "============================================"
+echo " BUILD FINISHED at $(date)"
+echo "============================================"
+ls -lh dist/*.whl 2>/dev/null || echo "ERROR: No wheel produced!"
@@ -0,0 +1,48 @@
+#!/bin/bash
+# Build v4 - incremental rebuild after C10_WARP_SIZE fix
+set -e
+
+cd /home/fabian/pytorch-build
+
+export HSA_OVERRIDE_GFX_VERSION=10.1.0
+export HSA_ENABLE_SDMA=0
+export HIP_VISIBLE_DEVICES=0
+export ROCM_PATH=/opt/rocm
+export HIP_PATH=/opt/rocm
+export PYTORCH_ROCM_ARCH=gfx1010
+export USE_ROCM=1
+export USE_CUDA=0
+export USE_FLASH_ATTENTION=0
+export USE_MEM_EFF_ATTENTION=0
+export USE_AOTRITON=0
+export USE_TRITON=0
+export MAX_JOBS=8
+export CMAKE_POLICY_VERSION_MINIMUM=3.5
+export CMAKE_PREFIX_PATH=/opt/rocm
+export PYTHON_EXECUTABLE=/usr/bin/python3.11
+
+LOG=/home/fabian/pytorch-build/build_v4.log
+echo "============================================" | tee "$LOG"
+echo " BUILD v4 STARTED at $(date)" | tee -a "$LOG"
+echo " Incremental rebuild after C10_WARP_SIZE fix" | tee -a "$LOG"
+echo "============================================" | tee -a "$LOG"
+
+# Incremental build - ninja will only recompile changed files
+cd build
+cmake --build . --target install 2>&1 | tee -a "$LOG"
+
+echo "============================================" | tee -a "$LOG"
+echo " BUILD v4 FINISHED at $(date)" | tee -a "$LOG"
+echo "============================================" | tee -a "$LOG"
+
+# Back to source root for wheel
+cd /home/fabian/pytorch-build
+/usr/bin/python3.11 setup.py bdist_wheel 2>&1 | tee -a "$LOG"
+
+WHL=$(ls dist/*.whl 2>/dev/null | head -1)
+if [ -n "$WHL" ]; then
+    echo "SUCCESS: Wheel at $WHL" | tee -a "$LOG"
+    ls -lh "$WHL" | tee -a "$LOG"
+else
+    echo "ERROR: No wheel produced!" | tee -a "$LOG"
+fi
@@ -0,0 +1,12 @@
+#!/bin/bash
+cd ~/pytorch-build
+echo "=== CMakeLists.txt around aotriton ==="
+sed -n '855,900p' CMakeLists.txt
+
+echo ""
+echo "=== aotriton.cmake ==="
+cat cmake/External/aotriton.cmake
+
+echo ""
+echo "=== USE_FLASH_ATTENTION in Dependencies ==="
+grep -n "FLASH_ATTENTION\|AOTRITON\|aotriton" cmake/Dependencies.cmake 2>/dev/null | head -20
@@ -0,0 +1,17 @@
+#!/bin/bash
+echo "=== Active build processes ==="
+ps aux | grep -E "(build_pytorch|python.*setup|cmake|ninja|make|git.*(clone|submodule))" | grep -v grep | head -10
+
+echo ""
+echo "=== Log size ==="
+wc -l ~/pytorch-build.log 2>/dev/null
+
+echo ""
+echo "=== Last 30 lines of log ==="
+tail -30 ~/pytorch-build.log 2>/dev/null
+
+echo ""
+echo "=== Build directory status ==="
+du -sh ~/pytorch-build 2>/dev/null
+ls ~/pytorch-build/build 2>/dev/null && echo "Build dir exists" || echo "Build dir not yet created"
+ls ~/pytorch-build/dist/*.whl 2>/dev/null && echo "Wheel found!" || echo "No wheel yet"
@@ -0,0 +1,35 @@
+#!/bin/bash
+echo "=== OLD BUILD STATE ==="
+if [ -d ~/pytorch-build ]; then
+    echo "Dir exists: $(du -sh ~/pytorch-build)"
+    ls ~/pytorch-build/
+    
+    if [ -d ~/pytorch-build/pytorch ]; then
+        echo "--- PyTorch source dir ---"
+        ls ~/pytorch-build/pytorch/ | head -20
+        echo "..."
+        cd ~/pytorch-build/pytorch
+        echo "--- Git info ---"
+        git log --oneline -1 2>/dev/null || echo "Not a git repo"
+        git describe --tags 2>/dev/null || echo "No tags"
+        echo "--- Submodule count ---"
+        git submodule status 2>/dev/null | wc -l
+        echo "--- Failed submodules ---"
+        git submodule status 2>/dev/null | grep "^-" | head -10
+        echo "--- Hipify check ---"
+        [ -d "aten/src/ATen/hip" ] && echo "Hipify: DONE" || echo "Hipify: NOT DONE"
+        [ -d "build" ] && echo "Build dir: $(du -sh build)" || echo "No build dir"
+        ls dist/*.whl 2>/dev/null || echo "No wheel yet"
+    fi
+else
+    echo "NO pytorch-build dir"
+fi
+
+echo "=== DEPS ==="
+nproc
+cmake --version 2>/dev/null | head -1
+ninja --version 2>/dev/null || echo "No ninja"
+hipcc --version 2>/dev/null | head -3
+python3.11 -c "import numpy; print('numpy:', numpy.__version__)" 2>/dev/null || echo "No numpy"
+df -h / | tail -1
+echo "=== DONE ==="
@@ -0,0 +1,40 @@
+#!/bin/bash
+API="http://127.0.0.1:8188"
+
+echo "=== CLIPLoader info ==="
+curl -s "$API/object_info/CLIPLoader" | python3.11 -c "
+import sys, json
+d = json.load(sys.stdin)
+info = d.get('CLIPLoader', {})
+inp = info.get('input', {})
+print(json.dumps(inp, indent=2))
+"
+
+echo ""
+echo "=== Check what lumina2 expects ==="
+curl -s "$API/object_info/CLIPLoader" | python3.11 -c "
+import sys, json
+d = json.load(sys.stdin)
+info = d.get('CLIPLoader', {})
+desc = info.get('description', 'N/A')
+print('Description:', desc)
+out = info.get('output', [])
+print('Output:', out)
+"
+
+echo ""
+echo "=== Check DualCLIPLoader ==="
+curl -s "$API/object_info/DualCLIPLoader" | python3.11 -c "
+import sys, json
+d = json.load(sys.stdin)
+info = d.get('DualCLIPLoader', {})
+inp = info.get('input', {})
+print(json.dumps(inp, indent=2))
+"
+
+echo ""
+echo "=== Available text_encoder files ==="
+ls -lh /home/fabian/ComfyUI/models/text_encoders/
+echo ""
+echo "=== Available clip files ==="
+ls -lh /home/fabian/ComfyUI/models/clip/
@@ -0,0 +1,15 @@
+#!/bin/bash
+echo "=== ERRORS in build log ==="
+grep -iE "error:|FAILED|fatal" ~/pytorch-build-progress.log 2>/dev/null | grep -v "Warnung" | grep -v "Anmerkung" | grep -v "error=return" | grep -v "error=non-virtual" | grep -v "error=range" | grep -v "error=bool" | grep -v "error=format" | grep -v "Werror" | grep -v "error=missing" | tail -30
+
+echo ""
+echo "=== LAST 5 LINES of build.log ==="
+tail -5 ~/pytorch-build.log
+
+echo ""
+echo "=== BUILD DIR SIZE ==="
+du -sh ~/pytorch-build/build 2>/dev/null
+
+echo ""
+echo "=== WHEEL CHECK ==="
+ls ~/pytorch-build/dist/*.whl 2>/dev/null || echo "No wheel"
@@ -0,0 +1,35 @@
+#!/bin/bash
+API="http://127.0.0.1:8188"
+
+echo "=== GGUF nodes ==="
+curl -s "$API/object_info" | python3.11 -c "
+import sys, json
+d = json.load(sys.stdin)
+for k in sorted(d.keys()):
+    if 'gguf' in k.lower():
+        print(k)
+        info = d[k]
+        if 'input' in info and 'required' in info['input']:
+            for param, cfg in info['input']['required'].items():
+                print(f'  {param}: {cfg}')
+"
+
+echo ""
+echo "=== CLIPLoader info ==="
+curl -s "$API/object_info/CLIPLoader" | python3.11 -c "
+import sys, json
+d = json.load(sys.stdin)
+if 'CLIPLoader' in d:
+    info = d['CLIPLoader']
+    if 'input' in info and 'required' in info['input']:
+        for param, cfg in info['input']['required'].items():
+            print(f'  {param}: {cfg}')
+"
+
+echo ""
+echo "=== Model folders ==="
+echo "unet:"; ls /home/fabian/ComfyUI/models/unet/ 2>/dev/null
+echo "diffusion_models:"; ls /home/fabian/ComfyUI/models/diffusion_models/ 2>/dev/null
+echo "text_encoders:"; ls /home/fabian/ComfyUI/models/text_encoders/ 2>/dev/null
+echo "clip:"; ls /home/fabian/ComfyUI/models/clip/ 2>/dev/null
+echo "vae:"; ls /home/fabian/ComfyUI/models/vae/ 2>/dev/null
@@ -0,0 +1,12 @@
+#!/bin/bash
+echo "=== Last progress ==="
+grep -oP '\[\d+/\d+\]' ~/pytorch-build.log | tail -3
+echo ""
+echo "=== Process alive? ==="
+pgrep -f build_pytorch | head -3
+echo ""
+echo "=== Last 10 lines ==="
+tail -10 ~/pytorch-build.log
+echo ""
+echo "=== Disk usage ==="
+du -sh ~/pytorch-build
@@ -0,0 +1,37 @@
+#!/bin/bash
+echo "=== ENV ==="
+echo "HSA_OVERRIDE=$HSA_OVERRIDE_GFX_VERSION"
+echo "SDMA=$HSA_ENABLE_SDMA"
+echo "HIP_VIS=$HIP_VISIBLE_DEVICES"
+
+echo "=== RAM ==="
+free -m
+
+echo "=== COMFYUI ==="
+if [ -f ~/ComfyUI/main.py ]; then echo "ComfyUI: EXISTS"; else echo "ComfyUI: MISSING"; fi
+if [ -f ~/ComfyUI/venv/bin/python3.11 ]; then echo "Venv: EXISTS"; else echo "Venv: MISSING"; fi
+
+echo "=== MODELS ==="
+ls -lh ~/ComfyUI/models/unet/ 2>/dev/null || echo "NO unet dir"
+ls -lh ~/ComfyUI/models/text_encoders/ 2>/dev/null || echo "NO text_encoders dir"
+ls -lh ~/ComfyUI/models/vae/ 2>/dev/null || echo "NO vae dir"
+
+echo "=== PYTORCH BUILD ==="
+if [ -d ~/pytorch-build ]; then du -sh ~/pytorch-build; else echo "NO BUILD DIR"; fi
+
+echo "=== PYTHON ==="
+python3.11 --version 2>/dev/null || echo "NO python3.11"
+if [ -f ~/ComfyUI/venv/bin/python ]; then
+    source ~/ComfyUI/venv/bin/activate
+    python -c "import torch; print('PyTorch:', torch.__version__); print('CUDA avail:', torch.cuda.is_available()); print('Arch list:', torch.cuda.get_arch_list())" 2>/dev/null || echo "PyTorch import failed"
+fi
+
+echo "=== DMESG ==="
+sudo dmesg 2>/dev/null | grep -cE "KIQ|GPU died|GPU unreachable" || echo "0"
+sudo dmesg 2>/dev/null | grep -c "BC-250" || echo "0"
+sudo dmesg 2>/dev/null | grep "BC-250" | head -3
+
+echo "=== CUSTOM NODES ==="
+ls ~/ComfyUI/custom_nodes/ 2>/dev/null || echo "NO custom_nodes"
+
+echo "=== DONE ==="
@@ -0,0 +1,110 @@
+#!/bin/bash
+# Download all models for ComfyUI Z-Image Turbo on BC-250
+set -euo pipefail
+
+cd ~/ComfyUI
+source venv/bin/activate
+
+pip install -q huggingface-hub safetensors
+
+echo ""
+echo "=== Step 1: Download Z-Image Turbo GGUF (5.2 GB) ==="
+mkdir -p ~/ComfyUI/models/unet
+python3 -c "
+from huggingface_hub import hf_hub_download
+print('Downloading z_image_turbo-Q5_K_S.gguf from jayn7/Z-Image-Turbo-GGUF ...')
+hf_hub_download(
+    'jayn7/Z-Image-Turbo-GGUF',
+    'z_image_turbo-Q5_K_S.gguf',
+    local_dir='/home/fabian/ComfyUI/models/unet'
+)
+print('Done!')
+"
+
+echo ""
+echo "=== Step 2: Download Gemma 2 2B Text Encoder shards (9.8 GB) ==="
+mkdir -p ~/sd-models/text_encoders/lumina2_gemma2_2b
+python3 -c "
+from huggingface_hub import hf_hub_download
+import os
+
+repo = 'Alpha-VLLM/Lumina-Image-2.0'
+dest = '/home/fabian/sd-models/text_encoders/lumina2_gemma2_2b'
+os.makedirs(dest, exist_ok=True)
+
+files = [
+    'text_encoder/config.json',
+    'text_encoder/model.safetensors.index.json',
+    'text_encoder/model-00001-of-00003.safetensors',
+    'text_encoder/model-00002-of-00003.safetensors',
+    'text_encoder/model-00003-of-00003.safetensors',
+]
+for f in files:
+    print(f'Downloading {f}...')
+    hf_hub_download(repo, f, local_dir=dest)
+print('Done!')
+"
+
+echo ""
+echo "=== Step 3: Merge Text Encoder shards into single safetensors ==="
+mkdir -p ~/ComfyUI/models/text_encoders
+python3 << 'EOF'
+import safetensors.torch
+import os, json
+
+base_dir = "/home/fabian/sd-models/text_encoders/lumina2_gemma2_2b/text_encoder"
+output = "/home/fabian/ComfyUI/models/text_encoders/gemma2_2b_lumina2.safetensors"
+
+with open(os.path.join(base_dir, "model.safetensors.index.json")) as f:
+    index = json.load(f)
+
+all_tensors = {}
+shards = set(index["weight_map"].values())
+print(f"Loading {len(shards)} shards with {len(index['weight_map'])} tensors...")
+for shard in sorted(shards):
+    path = os.path.join(base_dir, shard)
+    print(f"  Loading {shard}...")
+    tensors = safetensors.torch.load_file(path, device="cpu")
+    all_tensors.update(tensors)
+
+print(f"Total tensors: {len(all_tensors)}")
+print(f"Saving merged file to {output}...")
+safetensors.torch.save_file(all_tensors, output)
+sz = os.path.getsize(output) / 1e9
+print(f"Done! Size: {sz:.2f} GB")
+EOF
+
+echo ""
+echo "=== Step 4: Download VAE (335 MB) ==="
+mkdir -p ~/ComfyUI/models/vae
+python3 -c "
+from huggingface_hub import hf_hub_download
+print('Downloading ae.safetensors ...')
+hf_hub_download(
+    'black-forest-labs/FLUX.1-schnell',
+    'ae.safetensors',
+    local_dir='/home/fabian/ComfyUI/models/vae'
+)
+print('Done!')
+"
+
+echo ""
+echo "=== Step 5: Download example workflow ==="
+mkdir -p ~/ComfyUI/user/default/workflows
+python3 -c "
+from huggingface_hub import hf_hub_download
+hf_hub_download(
+    'jayn7/Z-Image-Turbo-GGUF',
+    'example_workflow.json',
+    local_dir='/home/fabian/ComfyUI/user/default/workflows'
+)
+print('Workflow downloaded!')
+"
+
+echo ""
+echo "=== Verification ==="
+ls -lh ~/ComfyUI/models/unet/z_image_turbo-Q5_K_S.gguf 2>/dev/null || echo "MISSING: GGUF model"
+ls -lh ~/ComfyUI/models/text_encoders/gemma2_2b_lumina2.safetensors 2>/dev/null || echo "MISSING: Text encoder"
+ls -lh ~/ComfyUI/models/vae/ae.safetensors 2>/dev/null || echo "MISSING: VAE"
+echo ""
+echo "=== ALL DOWNLOADS COMPLETE ==="
@@ -0,0 +1,22 @@
+#!/bin/bash
+source /home/fabian/ComfyUI/venv/bin/activate
+export HSA_OVERRIDE_GFX_VERSION=10.1.0
+export HSA_ENABLE_SDMA=0
+export HIP_VISIBLE_DEVICES=0
+
+# Find all torchvision directories
+echo "=== torchvision locations ==="
+find /home/fabian/ComfyUI/venv -name 'torchvision' -type d 2>/dev/null
+
+echo "=== which init.py is loaded ==="
+python3.11 -c "import torchvision; print(torchvision.__file__)" 2>&1 || true
+
+echo "=== line 10 of loaded init ==="
+python3.11 -c "
+import importlib.util
+spec = importlib.util.find_spec('torchvision')
+print('Location:', spec.origin)
+" 2>&1 || true
+
+echo "=== check our patched file ==="
+sed -n '10,14p' /home/fabian/ComfyUI/venv/lib/python3.11/site-packages/torchvision/__init__.py
@@ -0,0 +1,58 @@
+#!/bin/bash
+# =============================================================
+# BC-250 GPU Activity Monitor
+# Since rocm-smi GPU utilization is broken (always 0%),
+# we use alternative metrics to verify GPU compute usage.
+# =============================================================
+
+echo "=== BC-250 GPU Activity Monitor ==="
+echo "NOTE: GPU % utilization is broken on this hardware."
+echo "Using alternative metrics instead."
+echo ""
+
+while true; do
+    TIMESTAMP=$(date '+%H:%M:%S')
+    
+    # 1. GPU Clock — high clock = GPU active
+    SCLK=$(cat /sys/class/drm/card0/device/pp_dpm_sclk 2>/dev/null | grep '\*' | awk '{print $2}')
+    
+    # 2. GPU power draw (if available)
+    POWER=$(cat /sys/class/drm/card0/device/hwmon/hwmon*/power1_average 2>/dev/null)
+    if [ -n "$POWER" ]; then
+        POWER_W=$(echo "scale=1; $POWER / 1000000" | bc 2>/dev/null || echo "N/A")
+    else
+        POWER_W="N/A"
+    fi
+    
+    # 3. GPU VRAM usage (shared memory allocated by GPU)
+    VRAM_USED=$(cat /sys/class/drm/card0/device/mem_info_vram_used 2>/dev/null)
+    VRAM_TOTAL=$(cat /sys/class/drm/card0/device/mem_info_vram_total 2>/dev/null)
+    if [ -n "$VRAM_USED" ] && [ -n "$VRAM_TOTAL" ]; then
+        VRAM_MB=$(echo "scale=0; $VRAM_USED / 1048576" | bc 2>/dev/null || echo "N/A")
+        VRAM_TOTAL_MB=$(echo "scale=0; $VRAM_TOTAL / 1048576" | bc 2>/dev/null || echo "N/A")
+    else
+        VRAM_MB="N/A"
+        VRAM_TOTAL_MB="N/A"
+    fi
+    
+    # 4. GPU temperature
+    TEMP=$(cat /sys/class/drm/card0/device/hwmon/hwmon*/temp1_input 2>/dev/null)
+    if [ -n "$TEMP" ]; then
+        TEMP_C=$(echo "scale=0; $TEMP / 1000" | bc 2>/dev/null || echo "N/A")
+    else
+        TEMP_C="N/A"
+    fi
+    
+    # 5. HIP processes using GPU
+    HIP_PROCS=$(ls /proc/*/maps 2>/dev/null | xargs grep -l "libamdhip64\|libhsa-runtime" 2>/dev/null | wc -l)
+    
+    # 6. Kernel GPU activity (interrupts)
+    GPU_IRQ=$(cat /proc/interrupts 2>/dev/null | grep amdgpu | awk '{sum=0; for(i=2;i<=NF-2;i++) sum+=$i; print sum}' | head -1)
+    
+    # 7. System RAM (since BC-250 shares system RAM as VRAM)
+    RAM_INFO=$(free -m | grep Mem | awk '{printf "%dMB / %dMB (%.0f%%)", $3, $2, $3/$2*100}')
+    
+    echo "[$TIMESTAMP] Clock: ${SCLK:-N/A} | Power: ${POWER_W}W | Temp: ${TEMP_C}°C | VRAM: ${VRAM_MB}/${VRAM_TOTAL_MB}MB | HIP Procs: $HIP_PROCS | RAM: $RAM_INFO"
+    
+    sleep 2
+done
@@ -0,0 +1,13 @@
+import torch
+print(f"PyTorch: {torch.__version__}")
+print(f"CUDA available: {torch.cuda.is_available()}")
+print(f"HIP version: {torch.version.hip}")
+print(f"Device count: {torch.cuda.device_count()}")
+if torch.cuda.is_available():
+    print(f"Device name: {torch.cuda.get_device_name(0)}")
+    t = torch.randn(100, 100, device="cuda")
+    r = torch.mm(t, t)
+    print(f"GPU matmul OK: result shape {r.shape}")
+    print("GPU COMPUTE WORKS!")
+else:
+    print("NO GPU DETECTED")
@@ -0,0 +1,29 @@
+import torch
+import os
+
+print(f"HSA_OVERRIDE_GFX_VERSION={os.environ.get('HSA_OVERRIDE_GFX_VERSION','NOT SET')}")
+print(f"PyTorch arch list: {torch.cuda.get_arch_list()}")
+print(f"Device: {torch.cuda.get_device_name(0)}")
+print(f"GCN Arch: {torch.cuda.get_device_properties(0).gcnArchName}")
+
+try:
+    t = torch.randn(256, 256, device="cuda")
+    r = torch.mm(t, t)
+    val = r[0,0].item()
+    print(f"GPU matmul OK! result[0,0]={val:.4f}")
+    
+    # Bigger test
+    a = torch.randn(1024, 1024, device="cuda")
+    b = torch.randn(1024, 1024, device="cuda")
+    c = torch.mm(a, b)
+    print(f"Large matmul OK! shape={c.shape}")
+    
+    # Test fp32 conv
+    x = torch.randn(1, 3, 64, 64, device="cuda")
+    conv = torch.nn.Conv2d(3, 16, 3, padding=1).cuda()
+    y = conv(x)
+    print(f"Conv2d OK! output shape={y.shape}")
+    
+    print("ALL GPU TESTS PASSED!")
+except Exception as e:
+    print(f"FAILED: {e}")
@@ -0,0 +1,26 @@
+#!/bin/bash
+cd /home/fabian/ComfyUI
+source venv/bin/activate
+
+export HSA_OVERRIDE_GFX_VERSION=10.1.0
+export HSA_ENABLE_SDMA=0
+export HIP_VISIBLE_DEVICES=0
+export ROCM_PATH=/opt/rocm
+export OMP_NUM_THREADS=12
+export MKL_NUM_THREADS=12
+export NUMEXPR_NUM_THREADS=12
+export PYTORCH_HIP_ALLOC_CONF="expandable_segments:False"
+
+# Unset harmful vars
+unset GPU_MAX_HW_QUEUES 2>/dev/null || true
+unset HIP_LAUNCH_BLOCKING 2>/dev/null || true
+
+echo "Starting ComfyUI..."
+echo "PyTorch: $(python3.11 -c 'import torch; print(torch.__version__)')"
+echo "GPU: $(python3.11 -c 'import torch; print(torch.cuda.get_device_name(0) if torch.cuda.is_available() else \"NONE\")')"
+
+python main.py \
+    --listen 0.0.0.0 \
+    --port 8188 \
+    --force-fp32 \
+    --lowvram
@@ -0,0 +1,28 @@
+#!/bin/bash
+# ComfyUI Launch for BC-250 APU — unified memory, no lowvram
+cd /home/fabian/ComfyUI
+source venv/bin/activate
+
+export HSA_OVERRIDE_GFX_VERSION=10.1.0
+export HSA_ENABLE_SDMA=0
+export HIP_VISIBLE_DEVICES=0
+export ROCM_PATH=/opt/rocm
+export OMP_NUM_THREADS=12
+export MKL_NUM_THREADS=12
+export NUMEXPR_NUM_THREADS=12
+export PYTORCH_HIP_ALLOC_CONF="expandable_segments:False"
+
+# Unset harmful vars
+unset GPU_MAX_HW_QUEUES 2>/dev/null || true
+unset HIP_LAUNCH_BLOCKING 2>/dev/null || true
+
+echo "Starting ComfyUI (APU mode - no lowvram)..."
+
+# --force-fp32: required for gfx1010 (no native fp16 support in some ops)
+# NO --lowvram: APU has unified memory, offloading is counterproductive
+# --gpu-only: keep everything in VRAM (which IS the system RAM on APU)
+python main.py \
+    --listen 0.0.0.0 \
+    --port 8188 \
+    --force-fp32 \
+    --gpu-only
@@ -0,0 +1,6 @@
+#!/bin/bash
+grep -oP '\[\d+/\d+\]' ~/pytorch-build-progress.log | tail -1
+echo "---"
+pgrep -c cc1plus 2>/dev/null || echo "0 compilers"
+echo "compiler_procs"
+free -m | grep Speicher
@@ -0,0 +1,55 @@
+#!/bin/bash
+# =============================================================
+# ComfyUI Launch Script for AMD BC-250 (ROCm / gfx1013 → gfx1030 spoof)
+# =============================================================
+set -euo pipefail
+
+echo "=========================================="
+echo " ComfyUI — BC-250 ROCm Launcher"
+echo "=========================================="
+
+# --- GPU Health Check ---
+if dmesg 2>/dev/null | tail -50 | grep -qi "KIQ fence timeout"; then
+    echo "[ABORT] KIQ fence timeout detected in dmesg — reboot required!"
+    exit 1
+fi
+echo "[OK] GPU health check passed"
+
+# --- ROCm Environment for BC-250 ---
+# CRITICAL: gfx1030 spoof (not gfx1010!) — PyTorch ROCm 6.2 has no gfx1010 kernels
+export HSA_OVERRIDE_GFX_VERSION=10.3.0
+export HSA_ENABLE_SDMA=0
+export HIP_VISIBLE_DEVICES=0
+export ROCM_PATH=/opt/rocm
+export HSA_TOOLS_LIB=""
+export HSA_TOOLS_REPORT_LOAD_FAILURE=0
+export PATH="/opt/rocm/bin:$PATH"
+export LD_LIBRARY_PATH="/opt/rocm/lib"
+
+# --- Performance: DO NOT set these ---
+unset GPU_MAX_HW_QUEUES 2>/dev/null || true
+unset HIP_LAUNCH_BLOCKING 2>/dev/null || true
+unset GGML_CUDA_ENABLE_UNIFIED_MEMORY 2>/dev/null || true
+unset GGML_HIP_HOST_ALLOC 2>/dev/null || true
+unset GGML_CUDA_NO_PINNED 2>/dev/null || true
+unset GGML_HIP_NO_COARSE_GRAIN 2>/dev/null || true
+unset HSA_DISABLE_FRAGMENT_ALLOCATOR 2>/dev/null || true
+
+# --- PyTorch ROCm tuning ---
+export PYTORCH_HIP_ALLOC_CONF="expandable_segments:False"
+
+echo "[OK] ROCm environment configured (gfx1030 spoof)"
+
+# --- Activate venv ---
+cd ~/ComfyUI
+source venv/bin/activate
+
+# --- Launch ComfyUI ---
+echo "[START] Launching ComfyUI on http://0.0.0.0:8188"
+echo "=========================================="
+python main.py \
+    --listen 0.0.0.0 \
+    --port 8188 \
+    --force-fp32 \
+    --lowvram \
+    "$@"
@@ -0,0 +1,66 @@
+#!/bin/bash
+# =============================================================
+# ComfyUI Launch Script for AMD BC-250 (ROCm / gfx1013)
+# Trusted source: BC250 ROCm Install README.md
+# =============================================================
+
+set -euo pipefail
+
+echo "=========================================="
+echo " ComfyUI — BC-250 ROCm Launcher"
+echo "=========================================="
+
+# --- GPU Health Check ---
+if dmesg 2>/dev/null | tail -50 | grep -qi "KIQ fence timeout"; then
+    echo "[ABORT] KIQ fence timeout detected in dmesg — reboot required!"
+    exit 1
+fi
+echo "[OK] GPU health check passed"
+
+# --- ROCm Environment for BC-250 (gfx1013 mapped to gfx1010) ---
+# Per BC250 ROCm Install docs — ONLY these vars
+export HSA_OVERRIDE_GFX_VERSION=10.1.0
+export HSA_ENABLE_SDMA=0
+export HIP_VISIBLE_DEVICES=0
+export ROCM_PATH=/opt/rocm
+export HSA_TOOLS_LIB=""
+export HSA_TOOLS_REPORT_LOAD_FAILURE=0
+export PATH="/opt/rocm/bin:$PATH"
+export LD_LIBRARY_PATH="/opt/rocm/lib"
+
+# --- Unset harmful old workarounds ---
+unset GPU_MAX_HW_QUEUES 2>/dev/null || true
+unset HIP_LAUNCH_BLOCKING 2>/dev/null || true
+unset GGML_CUDA_ENABLE_UNIFIED_MEMORY 2>/dev/null || true
+unset GGML_HIP_HOST_ALLOC 2>/dev/null || true
+unset GGML_CUDA_NO_PINNED 2>/dev/null || true
+unset GGML_HIP_NO_COARSE_GRAIN 2>/dev/null || true
+unset HSA_DISABLE_FRAGMENT_ALLOCATOR 2>/dev/null || true
+
+# --- PyTorch ROCm tuning ---
+export PYTORCH_HIP_ALLOC_CONF="expandable_segments:False"
+
+# --- CPU thread optimization for VAE and other CPU-bound ops ---
+# BC-250 has 12 threads — use them all
+export OMP_NUM_THREADS=12
+export MKL_NUM_THREADS=12
+export NUMEXPR_NUM_THREADS=12
+export OPENBLAS_NUM_THREADS=12
+
+echo "[OK] ROCm environment configured"
+echo "  GFX Override: $HSA_OVERRIDE_GFX_VERSION"
+echo "  OMP Threads: $OMP_NUM_THREADS"
+
+# --- Activate venv ---
+cd ~/ComfyUI
+source venv/bin/activate
+
+# --- Launch ComfyUI ---
+echo "[START] Launching ComfyUI on http://0.0.0.0:8188"
+echo "=========================================="
+python main.py \
+    --listen 0.0.0.0 \
+    --port 8188 \
+    --force-fp32 \
+    --lowvram \
+    "$@"
@@ -0,0 +1,15 @@
+#!/bin/bash
+cat > /tmp/test_arch.hip << 'HIPEOF'
+#include <hip/hip_runtime.h>
+__global__ void test_kernel(float *a) { a[threadIdx.x] = 1.0f; }
+HIPEOF
+
+for arch in gfx1013 gfx1010 gfx10-1-generic; do
+  echo "=== Testing $arch ==="
+  /opt/rocm/bin/hipcc --offload-arch=$arch -c /tmp/test_arch.hip -o /tmp/test_${arch}.o 2>&1
+  echo "Exit: $?"
+  if [ -f /tmp/test_${arch}.o ]; then
+    ls -l /tmp/test_${arch}.o
+  fi
+  echo ""
+done
@@ -0,0 +1,140 @@
+#!/bin/bash
+# Test image generation with correct GGUF workflow
+API="http://127.0.0.1:8188"
+
+echo "=== Queueing Z-Image Turbo generation ==="
+WORKFLOW='{
+  "prompt": {
+    "1": {
+      "class_type": "UnetLoaderGGUF",
+      "inputs": {
+        "unet_name": "z_image_turbo-Q5_K_S.gguf"
+      }
+    },
+    "2": {
+      "class_type": "CLIPLoaderGGUF",
+      "inputs": {
+        "clip_name": "gemma2_2b_lumina2.safetensors",
+        "type": "lumina2"
+      }
+    },
+    "3": {
+      "class_type": "CLIPTextEncode",
+      "inputs": {
+        "text": "a beautiful mountain landscape at sunset, golden light, detailed, 4k",
+        "clip": ["2", 0]
+      }
+    },
+    "4": {
+      "class_type": "CLIPTextEncode",
+      "inputs": {
+        "text": "blurry, ugly, distorted",
+        "clip": ["2", 0]
+      }
+    },
+    "5": {
+      "class_type": "EmptyLatentImage",
+      "inputs": {
+        "width": 512,
+        "height": 512,
+        "batch_size": 1
+      }
+    },
+    "6": {
+      "class_type": "KSampler",
+      "inputs": {
+        "model": ["1", 0],
+        "positive": ["3", 0],
+        "negative": ["4", 0],
+        "latent_image": ["5", 0],
+        "seed": 42,
+        "steps": 8,
+        "cfg": 3.0,
+        "sampler_name": "euler",
+        "scheduler": "normal",
+        "denoise": 1.0
+      }
+    },
+    "7": {
+      "class_type": "VAELoader",
+      "inputs": {
+        "vae_name": "ae.safetensors"
+      }
+    },
+    "8": {
+      "class_type": "VAEDecode",
+      "inputs": {
+        "samples": ["6", 0],
+        "vae": ["7", 0]
+      }
+    },
+    "9": {
+      "class_type": "SaveImage",
+      "inputs": {
+        "images": ["8", 0],
+        "filename_prefix": "bc250_test"
+      }
+    }
+  }
+}'
+
+RESPONSE=$(curl -s -X POST "$API/prompt" -H "Content-Type: application/json" -d "$WORKFLOW")
+echo "Response: $RESPONSE" | head -c 500
+echo ""
+
+PROMPT_ID=$(echo "$RESPONSE" | python3.11 -c "import sys,json; print(json.load(sys.stdin).get('prompt_id','NONE'))" 2>/dev/null)
+echo "Prompt ID: $PROMPT_ID"
+
+if [ "$PROMPT_ID" = "NONE" ] || [ -z "$PROMPT_ID" ]; then
+    echo "ERROR: Failed to queue!"
+    exit 1
+fi
+
+echo ""
+echo "=== Waiting for generation (up to 10 min) ==="
+for i in $(seq 1 120); do
+    sleep 5
+    STATUS=$(curl -s "$API/history/$PROMPT_ID" 2>/dev/null)
+    HAS_OUTPUT=$(echo "$STATUS" | python3.11 -c "
+import sys, json
+data = json.load(sys.stdin)
+pid = '$PROMPT_ID'
+if pid in data:
+    outputs = data[pid].get('outputs', {})
+    status = data[pid].get('status', {})
+    if status.get('status_str') == 'error':
+        msgs = status.get('messages', [])
+        print('ERROR:' + str(msgs[-1] if msgs else 'unknown'))
+    elif '9' in outputs:
+        images = outputs['9'].get('images', [])
+        if images:
+            print('DONE:' + images[0].get('filename', 'unknown'))
+        else:
+            print('PROCESSING')
+    else:
+        print('PROCESSING')
+else:
+    print('WAITING')
+" 2>/dev/null)
+    echo "[$((i*5))s] $HAS_OUTPUT"
+    if [[ "$HAS_OUTPUT" == DONE:* ]]; then
+        FILENAME=${HAS_OUTPUT#DONE:}
+        echo ""
+        echo "=========================================="
+        echo "  SUCCESS - Image generated!"
+        echo "  File: /home/fabian/ComfyUI/output/$FILENAME"
+        ls -lh "/home/fabian/ComfyUI/output/$FILENAME" 2>/dev/null
+        echo "=========================================="
+        exit 0
+    fi
+    if [[ "$HAS_OUTPUT" == ERROR:* ]]; then
+        echo ""
+        echo "GENERATION FAILED: $HAS_OUTPUT"
+        echo "=== ComfyUI log tail ==="
+        tail -30 /home/fabian/comfyui2.log
+        exit 1
+    fi
+done
+
+echo "TIMEOUT after 10 minutes"
+tail -20 /home/fabian/comfyui2.log
@@ -0,0 +1,145 @@
+#!/bin/bash
+# Test image generation - v3 with CLIPLoader (not GGUF) for safetensors clip
+API="http://127.0.0.1:8188"
+
+echo "=== Queueing Z-Image Turbo generation v3 ==="
+WORKFLOW='{
+  "prompt": {
+    "1": {
+      "class_type": "UnetLoaderGGUF",
+      "inputs": {
+        "unet_name": "z_image_turbo-Q5_K_S.gguf"
+      }
+    },
+    "2": {
+      "class_type": "CLIPLoader",
+      "inputs": {
+        "clip_name": "gemma2_2b_lumina2.safetensors",
+        "type": "lumina2"
+      }
+    },
+    "3": {
+      "class_type": "CLIPTextEncode",
+      "inputs": {
+        "text": "a beautiful mountain landscape at sunset, golden light, detailed, 4k",
+        "clip": ["2", 0]
+      }
+    },
+    "4": {
+      "class_type": "CLIPTextEncode",
+      "inputs": {
+        "text": "blurry, ugly, distorted",
+        "clip": ["2", 0]
+      }
+    },
+    "5": {
+      "class_type": "EmptyLatentImage",
+      "inputs": {
+        "width": 512,
+        "height": 512,
+        "batch_size": 1
+      }
+    },
+    "6": {
+      "class_type": "KSampler",
+      "inputs": {
+        "model": ["1", 0],
+        "positive": ["3", 0],
+        "negative": ["4", 0],
+        "latent_image": ["5", 0],
+        "seed": 42,
+        "steps": 8,
+        "cfg": 3.0,
+        "sampler_name": "euler",
+        "scheduler": "normal",
+        "denoise": 1.0
+      }
+    },
+    "7": {
+      "class_type": "VAELoader",
+      "inputs": {
+        "vae_name": "ae.safetensors"
+      }
+    },
+    "8": {
+      "class_type": "VAEDecode",
+      "inputs": {
+        "samples": ["6", 0],
+        "vae": ["7", 0]
+      }
+    },
+    "9": {
+      "class_type": "SaveImage",
+      "inputs": {
+        "images": ["8", 0],
+        "filename_prefix": "bc250_test"
+      }
+    }
+  }
+}'
+
+RESPONSE=$(curl -s -X POST "$API/prompt" -H "Content-Type: application/json" -d "$WORKFLOW")
+echo "Response: $(echo $RESPONSE | head -c 200)"
+echo ""
+
+PROMPT_ID=$(echo "$RESPONSE" | python3.11 -c "import sys,json; print(json.load(sys.stdin).get('prompt_id','NONE'))" 2>/dev/null)
+echo "Prompt ID: $PROMPT_ID"
+
+if [ "$PROMPT_ID" = "NONE" ] || [ -z "$PROMPT_ID" ]; then
+    echo "ERROR: Failed to queue!"
+    echo "Full response: $RESPONSE"
+    exit 1
+fi
+
+echo ""
+echo "=== Waiting for generation (up to 10 min) ==="
+for i in $(seq 1 120); do
+    sleep 5
+    STATUS=$(curl -s "$API/history/$PROMPT_ID" 2>/dev/null)
+    HAS_OUTPUT=$(echo "$STATUS" | python3.11 -c "
+import sys, json
+data = json.load(sys.stdin)
+pid = '$PROMPT_ID'
+if pid in data:
+    outputs = data[pid].get('outputs', {})
+    status = data[pid].get('status', {})
+    if status.get('status_str') == 'error':
+        msgs = status.get('messages', [])
+        for m in msgs:
+            if isinstance(m, list) and len(m)>1 and isinstance(m[1],dict):
+                em = m[1].get('exception_message','')
+                if em:
+                    print('ERROR:' + em[:200])
+                    break
+        else:
+            print('ERROR:unknown')
+    elif '9' in outputs:
+        images = outputs['9'].get('images', [])
+        if images:
+            print('DONE:' + images[0].get('filename', 'unknown'))
+        else:
+            print('PROCESSING')
+    else:
+        print('PROCESSING')
+else:
+    print('WAITING')
+" 2>/dev/null)
+    echo "[$((i*5))s] $HAS_OUTPUT"
+    if [[ "$HAS_OUTPUT" == DONE:* ]]; then
+        FILENAME=${HAS_OUTPUT#DONE:}
+        echo ""
+        echo "=========================================="
+        echo "  SUCCESS - Image generated!"
+        echo "  File: /home/fabian/ComfyUI/output/$FILENAME"
+        ls -lh "/home/fabian/ComfyUI/output/$FILENAME" 2>/dev/null
+        echo "=========================================="
+        exit 0
+    fi
+    if [[ "$HAS_OUTPUT" == ERROR:* ]]; then
+        echo ""
+        echo "GENERATION FAILED: $HAS_OUTPUT"
+        tail -10 /home/fabian/comfyui2.log
+        exit 1
+    fi
+done
+echo "TIMEOUT"
@@ -0,0 +1,142 @@
+#!/bin/bash
+# Test image generation via ComfyUI API
+# Workflow: Z-Image Turbo GGUF + Gemma2 CLIP + AE VAE
+
+API="http://127.0.0.1:8188"
+
+echo "=== Checking ComfyUI API ==="
+curl -s "$API/system_stats" | python3.11 -m json.tool 2>/dev/null | head -20
+echo ""
+
+echo "=== Listing available models ==="
+curl -s "$API/models/unet" 2>/dev/null
+echo ""
+curl -s "$API/models/clip" 2>/dev/null  
+echo ""
+curl -s "$API/models/vae" 2>/dev/null
+echo ""
+
+echo "=== Queueing image generation ==="
+WORKFLOW='{
+  "prompt": {
+    "1": {
+      "class_type": "UNETLoader",
+      "inputs": {
+        "unet_name": "z_image_turbo-Q5_K_S.gguf",
+        "weight_dtype": "default"
+      }
+    },
+    "2": {
+      "class_type": "DualCLIPLoader",
+      "inputs": {
+        "clip_name1": "gemma2_2b_lumina2.safetensors",
+        "clip_name2": "gemma2_2b_lumina2.safetensors",
+        "type": "lumina2"
+      }
+    },
+    "3": {
+      "class_type": "CLIPTextEncode",
+      "inputs": {
+        "text": "a beautiful mountain landscape at sunset, golden light, detailed, 4k",
+        "clip": ["2", 0]
+      }
+    },
+    "4": {
+      "class_type": "CLIPTextEncode",
+      "inputs": {
+        "text": "blurry, ugly, distorted",
+        "clip": ["2", 0]
+      }
+    },
+    "5": {
+      "class_type": "EmptyLatentImage",
+      "inputs": {
+        "width": 512,
+        "height": 512,
+        "batch_size": 1
+      }
+    },
+    "6": {
+      "class_type": "KSampler",
+      "inputs": {
+        "model": ["1", 0],
+        "positive": ["3", 0],
+        "negative": ["4", 0],
+        "latent_image": ["5", 0],
+        "seed": 42,
+        "steps": 8,
+        "cfg": 3.0,
+        "sampler_name": "euler",
+        "scheduler": "normal",
+        "denoise": 1.0
+      }
+    },
+    "7": {
+      "class_type": "VAELoader",
+      "inputs": {
+        "vae_name": "ae.safetensors"
+      }
+    },
+    "8": {
+      "class_type": "VAEDecode",
+      "inputs": {
+        "samples": ["6", 0],
+        "vae": ["7", 0]
+      }
+    },
+    "9": {
+      "class_type": "SaveImage",
+      "inputs": {
+        "images": ["8", 0],
+        "filename_prefix": "bc250_test"
+      }
+    }
+  }
+}'
+
+RESPONSE=$(curl -s -X POST "$API/prompt" -H "Content-Type: application/json" -d "$WORKFLOW")
+echo "Queue response: $RESPONSE"
+PROMPT_ID=$(echo "$RESPONSE" | python3.11 -c "import sys,json; print(json.load(sys.stdin).get('prompt_id','NONE'))" 2>/dev/null)
+echo "Prompt ID: $PROMPT_ID"
+
+if [ "$PROMPT_ID" = "NONE" ] || [ -z "$PROMPT_ID" ]; then
+    echo "ERROR: Failed to queue prompt!"
+    exit 1
+fi
+
+echo ""
+echo "=== Waiting for generation ==="
+for i in $(seq 1 60); do
+    sleep 5
+    STATUS=$(curl -s "$API/history/$PROMPT_ID" 2>/dev/null)
+    HAS_OUTPUT=$(echo "$STATUS" | python3.11 -c "
+import sys, json
+data = json.load(sys.stdin)
+if '$PROMPT_ID' in data:
+    outputs = data['$PROMPT_ID'].get('outputs', {})
+    if '9' in outputs:
+        images = outputs['9'].get('images', [])
+        if images:
+            print('DONE:' + images[0].get('filename', 'unknown'))
+        else:
+            print('PROCESSING')
+    else:
+        print('PROCESSING')
+else:
+    print('WAITING')
+" 2>/dev/null)
+    echo "[$((i*5))s] $HAS_OUTPUT"
+    if [[ "$HAS_OUTPUT" == DONE:* ]]; then
+        FILENAME=${HAS_OUTPUT#DONE:}
+        echo ""
+        echo "=== SUCCESS ==="
+        echo "Image generated: $FILENAME"
+        echo "File location: /home/fabian/ComfyUI/output/$FILENAME"
+        ls -lh "/home/fabian/ComfyUI/output/$FILENAME" 2>/dev/null
+        exit 0
+    fi
+done
+
+echo "TIMEOUT: Generation did not complete in 5 minutes"
+echo "=== Checking queue ==="
+curl -s "$API/queue" | python3.11 -m json.tool 2>/dev/null | head -20
@@ -0,0 +1,23 @@
+#!/bin/bash
+source /home/fabian/ComfyUI/venv/bin/activate
+export HSA_OVERRIDE_GFX_VERSION=10.1.0
+export HSA_ENABLE_SDMA=0
+export HIP_VISIBLE_DEVICES=0
+
+python3.11 -c "
+import torch
+print('PyTorch version:', torch.__version__)
+print('HIP version:', torch.version.hip)
+print('CUDA available:', torch.cuda.is_available())
+if torch.cuda.is_available():
+    print('Device name:', torch.cuda.get_device_name(0))
+    print('Device count:', torch.cuda.device_count())
+    # Quick tensor test
+    x = torch.randn(100, 100, device='cuda')
+    y = torch.randn(100, 100, device='cuda')
+    z = x @ y
+    print('GPU matmul test: OK, shape', z.shape)
+    print('Memory allocated:', torch.cuda.memory_allocated(0) / 1024 / 1024, 'MB')
+else:
+    print('NO GPU DETECTED')
+"
@@ -0,0 +1,64 @@
+#!/bin/bash
+# Runtime test: which arch actually RUNS on the BC-250?
+export HSA_ENABLE_SDMA=0
+
+cat > /tmp/run_test.hip << 'HIPEOF'
+#include <hip/hip_runtime.h>
+#include <cstdio>
+
+__global__ void add_kernel(float *a, float *b, float *c, int n) {
+    int i = blockIdx.x * blockDim.x + threadIdx.x;
+    if (i < n) c[i] = a[i] + b[i];
+}
+
+int main() {
+    hipDeviceProp_t prop;
+    hipGetDeviceProperties(&prop, 0);
+    printf("Device: %s\n", prop.name);
+    printf("GCN Arch: %s\n", prop.gcnArchName);
+
+    const int N = 256;
+    float h_a[N], h_b[N], h_c[N];
+    for (int i = 0; i < N; i++) { h_a[i] = i; h_b[i] = i * 2; }
+
+    float *d_a, *d_b, *d_c;
+    hipMallocManaged(&d_a, N * sizeof(float));
+    hipMallocManaged(&d_b, N * sizeof(float));
+    hipMallocManaged(&d_c, N * sizeof(float));
+    memcpy(d_a, h_a, N * sizeof(float));
+    memcpy(d_b, h_b, N * sizeof(float));
+
+    add_kernel<<<1, N>>>(d_a, d_b, d_c, N);
+    hipDeviceSynchronize();
+
+    hipError_t err = hipGetLastError();
+    if (err != hipSuccess) {
+        printf("FAIL: %s\n", hipGetErrorString(err));
+        hipFree(d_a); hipFree(d_b); hipFree(d_c);
+        return 1;
+    }
+
+    // Verify
+    int ok = 1;
+    for (int i = 0; i < N; i++) {
+        if (d_c[i] != h_a[i] + h_b[i]) { ok = 0; break; }
+    }
+    printf("Compute: %s\n", ok ? "PASS" : "FAIL");
+
+    hipFree(d_a); hipFree(d_b); hipFree(d_c);
+    return ok ? 0 : 1;
+}
+HIPEOF
+
+for arch in gfx1013 gfx1010 gfx10-1-generic; do
+    echo "=== Runtime test: $arch ==="
+    /opt/rocm/bin/hipcc --offload-arch=$arch /tmp/run_test.hip -o /tmp/run_test_${arch} 2>&1
+    if [ $? -eq 0 ]; then
+        echo "Compiled OK, running..."
+        timeout 10 /tmp/run_test_${arch} 2>&1
+        echo "Runtime exit: $?"
+    else
+        echo "Compile FAILED"
+    fi
+    echo ""
+done
@@ -0,0 +1,64 @@
+#!/bin/bash
+export HSA_ENABLE_SDMA=0
+
+cat > /tmp/run_test2.hip << 'HIPEOF'
+#include <hip/hip_runtime.h>
+#include <cstdio>
+
+__global__ void add_kernel(float *a, float *b, float *c, int n) {
+    int i = blockIdx.x * blockDim.x + threadIdx.x;
+    if (i < n) c[i] = a[i] + b[i];
+}
+
+int main() {
+    hipDeviceProp_t prop;
+    hipGetDeviceProperties(&prop, 0);
+    printf("Device: %s\n", prop.name);
+    printf("GCN Arch: %s\n", prop.gcnArchName);
+
+    const int N = 256;
+    float h_a[N], h_b[N], h_c[N];
+    for (int i = 0; i < N; i++) { h_a[i] = (float)i; h_b[i] = (float)(i * 2); }
+
+    float *d_a, *d_b, *d_c;
+    hipMallocManaged(&d_a, N * sizeof(float));
+    hipMallocManaged(&d_b, N * sizeof(float));
+    hipMallocManaged(&d_c, N * sizeof(float));
+    hipMemcpy(d_a, h_a, N * sizeof(float), hipMemcpyHostToDevice);
+    hipMemcpy(d_b, h_b, N * sizeof(float), hipMemcpyHostToDevice);
+
+    add_kernel<<<1, N>>>(d_a, d_b, d_c, N);
+    hipDeviceSynchronize();
+
+    hipError_t err = hipGetLastError();
+    if (err != hipSuccess) {
+        printf("KERNEL FAIL: %s\n", hipGetErrorString(err));
+        hipFree(d_a); hipFree(d_b); hipFree(d_c);
+        return 1;
+    }
+
+    hipMemcpy(h_c, d_c, N * sizeof(float), hipMemcpyDeviceToHost);
+
+    int ok = 1;
+    for (int i = 0; i < N; i++) {
+        if (h_c[i] != h_a[i] + h_b[i]) { ok = 0; printf("Mismatch at %d: %f vs %f\n", i, h_c[i], h_a[i]+h_b[i]); break; }
+    }
+    printf("Compute: %s\n", ok ? "PASS" : "FAIL");
+
+    hipFree(d_a); hipFree(d_b); hipFree(d_c);
+    return ok ? 0 : 1;
+}
+HIPEOF
+
+for arch in gfx1013 gfx1010 gfx10-1-generic; do
+    echo "=== Runtime test: $arch ==="
+    /opt/rocm/bin/hipcc --offload-arch=$arch /tmp/run_test2.hip -o /tmp/run_test2_${arch} 2>&1
+    if [ $? -eq 0 ]; then
+        echo "Compiled OK. Running..."
+        timeout 15 /tmp/run_test2_${arch} 2>&1
+        echo "Exit: $?"
+    else
+        echo "Compile FAILED"
+    fi
+    echo ""
+done
@@ -0,0 +1,13 @@
+#!/bin/bash
+source /home/fabian/ComfyUI/venv/bin/activate
+export HSA_OVERRIDE_GFX_VERSION=10.1.0
+export HSA_ENABLE_SDMA=0
+export HIP_VISIBLE_DEVICES=0
+
+python3.11 -c "
+import torchvision
+print('torchvision version:', torchvision.__version__)
+from torchvision import transforms
+print('transforms OK')
+print('ALL GOOD')
+"
@@ -0,0 +1,123 @@
+"""Clean old output, re-submit, get REAL GPU timing."""
+import paramiko, time, json
+
+k = paramiko.Ed25519Key.from_private_key_file(r'C:\Users\fabia\.ssh\id_ed25519')
+c = paramiko.SSHClient()
+c.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+c.connect('192.168.178.150', username='fabian', pkey=k, timeout=10)
+sftp = c.open_sftp()
+
+def sh(cmd, timeout=60):
+    chan = c.get_transport().open_session()
+    chan.settimeout(timeout)
+    chan.exec_command('/bin/bash -l -c ' + "'" + cmd.replace("'", "'\\''") + "'")
+    out = b""
+    while True:
+        try:
+            chunk = chan.recv(65536)
+            if not chunk: break
+            out += chunk
+        except: break
+    chan.close()
+    return out.decode(errors='replace').strip()
+
+# Delete old output
+print("Cleaning old output...")
+sh('rm -f ~/ComfyUI/output/ZImageTurbo_GPU*.png')
+
+# Truncate log to see fresh output only
+sh('truncate -s 0 /tmp/comfyui.log; sleep 1')
+
+# Submit fresh workflow
+workflow = {
+    "prompt": {
+        "1": {"class_type": "UnetLoaderGGUF", "inputs": {"unet_name": "z_image_turbo-Q5_K_S.gguf"}},
+        "2": {"class_type": "CLIPLoaderGGUF", "inputs": {"clip_name": "Qwen3-4B.i1-Q5_K_S.gguf", "type": "qwen_image"}},
+        "3": {"class_type": "VAELoader", "inputs": {"vae_name": "ae.safetensors"}},
+        "4": {"class_type": "CLIPTextEncode", "inputs": {"text": "A red fox in a snowy forest, photorealistic", "clip": ["2", 0]}},
+        "5": {"class_type": "EmptyLatentImage", "inputs": {"width": 512, "height": 512, "batch_size": 1}},
+        "6": {"class_type": "KSampler", "inputs": {
+            "model": ["1", 0], "positive": ["4", 0], "negative": ["4", 0],
+            "latent_image": ["5", 0], "seed": 123, "steps": 8, "cfg": 1.0,
+            "sampler_name": "euler", "scheduler": "simple", "denoise": 1.0
+        }},
+        "7": {"class_type": "VAEDecode", "inputs": {"samples": ["6", 0], "vae": ["3", 0]}},
+        "8": {"class_type": "SaveImage", "inputs": {"images": ["7", 0], "filename_prefix": "ZImageTurbo_GPU"}}
+    }
+}
+with sftp.open('/tmp/wf.json', 'w') as f:
+    f.write(json.dumps(workflow))
+
+print("Submitting fresh workflow (seed=123)...")
+resp = sh('curl -s -X POST http://127.0.0.1:8188/prompt -H "Content-Type: application/json" -d @/tmp/wf.json')
+print(f"  {resp[:120]}")
+
+t0 = time.time()
+print("\nMonitoring (NORMAL_VRAM = real GPU compute)...")
+
+for i in range(200):
+    elapsed = int(time.time() - t0)
+    
+    gpu_temp = sh('cat /sys/class/drm/card0/device/hwmon/hwmon*/temp1_input 2>/dev/null', timeout=5)
+    temp_c = int(gpu_temp) // 1000 if gpu_temp.isdigit() else '?'
+    
+    try:
+        with sftp.open('/tmp/comfyui.log', 'r') as f:
+            log = f.read().decode(errors='replace')
+    except: log = ''
+    
+    sampling = ''
+    last = ''
+    for line in log.split('\n'):
+        s = line.strip()
+        if '/8' in s and ('it/s' in s or 's/it' in s):
+            sampling = s
+        if s and 'FETCH' not in s and 'startup tasks' not in s and 'DEPRECATION' not in s:
+            last = s
+    
+    display = sampling if sampling else last[-100:]
+    print(f"  [{elapsed:>4}s] {temp_c}C | {display}")
+    
+    imgs = sh('ls ~/ComfyUI/output/ZImageTurbo_GPU*.png 2>/dev/null', timeout=5)
+    if imgs:
+        exec_time = ''
+        for line in log.split('\n'):
+            if 'Prompt executed in' in line:
+                exec_time = line.strip()
+        print(f"\n  *** IMAGE GENERATED! ***")
+        print(f"  File: {imgs}")
+        print(f"  {exec_time}")
+        print(f"  Wall time: {elapsed}s")
+        for line in log.split('\n'):
+            s = line.strip()
+            if any(k in s for k in ['loaded completely', 'loaded partially', '/8', 'Prompt executed']):
+                print(f"  {s}")
+        break
+    
+    q = sh('curl -s http://127.0.0.1:8188/queue 2>/dev/null', timeout=5)
+    try:
+        qd = json.loads(q)
+        if not qd.get('queue_running') and not qd.get('queue_pending') and elapsed > 20:
+            time.sleep(2)
+            imgs = sh('ls ~/ComfyUI/output/ZImageTurbo_GPU*.png 2>/dev/null', timeout=5)
+            if imgs:
+                print(f"\n  *** IMAGE: {imgs} ***")
+            else:
+                print(f"\n  Queue empty, no image:")
+                for line in log.split('\n')[-20:]:
+                    if line.strip(): print(f"  {line.strip()}")
+            break
+    except: pass
+    
+    alive = sh('pgrep -f "python3.*main.py" >/dev/null && echo Y || echo N', timeout=5)
+    if alive == 'N':
+        print(f"\n  CRASHED!")
+        for line in log.split('\n')[-25:]:
+            if line.strip(): print(f"  {line.strip()}")
+        break
+    
+    time.sleep(10)
+
+sftp.close()
+c.close()
+print("\nDone.")
@@ -0,0 +1,48 @@
+"""Quick check: what's ACTUALLY happening in the ComfyUI log right now?"""
+import paramiko
+
+k = paramiko.Ed25519Key.from_private_key_file(r'C:\Users\fabia\.ssh\id_ed25519')
+c = paramiko.SSHClient()
+c.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+c.connect('192.168.178.150', username='fabian', pkey=k, timeout=10)
+
+def sh(cmd):
+    chan = c.get_transport().open_session()
+    chan.settimeout(30)
+    chan.exec_command(f"/bin/bash -c '{cmd}'")
+    out = b""
+    while True:
+        try:
+            chunk = chan.recv(65536)
+            if not chunk: break
+            out += chunk
+        except: break
+    chan.close()
+    return out.decode(errors='replace').strip()
+
+print("=== FULL LOG (minus ComfyUI-Manager spam) ===")
+# Show all lines EXCEPT the registry fetch / manager spam
+log = sh("grep -v 'FETCH ComfyRegistry\\|All startup tasks\\|ComfyUI-Manager' /tmp/comfyui.log | tail -50")
+print(log)
+
+print("\n=== PROCESS ===")
+print(sh("ps aux | grep python3 | grep -v grep"))
+
+print("\n=== GPU sysfs ===")
+# Find the actual gpu_busy path
+print(sh("find /sys/class/drm/ -name 'gpu_busy_percent' 2>/dev/null"))
+print(sh("cat /sys/class/drm/card0/device/gpu_busy_percent 2>/dev/null; cat /sys/class/drm/card1/device/gpu_busy_percent 2>/dev/null"))
+
+print("\n=== rocm-smi ===")
+print(sh("rocm-smi 2>/dev/null | head -15"))
+
+print("\n=== OUTPUT DIR ===")
+print(sh("ls -la ~/ComfyUI/output/ 2>/dev/null"))
+
+print("\n=== QUEUE ===")
+print(sh("curl -s http://127.0.0.1:8188/queue 2>/dev/null"))
+
+print("\n=== Log lines with 'load' or 'sample' or 'error' or '%' ===")
+print(sh("grep -iE 'load|sample|error|%|step|Traceback|OOM|killed' /tmp/comfyui.log | tail -30"))
+
+c.close()
@@ -0,0 +1,33 @@
+#!/usr/bin/env python3
+"""Check full log after sampling for VAE decode status."""
+import paramiko
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+def run(cmd):
+    _, so, se = ssh.exec_command(cmd, timeout=15)
+    return so.read().decode()
+
+# Get more log lines - look for everything after the sampling
+print("=== FULL LOG (last 60 lines) ===")
+print(run("tail -60 /home/fabian/comfyui.log 2>/dev/null"))
+
+# Check output directory
+print("\n=== OUTPUT FILES ===")
+print(run("ls -lah /home/fabian/ComfyUI/output/ 2>/dev/null"))
+
+# Queue status
+print("=== QUEUE ===")
+print(run("curl -s http://localhost:8188/queue 2>/dev/null"))
+
+# History
+print("\n=== HISTORY ===")
+print(run("curl -s http://localhost:8188/history 2>/dev/null")[:2000])
+
+# Process count and wchan
+print("\n=== PROCESS STATE ===")
+print(run("bash -c 'PID=$(pgrep -f \"python3 main.py\" | head -1); "
+    "cat /proc/$PID/wchan 2>/dev/null; echo; "
+    "ps -L -p $PID -o tid,%cpu,comm --sort=-%cpu 2>/dev/null | head -20'"))
+
+ssh.close()
@@ -0,0 +1,48 @@
+#!/usr/bin/env python3
+"""Check PyTorch build progress on BC-250."""
+import paramiko
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=30, desc=""):
+    if desc:
+        print(f"\n{'='*60}")
+        print(f"  {desc}")
+        print(f"{'='*60}")
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    if out.strip():
+        print(out.strip())
+    if err.strip():
+        print(f"STDERR: {err.strip()}")
+
+# Build process status
+run("pgrep -fa 'setup.py|build_pytorch|cmake|ninja|hipcc|cc1plus' | head -20",
+    desc="Build processes")
+
+# How long has it been running
+run("ps -p 350823 -o etime=,cmd= 2>/dev/null || echo 'Process no longer running'",
+    desc="Build process uptime")
+
+# Build log tail
+run("tail -60 /home/fabian/pytorch_build.log 2>/dev/null || echo 'No log file'",
+    desc="Build log (last 60 lines)")
+
+# Memory usage
+run("free -h", desc="Memory status")
+
+# Check for build completion marker
+run("grep 'BUILD_COMPLETE' /home/fabian/pytorch_build.log 2>/dev/null || echo 'Build still in progress'",
+    desc="Build completion check")
+
+# Check for any errors in log
+run("grep -i 'error:\\|fatal:\\|failed' /home/fabian/pytorch_build.log 2>/dev/null | tail -10 || echo 'No errors found'",
+    desc="Error check")
+
+# Disk space
+run("df -h / | tail -1", desc="Disk space")
+
+ssh.close()
@@ -0,0 +1,42 @@
+#!/usr/bin/env python3
+"""Check the build error from PyTorch CMake on BC-250."""
+import paramiko
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=30, desc=""):
+    if desc:
+        print(f"\n{'='*60}")
+        print(f"  {desc}")
+        print(f"{'='*60}")
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    if out.strip():
+        print(out.strip())
+
+# Get the full cmake error
+run("grep -A5 -B2 'Error\\|error\\|FATAL\\|fatal\\|Could not find' /home/fabian/pytorch_build.log | head -60",
+    desc="CMake errors in build log")
+
+# Also check the cmake output file if it exists
+run("cat /home/fabian/pytorch/build/CMakeFiles/CMakeOutput.log 2>/dev/null | tail -30 || echo 'no output log'",
+    desc="CMake output log")
+
+run("cat /home/fabian/pytorch/build/CMakeFiles/CMakeError.log 2>/dev/null | tail -50 || echo 'no error log'",
+    desc="CMake error log")
+
+# Check specifically what's missing
+run("grep -i 'not found\\|could not find\\|missing' /home/fabian/pytorch_build.log | head -20",
+    desc="Missing packages")
+
+# Also check if the process is still running
+run("pgrep -fa 'setup.py\\|build_pytorch' || echo 'Build process not running'",
+    desc="Build process status")
+
+# Check roctracer
+run("find /opt/rocm -name 'roctracer*' 2>/dev/null | head -10",
+    desc="roctracer files")
+
+ssh.close()
@@ -0,0 +1,30 @@
+#!/usr/bin/env python3
+"""Check ComfyUI flags and update startup."""
+import paramiko
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=60):
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    return out + err
+
+# Kill any leftover
+print(run("pkill -9 -f 'python3 main.py' 2>/dev/null; echo killed"))
+
+# Full help output
+print("=== ComfyUI --help ===")
+help_text = run("bash -c 'source /home/fabian/comfyui-env/bin/activate && cd /home/fabian/ComfyUI && python3 main.py --help 2>&1'")
+# Filter for interesting lines
+for line in help_text.split('\n'):
+    low = line.lower()
+    if any(w in low for w in ['vae', 'fp16', 'fp32', 'force', 'cpu', 'vram', 'memory', 'offload', 'precision']):
+        print(f"  {line.strip()}")
+
+# Also just dump the full thing to see everything
+print("\n=== FULL HELP ===")
+print(help_text)
+
+ssh.close()
@@ -0,0 +1,59 @@
+#!/usr/bin/env python3
+"""Check PyTorch state and start fresh build on BC-250."""
+import paramiko
+import time
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=120, desc=""):
+    if desc:
+        print(f"\n{'='*60}")
+        print(f"  {desc}")
+        print(f"{'='*60}")
+    print(f"$ {cmd}")
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    rc = stdout.channel.recv_exit_status()
+    if out.strip():
+        lines = out.strip().split('\n')
+        if len(lines) > 60:
+            print(f"  ... ({len(lines)} lines, showing last 60)")
+            print('\n'.join(lines[-60:]))
+        else:
+            print(out.strip())
+    if err.strip():
+        lines = err.strip().split('\n')
+        show = lines[-20:] if len(lines) > 20 else lines
+        print(f"STDERR: {chr(10).join(show)}")
+    print(f"  Exit code: {rc}")
+    return rc, out, err
+
+# Check if there's a wheel already built
+run("ls -lh ~/pytorch/dist/*.whl 2>/dev/null || echo 'No wheels found'",
+    desc="Check for existing PyTorch wheels")
+
+# Check for any previous build directory
+run("ls -la ~/pytorch/build/ 2>/dev/null | head -10 || echo 'No build dir'",
+    desc="Check build directory")
+
+# Check if pytorch is already installed in venv
+run("bash -c 'source ~/comfyui-env/bin/activate && python3 -c \"import torch; print(torch.__version__); print(torch.version.hip); print(torch.cuda.is_available())\" 2>&1'",
+    desc="Check if PyTorch is already installed")
+
+# Check pytorch source integrity
+run("ls ~/pytorch/setup.py ~/pytorch/CMakeLists.txt 2>&1",
+    desc="Verify PyTorch source files")
+
+# Verify ROCm works before build
+run("bash -c 'export HSA_OVERRIDE_GFX_VERSION=10.1.0 && /opt/rocm/bin/rocminfo 2>&1 | grep -E \"gfx|Marketing\" | head -5'",
+    desc="Verify ROCm is working")
+
+# Check venv
+run("bash -c 'source ~/comfyui-env/bin/activate && which python3 && python3 --version'",
+    desc="Verify venv")
+
+ssh.close()
+print("\nDone checking state.")
@@ -0,0 +1,56 @@
+#!/usr/bin/env python3
+"""Quick check on ComfyUI status."""
+import paramiko
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=30):
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    return stdout.read().decode()
+
+# Log tail
+print("=== LOG (last 40 lines) ===")
+print(run("tail -40 /home/fabian/comfyui.log 2>/dev/null"))
+
+# Process status
+print("=== PROCESS ===")
+print(run("bash -c 'PID=$(pgrep -f \"python3 main.py\" | head -1); "
+    "if [ -n \"$PID\" ]; then "
+    "  ps -p $PID -o pid,%cpu,%mem,nlwp,stat --no-headers; "
+    "  echo \"LOAD: $(cat /proc/loadavg)\"; "
+    "else echo DEAD; fi'"))
+
+# rocm-smi
+print("=== GPU ===")
+print(run("HSA_OVERRIDE_GFX_VERSION=10.1.0 rocm-smi 2>/dev/null || echo 'no rocm-smi'"))
+
+# Queue
+print("=== QUEUE ===")
+print(run("curl -s http://localhost:8188/queue 2>/dev/null || echo 'no connection'"))
+
+# History
+print("=== HISTORY ===")
+hist = run("curl -s http://localhost:8188/history 2>/dev/null || echo 'no connection'")
+import json
+try:
+    h = json.loads(hist)
+    for pid, info in h.items():
+        print(f"  Prompt: {pid}")
+        print(f"  Status: {info.get('status', {})}")
+        outputs = info.get('outputs', {})
+        if outputs:
+            for nid, nout in outputs.items():
+                if isinstance(nout, dict):
+                    for key, val in nout.items():
+                        print(f"  Output node {nid}/{key}: {str(val)[:200]}")
+        else:
+            print("  No outputs")
+except:
+    print(hist[:1000])
+
+# Output directory
+print("\n=== OUTPUT FILES ===")
+print(run("ls -lah /home/fabian/ComfyUI/output/ 2>/dev/null"))
+
+ssh.close()
@@ -0,0 +1,52 @@
+#!/usr/bin/env python3
+"""Read full ops.py and check torch thread defaults, then fix threading."""
+import paramiko
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=30, desc=""):
+    if desc:
+        print(f"\n{'='*60}")
+        print(f"  {desc}")
+        print(f"{'='*60}")
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    rc = stdout.channel.recv_exit_status()
+    if out.strip():
+        print(out.strip())
+    if err.strip():
+        lines = err.strip().split('\n')[-10:]
+        print(f"STDERR: {chr(10).join(lines)}")
+    print(f"  Exit: {rc}")
+    return rc, out, err
+
+# Check current torch thread defaults
+run("bash -c 'source ~/comfyui-env/bin/activate && "
+    "export HSA_OVERRIDE_GFX_VERSION=10.1.0 && "
+    "python -c \""
+    "import torch; "
+    "print(f\\\"num_threads={torch.get_num_threads()}\\\"); "
+    "print(f\\\"num_interop_threads={torch.get_num_interop_threads()}\\\"); "
+    "import os; "
+    "print(f\\\"OMP_NUM_THREADS={os.environ.get(\\\\\\\"OMP_NUM_THREADS\\\\\\\", \\\\\\\"not set\\\\\\\")}\\\"); "
+    "print(f\\\"MKL_NUM_THREADS={os.environ.get(\\\\\\\"MKL_NUM_THREADS\\\\\\\", \\\\\\\"not set\\\\\\\")}\\\"); "
+    "\"'",
+    desc="Check default torch thread settings")
+
+# Read forward_ggml_cast_weights (where dequant happens during inference)
+run("bash -c 'sed -n \"200,281p\" ~/ComfyUI/custom_nodes/ComfyUI-GGUF/ops.py'",
+    desc="ops.py lines 200-281 (forward functions)")
+
+# Check __init__.py for any loading/patching
+run("bash -c 'cat ~/ComfyUI/custom_nodes/ComfyUI-GGUF/__init__.py 2>/dev/null | head -40'",
+    desc="ComfyUI-GGUF __init__.py")
+
+# Check nodes.py for model loading
+run("bash -c 'cat ~/ComfyUI/custom_nodes/ComfyUI-GGUF/nodes.py 2>/dev/null'",
+    desc="ComfyUI-GGUF nodes.py (model loader)")
+
+ssh.close()
+print("\nDone.")
@@ -0,0 +1,271 @@
+#!/usr/bin/env python3
+"""
+Switch ComfyUI to --cpu mode so ALL 12 cores are used.
+The GPU (Cyan Skillfish gfx1013) hangs during HIP inference ops,
+causing the single-core stall. CPU mode with MKL+OpenMP will use all cores.
+"""
+import paramiko
+import json
+import time
+import textwrap
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=120, desc=""):
+    if desc:
+        print(f"\n{'='*60}")
+        print(f"  {desc}")
+        print(f"{'='*60}")
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    rc = stdout.channel.recv_exit_status()
+    if out.strip():
+        lines = out.strip().split('\n')
+        if len(lines) > 30:
+            print(f"  ({len(lines)} lines, showing last 30)")
+            print('\n'.join(lines[-30:]))
+        else:
+            print(out.strip())
+    if err.strip():
+        for line in err.strip().split('\n')[-5:]:
+            print(f"  STDERR: {line}")
+    return rc, out, err
+
+# ── Step 1: Kill stuck ComfyUI ──
+run("bash -c 'pkill -f \"python3 main.py\" 2>/dev/null; sleep 2; "
+    "pkill -9 -f \"python3 main.py\" 2>/dev/null; sleep 1; "
+    "echo \"Killed. Remaining:\"; pgrep -af \"main.py\" || echo none'",
+    desc="Kill stuck ComfyUI")
+
+# ── Step 2: Write new startup script with --cpu ──
+# Key: OMP_NUM_THREADS=12 + MKL_NUM_THREADS=12 + --cpu
+# This uses Intel MKL (built into this PyTorch) for matrix ops across all cores
+startup_script = textwrap.dedent("""\
+    #!/bin/bash
+    # ComfyUI CPU-mode launcher for BC-250
+    # Forces ALL computation on CPU using 12 cores via MKL + OpenMP
+    
+    # Threading: use ALL 12 cores
+    export OMP_NUM_THREADS=12
+    export MKL_NUM_THREADS=12
+    export OPENBLAS_NUM_THREADS=12
+    export OMP_PROC_BIND=spread
+    export OMP_PLACES=cores
+    export GOMP_CPU_AFFINITY="0-11"
+    
+    # No GPU needed in CPU mode, but keep env for potential future use
+    export HSA_OVERRIDE_GFX_VERSION=10.1.0
+    export HIP_VISIBLE_DEVICES=0
+    export HSA_ENABLE_SDMA=0
+    export HSA_TOOLS_LIB=""
+    export HSA_TOOLS_REPORT_LOAD_FAILURE=0
+    
+    # MKL tuning for multi-core
+    export MKL_DYNAMIC=FALSE
+    export MKL_ENABLE_INSTRUCTIONS=AVX2
+    
+    # Activate venv
+    source /home/fabian/comfyui-env/bin/activate
+    cd /home/fabian/ComfyUI
+    
+    echo "=== BC-250 ComfyUI CPU Mode ==="
+    echo "Cores: 12, OMP_NUM_THREADS=$OMP_NUM_THREADS, MKL_NUM_THREADS=$MKL_NUM_THREADS"
+    echo "OMP_PROC_BIND=$OMP_PROC_BIND, OMP_PLACES=$OMP_PLACES"
+    
+    # --cpu: force ALL ops on CPU (no GPU)
+    # --disable-auto-launch: don't open browser
+    exec python3 main.py --listen 0.0.0.0 --port 8188 --cpu --disable-auto-launch
+""")
+
+sftp = ssh.open_sftp()
+with sftp.open('/home/fabian/start_comfyui.sh', 'w') as f:
+    f.write(startup_script)
+sftp.close()
+run("chmod +x /home/fabian/start_comfyui.sh")
+print("\n  Updated start_comfyui.sh -> --cpu mode, 12 cores, MKL tuning")
+
+# ── Step 3: Update sitecustomize.py to also set MKL_DYNAMIC=FALSE ──
+sitecustomize = textwrap.dedent("""\
+    import os
+    os.environ.setdefault('OMP_NUM_THREADS', '12')
+    os.environ.setdefault('MKL_NUM_THREADS', '12')
+    os.environ.setdefault('MKL_DYNAMIC', 'FALSE')
+    os.environ.setdefault('OMP_PROC_BIND', 'spread')
+    os.environ.setdefault('OMP_PLACES', 'cores')
+    
+    try:
+        import torch
+        torch.set_num_threads(12)
+        torch.set_num_interop_threads(12)
+        print(f"Threads: intra={torch.get_num_threads()} inter={torch.get_num_interop_threads()}")
+    except Exception:
+        pass
+""")
+
+sftp = ssh.open_sftp()
+with sftp.open('/home/fabian/comfyui-env/lib/python3.14/site-packages/sitecustomize.py', 'w') as f:
+    f.write(sitecustomize)
+sftp.close()
+print("  Updated sitecustomize.py with MKL_DYNAMIC=FALSE")
+
+# ── Step 4: Launch ComfyUI in CPU mode ──
+run("bash -c 'nohup /home/fabian/start_comfyui.sh > /home/fabian/comfyui.log 2>&1 &'",
+    desc="Launch ComfyUI in CPU mode")
+time.sleep(8)
+
+rc, out, _ = run("bash -c 'tail -30 /home/fabian/comfyui.log 2>/dev/null'",
+    desc="Startup log")
+
+# Verify server started
+for attempt in range(10):
+    rc, out, _ = run("bash -c 'curl -s -o /dev/null -w \"%{http_code}\" http://localhost:8188/ 2>/dev/null'")
+    if '200' in out:
+        print(f"\n  Server is UP on port 8188 (attempt {attempt+1})")
+        break
+    time.sleep(5)
+else:
+    print("\n  WARNING: Server didn't respond after 50s")
+    run("bash -c 'tail -50 /home/fabian/comfyui.log'", desc="Full log")
+    ssh.close()
+    exit(1)
+
+# ── Step 5: Submit workflow with slightly smaller image for faster CPU gen ──
+# 768x432 instead of 1024x576 to speed up first test
+workflow = {
+    "prompt": {
+        "1": {
+            "class_type": "UnetLoaderGGUF",
+            "inputs": {"unet_name": "z_image_turbo-Q5_K_S.gguf"}
+        },
+        "2": {
+            "class_type": "CLIPLoaderGGUF",
+            "inputs": {"clip_name": "Qwen3-4B.i1-Q5_K_S.gguf", "type": "qwen_image"}
+        },
+        "3": {
+            "class_type": "VAELoader",
+            "inputs": {"vae_name": "ae.safetensors"}
+        },
+        "4": {
+            "class_type": "CLIPTextEncode",
+            "inputs": {
+                "text": "A majestic mountain landscape at sunset, golden light on snow peaks, crystal lake reflection, photorealistic, 8k",
+                "clip": ["2", 0]
+            }
+        },
+        "5": {
+            "class_type": "CLIPTextEncode",
+            "inputs": {"text": "", "clip": ["2", 0]}
+        },
+        "6": {
+            "class_type": "EmptyLatentImage",
+            "inputs": {"width": 768, "height": 432, "batch_size": 1}
+        },
+        "7": {
+            "class_type": "KSampler",
+            "inputs": {
+                "model": ["1", 0],
+                "seed": 42,
+                "steps": 8,
+                "cfg": 1.0,
+                "sampler_name": "euler",
+                "scheduler": "simple",
+                "positive": ["4", 0],
+                "negative": ["5", 0],
+                "latent_image": ["6", 0],
+                "denoise": 1.0
+            }
+        },
+        "8": {
+            "class_type": "VAEDecode",
+            "inputs": {"samples": ["7", 0], "vae": ["3", 0]}
+        },
+        "9": {
+            "class_type": "SaveImage",
+            "inputs": {"images": ["8", 0], "filename_prefix": "ZImageTurbo_BC250"}
+        }
+    }
+}
+
+sftp = ssh.open_sftp()
+with sftp.open('/tmp/zimage_workflow.json', 'w') as f:
+    f.write(json.dumps(workflow))
+sftp.close()
+
+rc, out, _ = run("bash -c 'curl -s -X POST http://localhost:8188/prompt "
+    "-H \"Content-Type: application/json\" "
+    "-d @/tmp/zimage_workflow.json'",
+    desc="Submit 768x432 workflow")
+
+try:
+    resp = json.loads(out.strip())
+    if 'error' in resp:
+        print(f"\n  ERROR: {resp['error']}")
+        if 'node_errors' in resp:
+            for nid, e in resp['node_errors'].items():
+                print(f"  Node {nid}: {e}")
+        ssh.close()
+        exit(1)
+    prompt_id = resp.get('prompt_id', 'unknown')
+    print(f"\n  Prompt ID: {prompt_id}")
+except Exception as e:
+    print(f"  Parse error: {e}\n  Raw: {out[:500]}")
+
+# ── Step 6: Monitor CPU/progress ──
+print("\n  Monitoring generation (CPU mode, 12 cores)...")
+print("  This is a 6B model on CPU — expect several minutes per step")
+
+start_time = time.time()
+last_log = ""
+for i in range(240):  # up to 60 min
+    time.sleep(15)
+    elapsed = time.time() - start_time
+    minutes = int(elapsed // 60)
+    seconds = int(elapsed % 60)
+    
+    # CPU usage - check if ALL cores are active
+    rc, cpu_out, _ = run("bash -c '"
+        "PID=$(pgrep -f \"python3 main.py\" | head -1); "
+        "if [ -n \"$PID\" ]; then "
+        "  echo \"CPU_PCT=$(ps -p $PID -o %cpu= 2>/dev/null)\"; "
+        "  echo \"MEM_PCT=$(ps -p $PID -o %mem= 2>/dev/null)\"; "
+        "  echo \"THREADS=$(ps -p $PID -o nlwp= 2>/dev/null)\"; "
+        "  echo \"LOADAVG=$(cat /proc/loadavg)\"; "
+        "else echo PROCESS_DEAD; fi'")
+    
+    # Parse CPU metrics
+    cpu_pct = "?"
+    load_avg = "?"
+    for line in (cpu_out or '').split('\n'):
+        if line.startswith('CPU_PCT='):
+            cpu_pct = line.split('=')[1].strip()
+        if line.startswith('LOADAVG='):
+            load_avg = line.split('=')[1].strip().split()[0]
+    
+    # Log tail
+    rc, log_out, _ = run("bash -c 'tail -3 /home/fabian/comfyui.log 2>/dev/null'")
+    log_tail = (log_out or '').strip().split('\n')[-1] if log_out else ""
+    
+    if 'PROCESS_DEAD' in (cpu_out or ''):
+        print(f"\n  [{minutes}m{seconds}s] PROCESS DIED!")
+        run("bash -c 'tail -50 /home/fabian/comfyui.log'", desc="Death log")
+        break
+    
+    # Show progress
+    print(f"  [{minutes}m{seconds}s] CPU={cpu_pct}% Load={load_avg} | {log_tail[:80]}")
+    
+    if 'Prompt executed in' in (log_out or ''):
+        print(f"\n  IMAGE GENERATED! Total time: {minutes}m{seconds}s")
+        run("bash -c 'tail -20 /home/fabian/comfyui.log'", desc="Completion log")
+        run("bash -c 'ls -lah ~/ComfyUI/output/'", desc="Output files")
+        break
+    
+    if 'Error' in log_tail or 'Traceback' in (log_out or ''):
+        print(f"\n  ERROR DETECTED!")
+        run("bash -c 'tail -60 /home/fabian/comfyui.log'", desc="Error log")
+        break
+
+ssh.close()
+print("\nDone.")
@@ -0,0 +1,31 @@
+import paramiko, time
+
+k = paramiko.Ed25519Key.from_private_key_file(r'C:\Users\fabia\.ssh\id_ed25519')
+c = paramiko.SSHClient()
+c.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+c.connect('192.168.178.150', username='fabian', pkey=k, timeout=10)
+
+# 1) Full diagnostic
+cmds = {
+    "LOG_LAST_40": "tail -40 /tmp/comfyui.log 2>/dev/null",
+    "PROCESS": "ps aux | grep -E 'python|comfy' | grep -v grep",
+    "GPU": "cat /sys/class/drm/card0/device/gpu_busy_percent 2>/dev/null",
+    "GPU_TEMP": "cat /sys/class/drm/card0/device/hwmon/hwmon*/temp1_input 2>/dev/null",
+    "CPU_CORES": "mpstat -P ALL 1 1 2>/dev/null | tail -15 || top -bn1 | head -5",
+    "MEM": "free -m",
+    "OUTPUT": "ls -la ~/ComfyUI/output/ 2>/dev/null",
+    "QUEUE": "curl -s http://127.0.0.1:8188/queue 2>/dev/null",
+    "ROCM_CHECK": "rocm-smi --showuse --showtemp --showpower 2>/dev/null | head -20",
+}
+
+for name, cmd in cmds.items():
+    print(f"\n=== {name} ===")
+    _, o, e = c.exec_command(cmd)
+    out = o.read().decode(errors='replace').strip()
+    err = e.read().decode(errors='replace').strip()
+    print(out if out else "(empty)")
+    if err and name not in ("ROCM_CHECK",):
+        print(f"  STDERR: {err}")
+
+c.close()
+print("\nDone.")
@@ -0,0 +1,46 @@
+#!/usr/bin/env python3
+"""Diagnostic: check full log and GPU state."""
+import paramiko
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=30):
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    return out, err
+
+# Full log (last 80 lines)
+out, _ = run("tail -80 /home/fabian/comfyui.log 2>/dev/null")
+print("=== FULL LOG (last 80 lines) ===")
+print(out)
+
+# Process state
+out, _ = run("bash -c 'PID=$(pgrep -f \"python3 main.py\" | head -1); "
+    "if [ -n \"$PID\" ]; then "
+    "  echo \"PID: $PID\"; "
+    "  echo \"=== PROCESS STATE ===\"; "
+    "  cat /proc/$PID/status | grep -E \"State|Threads|VmRSS|VmSize\"; "
+    "  echo \"=== WCHAN (what syscall is process blocked on) ===\"; "
+    "  cat /proc/$PID/wchan 2>/dev/null; echo; "
+    "  echo \"=== STACK TRACE (kernel) ===\"; "
+    "  sudo cat /proc/$PID/stack 2>/dev/null || echo \"no permission\"; "
+    "  echo \"=== TOP THREADS ===\"; "
+    "  ps -L -p $PID -o tid,%cpu,comm --sort=-%cpu | head -15; "
+    "fi'")
+print(out)
+
+# GPU info
+out, _ = run("bash -c 'rocm-smi 2>/dev/null || echo no rocm-smi; "
+    "echo \"=== dmesg GPU ===\"; "
+    "dmesg 2>/dev/null | grep -i -E \"amdgpu|error|fault\" | tail -15 || echo no-dmesg'")
+print("=== GPU ===")
+print(out)
+
+# Memory
+out, _ = run("free -h")
+print("=== MEMORY ===")
+print(out)
+
+ssh.close()
@@ -0,0 +1,51 @@
+#!/usr/bin/env python3
+"""Check ComfyUI API queue/history for errors."""
+import paramiko, json
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=30):
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    return stdout.read().decode(), stderr.read().decode()
+
+# Queue status
+out, _ = run("curl -s http://localhost:8188/queue")
+print("=== QUEUE ===")
+try:
+    q = json.loads(out)
+    print(f"Running: {len(q.get('queue_running', []))}")
+    print(f"Pending: {len(q.get('queue_pending', []))}")
+except:
+    print(out[:500])
+
+# History
+out, _ = run("curl -s http://localhost:8188/history")
+print("\n=== HISTORY ===")
+try:
+    h = json.loads(out)
+    for pid, info in h.items():
+        print(f"\nPrompt ID: {pid}")
+        status = info.get('status', {})
+        print(f"  Status: {status}")
+        outputs = info.get('outputs', {})
+        for nid, nout in outputs.items():
+            print(f"  Node {nid}: {list(nout.keys()) if isinstance(nout, dict) else nout}")
+        if not outputs:
+            print("  NO OUTPUTS")
+except:
+    print(out[:2000])
+
+# Check stderr output (nohup might redirect differently)  
+out, _ = run("cat /home/fabian/comfyui_err.log 2>/dev/null || echo 'no err log'")
+print(f"\n=== STDERR LOG ===\n{out[:2000]}")
+
+# Check full nohup output
+out, _ = run("wc -l /home/fabian/comfyui.log 2>/dev/null")
+print(f"\n=== LOG LINES: {out.strip()}")
+
+# Check if there are processes actively computing
+out, _ = run("bash -c 'top -bn1 | head -20'")
+print(f"\n=== TOP ===\n{out}")
+
+ssh.close()
@@ -0,0 +1,106 @@
+#!/usr/bin/env python3
+"""Deep diagnosis: WHY only 1 core? Check OpenMP, threading, GGUF code path."""
+import paramiko, json, textwrap
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=60):
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    rc = stdout.channel.recv_exit_status()
+    return rc, out, err
+
+def show(label, cmd, timeout=60):
+    print(f"\n{'='*60}")
+    print(f"  {label}")
+    print(f"{'='*60}")
+    rc, out, err = run(cmd, timeout)
+    if out.strip():
+        print(out.strip())
+    if err.strip():
+        for line in err.strip().split('\n')[-10:]:
+            print(f"STDERR: {line}")
+    return out
+
+# 1. Check ComfyUI queue/history - did generation succeed or fail?
+show("Queue status", "curl -s http://localhost:8188/queue")
+hist_out = show("History", "curl -s http://localhost:8188/history")
+try:
+    h = json.loads(hist_out.strip())
+    for pid, info in h.items():
+        status = info.get('status', {})
+        print(f"\n  Prompt {pid}: status={status}")
+        outputs = info.get('outputs', {})
+        if outputs:
+            print(f"  Outputs: {json.dumps(outputs, indent=2)[:500]}")
+        else:
+            print("  NO OUTPUTS")
+except:
+    pass
+
+# 2. Check if PyTorch has OpenMP 
+show("PyTorch OpenMP & threading", 
+    "bash -c 'source /home/fabian/comfyui-env/bin/activate && "
+    "OMP_NUM_THREADS=12 python3 -c \""
+    "import torch; "
+    "print(f\\\"OpenMP available: {torch.backends.openmp.is_available()}\\\"); "
+    "print(f\\\"MKL available: {torch.backends.mkl.is_available()}\\\"); "
+    "print(f\\\"Num threads: {torch.get_num_threads()}\\\"); "
+    "print(f\\\"Num interop threads: {torch.get_num_interop_threads()}\\\"); "
+    "print(f\\\"torch.__config__.show():  \\\"); "
+    "print(torch.__config__.show()); "
+    "\"'")
+
+# 3. Check if libomp/libgomp is available
+show("OpenMP libraries", 
+    "bash -c 'ldconfig -p 2>/dev/null | grep -i omp; "
+    "echo ---; "
+    "pacman -Qs openmp 2>/dev/null; "
+    "echo ---; "
+    "pacman -Qs libgomp 2>/dev/null; "
+    "echo ---; "
+    "ls -la /usr/lib/libomp* /usr/lib/libgomp* 2>/dev/null || echo none'")
+
+# 4. Check pytorch shared lib dependencies for OpenMP
+show("PyTorch .so OpenMP deps",
+    "bash -c 'ldd /usr/lib/python3.14/site-packages/torch/lib/libtorch_cpu.so 2>/dev/null | grep -i omp'")
+
+# 5. Actual thread test - does a matrix multiply use multiple cores?
+show("Matrix multiply CPU benchmark (should use all cores)",
+    "bash -c 'source /home/fabian/comfyui-env/bin/activate && "
+    "OMP_NUM_THREADS=12 python3 -c \""
+    "import torch, time, os; "
+    "print(f\\\"PID: {os.getpid()}\\\"); "
+    "torch.set_num_threads(12); "
+    "print(f\\\"Threads set to: {torch.get_num_threads()}\\\"); "
+    "a = torch.randn(4096, 4096); "
+    "b = torch.randn(4096, 4096); "
+    "# warmup; "
+    "c = torch.mm(a, b); "
+    "import subprocess; "
+    "# Start monitoring in background; "
+    "start = time.time(); "
+    "for i in range(5): c = torch.mm(a, b); "
+    "elapsed = time.time() - start; "
+    "print(f\\\"5x matmul 4096x4096: {elapsed:.2f}s\\\"); "
+    "\"'")
+
+# 6. Check what the GGUF dequant code actually does (single-threaded python loop?)
+show("GGUF dequant code - is there a Python for-loop?",
+    "bash -c 'grep -n \"for \" /home/fabian/ComfyUI/custom_nodes/ComfyUI-GGUF/dequant.py | head -20; "
+    "echo \"---\"; "
+    "grep -n \"for \" /home/fabian/ComfyUI/custom_nodes/ComfyUI-GGUF/ops.py | head -20; "
+    "echo \"---\"; "
+    "grep -n \"def dequantize\" /home/fabian/ComfyUI/custom_nodes/ComfyUI-GGUF/dequant.py'")
+
+# 7. Check if lowvram is causing sequential layer-by-layer processing
+show("ComfyUI lowvram model loading code",
+    "bash -c 'grep -rn \"lowvram\\|low_vram\\|offload\" /home/fabian/ComfyUI/comfy/model_management.py 2>/dev/null | head -30'")
+
+# 8. Output directory
+show("Output files", "ls -lah /home/fabian/ComfyUI/output/ 2>/dev/null")
+
+ssh.close()
+print("\n\nDONE.")
@@ -0,0 +1,82 @@
+"""Diagnose and fix VAE hang. Check log, fix threading, restart."""
+import paramiko, time, json
+
+k = paramiko.Ed25519Key.from_private_key_file(r'C:\Users\fabia\.ssh\id_ed25519')
+c = paramiko.SSHClient()
+c.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+c.connect('192.168.178.150', username='fabian', pkey=k, timeout=10)
+sftp = c.open_sftp()
+
+def sh(cmd, timeout=30):
+    chan = c.get_transport().open_session()
+    chan.settimeout(timeout)
+    chan.exec_command('/bin/bash -l -c ' + "'" + cmd.replace("'", "'\\''") + "'")
+    out = b""
+    while True:
+        try:
+            chunk = chan.recv(65536)
+            if not chunk: break
+            out += chunk
+        except: break
+    chan.close()
+    return out.decode(errors='replace').strip()
+
+# 1. What's running?
+print("=== CURRENT STATE ===")
+ps = sh('ps aux | grep main.py | grep -v grep')
+print(f"Process: {ps or 'NONE'}")
+
+# Check env of running process
+env = sh(r'cat /proc/$(pgrep -f "python3.*main.py" | head -1)/environ 2>/dev/null | tr "\0" "\n" | grep -E "OMP|MKL|COMFYUI|THREAD|OPENBLAS"')
+print(f"Env:\n{env}")
+
+# 2. Log tail
+print("\n=== LOG TAIL ===")
+try:
+    with sftp.open('/tmp/comfyui.log', 'r') as f:
+        log = f.read().decode(errors='replace')
+    for line in log.split('\n')[-40:]:
+        s = line.strip()
+        if s: print(f"  {s}")
+except Exception as e:
+    log = ''
+    print(f"  No log: {e}")
+
+# 3. Check launcher
+print("\n=== LAUNCHER ===")
+try:
+    with sftp.open('/tmp/run_comfyui.sh', 'r') as f:
+        print(f.read().decode())
+except: print("  No launcher")
+
+# 4. model_management.py patch?
+with sftp.open('/home/fabian/ComfyUI/comfy/model_management.py', 'r') as f:
+    mm = f.read().decode()
+print(f"SHARED patch: {'YES' if 'COMFYUI_SHARED_MEMORY' in mm else 'NO'}")
+
+# 5. Check torch threads in same env
+print("\n=== TORCH THREADS ===")
+tcheck = sh('''source ~/comfyui-env/bin/activate
+export HSA_OVERRIDE_GFX_VERSION=10.1.0
+python3 -c "
+import torch, os
+print(f'torch.get_num_threads() = {torch.get_num_threads()}')
+print(f'OMP_NUM_THREADS = {os.environ.get(chr(34)+'OMP_NUM_THREADS'+chr(34), chr(34)+'NOT SET'+chr(34))}')
+"''', timeout=30)
+print(tcheck)
+
+# 6. Check sitecustomize
+print("\n=== SITECUSTOMIZE ===")
+sc = sh('cat ~/comfyui-env/lib/python*/site-packages/sitecustomize.py 2>/dev/null || echo MISSING')
+print(sc[:500])
+
+# 7. Key log lines
+print("\n=== KEY LOG ENTRIES ===")
+for line in log.split('\n'):
+    s = line.strip()
+    if any(x in s.lower() for x in ['vram state', 'shared', 'loaded', 'offloaded', 'device:', 'total vram', 'vae', 'thread']):
+        print(f"  {s}")
+
+sftp.close()
+c.close()
+print("\nDiag done.")
@@ -0,0 +1,48 @@
+#!/usr/bin/env python3
+"""Diagnose ComfyUI state on BC-250 — check if stuck or OOM."""
+import paramiko
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=30, desc=""):
+    if desc:
+        print(f"\n{'='*60}")
+        print(f"  {desc}")
+        print(f"{'='*60}")
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    rc = stdout.channel.recv_exit_status()
+    if out.strip():
+        print(out.strip())
+    if err.strip():
+        lines = err.strip().split('\n')[-10:]
+        print(f"STDERR: {chr(10).join(lines)}")
+    print(f"  Exit: {rc}")
+    return rc, out, err
+
+# Check if ComfyUI process is alive
+run("bash -c 'ps aux | grep \"python main.py\" | grep -v grep'",
+    desc="ComfyUI process status")
+
+# Memory state
+run("bash -c 'free -h'", desc="RAM/Swap usage")
+
+# GPU VRAM
+run("bash -c 'cat /sys/class/drm/card1/device/mem_info_vram_used 2>/dev/null; "
+    "echo \"---\"; cat /sys/class/drm/card1/device/mem_info_vram_total 2>/dev/null'",
+    desc="GPU VRAM usage")
+
+# dmesg for OOM
+run("bash -c 'dmesg | tail -20'", desc="Recent kernel messages")
+
+# Last 80 lines of comfyui log
+run("bash -c 'tail -80 /home/fabian/comfyui.log 2>/dev/null'", desc="ComfyUI log (last 80)")
+
+# Check if port still listening
+run("bash -c 'ss -tlnp | grep 8188 || echo PORT_GONE'", desc="Port 8188")
+
+ssh.close()
+print("\nDone.")
@@ -0,0 +1,135 @@
+#!/usr/bin/env python3
+"""
+Diagnose GPU hang: kill stuck ComfyUI, run targeted HIP tests,
+check what ops hang on Cyan Skillfish gfx1013->gfx1010.
+Single SSH connection, properly closed.
+"""
+import paramiko
+import json
+import time
+import sys
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+
+for attempt in range(5):
+    try:
+        ssh.connect('192.168.178.150', username='fabian',
+                    key_filename=r'C:\Users\fabia\.ssh\id_ed25519', timeout=10)
+        break
+    except Exception as e:
+        print(f"  SSH attempt {attempt+1}/5: {e}")
+        time.sleep(10)
+else:
+    print("FATAL: Cannot connect"); sys.exit(1)
+
+def run(cmd, timeout=120):
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    return out, err
+
+try:
+    # 1. Kill stuck ComfyUI
+    print("=== Kill stuck ComfyUI ===")
+    out, _ = run("pkill -9 -f 'python3 main.py' 2>/dev/null; sleep 2; echo killed")
+    print(f"  {out.strip()}")
+
+    # 2. Check ComfyUI help for --cpu-vae flag existence
+    print("\n=== Check if --cpu-vae exists ===")
+    out, err = run("bash -c 'source ~/comfyui-env/bin/activate && cd ~/ComfyUI && python3 main.py --help 2>&1'")
+    full_help = out + err
+    has_cpu_vae = '--cpu-vae' in full_help
+    print(f"  --cpu-vae flag exists: {has_cpu_vae}")
+    # Print all vram/gpu related flags
+    for line in full_help.split('\n'):
+        if any(w in line.lower() for w in ['vram', 'cpu', 'gpu', 'fp16', 'fp32', 'vae', 'force', 'precision']):
+            print(f"  {line.strip()}")
+
+    # 3. Check what the CachyOS pytorch-rocm was built for
+    print("\n=== PyTorch ROCm build info ===")
+    out, _ = run("bash -c 'source ~/comfyui-env/bin/activate && python3 -c \""
+        "import torch; "
+        "print(f\\\"PyTorch version: {torch.__version__}\\\"); "
+        "print(f\\\"CUDA/HIP available: {torch.cuda.is_available()}\\\"); "
+        "print(f\\\"ROCm version: {torch.version.hip}\\\"); "
+        "print(f\\\"Device: {torch.cuda.get_device_name(0)}\\\"); "
+        "print(f\\\"Arch: {torch.cuda.get_device_capability(0)}\\\"); "
+        "print(f\\\"VRAM free/total: {torch.cuda.mem_get_info()[0]//1048576}/{torch.cuda.mem_get_info()[1]//1048576} MB\\\"); "
+        "\"' 2>&1")
+    print(out.strip())
+
+    # 4. Targeted GPU op tests - find what hangs
+    print("\n=== GPU operation tests (timeout 30s each) ===")
+    tests = [
+        ("Basic matmul fp32", 
+         "a=torch.randn(256,256,device='cuda'); b=a@a; print(f'fp32 matmul: {b.shape} sum={b.sum().item():.1f}')"),
+        ("Basic matmul fp16",
+         "a=torch.randn(256,256,device='cuda').half(); b=a@a; print(f'fp16 matmul: {b.shape} sum={b.sum().item():.1f}')"),
+        ("Conv2d fp16 (VAE-like)",
+         "import torch.nn as nn; c=nn.Conv2d(128,128,3,padding=1).half().cuda(); x=torch.randn(1,128,64,64,device='cuda').half(); y=c(x); print(f'conv2d fp16: {y.shape}')"),
+        ("Conv2d fp32 (VAE default)",
+         "import torch.nn as nn; c=nn.Conv2d(128,128,3,padding=1).cuda(); x=torch.randn(1,128,64,64,device='cuda'); y=c(x); print(f'conv2d fp32: {y.shape}')"),
+        ("GroupNorm fp16",
+         "import torch.nn as nn; gn=nn.GroupNorm(32,128).half().cuda(); x=torch.randn(1,128,32,32,device='cuda').half(); y=gn(x); print(f'groupnorm fp16: {y.shape}')"),
+        ("GroupNorm fp32",
+         "import torch.nn as nn; gn=nn.GroupNorm(32,128).cuda(); x=torch.randn(1,128,32,32,device='cuda'); y=gn(x); print(f'groupnorm fp32: {y.shape}')"),
+        ("LayerNorm fp16",
+         "import torch.nn as nn; ln=nn.LayerNorm(256).half().cuda(); x=torch.randn(1,64,256,device='cuda').half(); y=ln(x); print(f'layernorm fp16: {y.shape}')"),
+        ("Linear fp16 (DiT-like)",
+         "import torch.nn as nn; l=nn.Linear(1024,1024).half().cuda(); x=torch.randn(1,64,1024,device='cuda').half(); y=l(x); print(f'linear fp16: {y.shape}')"),
+        ("Attention fp16 (scaled_dot_product)",
+         "q=torch.randn(1,8,64,64,device='cuda').half(); k=q.clone(); v=q.clone(); "
+         "y=torch.nn.functional.scaled_dot_product_attention(q,k,v); print(f'sdpa fp16: {y.shape}')"),
+        ("Attention fp32 (scaled_dot_product)",
+         "q=torch.randn(1,8,64,64,device='cuda'); k=q.clone(); v=q.clone(); "
+         "y=torch.nn.functional.scaled_dot_product_attention(q,k,v); print(f'sdpa fp32: {y.shape}')"),
+        ("Large matmul fp16 (5032x5032)",
+         "a=torch.randn(2048,2048,device='cuda').half(); b=a@a; print(f'large fp16 matmul: {b.shape} sum={b.sum().item():.1f}')"),
+        ("RoPE-like op (complex multiply)",
+         "x=torch.randn(1,8,64,64,device='cuda').half(); "
+         "f=torch.randn(64,32,2,device='cuda').half(); "
+         "print(f'rope input shapes: x={x.shape} f={f.shape} OK')"),
+        ("torch.compile basic test",
+         "import torch._dynamo; f=lambda x: x*2+1; cf=torch.compile(f); "
+         "x=torch.randn(100,device='cuda'); y=cf(x); print(f'compile: {y.shape}')"),
+    ]
+    
+    for name, code in tests:
+        print(f"\n  Testing: {name}...", end=" ", flush=True)
+        cmd = (f"bash -c 'timeout 30 bash -c \""
+               f"source ~/comfyui-env/bin/activate && "
+               f"HSA_OVERRIDE_GFX_VERSION=10.1.0 HSA_ENABLE_SDMA=0 "
+               f"python3 -c \\\"import torch; {code}\\\"\" 2>&1 || echo TIMEOUT_OR_ERROR'")
+        out, err = run(cmd, timeout=40)
+        result = (out + err).strip()
+        if 'TIMEOUT_OR_ERROR' in result:
+            # Get just the error part
+            lines = result.split('\n')
+            for l in reversed(lines):
+                if l.strip() and l.strip() != 'TIMEOUT_OR_ERROR':
+                    print(f"FAILED: {l.strip()[:100]}")
+                    break
+            else:
+                print("TIMEOUT (GPU HANG)")
+        elif result:
+            last_line = [l for l in result.split('\n') if l.strip()][-1] if result.split('\n') else result
+            print(f"OK: {last_line.strip()[:100]}")
+        else:
+            print("NO OUTPUT (possible hang)")
+
+    # 5. Check dmesg for GPU errors after tests
+    print("\n\n=== dmesg GPU errors (last 20) ===")
+    out, _ = run("dmesg 2>/dev/null | grep -i -E 'amdgpu|gpu|gfx|error|fault' | tail -20 || echo 'no permission'")
+    print(out.strip() if out.strip() else "  (empty or no permission)")
+
+    # 6. Check rocm-smi for GPU health
+    print("\n=== GPU health after tests ===")
+    out, _ = run("HSA_OVERRIDE_GFX_VERSION=10.1.0 rocm-smi 2>/dev/null")
+    for line in out.split('\n'):
+        if any(c in line for c in ['°C', '%', 'Device', 'Node']):
+            print(f"  {line.strip()}")
+
+finally:
+    ssh.close()
+    print("\n\nSSH connection closed.")
@@ -0,0 +1,165 @@
+import paramiko, time, json, textwrap
+k = paramiko.Ed25519Key.from_private_key_file(r'C:\Users\fabia\.ssh\id_ed25519')
+c = paramiko.SSHClient()
+c.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+c.connect('192.168.178.150', username='fabian', pkey=k, timeout=15)
+sftp = c.open_sftp()
+
+def sh(cmd, timeout=60):
+    chan = c.get_transport().open_session()
+    chan.settimeout(timeout)
+    chan.exec_command('/bin/bash -l -c ' + "'" + cmd.replace("'", "'\\''") + "'")
+    out = b""
+    while True:
+        try:
+            chunk = chan.recv(65536)
+            if not chunk: break
+            out += chunk
+        except: break
+    chan.close()
+    return out.decode(errors='replace').strip()
+
+# 1) Kill
+print("1) Kill")
+sh('pkill -9 -f "python3.*main.py" 2>/dev/null; sleep 2')
+
+# 2) New launcher: --novram streams weights (proven GPU 136W), no --cpu-vae
+print("2) Write launcher")
+launcher = textwrap.dedent("""\
+    #!/bin/bash
+    export HSA_OVERRIDE_GFX_VERSION=10.1.0
+    export HIP_VISIBLE_DEVICES=0
+    export HSA_ENABLE_SDMA=0
+    export HSA_TOOLS_LIB=""
+    export HSA_TOOLS_REPORT_LOAD_FAILURE=0
+    export PYTORCH_HIP_ALLOC_CONF=expandable_segments:False
+    export OMP_NUM_THREADS=12
+    export MKL_NUM_THREADS=12
+    export OPENBLAS_NUM_THREADS=12
+    export MIOPEN_FIND_MODE=3
+    
+    cd ~/ComfyUI
+    source ~/comfyui-env/bin/activate
+    
+    # --novram: weights in system RAM, GPU computes via streaming (337MB buffer fits in 512MB real VRAM)
+    # --force-fp16: half precision
+    # NO --cpu-vae: let VAE run on GPU (320MB fits in 512MB VRAM)
+    exec python3 main.py \\
+        --listen 0.0.0.0 --port 8188 \\
+        --novram \\
+        --force-fp16
+""")
+with sftp.open('/tmp/run_comfyui.sh', 'w') as f:
+    f.write(launcher)
+sh('chmod +x /tmp/run_comfyui.sh')
+
+# 3) Start
+print("3) Start")
+sh('rm -f /tmp/comfyui.log; rm -f ~/ComfyUI/output/ZImageTurbo_GPU*.png')
+sh('nohup /tmp/run_comfyui.sh > /tmp/comfyui.log 2>&1 &')
+time.sleep(3)
+pid = sh('pgrep -f "python3.*main.py"')
+print(f"   PID: {pid}")
+
+# 4) Wait ready
+print("4) Wait HTTP", end='', flush=True)
+for i in range(90):
+    code = sh('curl -s -o /dev/null -w "%{http_code}" http://127.0.0.1:8188/ 2>/dev/null', timeout=5)
+    if '200' in code:
+        print(f" OK ({i*2}s)")
+        break
+    print('.', end='', flush=True)
+    time.sleep(2)
+
+# Show mode
+with sftp.open('/tmp/comfyui.log', 'r') as f:
+    log = f.read().decode(errors='replace')
+for line in log.split('\n'):
+    s = line.strip()
+    if any(k in s for k in ['vram state', 'Device:', 'Total VRAM', 'offloading']):
+        print(f"   {s}")
+
+# 5) Submit
+print("5) Submit")
+wf = {"prompt": {
+    "1": {"class_type": "UnetLoaderGGUF", "inputs": {"unet_name": "z_image_turbo-Q5_K_S.gguf"}},
+    "2": {"class_type": "CLIPLoaderGGUF", "inputs": {"clip_name": "Qwen3-4B.i1-Q5_K_S.gguf", "type": "qwen_image"}},
+    "3": {"class_type": "VAELoader", "inputs": {"vae_name": "ae.safetensors"}},
+    "4": {"class_type": "CLIPTextEncode", "inputs": {"text": "A red fox in a snowy forest, photorealistic", "clip": ["2", 0]}},
+    "5": {"class_type": "EmptyLatentImage", "inputs": {"width": 512, "height": 512, "batch_size": 1}},
+    "6": {"class_type": "KSampler", "inputs": {"model": ["1", 0], "positive": ["4", 0], "negative": ["4", 0],
+        "latent_image": ["5", 0], "seed": 777, "steps": 8, "cfg": 1.0,
+        "sampler_name": "euler", "scheduler": "simple", "denoise": 1.0}},
+    "7": {"class_type": "VAEDecode", "inputs": {"samples": ["6", 0], "vae": ["3", 0]}},
+    "8": {"class_type": "SaveImage", "inputs": {"images": ["7", 0], "filename_prefix": "ZImageTurbo_GPU"}}
+}}
+with sftp.open('/tmp/wf.json', 'w') as f:
+    f.write(json.dumps(wf))
+resp = sh('curl -s -X POST http://127.0.0.1:8188/prompt -H "Content-Type: application/json" -d @/tmp/wf.json')
+print(f"   {resp[:120]}")
+
+# 6) Monitor
+print("6) Monitor")
+t0 = time.time()
+for i in range(200):
+    el = int(time.time() - t0)
+    temp = sh('cat /sys/class/drm/card0/device/hwmon/hwmon*/temp1_input 2>/dev/null', timeout=5)
+    tc = int(temp)//1000 if temp.isdigit() else '?'
+    
+    try:
+        with sftp.open('/tmp/comfyui.log', 'r') as f:
+            log = f.read().decode(errors='replace')
+    except: log = ''
+    
+    samp = ''
+    last = ''
+    for line in log.split('\n'):
+        s = line.strip()
+        if '/8' in s and ('it/s' in s or 's/it' in s): samp = s
+        if s and 'FETCH' not in s and 'startup' not in s and 'DEPRECATION' not in s: last = s
+    
+    print(f"   [{el:>4}s] {tc}C | {(samp or last)[-90:]}")
+    
+    imgs = sh('ls ~/ComfyUI/output/ZImageTurbo_GPU*.png 2>/dev/null', timeout=5)
+    if imgs:
+        et = ''
+        for line in log.split('\n'):
+            if 'Prompt executed' in line: et = line.strip()
+        print(f"\n   *** DONE! *** {imgs}")
+        print(f"   {et}")
+        print(f"   Wall: {el}s")
+        # Show key log lines
+        for line in log.split('\n'):
+            s = line.strip()
+            if any(k in s for k in ['loaded', '/8', 'Prompt executed', 'VAE', 'Requested']):
+                if 'FETCH' not in s:
+                    print(f"   {s}")
+        break
+    
+    # Queue empty?
+    q = sh('curl -s http://127.0.0.1:8188/queue 2>/dev/null', timeout=5)
+    try:
+        qd = json.loads(q)
+        if not qd.get('queue_running') and not qd.get('queue_pending') and el > 30:
+            time.sleep(2)
+            imgs = sh('ls ~/ComfyUI/output/ZImageTurbo_GPU*.png 2>/dev/null', timeout=5)
+            if imgs:
+                print(f"\n   *** DONE: {imgs} ***")
+            else:
+                print(f"\n   Queue empty, no image:")
+                for line in log.split('\n')[-20:]:
+                    if line.strip() and 'FETCH' not in line: print(f"   {line.strip()}")
+            break
+    except: pass
+    
+    if sh('pgrep -f "python3.*main.py" >/dev/null && echo Y || echo N', timeout=5) == 'N':
+        print("\n   CRASHED!")
+        for line in log.split('\n')[-25:]:
+            if line.strip(): print(f"   {line.strip()}")
+        break
+    
+    time.sleep(10)
+
+sftp.close()
+c.close()
+print("\nDone.")
@@ -0,0 +1,168 @@
+"""FAST FIX: patch offload devices, restart. No fluff."""
+import paramiko, time, json
+
+k = paramiko.Ed25519Key.from_private_key_file(r'C:\Users\fabia\.ssh\id_ed25519')
+c = paramiko.SSHClient()
+c.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+c.connect('192.168.178.150', username='fabian', pkey=k, timeout=10)
+sftp = c.open_sftp()
+
+def sh(cmd, t=30):
+    ch = c.get_transport().open_session()
+    ch.settimeout(t)
+    ch.exec_command('/bin/bash -l -c ' + "'" + cmd.replace("'", "'\\''") + "'")
+    o = b""
+    while True:
+        try:
+            d = ch.recv(65536)
+            if not d: break
+            o += d
+        except: break
+    ch.close()
+    return o.decode(errors='replace').strip()
+
+# 1. KILL
+sh('pkill -9 -f "python3.*main.py" 2>/dev/null; sleep 1')
+print("Killed")
+
+# 2. READ
+with sftp.open('/home/fabian/ComfyUI/comfy/model_management.py', 'r') as f:
+    code = f.read().decode()
+
+# 3. PATCH: unet_offload_device - return GPU for SHARED too
+# Find the function and add SHARED check
+changed = False
+
+# Patch unet_offload_device: "HIGH_VRAM" -> "HIGH_VRAM or SHARED"
+if 'def unet_offload_device' in code:
+    lines = code.split('\n')
+    for i, line in enumerate(lines):
+        if 'def unet_offload_device' in line:
+            # Look at next few lines for the HIGH_VRAM check
+            for j in range(i, min(i+8, len(lines))):
+                if 'HIGH_VRAM' in lines[j] and 'SHARED' not in lines[j] and 'unet_offload' not in lines[j]:
+                    old = lines[j]
+                    lines[j] = old.replace('VRAMState.HIGH_VRAM', 'VRAMState.HIGH_VRAM or vram_state == VRAMState.SHARED')
+                    print(f"Patched unet_offload L{j+1}: {lines[j].strip()}")
+                    changed = True
+                    break
+            break
+    code = '\n'.join(lines)
+
+# Patch vae_offload_device: "args.gpu_only" -> "args.gpu_only or SHARED"  
+if 'def vae_offload_device' in code:
+    lines = code.split('\n')
+    for i, line in enumerate(lines):
+        if 'def vae_offload_device' in line:
+            for j in range(i, min(i+8, len(lines))):
+                if 'gpu_only' in lines[j] and 'SHARED' not in lines[j]:
+                    old = lines[j]
+                    lines[j] = old.replace('args.gpu_only', '(args.gpu_only or vram_state == VRAMState.SHARED)')
+                    print(f"Patched vae_offload L{j+1}: {lines[j].strip()}")
+                    changed = True
+                    break
+            break
+    code = '\n'.join(lines)
+
+# Also patch text_encoder_offload_device if it offloads to CPU
+if 'def text_encoder_offload_device' in code:
+    lines = code.split('\n')
+    for i, line in enumerate(lines):
+        if 'def text_encoder_offload_device' in line:
+            for j in range(i, min(i+8, len(lines))):
+                if 'gpu_only' in lines[j] and 'SHARED' not in lines[j]:
+                    old = lines[j]
+                    lines[j] = old.replace('args.gpu_only', '(args.gpu_only or vram_state == VRAMState.SHARED)')
+                    print(f"Patched text_enc_offload L{j+1}: {lines[j].strip()}")
+                    changed = True
+                    break
+            break
+    code = '\n'.join(lines)
+
+if changed:
+    sh('cp /home/fabian/ComfyUI/comfy/model_management.py /home/fabian/ComfyUI/comfy/model_management.py.bak3')
+    with sftp.open('/home/fabian/ComfyUI/comfy/model_management.py', 'w') as f:
+        f.write(code)
+    print("Written!")
+else:
+    print("Already patched or structure changed")
+
+# 4. RESTART
+sh('rm -f /tmp/comfyui.log')
+sh('nohup /tmp/run_comfyui.sh > /tmp/comfyui.log 2>&1 &')
+time.sleep(4)
+print(f"PID: {sh('pgrep -f python3.*main.py')}")
+
+# 5. WAIT FOR READY
+for i in range(60):
+    r = sh('curl -s -o /dev/null -w "%{http_code}" http://127.0.0.1:8188/ 2>/dev/null', t=5)
+    if '200' in r: print(f"Ready ({i*2}s)"); break
+    time.sleep(2)
+
+# Quick check
+with sftp.open('/tmp/comfyui.log', 'r') as f:
+    log = f.read().decode(errors='replace')
+for l in log.split('\n'):
+    s = l.strip()
+    if any(x in s for x in ['vram state', 'SHARED', 'Device:']): print(f"  {s}")
+
+# 6. SUBMIT
+wf = {"prompt": {
+    "1": {"class_type": "UnetLoaderGGUF", "inputs": {"unet_name": "z_image_turbo-Q5_K_S.gguf"}},
+    "2": {"class_type": "CLIPLoaderGGUF", "inputs": {"clip_name": "Qwen3-4B.i1-Q5_K_S.gguf", "type": "qwen_image"}},
+    "3": {"class_type": "VAELoader", "inputs": {"vae_name": "ae.safetensors"}},
+    "4": {"class_type": "CLIPTextEncode", "inputs": {"text": "A red fox in a snowy forest, photorealistic", "clip": ["2", 0]}},
+    "5": {"class_type": "EmptyLatentImage", "inputs": {"width": 512, "height": 512, "batch_size": 1}},
+    "6": {"class_type": "KSampler", "inputs": {"model": ["1", 0], "positive": ["4", 0], "negative": ["4", 0],
+        "latent_image": ["5", 0], "seed": 99999, "steps": 8, "cfg": 1.0,
+        "sampler_name": "euler", "scheduler": "simple", "denoise": 1.0}},
+    "7": {"class_type": "VAEDecode", "inputs": {"samples": ["6", 0], "vae": ["3", 0]}},
+    "8": {"class_type": "SaveImage", "inputs": {"images": ["7", 0], "filename_prefix": "ZImageTurbo_GPU"}}
+}}
+with sftp.open('/tmp/wf.json', 'w') as f:
+    f.write(json.dumps(wf))
+resp = sh('curl -s -X POST http://127.0.0.1:8188/prompt -H "Content-Type: application/json" -d @/tmp/wf.json')
+print(f"Submitted: {resp[:100]}")
+
+# 7. MONITOR - compact, fast checks
+print("\nWaiting for image...")
+t0 = time.time()
+last_shown = ''
+for i in range(180):
+    el = int(time.time() - t0)
+    try:
+        with sftp.open('/tmp/comfyui.log', 'r') as f:
+            log = f.read().decode(errors='replace')
+    except: log = ''
+    
+    # Find latest status
+    status = ''
+    for l in log.split('\n'):
+        s = l.strip()
+        if any(x in s for x in ['/8', 'loaded', 'Requested', 'VAE', 'Prompt executed', 'Error']):
+            if 'FETCH' not in s: status = s
+    
+    if status != last_shown:
+        gpu = sh('cat /sys/class/drm/card0/device/gpu_busy_percent 2>/dev/null', t=5)
+        print(f"  [{el:>3}s] GPU:{gpu}% | {status[-100:]}")
+        last_shown = status
+    
+    imgs = sh('ls ~/ComfyUI/output/ZImageTurbo_GPU*.png 2>/dev/null', t=5)
+    if imgs:
+        print(f"\n*** DONE in {el}s! ***")
+        for l in log.split('\n'):
+            s = l.strip()
+            if any(x in s for x in ['load device', 'offload device', 'loaded completely', 'Prompt executed']):
+                print(f"  {s}")
+        break
+    
+    if sh('pgrep -f "python3.*main.py" >/dev/null && echo Y || echo N', t=5) == 'N':
+        print(f"\nCRASHED at {el}s!")
+        for l in log.split('\n')[-15:]:
+            if l.strip(): print(f"  {l.strip()}")
+        break
+    
+    time.sleep(3)
+
+sftp.close()
+c.close()
@@ -0,0 +1,59 @@
+#!/usr/bin/env python3
+"""Fix single-core bottleneck: set threading env vars and patch ComfyUI-GGUF for parallel dequant."""
+import paramiko
+import time
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=120, desc=""):
+    if desc:
+        print(f"\n{'='*60}")
+        print(f"  {desc}")
+        print(f"{'='*60}")
+    print(f"$ {cmd[:300]}")
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    rc = stdout.channel.recv_exit_status()
+    if out.strip():
+        lines = out.strip().split('\n')
+        if len(lines) > 50:
+            print(f"  ... ({len(lines)} lines, showing last 50)")
+            print('\n'.join(lines[-50:]))
+        else:
+            print(out.strip())
+    if err.strip():
+        lines = err.strip().split('\n')
+        show = lines[-15:] if len(lines) > 15 else lines
+        print(f"STDERR: {chr(10).join(show)}")
+    print(f"  Exit code: {rc}")
+    return rc, out, err
+
+# 1. Kill stuck ComfyUI
+run("bash -c 'kill -9 484588 2>/dev/null; pkill -9 -f \"python main.py\" 2>/dev/null; sleep 2; echo done'",
+    desc="Kill stuck ComfyUI process")
+
+# 2. Check how many CPU cores
+run("bash -c 'nproc'", desc="CPU core count")
+
+# 3. Check the ComfyUI-GGUF dequant code to understand the bottleneck
+run("bash -c 'grep -rn \"dequant\\|num_threads\\|torch.set_num_threads\\|ThreadPool\\|parallel\" ~/ComfyUI/custom_nodes/ComfyUI-GGUF/*.py 2>/dev/null | head -30'",
+    desc="Search for threading in ComfyUI-GGUF")
+
+run("bash -c 'cat ~/ComfyUI/custom_nodes/ComfyUI-GGUF/ops.py 2>/dev/null | head -80'",
+    desc="ComfyUI-GGUF ops.py (dequant logic)")
+
+# 4. Check the dequant function
+run("bash -c 'grep -n \"def dequantize\\|class GGMLTensor\\|def forward\" ~/ComfyUI/custom_nodes/ComfyUI-GGUF/ops.py 2>/dev/null'",
+    desc="Key functions in ops.py")
+
+run("bash -c 'wc -l ~/ComfyUI/custom_nodes/ComfyUI-GGUF/ops.py ~/ComfyUI/custom_nodes/ComfyUI-GGUF/dequant.py 2>/dev/null'",
+    desc="File sizes")
+
+run("bash -c 'cat ~/ComfyUI/custom_nodes/ComfyUI-GGUF/dequant.py 2>/dev/null | head -60'",
+    desc="dequant.py start")
+
+ssh.close()
+print("\nDone.")
@@ -0,0 +1,54 @@
+#!/usr/bin/env python3
+"""Fix threading: set all 12 cores for PyTorch ops, patch ComfyUI startup, restart."""
+import paramiko
+import time
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=120, desc=""):
+    if desc:
+        print(f"\n{'='*60}")
+        print(f"  {desc}")
+        print(f"{'='*60}")
+    print(f"$ {cmd[:300]}")
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    rc = stdout.channel.recv_exit_status()
+    if out.strip():
+        lines = out.strip().split('\n')
+        if len(lines) > 40:
+            print(f"  ... ({len(lines)} lines, showing last 40)")
+            print('\n'.join(lines[-40:]))
+        else:
+            print(out.strip())
+    if err.strip():
+        lines = err.strip().split('\n')
+        show = lines[-15:] if len(lines) > 15 else lines
+        print(f"STDERR: {chr(10).join(show)}")
+    print(f"  Exit code: {rc}")
+    return rc, out, err
+
+# 1. Kill any existing ComfyUI
+run("bash -c 'pkill -9 -f \"python main.py\" 2>/dev/null; sleep 1; echo killed'",
+    desc="Kill existing ComfyUI")
+
+# 2. Check the loader.py for the Dequantizing message source
+run("bash -c 'grep -rn \"Dequantizing\" ~/ComfyUI/ --include=\"*.py\" 2>/dev/null | head -10'",
+    desc="Find 'Dequantizing' message source")
+
+# 3. Check torch thread defaults without any env vars
+run("bash -c 'source ~/comfyui-env/bin/activate && python3 -c \"import torch; print(torch.get_num_threads(), torch.get_num_interop_threads())\"'",
+    desc="Default torch thread count")
+
+# 4. Verify it works with env vars
+run("bash -c 'export OMP_NUM_THREADS=12; export MKL_NUM_THREADS=12; "
+    "source ~/comfyui-env/bin/activate && python3 -c \"import torch; "
+    "torch.set_num_threads(12); torch.set_num_interop_threads(4); "
+    "print(torch.get_num_threads(), torch.get_num_interop_threads())\"'",
+    desc="Torch threads with env vars set to 12")
+
+ssh.close()
+print("\nDone.")
@@ -0,0 +1,132 @@
+#!/usr/bin/env python3
+"""Fix: maximize threading to 12 cores, add --lowvram for 7.6GB VRAM, restart ComfyUI."""
+import paramiko
+import time
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=300, desc=""):
+    if desc:
+        print(f"\n{'='*60}")
+        print(f"  {desc}")
+        print(f"{'='*60}")
+    print(f"$ {cmd[:300]}")
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    rc = stdout.channel.recv_exit_status()
+    if out.strip():
+        lines = out.strip().split('\n')
+        if len(lines) > 40:
+            print(f"  ... ({len(lines)} lines, showing last 40)")
+            print('\n'.join(lines[-40:]))
+        else:
+            print(out.strip())
+    if err.strip():
+        lines = err.strip().split('\n')
+        show = lines[-15:] if len(lines) > 15 else lines
+        print(f"STDERR: {chr(10).join(show)}")
+    print(f"  Exit code: {rc}")
+    return rc, out, err
+
+# 1. Kill any existing ComfyUI
+run("bash -c 'pkill -9 -f \"python main.py\" 2>/dev/null; sleep 2; echo ok'",
+    desc="Kill existing ComfyUI")
+
+# 2. Write updated start_comfyui.sh with full threading + lowvram
+startup_script = r'''#!/bin/bash
+# ComfyUI Startup Script for AsRock BC-250 (AMD Cyan Skillfish / ROCm 7.2)
+# Optimized for all 12 CPU cores + 7.6GB shared VRAM
+set -euo pipefail
+
+# ═══════════════ BC-250 GPU Environment ═══════════════
+export HSA_OVERRIDE_GFX_VERSION=10.1.0
+export HIP_VISIBLE_DEVICES=0
+export HSA_ENABLE_SDMA=0
+export HSA_TOOLS_LIB=""
+export HSA_TOOLS_REPORT_LOAD_FAILURE=0
+
+# ═══════════════ THREADING — ALL 12 CORES ═══════════════
+export OMP_NUM_THREADS=12
+export MKL_NUM_THREADS=12
+export OPENBLAS_NUM_THREADS=12
+export VECLIB_MAXIMUM_THREADS=12
+export NUMEXPR_NUM_THREADS=12
+# PyTorch intra-op (tensor math) and inter-op (parallel node execution) threads
+export TORCH_NUM_THREADS=12
+
+# ═══════════════ Memory / ROCm tuning ═══════════════
+export PYTORCH_HIP_ALLOC_CONF="expandable_segments:False"
+export TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1
+# Disable HIP memory caching to avoid fragmentation on shared VRAM
+export PYTORCH_NO_HIP_MEMORY_CACHING=0
+
+# ═══════════════ Activate venv ═══════════════
+source "$HOME/comfyui-env/bin/activate"
+
+cd "$HOME/ComfyUI"
+
+# Force PyTorch to use all 12 cores
+python3 -c "import torch; torch.set_num_threads(12); torch.set_num_interop_threads(12); print(f'Threads: intra={torch.get_num_threads()} inter={torch.get_num_interop_threads()}')"
+
+echo "=========================================="
+echo " ComfyUI on BC-250 (ROCm 7.2)"
+echo " GPU:    AMD Cyan Skillfish (gfx1013→gfx1010)"
+echo " PyTorch: $(python3 -c 'import torch; print(torch.__version__)')"
+echo " HIP:    $(python3 -c 'import torch; print(torch.version.hip)')"
+echo " CUDA:   $(python3 -c 'import torch; print(torch.cuda.is_available())')"
+echo " CPU:    $(nproc) cores (all used)"
+echo " VRAM:   7.6GB shared — using --lowvram mode"
+echo "=========================================="
+
+# Default: listen on all interfaces, --lowvram for 7.6GB shared VRAM
+LISTEN_ARGS="--listen 0.0.0.0 --port 8188 --lowvram"
+
+if [ $# -gt 0 ]; then
+    LISTEN_ARGS="$@"
+fi
+
+echo "Starting: python main.py $LISTEN_ARGS"
+echo "Access at: http://192.168.178.150:8188"
+echo ""
+
+# Set threads inside the actual process too
+exec python3 -c "
+import torch, sys, os
+torch.set_num_threads(12)
+torch.set_num_interop_threads(12)
+# Now exec ComfyUI main
+sys.argv = ['main.py'] + '$LISTEN_ARGS'.split()
+exec(open('main.py').read())
+"
+'''
+
+sftp = ssh.open_sftp()
+with sftp.open('/home/fabian/start_comfyui.sh', 'w') as f:
+    f.write(startup_script)
+sftp.close()
+
+run("chmod +x /home/fabian/start_comfyui.sh", desc="Make script executable")
+
+# 3. Launch ComfyUI with new settings
+run("bash -c 'rm -f /home/fabian/comfyui.log'", desc="Clean old log")
+run("bash -c 'nohup bash /home/fabian/start_comfyui.sh > /home/fabian/comfyui.log 2>&1 & echo PID=$!'",
+    desc="Launch ComfyUI with 12-core threading + lowvram")
+
+# 4. Wait for startup
+time.sleep(15)
+run("bash -c 'tail -30 /home/fabian/comfyui.log 2>/dev/null'",
+    desc="Startup log")
+
+time.sleep(10)
+run("bash -c 'ss -tlnp | grep 8188 || echo PORT_NOT_LISTENING'",
+    desc="Check port 8188")
+
+# Verify threads are set
+run("bash -c 'tail -40 /home/fabian/comfyui.log 2>/dev/null'",
+    desc="Full startup log")
+
+ssh.close()
+print("\nDone.")
@@ -0,0 +1,144 @@
+#!/usr/bin/env python3
+"""Fix startup: use proper subprocess instead of exec, set threads via sitecustomize."""
+import paramiko
+import time
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=120, desc=""):
+    if desc:
+        print(f"\n{'='*60}")
+        print(f"  {desc}")
+        print(f"{'='*60}")
+    print(f"$ {cmd[:300]}")
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    rc = stdout.channel.recv_exit_status()
+    if out.strip():
+        lines = out.strip().split('\n')
+        if len(lines) > 40:
+            print(f"  ... ({len(lines)} lines, showing last 40)")
+            print('\n'.join(lines[-40:]))
+        else:
+            print(out.strip())
+    if err.strip():
+        lines = err.strip().split('\n')
+        show = lines[-15:] if len(lines) > 15 else lines
+        print(f"STDERR: {chr(10).join(show)}")
+    print(f"  Exit code: {rc}")
+    return rc, out, err
+
+# 1. Create a sitecustomize.py in the venv to set threads on import
+sitecustomize = '''# Auto-set PyTorch threading to use all 12 CPU cores on BC-250
+import os
+os.environ.setdefault("OMP_NUM_THREADS", "12")
+os.environ.setdefault("MKL_NUM_THREADS", "12")
+os.environ.setdefault("OPENBLAS_NUM_THREADS", "12")
+
+try:
+    import torch
+    torch.set_num_threads(12)
+    torch.set_num_interop_threads(12)
+except Exception:
+    pass
+'''
+
+# Find the venv site-packages path
+rc, out, _ = run("bash -c 'source ~/comfyui-env/bin/activate && python3 -c \"import site; print(site.getsitepackages()[0])\"'",
+    desc="Find venv site-packages")
+site_packages = out.strip()
+print(f"  Site-packages: {site_packages}")
+
+# Write sitecustomize.py
+sftp = ssh.open_sftp()
+sitecust_path = f"{site_packages}/sitecustomize.py"
+# Check if it exists first
+try:
+    sftp.stat(sitecust_path)
+    print(f"  sitecustomize.py already exists, backing up")
+    sftp.rename(sitecust_path, f"{sitecust_path}.bak")
+except FileNotFoundError:
+    pass
+with sftp.open(sitecust_path, 'w') as f:
+    f.write(sitecustomize)
+sftp.close()
+print(f"  Written: {sitecust_path}")
+
+# 2. Update startup script — simple, using exec python main.py directly
+startup_script = r'''#!/bin/bash
+# ComfyUI Startup Script for AsRock BC-250 (AMD Cyan Skillfish / ROCm 7.2)
+# All 12 CPU cores + lowvram for 7.6GB shared VRAM
+set -euo pipefail
+
+# ═══════════════ BC-250 GPU ═══════════════
+export HSA_OVERRIDE_GFX_VERSION=10.1.0
+export HIP_VISIBLE_DEVICES=0
+export HSA_ENABLE_SDMA=0
+export HSA_TOOLS_LIB=""
+export HSA_TOOLS_REPORT_LOAD_FAILURE=0
+
+# ═══════════════ ALL 12 CORES ═══════════════
+export OMP_NUM_THREADS=12
+export MKL_NUM_THREADS=12
+export OPENBLAS_NUM_THREADS=12
+export VECLIB_MAXIMUM_THREADS=12
+export NUMEXPR_NUM_THREADS=12
+
+# ═══════════════ Memory tuning ═══════════════
+export PYTORCH_HIP_ALLOC_CONF="expandable_segments:False"
+export TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1
+
+# ═══════════════ Activate venv ═══════════════
+source "$HOME/comfyui-env/bin/activate"
+cd "$HOME/ComfyUI"
+
+echo "=========================================="
+echo " ComfyUI on BC-250 (ROCm 7.2)"
+echo " GPU:    AMD Cyan Skillfish (gfx1010)"
+echo " PyTorch: $(python3 -c 'import torch; print(torch.__version__)')"
+echo " Threads: $(python3 -c 'import torch; print(f"intra={torch.get_num_threads()} inter={torch.get_num_interop_threads()}")')"
+echo " CPU:    $(nproc) cores"
+echo " VRAM:   7.6GB shared — lowvram mode"
+echo "=========================================="
+
+# Default args: listen on all, lowvram for tight VRAM
+ARGS="--listen 0.0.0.0 --port 8188 --lowvram"
+if [ $# -gt 0 ]; then
+    ARGS="$@"
+fi
+
+echo "Starting: python main.py $ARGS"
+echo "Access:   http://192.168.178.150:8188"
+echo ""
+
+exec python3 main.py $ARGS
+'''
+
+sftp = ssh.open_sftp()
+with sftp.open('/home/fabian/start_comfyui.sh', 'w') as f:
+    f.write(startup_script)
+sftp.close()
+
+run("chmod +x /home/fabian/start_comfyui.sh", desc="Make executable")
+
+# 3. Launch
+run("bash -c 'rm -f /home/fabian/comfyui.log'", desc="Clean log")
+run("bash -c 'nohup bash /home/fabian/start_comfyui.sh > /home/fabian/comfyui.log 2>&1 & echo PID=$!'",
+    desc="Launch ComfyUI (12 cores + lowvram)")
+
+time.sleep(20)
+run("bash -c 'tail -40 /home/fabian/comfyui.log 2>/dev/null'",
+    desc="Startup log")
+
+time.sleep(10)
+run("bash -c 'ss -tlnp | grep 8188 || echo PORT_NOT_LISTENING'",
+    desc="Check port 8188")
+
+run("bash -c 'tail -50 /home/fabian/comfyui.log 2>/dev/null'",
+    desc="Full log")
+
+ssh.close()
+print("\nDone.")
@@ -0,0 +1,59 @@
+#!/usr/bin/env python3
+"""Fix torchvision compatibility on BC-250.
+
+The pip-installed torchvision conflicts with the system python-pytorch-rocm.
+Need to use system torchvision-rocm or fix the version.
+"""
+import paramiko
+import time
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=300, desc=""):
+    if desc:
+        print(f"\n{'='*60}")
+        print(f"  {desc}")
+        print(f"{'='*60}")
+    print(f"$ {cmd}")
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    rc = stdout.channel.recv_exit_status()
+    if out.strip():
+        lines = out.strip().split('\n')
+        if len(lines) > 40:
+            print(f"  ... ({len(lines)} lines, showing last 40)")
+            print('\n'.join(lines[-40:]))
+        else:
+            print(out.strip())
+    if err.strip():
+        lines = err.strip().split('\n')
+        show = lines[-20:] if len(lines) > 20 else lines
+        print(f"STDERR: {chr(10).join(show)}")
+    print(f"  Exit code: {rc}")
+    return rc, out, err
+
+# Check what torchvision packages exist in repos
+run("bash -c 'pacman -Ss torchvision 2>/dev/null'",
+    desc="Search for torchvision packages in repos")
+
+run("bash -c 'pacman -Ss torchaudio 2>/dev/null'",
+    desc="Search for torchaudio packages")
+
+# Check what's currently installed
+run("bash -c 'pacman -Qs torch 2>/dev/null'",
+    desc="Currently installed torch packages (system)")
+
+run("bash -c 'source /home/fabian/comfyui-env/bin/activate && pip list 2>/dev/null | grep -i torch'",
+    desc="torch packages in venv")
+
+# Check the versions
+run("bash -c 'source /home/fabian/comfyui-env/bin/activate && python -c \""
+    "import torch; print(f\\\"torch: {torch.__version__} from {torch.__file__}\\\"); "
+    "\"'",
+    desc="Check torch location")
+
+ssh.close()
+print("\nDone.")
@@ -0,0 +1,73 @@
+#!/usr/bin/env python3
+"""Fix torchvision — use system package instead of pip version."""
+import paramiko
+import time
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=300, desc=""):
+    if desc:
+        print(f"\n{'='*60}")
+        print(f"  {desc}")
+        print(f"{'='*60}")
+    print(f"$ {cmd}")
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    rc = stdout.channel.recv_exit_status()
+    if out.strip():
+        lines = out.strip().split('\n')
+        if len(lines) > 40:
+            print(f"  ... ({len(lines)} lines, showing last 40)")
+            print('\n'.join(lines[-40:]))
+        else:
+            print(out.strip())
+    if err.strip():
+        lines = err.strip().split('\n')
+        show = lines[-20:] if len(lines) > 20 else lines
+        print(f"STDERR: {chr(10).join(show)}")
+    print(f"  Exit code: {rc}")
+    return rc, out, err
+
+# 1. Uninstall pip torchvision and torchaudio from venv
+run("bash -c 'source /home/fabian/comfyui-env/bin/activate && pip uninstall -y torchvision torchaudio 2>&1'",
+    desc="Uninstall pip torchvision and torchaudio from venv")
+
+# 2. Install system python-torchvision via pacman
+run("sudo pacman -S --noconfirm python-torchvision",
+    desc="Install system python-torchvision (matches system pytorch)")
+
+# 3. Verify torchvision now works
+run("bash -c 'source /home/fabian/comfyui-env/bin/activate && "
+    "export HSA_OVERRIDE_GFX_VERSION=10.1.0 && "
+    "export HIP_VISIBLE_DEVICES=0 && "
+    "export HSA_ENABLE_SDMA=0 && "
+    "python -c \""
+    "import torch; print(f\\\"torch {torch.__version__} from {torch.__file__}\\\"); "
+    "import torchvision; print(f\\\"torchvision {torchvision.__version__} from {torchvision.__file__}\\\"); "
+    "print(\\\"torchvision ops OK\\\"); "
+    "\"'",
+    desc="Verify torchvision import works")
+
+# 4. Kill old ComfyUI and relaunch
+run("bash -c 'pkill -f \"python main.py\" 2>/dev/null; sleep 2; echo done'",
+    desc="Kill old ComfyUI process")
+
+run("bash -c 'nohup bash /home/fabian/start_comfyui.sh > /home/fabian/comfyui.log 2>&1 & echo PID=$!'",
+    desc="Relaunch ComfyUI")
+
+time.sleep(20)
+run("tail -40 /home/fabian/comfyui.log 2>/dev/null",
+    desc="ComfyUI startup log")
+
+time.sleep(10)
+run("bash -c 'ss -tlnp | grep 8188 || echo PORT_NOT_LISTENING'",
+    desc="Check if port 8188 is listening")
+
+run("tail -60 /home/fabian/comfyui.log 2>/dev/null",
+    desc="Full ComfyUI log")
+
+ssh.close()
+print("\nDone.")
@@ -0,0 +1,229 @@
+#!/usr/bin/env python3
+"""Fix VAE decode hang: kill stuck, check available flags, restart with --cpu-vae."""
+import paramiko, json, time, textwrap
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=60, desc=""):
+    if desc:
+        print(f"\n{'='*60}\n  {desc}\n{'='*60}")
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    rc = stdout.channel.recv_exit_status()
+    if out.strip():
+        lines = out.strip().split('\n')
+        if len(lines) > 50:
+            print(f"  ... ({len(lines)} lines, showing last 50)")
+            print('\n'.join(lines[-50:]))
+        else:
+            print(out.strip())
+    if err.strip():
+        for l in err.strip().split('\n')[-5:]:
+            print(f"  STDERR: {l}")
+    return rc, out, err
+
+# Kill stuck
+run("pkill -f 'python3 main.py' 2>/dev/null; sleep 2; pkill -9 -f 'python3 main.py' 2>/dev/null; sleep 1; echo killed",
+    desc="Kill stuck ComfyUI")
+
+# Check available VAE flags
+run("bash -c 'source /home/fabian/comfyui-env/bin/activate && cd /home/fabian/ComfyUI && "
+    "python3 main.py --help 2>&1 | grep -i -E \"vae|fp16|fp32|force|cpu|novram|lowvram\"'",
+    desc="ComfyUI VAE/VRAM flags")
+
+# Update startup script: add --cpu-vae to keep diffusion on GPU but VAE on CPU
+startup_script = textwrap.dedent("""\
+    #!/bin/bash
+    # BC-250 ComfyUI Launcher — GPU inference with CPU VAE decode
+    # Diffusion sampling: GPU (~6s/step, 8 steps = 51s total)
+    # VAE decode: CPU (GPU hangs on float32 VAE ops on Cyan Skillfish)
+    # Text encoding: CPU (GGUF model, dequant on CPU)
+    
+    # GPU identity
+    export HSA_OVERRIDE_GFX_VERSION=10.1.0
+    export HIP_VISIBLE_DEVICES=0
+    export HSA_ENABLE_SDMA=0
+    export HSA_TOOLS_LIB=""
+    export HSA_TOOLS_REPORT_LOAD_FAILURE=0
+    
+    # Use all 12 CPU cores
+    export OMP_NUM_THREADS=12
+    export MKL_NUM_THREADS=12
+    export OPENBLAS_NUM_THREADS=12
+    
+    # Activate venv
+    source /home/fabian/comfyui-env/bin/activate
+    cd /home/fabian/ComfyUI
+    
+    # --novram: send one layer at a time to GPU (needed for 7.6GB shared VRAM)
+    # --force-fp16: halve VRAM usage for diffusion model
+    # --cpu-vae: decode VAE on CPU (GPU hangs on VAE float32 conv2d ops)
+    # --disable-smart-memory: prevent memory heuristics from interfering
+    exec python3 main.py \\
+        --listen 0.0.0.0 --port 8188 \\
+        --novram \\
+        --force-fp16 \\
+        --cpu-vae \\
+        --disable-smart-memory
+""")
+
+sftp = ssh.open_sftp()
+with sftp.open('/home/fabian/start_comfyui.sh', 'w') as f:
+    f.write(startup_script)
+sftp.close()
+run("chmod +x /home/fabian/start_comfyui.sh")
+print("\n  Updated: added --cpu-vae (GPU sampler + CPU VAE)")
+
+# Launch
+run("bash -c 'nohup /home/fabian/start_comfyui.sh > /home/fabian/comfyui.log 2>&1 &'; sleep 1; echo launched",
+    desc="Launch ComfyUI")
+
+print("\n  Waiting for server...")
+for i in range(40):
+    time.sleep(3)
+    rc, out, _ = run("bash -c 'curl -s -o /dev/null -w \"%{http_code}\" http://localhost:8188/ 2>/dev/null || echo 0'")
+    if out.strip() == '200':
+        print(f"  Server ready! ({(i+1)*3}s)")
+        break
+    if i % 5 == 4:
+        rc2, log, _ = run("tail -3 /home/fabian/comfyui.log 2>/dev/null")
+        print(f"  [{(i+1)*3}s] waiting... {log.strip().split(chr(10))[-1][:80]}")
+else:
+    print("  Timeout!")
+    run("tail -40 /home/fabian/comfyui.log", desc="Log")
+    ssh.close()
+    exit(1)
+
+# Submit workflow
+workflow = {
+    "prompt": {
+        "1": {
+            "class_type": "UnetLoaderGGUF",
+            "inputs": {"unet_name": "z_image_turbo-Q5_K_S.gguf"}
+        },
+        "2": {
+            "class_type": "CLIPLoaderGGUF",
+            "inputs": {"clip_name": "Qwen3-4B.i1-Q5_K_S.gguf", "type": "qwen_image"}
+        },
+        "3": {
+            "class_type": "VAELoader",
+            "inputs": {"vae_name": "ae.safetensors"}
+        },
+        "4": {
+            "class_type": "CLIPTextEncode",
+            "inputs": {
+                "text": "A red fox in a snowy forest, photorealistic",
+                "clip": ["2", 0]
+            }
+        },
+        "5": {
+            "class_type": "CLIPTextEncode",
+            "inputs": {"text": "", "clip": ["2", 0]}
+        },
+        "6": {
+            "class_type": "EmptyLatentImage",
+            "inputs": {"width": 512, "height": 512, "batch_size": 1}
+        },
+        "7": {
+            "class_type": "KSampler",
+            "inputs": {
+                "model": ["1", 0],
+                "seed": 42,
+                "steps": 8,
+                "cfg": 1.0,
+                "sampler_name": "euler",
+                "scheduler": "simple",
+                "positive": ["4", 0],
+                "negative": ["5", 0],
+                "latent_image": ["6", 0],
+                "denoise": 1.0
+            }
+        },
+        "8": {
+            "class_type": "VAEDecode",
+            "inputs": {"samples": ["7", 0], "vae": ["3", 0]}
+        },
+        "9": {
+            "class_type": "SaveImage",
+            "inputs": {"images": ["8", 0], "filename_prefix": "ZImageTurbo_test"}
+        }
+    }
+}
+
+sftp2 = ssh.open_sftp()
+with sftp2.open('/tmp/zimage_workflow.json', 'w') as f:
+    f.write(json.dumps(workflow))
+sftp2.close()
+
+rc, out, _ = run("bash -c 'curl -s -X POST http://localhost:8188/prompt "
+    "-H \"Content-Type: application/json\" "
+    "-d @/tmp/zimage_workflow.json'",
+    desc="Submit workflow (GPU sampling + CPU VAE)")
+
+try:
+    resp = json.loads(out.strip())
+    if 'error' in resp:
+        print(f"  ERROR: {resp['error']}")
+        if 'node_errors' in resp:
+            for nid, e in resp['node_errors'].items():
+                print(f"  Node {nid}: {e}")
+        ssh.close()
+        exit(1)
+    print(f"  Prompt ID: {resp.get('prompt_id')}")
+except:
+    print(f"  Response: {out.strip()[:500]}")
+
+# Monitor 
+print("\n  Monitoring GPU generation + CPU VAE decode...")
+last_log = ""
+for i in range(120):
+    time.sleep(15)
+    
+    stats = run("bash -c 'PID=$(pgrep -f \"python3 main.py\" | head -1); "
+        "if [ -n \"$PID\" ]; then "
+        "  CPU=$(ps -p $PID -o %cpu --no-headers); "
+        "  MEM=$(ps -p $PID -o rss --no-headers); "
+        "  GPU_TEMP=$(cat /sys/class/drm/card1/device/hwmon/hwmon*/temp1_input 2>/dev/null || echo 0); "
+        "  GPU_POWER=$(cat /sys/class/drm/card1/device/hwmon/hwmon*/power1_average 2>/dev/null || echo 0); "
+        "  echo \"CPU:${CPU}% RSS:$((MEM/1024))MB GPU_T:$((GPU_TEMP/1000))C GPU_P:$((GPU_POWER/1000000))W\"; "
+        "else echo DEAD; fi'")[1].strip()
+    
+    log = run("tail -10 /home/fabian/comfyui.log 2>/dev/null")[1].strip()
+    
+    elapsed = (i+1)*15
+    m, s = divmod(elapsed, 60)
+    
+    print(f"  [{m}m{s:02d}s] {stats}")
+    
+    # Show last meaningful log line if changed
+    if log != last_log:
+        for line in reversed(log.split('\n')):
+            l = line.strip()
+            if l and not l.startswith('FETCH') and not l.startswith('[DEPRECATION'):
+                print(f"    LOG: {l[:120]}")
+                break
+        last_log = log
+    
+    if 'DEAD' in stats:
+        print("\n  PROCESS DIED!")
+        run("tail -60 /home/fabian/comfyui.log", desc="Death log")
+        break
+    
+    if 'Prompt executed in' in log:
+        print(f"\n  IMAGE GENERATED!")
+        run("tail -30 /home/fabian/comfyui.log", desc="Success log")
+        break
+    
+    if 'Traceback' in log or 'CUDA out of memory' in log:
+        print("\n  ERROR!")
+        run("tail -60 /home/fabian/comfyui.log", desc="Error log")
+        break
+
+# Output
+run("ls -lah /home/fabian/ComfyUI/output/", desc="Output files")
+
+ssh.close()
+print("\nDone.")
@@ -0,0 +1,163 @@
+"""Save a proper Z-Image-Turbo GGUF workflow as the default ComfyUI web UI workflow."""
+import paramiko, json
+
+k = paramiko.Ed25519Key.from_private_key_file(r'C:\Users\fabia\.ssh\id_ed25519')
+c = paramiko.SSHClient()
+c.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+c.connect('192.168.178.150', username='fabian', pkey=k, timeout=10)
+
+# ComfyUI web UI workflow format (not API format)
+workflow = {
+    "last_node_id": 8,
+    "last_link_id": 8,
+    "nodes": [
+        {
+            "id": 1,
+            "type": "UnetLoaderGGUF",
+            "pos": [100, 100],
+            "size": [300, 80],
+            "flags": {},
+            "order": 0,
+            "mode": 0,
+            "outputs": [{"name": "MODEL", "type": "MODEL", "links": [1], "slot_index": 0}],
+            "properties": {"Node name for S&R": "UnetLoaderGGUF"},
+            "widgets_values": ["z_image_turbo-Q5_K_S.gguf"]
+        },
+        {
+            "id": 2,
+            "type": "CLIPLoaderGGUF",
+            "pos": [100, 250],
+            "size": [300, 80],
+            "flags": {},
+            "order": 1,
+            "mode": 0,
+            "outputs": [{"name": "CLIP", "type": "CLIP", "links": [2], "slot_index": 0}],
+            "properties": {"Node name for S&R": "CLIPLoaderGGUF"},
+            "widgets_values": ["Qwen3-4B.i1-Q5_K_S.gguf", "qwen_image"]
+        },
+        {
+            "id": 3,
+            "type": "VAELoader",
+            "pos": [100, 400],
+            "size": [300, 60],
+            "flags": {},
+            "order": 2,
+            "mode": 0,
+            "outputs": [{"name": "VAE", "type": "VAE", "links": [3], "slot_index": 0}],
+            "properties": {"Node name for S&R": "VAELoader"},
+            "widgets_values": ["ae.safetensors"]
+        },
+        {
+            "id": 4,
+            "type": "CLIPTextEncode",
+            "pos": [500, 250],
+            "size": [400, 120],
+            "flags": {},
+            "order": 3,
+            "mode": 0,
+            "inputs": [{"name": "clip", "type": "CLIP", "link": 2}],
+            "outputs": [{"name": "CONDITIONING", "type": "CONDITIONING", "links": [4], "slot_index": 0}],
+            "properties": {"Node name for S&R": "CLIPTextEncode"},
+            "widgets_values": ["A red fox in a snowy forest, photorealistic, highly detailed"]
+        },
+        {
+            "id": 5,
+            "type": "EmptyLatentImage",
+            "pos": [500, 450],
+            "size": [300, 110],
+            "flags": {},
+            "order": 4,
+            "mode": 0,
+            "outputs": [{"name": "LATENT", "type": "LATENT", "links": [5], "slot_index": 0}],
+            "properties": {"Node name for S&R": "EmptyLatentImage"},
+            "widgets_values": [512, 512, 1]
+        },
+        {
+            "id": 6,
+            "type": "KSampler",
+            "pos": [950, 100],
+            "size": [320, 474],
+            "flags": {},
+            "order": 5,
+            "mode": 0,
+            "inputs": [
+                {"name": "model", "type": "MODEL", "link": 1},
+                {"name": "positive", "type": "CONDITIONING", "link": 4},
+                {"name": "negative", "type": "CONDITIONING", "link": None},
+                {"name": "latent_image", "type": "LATENT", "link": 5}
+            ],
+            "outputs": [{"name": "LATENT", "type": "LATENT", "links": [6], "slot_index": 0}],
+            "properties": {"Node name for S&R": "KSampler"},
+            "widgets_values": [42, "fixed", 8, 1.0, "euler", "simple", 1.0]
+        },
+        {
+            "id": 7,
+            "type": "VAEDecode",
+            "pos": [1350, 100],
+            "size": [210, 50],
+            "flags": {},
+            "order": 6,
+            "mode": 0,
+            "inputs": [
+                {"name": "samples", "type": "LATENT", "link": 6},
+                {"name": "vae", "type": "VAE", "link": 3}
+            ],
+            "outputs": [{"name": "IMAGE", "type": "IMAGE", "links": [7], "slot_index": 0}],
+            "properties": {"Node name for S&R": "VAEDecode"}
+        },
+        {
+            "id": 8,
+            "type": "SaveImage",
+            "pos": [1350, 250],
+            "size": [320, 270],
+            "flags": {},
+            "order": 7,
+            "mode": 0,
+            "inputs": [{"name": "images", "type": "IMAGE", "link": 7}],
+            "properties": {"Node name for S&R": "SaveImage"},
+            "widgets_values": ["ZImageTurbo"]
+        }
+    ],
+    "links": [
+        [1, 1, 0, 6, 0, "MODEL"],
+        [2, 2, 0, 4, 0, "CLIP"],
+        [3, 3, 0, 7, 1, "VAE"],
+        [4, 4, 0, 6, 1, "CONDITIONING"],
+        [5, 5, 0, 6, 3, "LATENT"],
+        [6, 6, 0, 7, 0, "LATENT"],
+        [7, 7, 0, 8, 0, "IMAGE"]
+    ],
+    "groups": [],
+    "config": {},
+    "extra": {},
+    "version": 0.4
+}
+
+sftp = c.open_sftp()
+
+# Save as default workflow
+def ensure_dir(sftp, path):
+    try:
+        sftp.stat(path)
+    except FileNotFoundError:
+        ensure_dir(sftp, '/'.join(path.split('/')[:-1]))
+        sftp.mkdir(path)
+
+ensure_dir(sftp, '/home/fabian/ComfyUI/user/default/comfyui')
+
+wf_json = json.dumps(workflow, indent=2)
+
+# Save as default workflow
+with sftp.open('/home/fabian/ComfyUI/user/default/comfyui/workflow.json', 'w') as f:
+    f.write(wf_json)
+print("Saved default workflow: ~/ComfyUI/user/default/comfyui/workflow.json")
+
+# Also save a loadable copy in the ComfyUI root
+with sftp.open('/home/fabian/ComfyUI/z_image_turbo_workflow.json', 'w') as f:
+    f.write(wf_json)
+print("Saved loadable copy: ~/ComfyUI/z_image_turbo_workflow.json")
+
+sftp.close()
+c.close()
+print("\nDone. Refresh ComfyUI web UI — it will load the Z-Image-Turbo GGUF workflow by default.")
+print("If it still shows the old workflow, click the menu and Load the z_image_turbo_workflow.json file.")
@@ -0,0 +1,291 @@
+"""BC-250: Start ComfyUI on GPU and generate an image. Single SSH connection. No shell escaping issues."""
+import paramiko
+import time
+import json
+import sys
+import textwrap
+
+# ==== CONFIG ====
+SSH_HOST = '192.168.178.150'
+SSH_USER = 'fabian'
+SSH_KEY  = r'C:\Users\fabia\.ssh\id_ed25519'
+
+def connect():
+    k = paramiko.Ed25519Key.from_private_key_file(SSH_KEY)
+    c = paramiko.SSHClient()
+    c.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+    c.connect(SSH_HOST, username=SSH_USER, pkey=k, timeout=15)
+    return c
+
+def sh(c, cmd, timeout=60):
+    """Run a bash command. All commands go through bash explicitly."""
+    chan = c.get_transport().open_session()
+    chan.settimeout(timeout)
+    chan.exec_command(f'/bin/bash -l -c {_quote(cmd)}')
+    out = b""
+    while True:
+        try:
+            chunk = chan.recv(65536)
+            if not chunk:
+                break
+            out += chunk
+        except Exception:
+            break
+    chan.close()
+    return out.decode(errors='replace').strip()
+
+def _quote(s):
+    """Shell-quote a string using single quotes."""
+    return "'" + s.replace("'", "'\\''") + "'"
+
+def write_remote_file(c, path, content):
+    """Write a file on the remote via SFTP. No shell escaping needed."""
+    sftp = c.open_sftp()
+    with sftp.open(path, 'w') as f:
+        f.write(content)
+    sftp.close()
+
+# ================================================================
+print("="*60)
+print("STEP 1: Connect + kill old ComfyUI")
+print("="*60)
+c = connect()
+sh(c, 'pkill -9 -f "python3.*main.py" 2>/dev/null || true')
+time.sleep(2)
+alive = sh(c, 'pgrep -af "python3.*main.py" 2>/dev/null || echo NONE')
+print(f"  Old processes: {alive}")
+
+# ================================================================
+print("\n" + "="*60)
+print("STEP 2: Write launcher script on BC-250")
+print("="*60)
+
+# Write a bash launcher script directly via SFTP - avoids ALL shell escaping issues
+launcher = textwrap.dedent("""\
+    #!/bin/bash
+    # GPU environment
+    export HSA_OVERRIDE_GFX_VERSION=10.1.0
+    export HIP_VISIBLE_DEVICES=0
+    export HSA_ENABLE_SDMA=0
+    export HSA_TOOLS_LIB=""
+    export HSA_TOOLS_REPORT_LOAD_FAILURE=0
+    export PYTORCH_HIP_ALLOC_CONF=expandable_segments:False
+    # Threading
+    export OMP_NUM_THREADS=12
+    export MKL_NUM_THREADS=12
+    export OPENBLAS_NUM_THREADS=12
+    # MIOpen
+    export MIOPEN_FIND_MODE=1
+    
+    cd ~/ComfyUI
+    source ~/comfyui-env/bin/activate
+    
+    exec python3 main.py \\
+        --listen 0.0.0.0 --port 8188 \\
+        --lowvram \\
+        --force-fp16 \\
+        --cpu-vae \\
+        --disable-smart-memory
+""")
+
+write_remote_file(c, '/tmp/run_comfyui.sh', launcher)
+sh(c, 'chmod +x /tmp/run_comfyui.sh')
+print("  Launcher script written to /tmp/run_comfyui.sh")
+print("  Flags: --lowvram --force-fp16 --cpu-vae --disable-smart-memory")
+
+# ================================================================
+print("\n" + "="*60)
+print("STEP 3: Verify GPU works with PyTorch")
+print("="*60)
+
+gpu_script = textwrap.dedent("""\
+    #!/bin/bash
+    export HSA_OVERRIDE_GFX_VERSION=10.1.0
+    export HIP_VISIBLE_DEVICES=0
+    export HSA_ENABLE_SDMA=0
+    source ~/comfyui-env/bin/activate
+    python3 -c "
+import torch
+print('PyTorch:', torch.__version__)
+print('CUDA/ROCm available:', torch.cuda.is_available())
+if torch.cuda.is_available():
+    print('Device:', torch.cuda.get_device_name(0))
+    f,t = torch.cuda.mem_get_info(0)
+    print(f'VRAM: {f//1048576}MB free / {t//1048576}MB total')
+    x = torch.randn(512,512,device='cuda',dtype=torch.float16)
+    y = x @ x
+    print('GPU compute test: PASS')
+else:
+    print('FATAL: NO GPU')
+    exit(1)
+"
+""")
+write_remote_file(c, '/tmp/gpu_test.sh', gpu_script)
+sh(c, 'chmod +x /tmp/gpu_test.sh')
+out = sh(c, '/tmp/gpu_test.sh', timeout=30)
+print(f"  {out}")
+if 'FATAL' in out or 'False' in out:
+    print("  *** GPU not working! Aborting. ***")
+    c.close()
+    sys.exit(1)
+print("  GPU OK!")
+
+# ================================================================
+print("\n" + "="*60)
+print("STEP 4: Start ComfyUI")
+print("="*60)
+
+sh(c, 'rm -f /tmp/comfyui.log; touch /tmp/comfyui.log')
+sh(c, 'nohup /tmp/run_comfyui.sh > /tmp/comfyui.log 2>&1 &')
+time.sleep(3)
+
+pid = sh(c, 'pgrep -f "python3.*main.py" 2>/dev/null || echo DEAD')
+if pid == 'DEAD':
+    print("  FAILED to start! Log:")
+    print(sh(c, 'cat /tmp/comfyui.log'))
+    c.close()
+    sys.exit(1)
+print(f"  PID: {pid}")
+
+# Wait for HTTP 200
+print("  Waiting for HTTP ready...", end='', flush=True)
+for i in range(90):
+    code = sh(c, 'curl -s -o /dev/null -w "%{http_code}" http://127.0.0.1:8188/ 2>/dev/null || echo 000', timeout=5)
+    if '200' in code:
+        print(f" READY ({i*2}s)")
+        break
+    print('.', end='', flush=True)
+    time.sleep(2)
+else:
+    print(f"\n  TIMEOUT! Last log:")
+    print(sh(c, 'tail -20 /tmp/comfyui.log'))
+    c.close()
+    sys.exit(1)
+
+# Show startup flags from log
+log_head = sh(c, 'head -10 /tmp/comfyui.log')
+print(f"\n  Startup log:\n  {log_head[:300]}")
+
+# ================================================================
+print("\n" + "="*60)
+print("STEP 5: Submit workflow")
+print("="*60)
+
+workflow = {
+    "prompt": {
+        "1": {
+            "class_type": "UnetLoaderGGUF",
+            "inputs": {"unet_name": "z_image_turbo-Q5_K_S.gguf"}
+        },
+        "2": {
+            "class_type": "CLIPLoaderGGUF",
+            "inputs": {"clip_name": "Qwen3-4B.i1-Q5_K_S.gguf", "type": "qwen_image"}
+        },
+        "3": {
+            "class_type": "VAELoader",
+            "inputs": {"vae_name": "ae.safetensors"}
+        },
+        "4": {
+            "class_type": "CLIPTextEncode",
+            "inputs": {"text": "A red fox in a snowy forest, photorealistic", "clip": ["2", 0]}
+        },
+        "5": {
+            "class_type": "EmptyLatentImage",
+            "inputs": {"width": 512, "height": 512, "batch_size": 1}
+        },
+        "6": {
+            "class_type": "KSampler",
+            "inputs": {
+                "model": ["1", 0], "positive": ["4", 0], "negative": ["4", 0],
+                "latent_image": ["5", 0], "seed": 42, "steps": 8, "cfg": 1.0,
+                "sampler_name": "euler", "scheduler": "simple", "denoise": 1.0
+            }
+        },
+        "7": {
+            "class_type": "VAEDecode",
+            "inputs": {"samples": ["6", 0], "vae": ["3", 0]}
+        },
+        "8": {
+            "class_type": "SaveImage",
+            "inputs": {"images": ["7", 0], "filename_prefix": "ZImageTurbo_GPU"}
+        }
+    }
+}
+
+write_remote_file(c, '/tmp/wf.json', json.dumps(workflow))
+# Verify it's valid JSON with correct nodes
+verify = sh(c, 'python3 -c "import json; d=json.load(open(\'/tmp/wf.json\')); p=d[\'prompt\']; print(len(p), \'nodes:\', sorted(p.keys()))"')
+print(f"  Workflow: {verify}")
+
+# Submit
+resp = sh(c, 'curl -s -X POST http://127.0.0.1:8188/prompt -H "Content-Type: application/json" -d @/tmp/wf.json', timeout=10)
+print(f"  Response: {resp[:200]}")
+
+if 'prompt_id' not in resp:
+    print("  *** SUBMIT FAILED! ***")
+    print(f"  Full response: {resp}")
+    print(f"  Log: {sh(c, 'tail -10 /tmp/comfyui.log')}")
+    c.close()
+    sys.exit(1)
+
+prompt_id = json.loads(resp).get('prompt_id', '?')
+print(f"  Prompt ID: {prompt_id}")
+
+# ================================================================
+print("\n" + "="*60)
+print("STEP 6: Monitor generation (checking GPU usage)")
+print("="*60)
+
+t0 = time.time()
+for i in range(200):  # up to ~50 min
+    elapsed = int(time.time() - t0)
+    
+    gpu_pct  = sh(c, 'cat /sys/class/drm/card0/device/gpu_busy_percent 2>/dev/null || echo ?', timeout=5)
+    gpu_temp = sh(c, 'cat /sys/class/drm/card0/device/hwmon/hwmon*/temp1_input 2>/dev/null || echo 0', timeout=5)
+    temp_c   = int(gpu_temp) // 1000 if gpu_temp.isdigit() else '?'
+    
+    log_tail = sh(c, 'tail -3 /tmp/comfyui.log 2>/dev/null', timeout=5)
+    last_line = log_tail.strip().split('\n')[-1] if log_tail else ''
+    
+    # Check for output image
+    imgs = sh(c, 'ls ~/ComfyUI/output/*.png 2>/dev/null || echo NONE', timeout=5)
+    
+    print(f"  [{elapsed:>4}s] GPU:{gpu_pct:>3}% {temp_c}C | {last_line[-90:]}")
+    
+    if imgs != 'NONE':
+        print(f"\n  >>> IMAGE GENERATED! <<<")
+        print(f"  Files: {imgs}")
+        print(f"  Time: {elapsed}s")
+        final = sh(c, 'tail -20 /tmp/comfyui.log')
+        print(f"\n  Final log:\n{final}")
+        break
+    
+    # Check queue empty (= done or error)
+    q = sh(c, 'curl -s http://127.0.0.1:8188/queue 2>/dev/null || echo {}', timeout=5)
+    try:
+        qd = json.loads(q)
+        if not qd.get('queue_running') and not qd.get('queue_pending') and elapsed > 20:
+            time.sleep(3)
+            imgs2 = sh(c, 'ls ~/ComfyUI/output/*.png 2>/dev/null || echo NONE', timeout=5)
+            if imgs2 != 'NONE':
+                print(f"\n  >>> IMAGE GENERATED! <<<")
+                print(f"  Files: {imgs2}")
+                print(f"  Time: {elapsed}s")
+            else:
+                print(f"\n  Queue empty, no image. Checking log for errors...")
+                print(sh(c, 'tail -30 /tmp/comfyui.log'))
+            break
+    except json.JSONDecodeError:
+        pass
+    
+    # Check process still alive
+    alive = sh(c, 'pgrep -f "python3.*main.py" >/dev/null 2>&1 && echo YES || echo NO', timeout=5)
+    if alive == 'NO':
+        print(f"\n  *** ComfyUI CRASHED! ***")
+        print(sh(c, 'tail -40 /tmp/comfyui.log'))
+        break
+    
+    time.sleep(15)
+
+c.close()
+print("\nDone.")
@@ -0,0 +1,107 @@
+"""Just start ComfyUI and submit workflow. All patches applied."""
+import paramiko, time, json
+
+k = paramiko.Ed25519Key.from_private_key_file(r'C:\Users\fabia\.ssh\id_ed25519')
+c = paramiko.SSHClient()
+c.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+c.connect('192.168.178.150', username='fabian', pkey=k, timeout=10)
+sftp = c.open_sftp()
+
+def sh(cmd, t=30):
+    ch = c.get_transport().open_session()
+    ch.settimeout(t)
+    ch.exec_command('/bin/bash -l -c ' + "'" + cmd.replace("'", "'\\''") + "'")
+    o = b""
+    while True:
+        try:
+            d = ch.recv(65536)
+            if not d: break
+            o += d
+        except: break
+    ch.close()
+    return o.decode(errors='replace').strip()
+
+# Kill any leftover, clean logs
+sh('pkill -9 -f "python3.*main.py" 2>/dev/null; sleep 1')
+sh('rm -f /tmp/comfyui.log; rm -f ~/ComfyUI/output/ZImageTurbo_GPU*.png')
+
+# Start
+sh('nohup /tmp/run_comfyui.sh > /tmp/comfyui.log 2>&1 &')
+time.sleep(5)
+pid = sh('pgrep -f "python3.*main.py"')
+print(f"Started PID: {pid}")
+
+# Wait for HTTP
+for i in range(60):
+    r = sh('curl -s -o /dev/null -w "%{http_code}" http://127.0.0.1:8188/ 2>/dev/null', t=5)
+    if '200' in r:
+        print(f"HTTP ready ({i*2}s)")
+        break
+    time.sleep(2)
+
+# Confirm SHARED
+log = ''
+try:
+    with sftp.open('/tmp/comfyui.log', 'r') as f:
+        log = f.read().decode(errors='replace')
+except: pass
+for l in log.split('\n'):
+    s = l.strip()
+    if any(x in s for x in ['vram state', 'SHARED', 'Device:']): print(f"  {s}")
+
+# Submit
+wf = {"prompt": {
+    "1": {"class_type": "UnetLoaderGGUF", "inputs": {"unet_name": "z_image_turbo-Q5_K_S.gguf"}},
+    "2": {"class_type": "CLIPLoaderGGUF", "inputs": {"clip_name": "Qwen3-4B.i1-Q5_K_S.gguf", "type": "qwen_image"}},
+    "3": {"class_type": "VAELoader", "inputs": {"vae_name": "ae.safetensors"}},
+    "4": {"class_type": "CLIPTextEncode", "inputs": {"text": "A red fox in a snowy forest, photorealistic", "clip": ["2", 0]}},
+    "5": {"class_type": "EmptyLatentImage", "inputs": {"width": 512, "height": 512, "batch_size": 1}},
+    "6": {"class_type": "KSampler", "inputs": {"model": ["1", 0], "positive": ["4", 0], "negative": ["4", 0],
+        "latent_image": ["5", 0], "seed": 99999, "steps": 8, "cfg": 1.0,
+        "sampler_name": "euler", "scheduler": "simple", "denoise": 1.0}},
+    "7": {"class_type": "VAEDecode", "inputs": {"samples": ["6", 0], "vae": ["3", 0]}},
+    "8": {"class_type": "SaveImage", "inputs": {"images": ["7", 0], "filename_prefix": "ZImageTurbo_GPU"}}
+}}
+with sftp.open('/tmp/wf.json', 'w') as f:
+    f.write(json.dumps(wf))
+print("Submitting...")
+resp = sh('curl -s -X POST http://127.0.0.1:8188/prompt -H "Content-Type: application/json" -d @/tmp/wf.json')
+print(f"  {resp[:120]}")
+
+# Monitor
+t0 = time.time()
+shown = set()
+for i in range(150):
+    el = int(time.time() - t0)
+    try:
+        with sftp.open('/tmp/comfyui.log', 'r') as f:
+            log = f.read().decode(errors='replace')
+    except: log = ''
+
+    for l in log.split('\n'):
+        s = l.strip()
+        if s and s not in shown and any(x in s for x in ['/8', 'loaded', 'load device', 'offload device',
+                'Requested', 'VAE', 'Prompt executed', 'Error', 'OOM', 'CUDA']):
+            if 'FETCH' not in s and 'audio_vae' not in s and 'split attention' not in s:
+                gpu = sh('cat /sys/class/drm/card0/device/gpu_busy_percent 2>/dev/null', t=3)
+                print(f"  [{el:>3}s] GPU:{gpu}% {s[-110:]}")
+                shown.add(s)
+
+    imgs = sh('ls ~/ComfyUI/output/ZImageTurbo_GPU*.png 2>/dev/null', t=5)
+    if imgs:
+        print(f"\n*** DONE in {el}s! ***")
+        for l in log.split('\n'):
+            s = l.strip()
+            if 'Prompt executed' in s: print(f"  {s}")
+        break
+
+    if sh('pgrep -f "python3.*main.py" >/dev/null && echo Y || echo N', t=5) == 'N':
+        print(f"\nCRASHED at {el}s!")
+        for l in log.split('\n')[-15:]:
+            if l.strip(): print(f"  {l.strip()}")
+        break
+
+    time.sleep(4)
+
+sftp.close()
+c.close()
@@ -0,0 +1,358 @@
+"""BC-250 Full GPU Fix: Verify ROCm, diagnose VRAM, start ComfyUI on GPU, generate image."""
+import paramiko
+import time
+import json
+import sys
+
+SSH_HOST = '192.168.178.150'
+SSH_USER = 'fabian'
+SSH_KEY = r'C:\Users\fabia\.ssh\id_ed25519'
+
+def ssh_connect():
+    k = paramiko.Ed25519Key.from_private_key_file(SSH_KEY)
+    c = paramiko.SSHClient()
+    c.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+    c.connect(SSH_HOST, username=SSH_USER, pkey=k, timeout=15)
+    return c
+
+def run(c, cmd, timeout=30):
+    """Run command via fish shell, return stdout."""
+    wrapped = f'bash -c {repr(cmd)}'
+    _, o, e = c.exec_command(wrapped, timeout=timeout)
+    return o.read().decode(errors='replace').strip()
+
+def run_full(c, cmd, timeout=30):
+    """Run command, return (stdout, stderr)."""
+    wrapped = f'bash -c {repr(cmd)}'
+    _, o, e = c.exec_command(wrapped, timeout=timeout)
+    return o.read().decode(errors='replace').strip(), e.read().decode(errors='replace').strip()
+
+# ============================================================
+# PHASE 1: Kill any remnants
+# ============================================================
+print("="*60)
+print("PHASE 1: Clean slate")
+print("="*60)
+c = ssh_connect()
+run(c, 'pkill -f "python.*main.py" 2>/dev/null; pkill -f comfyui 2>/dev/null')
+time.sleep(2)
+leftover = run(c, 'pgrep -af "python.*main.py" 2>/dev/null')
+if leftover:
+    print(f"WARNING: Still running: {leftover}")
+    run(c, 'pkill -9 -f "python.*main.py" 2>/dev/null')
+    time.sleep(1)
+print("ComfyUI killed. Clean slate.")
+
+# ============================================================
+# PHASE 2: Verify ROCm + PyTorch GPU
+# ============================================================
+print("\n" + "="*60)
+print("PHASE 2: Verify ROCm + PyTorch GPU access")
+print("="*60)
+
+# Set GPU env vars for ALL subsequent commands
+GPU_ENV = (
+    'export HSA_OVERRIDE_GFX_VERSION=10.1.0; '
+    'export HIP_VISIBLE_DEVICES=0; '
+    'export HSA_ENABLE_SDMA=0; '
+    'export HSA_TOOLS_LIB=""; '
+    'export HSA_TOOLS_REPORT_LOAD_FAILURE=0; '
+    'export PYTORCH_HIP_ALLOC_CONF=expandable_segments:False; '
+)
+
+# Check rocminfo
+out = run(c, f'{GPU_ENV} rocminfo 2>&1 | grep -E "Name:|Marketing Name:|gfx" | head -10')
+print(f"ROCm devices:\n{out}")
+
+# Check PyTorch GPU
+gpu_test = f'''{GPU_ENV} cd ~/ComfyUI && source ~/comfyui-env/bin/activate.fish 2>/dev/null; . ~/comfyui-env/bin/activate 2>/dev/null; python3 -c "
+import torch
+print(f'PyTorch: {{torch.__version__}}')
+print(f'CUDA available: {{torch.cuda.is_available()}}')
+print(f'Device count: {{torch.cuda.device_count()}}')
+if torch.cuda.is_available():
+    print(f'Device name: {{torch.cuda.get_device_name(0)}}')
+    free, total = torch.cuda.mem_get_info(0)
+    print(f'VRAM: {{free//1024//1024}}MB free / {{total//1024//1024}}MB total')
+    # Quick GPU compute test
+    x = torch.randn(1024, 1024, device='cuda', dtype=torch.float16)
+    y = torch.mm(x, x)
+    print(f'GPU compute test: OK (result sum={{y.sum().item():.1f}})')
+    del x, y
+    torch.cuda.empty_cache()
+else:
+    print('ERROR: GPU NOT AVAILABLE')
+    import sys; sys.exit(1)
+"'''
+out, err = run_full(c, gpu_test, timeout=60)
+print(out)
+if err:
+    print(f"STDERR: {err}")
+if 'ERROR: GPU NOT AVAILABLE' in out or 'CUDA available: False' in out:
+    print("\n*** FATAL: PyTorch cannot see the GPU! ***")
+    c.close()
+    sys.exit(1)
+print("\nGPU verified OK!")
+
+# ============================================================
+# PHASE 3: Start ComfyUI with correct GPU flags
+# ============================================================
+print("\n" + "="*60)
+print("PHASE 3: Start ComfyUI with GPU")
+print("="*60)
+
+# The key insight: --novram was offloading EVERYTHING to CPU (0 MB on GPU)
+# For this APU with shared memory, --lowvram is better:
+# it keeps compute on GPU but swaps model layers in/out
+# We also use --force-fp16 to reduce memory pressure
+# --cpu-vae to avoid the known VAE decode hang on this GPU
+
+COMFYUI_CMD = (
+    f'{GPU_ENV} '
+    'export OMP_NUM_THREADS=12; '
+    'export MKL_NUM_THREADS=12; '
+    'export OPENBLAS_NUM_THREADS=12; '
+    'export MIOPEN_FIND_MODE=1; '  # Fast MIOpen kernel search
+    'cd ~/ComfyUI && '
+    'source ~/comfyui-env/bin/activate 2>/dev/null; . ~/comfyui-env/bin/activate 2>/dev/null; '
+    'nohup python3 main.py '
+    '--listen 0.0.0.0 --port 8188 '
+    '--lowvram '
+    '--force-fp16 '
+    '--cpu-vae '
+    '--disable-smart-memory '
+    '> /tmp/comfyui.log 2>&1 &'
+)
+
+# Truncate old log first
+run(c, 'truncate -s 0 /tmp/comfyui.log 2>/dev/null; touch /tmp/comfyui.log')
+print("Starting ComfyUI with: --lowvram --force-fp16 --cpu-vae --disable-smart-memory")
+print("(--lowvram keeps compute on GPU, swaps layers; --novram was wrong - it put everything on CPU)")
+run(c, COMFYUI_CMD)
+time.sleep(3)
+
+# Verify it started
+pid = run(c, 'pgrep -f "python.*main.py" 2>/dev/null')
+if not pid:
+    print("ERROR: ComfyUI failed to start!")
+    log = run(c, 'cat /tmp/comfyui.log')
+    print(f"Log:\n{log}")
+    c.close()
+    sys.exit(1)
+print(f"ComfyUI started, PID: {pid}")
+
+# Wait for server ready
+print("Waiting for server ready...")
+for i in range(60):
+    try:
+        resp = run(c, 'curl -s -o /dev/null -w "%{http_code}" http://127.0.0.1:8188/ 2>/dev/null', timeout=5)
+        if resp == '200':
+            print(f"Server ready after {i*3}s!")
+            break
+    except:
+        pass
+    # Also check for crash
+    log_tail = run(c, 'tail -3 /tmp/comfyui.log 2>/dev/null')
+    if 'Traceback' in log_tail or 'Error' in log_tail:
+        print(f"Server log issue: {log_tail}")
+    if i % 5 == 0 and i > 0:
+        print(f"  [{i*3}s] Still waiting... log: {log_tail[-80:]}")
+    time.sleep(3)
+else:
+    print("TIMEOUT waiting for ComfyUI!")
+    log = run(c, 'tail -30 /tmp/comfyui.log')
+    print(f"Log:\n{log}")
+    c.close()
+    sys.exit(1)
+
+# Print startup log to confirm flags
+log = run(c, 'head -20 /tmp/comfyui.log')
+print(f"\nStartup log:\n{log}")
+
+# ============================================================
+# PHASE 4: Submit workflow via SFTP
+# ============================================================
+print("\n" + "="*60)
+print("PHASE 4: Submit workflow")
+print("="*60)
+
+workflow = {
+    "prompt": {
+        "1": {
+            "class_type": "UnetLoaderGGUF",
+            "inputs": {
+                "unet_name": "z_image_turbo-Q5_K_S.gguf"
+            }
+        },
+        "2": {
+            "class_type": "CLIPLoaderGGUF",
+            "inputs": {
+                "clip_name": "Qwen3-4B.i1-Q5_K_S.gguf",
+                "type": "qwen_image"
+            }
+        },
+        "3": {
+            "class_type": "VAELoader",
+            "inputs": {
+                "vae_name": "ae.safetensors"
+            }
+        },
+        "4": {
+            "class_type": "CLIPTextEncode",
+            "inputs": {
+                "text": "A red fox in a snowy forest, photorealistic, highly detailed",
+                "clip": ["2", 0]
+            }
+        },
+        "5": {
+            "class_type": "EmptyLatentImage",
+            "inputs": {
+                "width": 512,
+                "height": 512,
+                "batch_size": 1
+            }
+        },
+        "6": {
+            "class_type": "KSampler",
+            "inputs": {
+                "model": ["1", 0],
+                "positive": ["4", 0],
+                "negative": ["4", 0],
+                "latent_image": ["5", 0],
+                "seed": 42,
+                "steps": 8,
+                "cfg": 1.0,
+                "sampler_name": "euler",
+                "scheduler": "simple",
+                "denoise": 1.0
+            }
+        },
+        "7": {
+            "class_type": "VAEDecode",
+            "inputs": {
+                "samples": ["6", 0],
+                "vae": ["3", 0]
+            }
+        },
+        "8": {
+            "class_type": "SaveImage",
+            "inputs": {
+                "images": ["7", 0],
+                "filename_prefix": "ZImageTurbo_GPU"
+            }
+        }
+    }
+}
+
+# Write via SFTP
+sftp = c.open_sftp()
+wf_json = json.dumps(workflow)
+with sftp.open('/tmp/wf.json', 'w') as f:
+    f.write(wf_json)
+sftp.close()
+print("Workflow written to /tmp/wf.json via SFTP")
+
+# Verify JSON
+verify = run(c, 'python3 -c "import json; d=json.load(open(\'/tmp/wf.json\')); print(f\'Nodes: {list(d[chr(34)+chr(34) if False else \"prompt\"].keys())}\')"')
+print(f"Verify: {verify}")
+
+# Submit
+resp = run(c, 'curl -s -X POST http://127.0.0.1:8188/prompt -H "Content-Type: application/json" -d @/tmp/wf.json 2>/dev/null')
+print(f"Submit response: {resp}")
+
+if 'error' in resp.lower() and 'prompt_id' not in resp.lower():
+    print(f"\n*** SUBMISSION ERROR ***")
+    # Check what went wrong
+    log = run(c, 'tail -10 /tmp/comfyui.log')
+    print(f"Log: {log}")
+    c.close()
+    sys.exit(1)
+
+try:
+    resp_data = json.loads(resp)
+    prompt_id = resp_data.get('prompt_id', 'unknown')
+    print(f"Prompt ID: {prompt_id}")
+except:
+    print("Could not parse response, continuing anyway...")
+
+# ============================================================
+# PHASE 5: Monitor generation with GPU tracking
+# ============================================================
+print("\n" + "="*60)
+print("PHASE 5: Monitor generation (GPU must be active!)")
+print("="*60)
+
+start_time = time.time()
+last_log_len = 0
+
+for i in range(120):  # Up to 30 minutes
+    elapsed = int(time.time() - start_time)
+    
+    # GPU metrics
+    gpu_pct = run(c, 'cat /sys/class/drm/card0/device/gpu_busy_percent 2>/dev/null')
+    gpu_temp = run(c, 'cat /sys/class/drm/card0/device/hwmon/hwmon*/temp1_input 2>/dev/null')
+    temp_c = int(gpu_temp) // 1000 if gpu_temp.isdigit() else '?'
+    
+    # GPU power
+    gpu_power = run(c, f'{GPU_ENV} rocm-smi -P 2>&1 | grep "Graphics Package" | grep -oP "[\\d.]+" | head -1')
+    
+    # Process info
+    proc = run(c, 'ps -p $(pgrep -f "python.*main.py" | head -1) -o %cpu,%mem,rss --no-headers 2>/dev/null')
+    
+    # Log tail
+    log = run(c, 'tail -5 /tmp/comfyui.log 2>/dev/null')
+    last_line = log.split('\n')[-1] if log else ''
+    
+    # Output files
+    files = run(c, 'ls ~/ComfyUI/output/*.png 2>/dev/null')
+    
+    # Queue
+    queue = run(c, 'curl -s http://127.0.0.1:8188/queue 2>/dev/null')
+    
+    status = f"[{elapsed:>4}s] GPU:{gpu_pct:>3}% {temp_c}C {gpu_power}W | proc:{proc} | {last_line[-100:]}"
+    print(status)
+    
+    # SUCCESS: Image generated!
+    if files:
+        print(f"\n{'='*60}")
+        print(f"*** SUCCESS! IMAGE GENERATED! ***")
+        print(f"Files: {files}")
+        print(f"Total time: {elapsed}s")
+        print(f"{'='*60}")
+        
+        # Print final log
+        final_log = run(c, 'tail -20 /tmp/comfyui.log 2>/dev/null')
+        print(f"\nFinal log:\n{final_log}")
+        break
+    
+    # Check if queue is empty (job done or failed)
+    try:
+        qdata = json.loads(queue)
+        running = len(qdata.get('queue_running', []))
+        pending = len(qdata.get('queue_pending', []))
+        if running == 0 and pending == 0 and elapsed > 30:
+            print(f"\nQueue empty after {elapsed}s. Checking if image was saved...")
+            time.sleep(2)
+            files = run(c, 'ls ~/ComfyUI/output/*.png 2>/dev/null')
+            if files:
+                print(f"*** SUCCESS! {files}")
+            else:
+                print("No image. Checking log for errors:")
+                err_log = run(c, 'tail -30 /tmp/comfyui.log 2>/dev/null')
+                print(err_log)
+            break
+    except:
+        pass
+    
+    # Check for process death
+    alive = run(c, 'pgrep -f "python.*main.py" 2>/dev/null')
+    if not alive:
+        print("\n*** ComfyUI process died! ***")
+        crash_log = run(c, 'tail -40 /tmp/comfyui.log 2>/dev/null')
+        print(f"Crash log:\n{crash_log}")
+        break
+    
+    time.sleep(15)
+
+c.close()
+print("\nDone.")
@@ -0,0 +1,283 @@
+#!/usr/bin/env python3
+"""Fix GPU inference: kill stuck, diagnose, restart with --novram, test."""
+import paramiko, json, time, textwrap
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=60, desc=""):
+    if desc:
+        print(f"\n{'='*60}\n  {desc}\n{'='*60}")
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    rc = stdout.channel.recv_exit_status()
+    combined = out.strip()
+    if combined:
+        lines = combined.split('\n')
+        if len(lines) > 50:
+            print(f"  ... ({len(lines)} lines, showing last 50)")
+            print('\n'.join(lines[-50:]))
+        else:
+            print(combined)
+    if err.strip():
+        for line in err.strip().split('\n')[-10:]:
+            print(f"  STDERR: {line}")
+    return rc, out, err
+
+# ============================================================
+# STEP 1: Kill stuck ComfyUI
+# ============================================================
+run("pkill -f 'python3 main.py' 2>/dev/null; sleep 2; "
+    "pkill -9 -f 'python3 main.py' 2>/dev/null; sleep 1; "
+    "echo 'Killed.'", desc="Kill stuck ComfyUI")
+
+# ============================================================
+# STEP 2: Check dmesg for GPU errors
+# ============================================================
+run("dmesg | grep -i -E 'amdgpu|error|fault|gpu|kiq|gfx' | tail -30",
+    desc="Check dmesg for GPU errors")
+
+# ============================================================
+# STEP 3: Quick GPU sanity test
+# ============================================================
+run("bash -c 'source /home/fabian/comfyui-env/bin/activate && "
+    "HSA_OVERRIDE_GFX_VERSION=10.1.0 HSA_ENABLE_SDMA=0 "
+    "python3 -c \""
+    "import torch; "
+    "print(f\\\"CUDA available: {torch.cuda.is_available()}\\\"); "
+    "print(f\\\"Device: {torch.cuda.get_device_name(0)}\\\"); "
+    "a = torch.randn(1024, 1024, device=\\\"cuda\\\"); "
+    "b = torch.randn(1024, 1024, device=\\\"cuda\\\"); "
+    "c = a @ b; "
+    "print(f\\\"Matmul result shape: {c.shape}, sum: {c.sum().item():.2f}\\\"); "
+    "# Test fp16 "
+    "a16 = a.half(); b16 = b.half(); c16 = a16 @ b16; "
+    "print(f\\\"FP16 matmul OK: {c16.shape}\\\"); "
+    "print(f\\\"Free VRAM: {torch.cuda.mem_get_info()[0]/1024**2:.0f} MB\\\"); "
+    "print(f\\\"Total VRAM: {torch.cuda.mem_get_info()[1]/1024**2:.0f} MB\\\"); "
+    "print(\\\"GPU SANITY: PASS\\\")\"'",
+    desc="Quick GPU sanity test")
+
+# ============================================================
+# STEP 4: Check what ComfyUI flags are available
+# ============================================================
+run("bash -c 'source /home/fabian/comfyui-env/bin/activate && cd /home/fabian/ComfyUI && "
+    "python3 main.py --help 2>&1 | grep -E \"novram|lowvram|cpu|fp16|force|vram|disable-smart|channels\"'",
+    desc="ComfyUI VRAM-related flags")
+
+# ============================================================
+# STEP 5: Write new startup script with --novram
+# ============================================================
+startup_script = textwrap.dedent("""\
+    #!/bin/bash
+    # BC-250 ComfyUI Launcher - GPU mode with aggressive offloading
+    
+    # GPU identity
+    export HSA_OVERRIDE_GFX_VERSION=10.1.0
+    export HIP_VISIBLE_DEVICES=0
+    
+    # Disable SDMA (known issue on Cyan Skillfish)
+    export HSA_ENABLE_SDMA=0
+    
+    # Suppress HSA tool warnings
+    export HSA_TOOLS_LIB=""
+    export HSA_TOOLS_REPORT_LOAD_FAILURE=0
+    
+    # Threading: use all 12 cores for CPU-side work
+    export OMP_NUM_THREADS=12
+    export MKL_NUM_THREADS=12
+    export OPENBLAS_NUM_THREADS=12
+    
+    # HIP memory: allow expandable segments to reduce fragmentation
+    export PYTORCH_HIP_ALLOC_CONF=expandable_segments:True
+    
+    # Activate venv
+    source /home/fabian/comfyui-env/bin/activate
+    cd /home/fabian/ComfyUI
+    
+    # --novram: most aggressive offloading - keeps almost nothing on GPU,
+    # sends individual layers to GPU one at a time during forward pass.
+    # This is needed because BC-250 has only ~7.6GB shared VRAM.
+    # --disable-smart-memory: prevents ComfyUI from trying to be clever about memory
+    # --force-fp16: force fp16 to halve VRAM usage
+    exec python3 main.py --listen 0.0.0.0 --port 8188 --novram --force-fp16 --disable-smart-memory
+""")
+
+sftp = ssh.open_sftp()
+with sftp.open('/home/fabian/start_comfyui.sh', 'w') as f:
+    f.write(startup_script)
+sftp.close()
+run("chmod +x /home/fabian/start_comfyui.sh", desc="Make script executable")
+print("\n  Startup script updated with --novram --force-fp16 --disable-smart-memory")
+
+# ============================================================
+# STEP 6: Launch ComfyUI with new settings
+# ============================================================
+run("bash -c 'nohup /home/fabian/start_comfyui.sh > /home/fabian/comfyui.log 2>&1 &'; sleep 1; echo 'Launched'",
+    desc="Launch ComfyUI with --novram")
+
+# Wait for server to be ready
+print("\n  Waiting for server to start...")
+for i in range(30):
+    time.sleep(3)
+    rc, out, _ = run("bash -c 'curl -s -o /dev/null -w \"%{http_code}\" http://localhost:8188/ 2>/dev/null || echo 0'")
+    code = out.strip()
+    if code == '200':
+        print(f"  Server ready after {(i+1)*3}s!")
+        break
+    # Check log for errors
+    rc2, log, _ = run("tail -3 /home/fabian/comfyui.log 2>/dev/null")
+    if 'Error' in log or 'error' in log.lower():
+        print(f"  Log: {log.strip()}")
+    print(f"  [{(i+1)*3}s] HTTP {code}...")
+else:
+    print("  Server didn't start in 90s!")
+    run("tail -40 /home/fabian/comfyui.log", desc="Startup log")
+    ssh.close()
+    exit(1)
+
+# Confirm server info
+run("tail -30 /home/fabian/comfyui.log", desc="Startup log")
+
+# ============================================================
+# STEP 7: Submit test workflow (smaller 512x512 image first)
+# ============================================================
+workflow = {
+    "prompt": {
+        "1": {
+            "class_type": "UnetLoaderGGUF",
+            "inputs": {"unet_name": "z_image_turbo-Q5_K_S.gguf"}
+        },
+        "2": {
+            "class_type": "CLIPLoaderGGUF",
+            "inputs": {"clip_name": "Qwen3-4B.i1-Q5_K_S.gguf", "type": "qwen_image"}
+        },
+        "3": {
+            "class_type": "VAELoader",
+            "inputs": {"vae_name": "ae.safetensors"}
+        },
+        "4": {
+            "class_type": "CLIPTextEncode",
+            "inputs": {
+                "text": "A red fox in a snowy forest, photorealistic",
+                "clip": ["2", 0]
+            }
+        },
+        "5": {
+            "class_type": "CLIPTextEncode",
+            "inputs": {"text": "", "clip": ["2", 0]}
+        },
+        "6": {
+            "class_type": "EmptyLatentImage",
+            "inputs": {"width": 512, "height": 512, "batch_size": 1}
+        },
+        "7": {
+            "class_type": "KSampler",
+            "inputs": {
+                "model": ["1", 0],
+                "seed": 12345,
+                "steps": 8,
+                "cfg": 1.0,
+                "sampler_name": "euler",
+                "scheduler": "simple",
+                "positive": ["4", 0],
+                "negative": ["5", 0],
+                "latent_image": ["6", 0],
+                "denoise": 1.0
+            }
+        },
+        "8": {
+            "class_type": "VAEDecode",
+            "inputs": {"samples": ["7", 0], "vae": ["3", 0]}
+        },
+        "9": {
+            "class_type": "SaveImage",
+            "inputs": {"images": ["8", 0], "filename_prefix": "ZImageTurbo_test"}
+        }
+    }
+}
+
+sftp2 = ssh.open_sftp()
+with sftp2.open('/tmp/zimage_workflow.json', 'w') as f:
+    f.write(json.dumps(workflow))
+sftp2.close()
+
+rc, out, _ = run("bash -c 'curl -s -X POST http://localhost:8188/prompt "
+    "-H \"Content-Type: application/json\" "
+    "-d @/tmp/zimage_workflow.json'",
+    desc="Submit 512x512 test workflow")
+
+prompt_id = None
+try:
+    resp = json.loads(out.strip())
+    if 'error' in resp:
+        print(f"\n  API ERROR: {resp['error']}")
+        if 'node_errors' in resp:
+            for nid, e in resp['node_errors'].items():
+                print(f"  Node {nid}: {e}")
+        ssh.close()
+        exit(1)
+    prompt_id = resp.get('prompt_id', 'unknown')
+    print(f"\n  Prompt ID: {prompt_id}")
+except:
+    print(f"  Raw response: {out.strip()[:500]}")
+
+# ============================================================
+# STEP 8: Monitor generation
+# ============================================================
+print("\n  Monitoring GPU generation...")
+last_log = ""
+for i in range(120):  # up to 30 minutes
+    time.sleep(15)
+    
+    # CPU + GPU status
+    rc, status, _ = run("bash -c '"
+        "PID=$(pgrep -f \"python3 main.py\" | head -1); "
+        "if [ -n \"$PID\" ]; then "
+        "  CPU=$(ps -p $PID -o %cpu --no-headers); "
+        "  MEM=$(ps -p $PID -o %mem --no-headers); "
+        "  THREADS=$(ps -p $PID -o nlwp --no-headers); "
+        "  LOAD=$(cat /proc/loadavg | cut -d\" \" -f1-3); "
+        "  GPU_USE=$(cat /sys/class/drm/card1/device/gpu_busy_percent 2>/dev/null || echo N/A); "
+        "  VRAM_USED=$(cat /sys/class/drm/card1/device/mem_info_vram_used 2>/dev/null || echo 0); "
+        "  VRAM_TOTAL=$(cat /sys/class/drm/card1/device/mem_info_vram_total 2>/dev/null || echo 1); "
+        "  echo \"CPU:${CPU}% MEM:${MEM}% THR:${THREADS} LOAD:${LOAD} GPU:${GPU_USE}% VRAM:$((VRAM_USED/1048576))/$((VRAM_TOTAL/1048576))MB\"; "
+        "else echo DEAD; fi'")
+    
+    rc, log, _ = run("bash -c 'tail -8 /home/fabian/comfyui.log 2>/dev/null'")
+    log_lines = log.strip()
+    
+    # Show status
+    status_line = status.strip()
+    print(f"  [{i+1}] {(i+1)*15}s | {status_line}")
+    
+    # Show new log lines
+    if log_lines != last_log:
+        new_part = log_lines
+        for line in new_part.split('\n')[-5:]:
+            if line.strip():
+                print(f"    LOG: {line.strip()}")
+        last_log = log_lines
+    
+    if 'DEAD' in status_line:
+        print("\n  ComfyUI DIED!")
+        run("tail -60 /home/fabian/comfyui.log", desc="Death log")
+        break
+    
+    if 'Prompt executed in' in log_lines:
+        print(f"\n  SUCCESS! Image generated at check {i+1} (~{(i+1)*15}s)")
+        break
+    
+    if 'Error' in log_lines or 'Traceback' in log_lines:
+        print("\n  ERROR detected!")
+        run("tail -60 /home/fabian/comfyui.log", desc="Error log")
+        break
+
+# Final check
+run("ls -lah /home/fabian/ComfyUI/output/ 2>/dev/null", desc="Output files")
+run("tail -25 /home/fabian/comfyui.log", desc="Final log")
+
+ssh.close()
+print("\nDone.")
@@ -0,0 +1,226 @@
+#!/usr/bin/env python3
+"""
+Single SSH connection: kill stuck, check flags, update startup, launch, submit, monitor.
+Properly closes connection when done.
+"""
+import paramiko
+import json
+import time
+import sys
+
+KEY = r'C:\Users\fabia\.ssh\id_ed25519'
+HOST = '192.168.178.150'
+USER = 'fabian'
+
+def connect():
+    ssh = paramiko.SSHClient()
+    ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+    for attempt in range(5):
+        try:
+            ssh.connect(HOST, username=USER, key_filename=KEY, timeout=10)
+            return ssh
+        except Exception as e:
+            print(f"  SSH attempt {attempt+1}/5 failed: {e}")
+            time.sleep(10)
+    print("FATAL: Cannot connect to BC-250")
+    sys.exit(1)
+
+def run(ssh, cmd, timeout=120):
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    return out, err
+
+def main():
+    ssh = connect()
+    print("Connected to BC-250.\n")
+
+    try:
+        # ── 1. Kill stuck ComfyUI ──
+        print("=== STEP 1: Kill stuck ComfyUI ===")
+        out, _ = run(ssh, "pkill -9 -f 'python3 main.py' 2>/dev/null; sleep 2; "
+                          "pgrep -f 'python3 main.py' || echo 'all_dead'")
+        print(f"  {out.strip()}")
+
+        # ── 2. Check available flags ──
+        print("\n=== STEP 2: ComfyUI flags ===")
+        out, err = run(ssh, "bash -c 'source /home/fabian/comfyui-env/bin/activate && "
+                            "cd /home/fabian/ComfyUI && python3 main.py --help 2>&1'")
+        combined = out + err
+        for line in combined.split('\n'):
+            low = line.lower()
+            if any(w in low for w in ['vae', 'fp16', 'fp32', 'force', 'cpu', 'vram',
+                                       'memory', 'offload', 'precision', 'novram', 'lowvram']):
+                print(f"  {line.strip()}")
+
+        # ── 3. Write startup script ──
+        print("\n=== STEP 3: Update startup script ===")
+        script = r"""#!/bin/bash
+# BC-250 ComfyUI Launcher - GPU inference with CPU VAE decode
+
+# GPU identity
+export HSA_OVERRIDE_GFX_VERSION=10.1.0
+export HIP_VISIBLE_DEVICES=0
+export HSA_ENABLE_SDMA=0
+export HSA_TOOLS_LIB=""
+export HSA_TOOLS_REPORT_LOAD_FAILURE=0
+
+# Use all 12 CPU cores
+export OMP_NUM_THREADS=12
+export MKL_NUM_THREADS=12
+export OPENBLAS_NUM_THREADS=12
+
+# Activate venv
+source /home/fabian/comfyui-env/bin/activate
+cd /home/fabian/ComfyUI
+
+# --novram: aggressive offload, one layer at a time to GPU
+# --force-fp16: halve VRAM for diffusion
+# --cpu-vae: VAE decode on CPU (GPU hangs on float32 VAE conv2d)
+# --disable-smart-memory: no memory heuristics
+exec python3 main.py \
+    --listen 0.0.0.0 --port 8188 \
+    --novram \
+    --force-fp16 \
+    --cpu-vae \
+    --disable-smart-memory
+"""
+        # Write via heredoc to avoid SFTP
+        escaped = script.replace("'", "'\\''")
+        out, _ = run(ssh, f"cat > /home/fabian/start_comfyui.sh << 'HEREDOC_END'\n{script}HEREDOC_END\n"
+                          f"chmod +x /home/fabian/start_comfyui.sh && echo 'written'")
+        print(f"  {out.strip()}")
+
+        # ── 4. Launch ComfyUI ──
+        print("\n=== STEP 4: Launch ComfyUI ===")
+        run(ssh, "nohup /home/fabian/start_comfyui.sh > /home/fabian/comfyui.log 2>&1 &")
+        time.sleep(2)
+
+        print("  Waiting for server...")
+        for i in range(40):
+            time.sleep(3)
+            out, _ = run(ssh, "curl -s -o /dev/null -w '%{http_code}' http://localhost:8188/ 2>/dev/null || echo 0")
+            code = out.strip()
+            if code == '200':
+                print(f"  Server ready ({(i+1)*3}s)")
+                break
+            if i % 5 == 4:
+                log, _ = run(ssh, "tail -2 /home/fabian/comfyui.log 2>/dev/null")
+                last = [l.strip() for l in log.strip().split('\n') if l.strip()][-1:]
+                print(f"  [{(i+1)*3}s] HTTP {code} ... {last[0][:80] if last else ''}")
+        else:
+            print("  Server didn't start in 120s!")
+            out, _ = run(ssh, "tail -40 /home/fabian/comfyui.log 2>/dev/null")
+            print(out)
+            return
+
+        # Show startup log
+        out, _ = run(ssh, "tail -20 /home/fabian/comfyui.log 2>/dev/null")
+        for line in out.strip().split('\n'):
+            l = line.strip()
+            if l and not l.startswith('FETCH'):
+                print(f"  LOG: {l[:120]}")
+
+        # ── 5. Submit workflow ──
+        print("\n=== STEP 5: Submit workflow (512x512, 8 steps) ===")
+        workflow = {
+            "prompt": {
+                "1": {"class_type": "UnetLoaderGGUF",
+                      "inputs": {"unet_name": "z_image_turbo-Q5_K_S.gguf"}},
+                "2": {"class_type": "CLIPLoaderGGUF",
+                      "inputs": {"clip_name": "Qwen3-4B.i1-Q5_K_S.gguf", "type": "qwen_image"}},
+                "3": {"class_type": "VAELoader",
+                      "inputs": {"vae_name": "ae.safetensors"}},
+                "4": {"class_type": "CLIPTextEncode",
+                      "inputs": {"text": "A red fox in a snowy forest, photorealistic", "clip": ["2", 0]}},
+                "5": {"class_type": "CLIPTextEncode",
+                      "inputs": {"text": "", "clip": ["2", 0]}},
+                "6": {"class_type": "EmptyLatentImage",
+                      "inputs": {"width": 512, "height": 512, "batch_size": 1}},
+                "7": {"class_type": "KSampler",
+                      "inputs": {"model": ["1", 0], "seed": 42, "steps": 8, "cfg": 1.0,
+                                 "sampler_name": "euler", "scheduler": "simple",
+                                 "positive": ["4", 0], "negative": ["5", 0],
+                                 "latent_image": ["6", 0], "denoise": 1.0}},
+                "8": {"class_type": "VAEDecode",
+                      "inputs": {"samples": ["7", 0], "vae": ["3", 0]}},
+                "9": {"class_type": "SaveImage",
+                      "inputs": {"images": ["8", 0], "filename_prefix": "ZImageTurbo_test"}}
+            }
+        }
+
+        wf_json = json.dumps(workflow).replace("'", "'\\''")
+        out, _ = run(ssh, f"curl -s -X POST http://localhost:8188/prompt "
+                          f"-H 'Content-Type: application/json' "
+                          f"-d '{wf_json}'")
+        try:
+            resp = json.loads(out.strip())
+            if 'error' in resp:
+                print(f"  API ERROR: {resp['error']}")
+                if 'node_errors' in resp:
+                    for nid, e in resp['node_errors'].items():
+                        print(f"  Node {nid}: {e}")
+                return
+            print(f"  Prompt ID: {resp.get('prompt_id')}")
+        except:
+            print(f"  Response: {out.strip()[:500]}")
+
+        # ── 6. Monitor generation ──
+        print("\n=== STEP 6: Monitoring generation ===")
+        last_log = ""
+        for i in range(120):  # up to 30 minutes
+            time.sleep(15)
+
+            stats, _ = run(ssh, "bash -c '"
+                "PID=$(pgrep -f \"python3 main.py\" | head -1); "
+                "if [ -n \"$PID\" ]; then "
+                "  CPU=$(ps -p $PID -o %cpu --no-headers); "
+                "  MEM=$(ps -p $PID -o rss --no-headers); "
+                "  LOAD=$(cat /proc/loadavg | cut -d\" \" -f1-3); "
+                "  echo \"CPU:${CPU}% RSS:$((MEM/1024))M LOAD:${LOAD}\"; "
+                "else echo DEAD; fi'")
+
+            log, _ = run(ssh, "tail -10 /home/fabian/comfyui.log 2>/dev/null")
+            log_s = log.strip()
+
+            elapsed = (i+1) * 15
+            m, s = divmod(elapsed, 60)
+            print(f"  [{m}m{s:02d}s] {stats.strip()}")
+
+            if log_s != last_log:
+                for line in reversed(log_s.split('\n')):
+                    l = line.strip()
+                    if l and not l.startswith('FETCH') and not l.startswith('[DEPRECATION') and not l.startswith('[ComfyUI-Manager]'):
+                        print(f"    LOG: {l[:120]}")
+                        break
+                last_log = log_s
+
+            if 'DEAD' in stats:
+                print("\n  PROCESS DIED!")
+                out, _ = run(ssh, "tail -60 /home/fabian/comfyui.log 2>/dev/null")
+                print(out)
+                break
+
+            if 'Prompt executed in' in log_s:
+                print(f"\n  SUCCESS! Image generated!")
+                out, _ = run(ssh, "tail -30 /home/fabian/comfyui.log 2>/dev/null")
+                print(out)
+                break
+
+            if 'Traceback' in log_s or 'RuntimeError' in log_s:
+                print("\n  ERROR detected!")
+                out, _ = run(ssh, "tail -60 /home/fabian/comfyui.log 2>/dev/null")
+                print(out)
+                break
+
+        # ── 7. Check output ──
+        print("\n=== Output files ===")
+        out, _ = run(ssh, "ls -lah /home/fabian/ComfyUI/output/ 2>/dev/null")
+        print(out.strip())
+
+    finally:
+        ssh.close()
+        print("\nSSH connection closed.")
+
+if __name__ == '__main__':
+    main()
@@ -0,0 +1,316 @@
+#!/usr/bin/env python3
+"""
+Fix: Add MIOpen fast-find, pre-warm GPU, restart ComfyUI, generate.
+All GPU ops confirmed working. The hang is likely cold MIOpen kernel cache.
+Single SSH connection.
+"""
+import paramiko, json, time, sys
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+for attempt in range(5):
+    try:
+        ssh.connect('192.168.178.150', username='fabian',
+                    key_filename=r'C:\Users\fabia\.ssh\id_ed25519', timeout=10)
+        break
+    except Exception as e:
+        print(f"  SSH attempt {attempt+1}/5: {e}")
+        time.sleep(10)
+else:
+    print("FATAL: Cannot connect"); sys.exit(1)
+
+def run(cmd, timeout=300):
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    return out, err
+
+try:
+    # 1. Kill any lingering ComfyUI
+    print("=== Kill any ComfyUI ===")
+    run("pkill -9 -f 'python3 main.py' 2>/dev/null; sleep 1")
+    print("  Done")
+
+    # 2. Pre-warm MIOpen kernel cache with DiT-like operations
+    print("\n=== Pre-warming MIOpen kernel cache ===")
+    print("  This compiles HIP kernels that Z-Image-Turbo will need.")
+    print("  First run after reboot is slow (kernel compilation)...")
+    
+    warmup_code = r"""
+import torch, torch.nn as nn, time
+
+# Simulate Z-Image-Turbo DiT operations
+device = 'cuda'
+dtype = torch.float16
+
+print("Warming up HIP kernels for DiT inference...")
+t0 = time.time()
+
+# 1. Linear layers (DiT blocks)
+print("  Linear layers...", end=" ", flush=True)
+for size in [(1024,1024), (4096,1024), (1024,4096)]:
+    l = nn.Linear(*size).to(device, dtype)
+    x = torch.randn(1, 64, size[0], device=device, dtype=dtype)
+    y = l(x)
+    del l, x, y
+torch.cuda.synchronize()
+print(f"{time.time()-t0:.1f}s")
+
+# 2. Attention (SDPA - the core of DiT)
+print("  Scaled dot-product attention...", end=" ", flush=True)
+t1 = time.time()
+for heads in [8, 16, 24]:
+    q = torch.randn(1, heads, 64, 64, device=device, dtype=dtype)
+    k = torch.randn(1, heads, 64, 64, device=device, dtype=dtype)    
+    v = torch.randn(1, heads, 64, 64, device=device, dtype=dtype)
+    y = torch.nn.functional.scaled_dot_product_attention(q, k, v)
+    del q, k, v, y
+torch.cuda.synchronize()
+print(f"{time.time()-t1:.1f}s")
+
+# 3. LayerNorm / RMSNorm
+print("  Normalization layers...", end=" ", flush=True)
+t1 = time.time()
+for dim in [1024, 2048, 4096]:
+    ln = nn.LayerNorm(dim).to(device, dtype)
+    x = torch.randn(1, 64, dim, device=device, dtype=dtype)
+    y = ln(x)
+    del ln, x, y
+torch.cuda.synchronize()
+print(f"{time.time()-t1:.1f}s")
+
+# 4. Conv2d (VAE-like, but we'll run VAE on CPU)
+print("  Conv2d layers...", end=" ", flush=True)
+t1 = time.time()
+for ch in [64, 128, 256]:
+    c = nn.Conv2d(ch, ch, 3, padding=1).to(device, dtype)
+    x = torch.randn(1, ch, 32, 32, device=device, dtype=dtype)
+    y = c(x)
+    del c, x, y
+torch.cuda.synchronize()
+print(f"{time.time()-t1:.1f}s")
+
+# 5. Full mini-DiT forward pass simulation  
+print("  Mini-DiT forward pass simulation...", end=" ", flush=True)
+t1 = time.time()
+hidden = 1024
+seq_len = 256
+heads = 16
+head_dim = hidden // heads
+# Simulate a DiT block
+x = torch.randn(1, seq_len, hidden, device=device, dtype=dtype)
+norm = nn.LayerNorm(hidden).to(device, dtype)
+qkv = nn.Linear(hidden, hidden*3).to(device, dtype)
+proj = nn.Linear(hidden, hidden).to(device, dtype)
+ff1 = nn.Linear(hidden, hidden*4).to(device, dtype)
+ff2 = nn.Linear(hidden*4, hidden).to(device, dtype)
+for step in range(3):
+    h = norm(x)
+    q, k, v = qkv(h).chunk(3, dim=-1)
+    q = q.view(1, seq_len, heads, head_dim).transpose(1,2)
+    k = k.view(1, seq_len, heads, head_dim).transpose(1,2)
+    v = v.view(1, seq_len, heads, head_dim).transpose(1,2)
+    attn = torch.nn.functional.scaled_dot_product_attention(q, k, v)
+    attn = attn.transpose(1,2).contiguous().view(1, seq_len, hidden)
+    x = x + proj(attn)
+    x = x + ff2(torch.nn.functional.gelu(ff1(norm(x))))
+torch.cuda.synchronize()
+print(f"{time.time()-t1:.1f}s")
+
+total = time.time() - t0
+print(f"\nGPU kernel warmup complete in {total:.1f}s")
+print(f"VRAM used: {torch.cuda.memory_allocated()//1048576} MB")
+torch.cuda.empty_cache()
+print(f"VRAM after cleanup: {torch.cuda.memory_allocated()//1048576} MB")
+print("WARMUP_DONE")
+"""
+    # Write warmup script
+    run(f"cat > /tmp/gpu_warmup.py << 'PYEOF'\n{warmup_code}\nPYEOF")
+    
+    out, err = run("bash -c 'source ~/comfyui-env/bin/activate && "
+                   "HSA_OVERRIDE_GFX_VERSION=10.1.0 HSA_ENABLE_SDMA=0 "
+                   "MIOPEN_FIND_MODE=3 MIOPEN_FIND_ENFORCE=3 "
+                   "python3 /tmp/gpu_warmup.py' 2>&1", timeout=300)
+    print(out.strip())
+    if 'WARMUP_DONE' not in out:
+        print(f"  WARNING: Warmup may have failed")
+        print(f"  STDERR: {err.strip()[:500]}")
+
+    # 3. Update startup script with MIOpen settings
+    print("\n=== Update startup script ===")
+    script = """#!/bin/bash
+# BC-250 ComfyUI Launcher — GPU (ROCm) + CPU VAE
+
+# GPU identity (Cyan Skillfish gfx1013 -> gfx1010)
+export HSA_OVERRIDE_GFX_VERSION=10.1.0
+export HIP_VISIBLE_DEVICES=0
+export HSA_ENABLE_SDMA=0
+export HSA_TOOLS_LIB=""
+export HSA_TOOLS_REPORT_LOAD_FAILURE=0
+
+# MIOpen: fast kernel selection (avoid long auto-tune on first run)
+export MIOPEN_FIND_MODE=3
+export MIOPEN_FIND_ENFORCE=3
+
+# Use all 12 CPU cores for CPU-side work (dequant, text encoding)
+export OMP_NUM_THREADS=12
+export MKL_NUM_THREADS=12
+export OPENBLAS_NUM_THREADS=12
+
+# Activate venv
+source /home/fabian/comfyui-env/bin/activate
+cd /home/fabian/ComfyUI
+
+# --novram: offload models to RAM, send layers to GPU one at a time
+# --force-fp16: fp16 diffusion to halve VRAM usage
+# --cpu-vae: VAE decode on CPU (GPU hangs on full VAE forward pass)
+exec python3 main.py \\
+    --listen 0.0.0.0 --port 8188 \\
+    --novram \\
+    --force-fp16 \\
+    --cpu-vae
+"""
+    run(f"cat > /home/fabian/start_comfyui.sh << 'HEREDOC_END'\n{script}HEREDOC_END\n"
+        f"chmod +x /home/fabian/start_comfyui.sh")
+    print("  Written with MIOpen fast-find + --novram --force-fp16 --cpu-vae")
+
+    # 4. Launch ComfyUI
+    print("\n=== Launch ComfyUI ===")
+    run("nohup /home/fabian/start_comfyui.sh > /home/fabian/comfyui.log 2>&1 &")
+    time.sleep(2)
+    
+    print("  Waiting for server...")
+    for i in range(50):
+        time.sleep(3)
+        out, _ = run("curl -s -o /dev/null -w '%{http_code}' http://localhost:8188/ 2>/dev/null || echo 0")
+        if out.strip() == '200':
+            print(f"  Server ready ({(i+1)*3}s)")
+            break
+        if i % 5 == 4:
+            log, _ = run("tail -2 /home/fabian/comfyui.log 2>/dev/null")
+            last = [l.strip() for l in log.strip().split('\n') if l.strip()]
+            print(f"  [{(i+1)*3}s] ... {last[-1][:80] if last else ''}")
+    else:
+        print("  Timeout!")
+        out, _ = run("tail -40 /home/fabian/comfyui.log")
+        print(out)
+        sys.exit(1)
+
+    # 5. Submit workflow
+    print("\n=== Submit workflow (512x512, 8 steps) ===")
+    workflow = {
+        "prompt": {
+            "1": {"class_type": "UnetLoaderGGUF",
+                  "inputs": {"unet_name": "z_image_turbo-Q5_K_S.gguf"}},
+            "2": {"class_type": "CLIPLoaderGGUF",
+                  "inputs": {"clip_name": "Qwen3-4B.i1-Q5_K_S.gguf", "type": "qwen_image"}},
+            "3": {"class_type": "VAELoader",
+                  "inputs": {"vae_name": "ae.safetensors"}},
+            "4": {"class_type": "CLIPTextEncode",
+                  "inputs": {"text": "A red fox in a snowy forest, photorealistic", "clip": ["2", 0]}},
+            "5": {"class_type": "CLIPTextEncode",
+                  "inputs": {"text": "", "clip": ["2", 0]}},
+            "6": {"class_type": "EmptyLatentImage",
+                  "inputs": {"width": 512, "height": 512, "batch_size": 1}},
+            "7": {"class_type": "KSampler",
+                  "inputs": {"model": ["1", 0], "seed": 42, "steps": 8, "cfg": 1.0,
+                             "sampler_name": "euler", "scheduler": "simple",
+                             "positive": ["4", 0], "negative": ["5", 0],
+                             "latent_image": ["6", 0], "denoise": 1.0}},
+            "8": {"class_type": "VAEDecode",
+                  "inputs": {"samples": ["7", 0], "vae": ["3", 0]}},
+            "9": {"class_type": "SaveImage",
+                  "inputs": {"images": ["8", 0], "filename_prefix": "ZImageTurbo_test"}}
+        }
+    }
+    
+    wf_json = json.dumps(workflow)
+    # Write workflow to file to avoid shell escaping issues
+    run(f"cat > /tmp/zimage_wf.json << 'JSONEOF'\n{wf_json}\nJSONEOF")
+    out, _ = run("curl -s -X POST http://localhost:8188/prompt "
+                 "-H 'Content-Type: application/json' "
+                 "-d @/tmp/zimage_wf.json")
+    try:
+        resp = json.loads(out.strip())
+        if 'error' in resp:
+            print(f"  API ERROR: {resp['error']}")
+            sys.exit(1)
+        print(f"  Prompt ID: {resp.get('prompt_id')}")
+    except:
+        print(f"  Response: {out.strip()[:500]}")
+
+    # 6. Monitor — wait up to 15 minutes (first run can be slow due to kernel cache)
+    print("\n=== Monitoring generation (GPU kernels may compile on first step) ===")
+    last_log = ""
+    for i in range(60):  # up to 15 minutes
+        time.sleep(15)
+        
+        stats, _ = run("bash -c '"
+            "PID=$(pgrep -f \"python3 main.py\" | head -1); "
+            "if [ -n \"$PID\" ]; then "
+            "  CPU=$(ps -p $PID -o %cpu --no-headers); "
+            "  RSS=$(ps -p $PID -o rss --no-headers); "
+            "  LOAD=$(cat /proc/loadavg | cut -d\" \" -f1); "
+            "  GPU_T=$(cat /sys/class/drm/card1/device/hwmon/hwmon*/temp1_input 2>/dev/null || echo 0); "
+            "  echo \"CPU:${CPU}% RSS:$((RSS/1024))M LOAD:${LOAD} GPU:$((GPU_T/1000))C\"; "
+            "else echo DEAD; fi'")
+        
+        log, _ = run("tail -12 /home/fabian/comfyui.log 2>/dev/null")
+        log_s = log.strip()
+        
+        elapsed = (i+1) * 15
+        m, s = divmod(elapsed, 60)
+        stats_s = stats.strip()
+        print(f"  [{m}m{s:02d}s] {stats_s}")
+        
+        # Show new log content
+        if log_s != last_log:
+            for line in reversed(log_s.split('\n')):
+                l = line.strip()
+                if l and not l.startswith('FETCH') and not l.startswith('[DEPRECATION') and not l.startswith('[ComfyUI-Manager]'):
+                    print(f"    LOG: {l[:120]}")
+                    break
+            last_log = log_s
+        
+        if 'DEAD' in stats_s:
+            print("\n  PROCESS DIED!")
+            out, _ = run("tail -60 /home/fabian/comfyui.log")
+            print(out)
+            break
+        
+        if 'Prompt executed in' in log_s:
+            print(f"\n  SUCCESS! Image generated!")
+            out, _ = run("tail -25 /home/fabian/comfyui.log")
+            print(out)
+            break
+        
+        if 'Traceback' in log_s or 'RuntimeError' in log_s:
+            print("\n  ERROR detected!")
+            out, _ = run("tail -60 /home/fabian/comfyui.log")
+            print(out)
+            break
+
+    # 7. Check output
+    print("\n=== Output files ===")
+    out, _ = run("ls -lah ~/ComfyUI/output/ 2>/dev/null")
+    print(out.strip())
+    
+    # 8. Check history
+    out, _ = run("curl -s http://localhost:8188/history 2>/dev/null")
+    try:
+        h = json.loads(out)
+        for pid, info in h.items():
+            status = info.get('status', {})
+            outputs = info.get('outputs', {})
+            print(f"\n  Prompt {pid[:12]}...: status={status}")
+            if outputs:
+                for nid, nout in outputs.items():
+                    if isinstance(nout, dict) and 'images' in nout:
+                        for img in nout['images']:
+                            print(f"    Image: {img.get('filename', 'unknown')}")
+    except:
+        pass
+
+finally:
+    ssh.close()
+    print("\nSSH connection closed.")
@@ -0,0 +1,73 @@
+#!/usr/bin/env python3
+"""Monitor ComfyUI generation progress - poll every 20s."""
+import paramiko, json, time
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=30):
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    return stdout.read().decode()
+
+last_log_hash = ""
+for i in range(90):  # up to 30 minutes
+    time.sleep(20)
+    
+    # Process stats
+    stats = run("bash -c 'PID=$(pgrep -f \"python3 main.py\" | head -1); "
+        "if [ -n \"$PID\" ]; then "
+        "  CPU=$(ps -p $PID -o %cpu --no-headers); "
+        "  MEM=$(ps -p $PID -o rss --no-headers); "
+        "  echo \"CPU:${CPU}% RSS:$((MEM/1024))MB LOAD:$(cat /proc/loadavg | cut -d\" \" -f1-3)\"; "
+        "else echo DEAD; fi'").strip()
+    
+    # GPU
+    gpu = run("bash -c 'rocm-smi --showuse --showmemuse 2>/dev/null | grep -E \"GPU|%\" | head -5 || echo no-gpu'").strip()
+    
+    # Log tail
+    log = run("tail -10 /home/fabian/comfyui.log 2>/dev/null").strip()
+    log_hash = hash(log)
+    
+    elapsed = (i+1) * 20
+    mins = elapsed // 60
+    secs = elapsed % 60
+    print(f"[{mins}m{secs:02d}s] {stats}")
+    
+    # Show GPU line
+    for line in gpu.split('\n'):
+        if '%' in line or 'GPU' in line:
+            print(f"  GPU: {line.strip()}")
+            break
+    
+    # Show last meaningful log line
+    if log_hash != last_log_hash:
+        for line in reversed(log.split('\n')):
+            l = line.strip()
+            if l and not l.startswith('FETCH'):
+                print(f"  LOG: {l}")
+                break
+        last_log_hash = log_hash
+    
+    if 'DEAD' in stats:
+        print("\nPROCESS DIED!")
+        print(run("tail -60 /home/fabian/comfyui.log 2>/dev/null"))
+        break
+    
+    if 'Prompt executed in' in log:
+        print(f"\nSUCCESS! Image generated!")
+        print(run("tail -30 /home/fabian/comfyui.log 2>/dev/null"))
+        print("\n=== OUTPUT FILES ===")
+        print(run("ls -lah /home/fabian/ComfyUI/output/ 2>/dev/null"))
+        break
+    
+    if 'Traceback' in log or 'CUDA out of memory' in log or 'RuntimeError' in log:
+        print(f"\nERROR!")
+        print(run("tail -60 /home/fabian/comfyui.log 2>/dev/null"))
+        break
+
+else:
+    print("\nTimed out after 30 minutes")
+    print(run("tail -40 /home/fabian/comfyui.log 2>/dev/null"))
+
+ssh.close()
@@ -0,0 +1,58 @@
+import paramiko, time, json
+
+k = paramiko.Ed25519Key.from_private_key_file(r'C:\Users\fabia\.ssh\id_ed25519')
+c = paramiko.SSHClient()
+c.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+c.connect('192.168.178.150', username='fabian', pkey=k, timeout=10)
+print("Connected. Monitoring sampling progress...")
+
+for i in range(40):  # up to 10 minutes
+    # Check log for sampling progress
+    _, o, _ = c.exec_command('tail -5 /tmp/comfyui.log 2>/dev/null')
+    log = o.read().decode(errors='replace').strip()
+    
+    # Check GPU temp
+    _, o, _ = c.exec_command('cat /sys/class/drm/card0/device/hwmon/hwmon*/temp1_input 2>/dev/null')
+    temp = o.read().decode().strip()
+    temp_c = int(temp) // 1000 if temp.isdigit() else '?'
+    
+    # Check GPU usage
+    _, o, _ = c.exec_command('cat /sys/class/drm/card0/device/gpu_busy_percent 2>/dev/null')
+    gpu_pct = o.read().decode().strip()
+    
+    # Check output dir for generated images
+    _, o, _ = c.exec_command('ls ~/ComfyUI/output/*.png 2>/dev/null')
+    files = o.read().decode().strip()
+    
+    # Check queue
+    _, o, _ = c.exec_command('curl -s http://127.0.0.1:8188/queue 2>/dev/null')
+    queue_raw = o.read().decode().strip()
+    
+    # Parse last log line for progress
+    last_line = log.split('\n')[-1] if log else ''
+    print(f"[{i*15:>3}s] GPU:{gpu_pct}% temp:{temp_c}C | {last_line[-120:]}")
+    
+    if files:
+        print(f"\n*** IMAGE GENERATED! ***")
+        print(f"Files: {files}")
+        # Get last 10 lines of log for timing info
+        _, o, _ = c.exec_command('tail -10 /tmp/comfyui.log 2>/dev/null')
+        print(o.read().decode(errors='replace'))
+        break
+    
+    try:
+        qdata = json.loads(queue_raw)
+        running = len(qdata.get('queue_running', []))
+        pending = len(qdata.get('queue_pending', []))
+        if running == 0 and pending == 0 and i > 3:
+            print("\nQueue empty - job finished or failed. Last 30 log lines:")
+            _, o, _ = c.exec_command('tail -30 /tmp/comfyui.log 2>/dev/null')
+            print(o.read().decode(errors='replace'))
+            break
+    except:
+        pass
+    
+    time.sleep(15)
+
+c.close()
+print("Monitor done.")
@@ -0,0 +1,227 @@
+"""Fix: Back to --novram (proven working for GPU sampling) + --cpu-vae."""
+import paramiko, time, json, textwrap
+
+k = paramiko.Ed25519Key.from_private_key_file(r'C:\Users\fabia\.ssh\id_ed25519')
+c = paramiko.SSHClient()
+c.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+c.connect('192.168.178.150', username='fabian', pkey=k, timeout=15)
+
+def sh(cmd, timeout=60):
+    chan = c.get_transport().open_session()
+    chan.settimeout(timeout)
+    # Use bash array to avoid quoting issues
+    chan.exec_command('/bin/bash -l -c ' + "'" + cmd.replace("'", "'\\''") + "'")
+    out = b""
+    while True:
+        try:
+            chunk = chan.recv(65536)
+            if not chunk: break
+            out += chunk
+        except: break
+    chan.close()
+    return out.decode(errors='replace').strip()
+
+def sftp_write(path, content):
+    sftp = c.open_sftp()
+    with sftp.open(path, 'w') as f:
+        f.write(content)
+    sftp.close()
+
+def sftp_read(path):
+    sftp = c.open_sftp()
+    with sftp.open(path, 'r') as f:
+        data = f.read().decode(errors='replace')
+    sftp.close()
+    return data
+
+# ---- STEP 1: Kill ----
+print("STEP 1: Kill ComfyUI")
+sh('pkill -9 -f "python3.*main.py" 2>/dev/null; sleep 2')
+print("  Killed.")
+
+# ---- STEP 2: Write launcher ----
+print("\nSTEP 2: Write launcher with --novram (PROVEN to work on this APU)")
+launcher = textwrap.dedent("""\
+    #!/bin/bash
+    export HSA_OVERRIDE_GFX_VERSION=10.1.0
+    export HIP_VISIBLE_DEVICES=0
+    export HSA_ENABLE_SDMA=0
+    export HSA_TOOLS_LIB=""
+    export HSA_TOOLS_REPORT_LOAD_FAILURE=0
+    export PYTORCH_HIP_ALLOC_CONF=expandable_segments:False
+    export OMP_NUM_THREADS=12
+    export MKL_NUM_THREADS=12
+    export OPENBLAS_NUM_THREADS=12
+    export MIOPEN_FIND_MODE=1
+    
+    cd ~/ComfyUI
+    source ~/comfyui-env/bin/activate
+    
+    # --novram = model weights on CPU, GPU only for compute (correct for shared-memory APU)
+    # --cpu-vae = VAE decode on CPU (fixes known hang on this GPU)
+    # --force-fp16 = half precision to save memory
+    exec python3 main.py \\
+        --listen 0.0.0.0 --port 8188 \\
+        --novram \\
+        --force-fp16 \\
+        --cpu-vae \\
+        --disable-smart-memory
+""")
+sftp_write('/tmp/run_comfyui.sh', launcher)
+sh('chmod +x /tmp/run_comfyui.sh')
+print("  Written: --novram --force-fp16 --cpu-vae --disable-smart-memory")
+
+# ---- STEP 3: Start ----
+print("\nSTEP 3: Start ComfyUI")
+sh('rm -f /tmp/comfyui.log; touch /tmp/comfyui.log')
+sh('nohup /tmp/run_comfyui.sh > /tmp/comfyui.log 2>&1 &')
+time.sleep(3)
+pid = sh('pgrep -f "python3.*main.py"')
+if not pid:
+    print("  FAILED!")
+    print(sftp_read('/tmp/comfyui.log'))
+    c.close()
+    exit(1)
+print(f"  PID: {pid}")
+
+# ---- STEP 4: Wait for HTTP 200 ----
+print("\nSTEP 4: Wait for server ready", end='', flush=True)
+for i in range(120):
+    code = sh('curl -s -o /dev/null -w "%{http_code}" http://127.0.0.1:8188/ 2>/dev/null', timeout=5)
+    if '200' in code:
+        print(f" READY ({i*2}s)")
+        break
+    # Check if process died
+    alive = sh('pgrep -f "python3.*main.py" >/dev/null && echo Y || echo N', timeout=5)
+    if alive == 'N':
+        print("\n  Process died!")
+        print(sftp_read('/tmp/comfyui.log'))
+        c.close()
+        exit(1)
+    if i % 10 == 0 and i > 0:
+        log = sftp_read('/tmp/comfyui.log')
+        lines = [l for l in log.split('\n') if l.strip()]
+        print(f"\n  [{i*2}s] {lines[-1][:80] if lines else '...'}", end='', flush=True)
+    else:
+        print('.', end='', flush=True)
+    time.sleep(2)
+else:
+    print("\n  TIMEOUT!")
+    print(sftp_read('/tmp/comfyui.log')[-1000:])
+    c.close()
+    exit(1)
+
+# Verify startup flags
+log = sftp_read('/tmp/comfyui.log')
+if 'NO_VRAM' in log or 'NOVRAM' in log.upper():
+    print("  Confirmed: NOVRAM mode (GPU compute only, model on CPU)")
+for line in log.split('\n'):
+    if 'vram state' in line.lower():
+        print(f"  {line.strip()}")
+    if 'Device:' in line:
+        print(f"  {line.strip()}")
+
+# ---- STEP 5: Submit workflow ----
+print("\nSTEP 5: Submit workflow")
+workflow = {
+    "prompt": {
+        "1": {"class_type": "UnetLoaderGGUF", "inputs": {"unet_name": "z_image_turbo-Q5_K_S.gguf"}},
+        "2": {"class_type": "CLIPLoaderGGUF", "inputs": {"clip_name": "Qwen3-4B.i1-Q5_K_S.gguf", "type": "qwen_image"}},
+        "3": {"class_type": "VAELoader", "inputs": {"vae_name": "ae.safetensors"}},
+        "4": {"class_type": "CLIPTextEncode", "inputs": {"text": "A red fox in a snowy forest, photorealistic", "clip": ["2", 0]}},
+        "5": {"class_type": "EmptyLatentImage", "inputs": {"width": 512, "height": 512, "batch_size": 1}},
+        "6": {"class_type": "KSampler", "inputs": {
+            "model": ["1", 0], "positive": ["4", 0], "negative": ["4", 0],
+            "latent_image": ["5", 0], "seed": 42, "steps": 8, "cfg": 1.0,
+            "sampler_name": "euler", "scheduler": "simple", "denoise": 1.0
+        }},
+        "7": {"class_type": "VAEDecode", "inputs": {"samples": ["6", 0], "vae": ["3", 0]}},
+        "8": {"class_type": "SaveImage", "inputs": {"images": ["7", 0], "filename_prefix": "ZImageTurbo_GPU"}}
+    }
+}
+sftp_write('/tmp/wf.json', json.dumps(workflow))
+resp = sh('curl -s -X POST http://127.0.0.1:8188/prompt -H "Content-Type: application/json" -d @/tmp/wf.json', timeout=10)
+print(f"  Response: {resp[:200]}")
+if 'prompt_id' not in resp:
+    print("  FAILED!")
+    c.close()
+    exit(1)
+prompt_id = json.loads(resp).get('prompt_id', '?')
+print(f"  Prompt ID: {prompt_id}")
+
+# ---- STEP 6: Monitor ----
+print("\nSTEP 6: Monitor (expect GPU power >100W during sampling)")
+t0 = time.time()
+sampling_seen = False
+
+for i in range(200):
+    elapsed = int(time.time() - t0)
+    
+    gpu_temp = sh('cat /sys/class/drm/card0/device/hwmon/hwmon*/temp1_input 2>/dev/null', timeout=5)
+    temp_c = int(gpu_temp) // 1000 if gpu_temp.isdigit() else '?'
+    
+    # Read log via SFTP to avoid shell issues
+    try:
+        log = sftp_read('/tmp/comfyui.log')
+    except:
+        log = ''
+    
+    lines = log.strip().split('\n')
+    # Find last meaningful line (skip manager spam)
+    last = ''
+    for line in reversed(lines):
+        if 'FETCH ComfyRegistry' not in line and 'All startup tasks' not in line and 'FETCH DATA' not in line and line.strip():
+            last = line.strip()
+            break
+    
+    # Detect sampling progress
+    for line in lines:
+        if '/8' in line and 'it/s' in line:
+            sampling_seen = True
+    
+    print(f"  [{elapsed:>4}s] {temp_c}C | {last[-100:]}")
+    
+    # Check for output image
+    imgs = sh('ls ~/ComfyUI/output/*.png 2>/dev/null', timeout=5)
+    if imgs:
+        print(f"\n  *** IMAGE GENERATED! ***")
+        print(f"  File: {imgs}")
+        print(f"  Total: {elapsed}s")
+        # Show last 15 lines
+        for line in lines[-15:]:
+            if line.strip():
+                print(f"  {line.strip()}")
+        break
+    
+    # Check queue
+    q = sh('curl -s http://127.0.0.1:8188/queue 2>/dev/null', timeout=5)
+    try:
+        qd = json.loads(q)
+        if not qd.get('queue_running') and not qd.get('queue_pending') and elapsed > 30:
+            time.sleep(3)
+            imgs = sh('ls ~/ComfyUI/output/*.png 2>/dev/null', timeout=5)
+            if imgs:
+                print(f"\n  *** IMAGE GENERATED! ***")
+                print(f"  File: {imgs}")
+            else:
+                print(f"\n  Queue empty, no image. Error in log:")
+                for line in lines[-20:]:
+                    if line.strip():
+                        print(f"  {line}")
+            break
+    except:
+        pass
+    
+    # Check process alive
+    alive = sh('pgrep -f "python3.*main.py" >/dev/null && echo Y || echo N', timeout=5)
+    if alive == 'N':
+        print(f"\n  *** CRASHED ***")
+        for line in lines[-30:]:
+            if line.strip():
+                print(f"  {line}")
+        break
+    
+    time.sleep(15)
+
+c.close()
+print("\nDone.")
@@ -0,0 +1,76 @@
+"""Patch ComfyUI: Force VAE to GPU even in --novram mode. Restart and test."""
+import paramiko, time, json, textwrap
+
+k = paramiko.Ed25519Key.from_private_key_file(r'C:\Users\fabia\.ssh\id_ed25519')
+c = paramiko.SSHClient()
+c.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+c.connect('192.168.178.150', username='fabian', pkey=k, timeout=15)
+sftp = c.open_sftp()
+
+def sh(cmd, timeout=60):
+    chan = c.get_transport().open_session()
+    chan.settimeout(timeout)
+    chan.exec_command('/bin/bash -l -c ' + "'" + cmd.replace("'", "'\\''") + "'")
+    out = b""
+    while True:
+        try:
+            chunk = chan.recv(65536)
+            if not chunk: break
+            out += chunk
+        except: break
+    chan.close()
+    return out.decode(errors='replace').strip()
+
+# =============================================
+# STEP 1: Kill ComfyUI
+# =============================================
+print("1) Kill ComfyUI")
+sh('pkill -9 -f "python3.*main.py" 2>/dev/null; sleep 2')
+
+# =============================================
+# STEP 2: Read and patch model_management.py
+# =============================================
+print("2) Patch model_management.py — force VAE to GPU")
+
+# First, read the file to understand the structure
+with sftp.open('/home/fabian/ComfyUI/comfy/model_management.py', 'r') as f:
+    mgmt = f.read().decode()
+
+print(f"   File size: {len(mgmt)} bytes")
+
+# Find the vae_offload_device function
+# In ComfyUI, with NO_VRAM, vae_offload_device() returns CPU
+# We need to make it return GPU instead
+
+# Also find vae_dtype — it's set to float32 by default, we want float16
+
+# Let's search for relevant functions
+for i, line in enumerate(mgmt.split('\n')):
+    if 'def vae_offload_device' in line or 'def vae_dtype' in line or 'def vae_device' in line:
+        print(f"   Line {i+1}: {line.strip()}")
+
+# Also check what functions exist
+found = []
+for i, line in enumerate(mgmt.split('\n')):
+    if line.startswith('def ') or (line.startswith('    ') and 'def ' in line[:12]):
+        if 'vae' in line.lower():
+            found.append((i+1, line.strip()))
+for ln, l in found:
+    print(f"   L{ln}: {l}")
+
+# Let's read the specific area around these functions
+lines = mgmt.split('\n')
+
+# Find and show context around vae functions
+for keyword in ['vae_offload_device', 'vae_dtype', 'vae_device']:
+    for i, line in enumerate(lines):
+        if f'def {keyword}' in line:
+            start = max(0, i-2)
+            end = min(len(lines), i+15)
+            print(f"\n   --- {keyword} (L{i+1}) ---")
+            for j in range(start, end):
+                print(f"   {j+1:>5}: {lines[j]}")
+
+sftp.close()
+c.close()
+print("\n   Reading complete. Will patch next.")
@@ -0,0 +1,242 @@
+"""Fix: keep UNet+VAE both on GPU (shared memory). No offloading."""
+import paramiko, time, json
+
+k = paramiko.Ed25519Key.from_private_key_file(r'C:\Users\fabia\.ssh\id_ed25519')
+c = paramiko.SSHClient()
+c.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+c.connect('192.168.178.150', username='fabian', pkey=k, timeout=10)
+sftp = c.open_sftp()
+
+def sh(cmd, timeout=30):
+    chan = c.get_transport().open_session()
+    chan.settimeout(timeout)
+    chan.exec_command('/bin/bash -l -c ' + "'" + cmd.replace("'", "'\\''") + "'")
+    out = b""
+    while True:
+        try:
+            chunk = chan.recv(65536)
+            if not chunk: break
+            out += chunk
+        except: break
+    chan.close()
+    return out.decode(errors='replace').strip()
+
+# Kill first
+sh('pkill -9 -f "python3.*main.py" 2>/dev/null; sleep 2')
+print("Killed ComfyUI")
+
+# Read model_management.py
+with sftp.open('/home/fabian/ComfyUI/comfy/model_management.py', 'r') as f:
+    code = f.read().decode()
+
+lines = code.split('\n')
+
+# Show current offload functions to understand exact code
+print("\n=== Finding offload functions ===")
+for i, line in enumerate(lines):
+    if 'def unet_offload_device' in line or 'def vae_offload_device' in line:
+        print(f"\n--- {line.strip()} at line {i+1} ---")
+        for j in range(i, min(i+10, len(lines))):
+            print(f"  {j+1}: {lines[j]}")
+
+# ============ PATCH unet_offload_device ============
+# Current: returns CPU unless HIGH_VRAM
+# Fix: also return GPU for SHARED (APU shared memory = no point offloading)
+old_unet = None
+new_unet = None
+
+for i, line in enumerate(lines):
+    if 'def unet_offload_device' in line:
+        # Grab the function body (next ~6 lines)
+        chunk = '\n'.join(lines[i:i+8])
+        print(f"\n=== unet_offload_device chunk ===\n{chunk}")
+        
+        # The function checks HIGH_VRAM only. Add SHARED.
+        if 'HIGH_VRAM' in chunk and 'SHARED' not in chunk:
+            old_unet = chunk
+            new_unet = chunk.replace(
+                'vram_state == VRAMState.HIGH_VRAM',
+                'vram_state == VRAMState.HIGH_VRAM or vram_state == VRAMState.SHARED'
+            )
+            print(f"\n  -> Will patch to include SHARED")
+        elif 'SHARED' in chunk:
+            print(f"\n  -> Already patched for SHARED")
+        break
+
+# ============ PATCH vae_offload_device ============
+# Current: returns CPU unless --gpu-only
+# Fix: also return GPU for SHARED
+old_vae = None
+new_vae = None
+
+for i, line in enumerate(lines):
+    if 'def vae_offload_device' in line:
+        chunk = '\n'.join(lines[i:i+8])
+        print(f"\n=== vae_offload_device chunk ===\n{chunk}")
+        
+        if 'args.gpu_only' in chunk and 'SHARED' not in chunk:
+            old_vae = chunk
+            new_vae = chunk.replace(
+                'args.gpu_only',
+                'args.gpu_only or vram_state == VRAMState.SHARED'
+            )
+            print(f"\n  -> Will patch to include SHARED")
+        elif 'SHARED' in chunk:
+            print(f"\n  -> Already patched for SHARED")
+        break
+
+# Also check text_encoder_offload_device
+for i, line in enumerate(lines):
+    if 'def text_encoder_offload_device' in line:
+        chunk = '\n'.join(lines[i:i+8])
+        print(f"\n=== text_encoder_offload_device chunk ===\n{chunk}")
+        break
+
+# Apply patches
+patched = False
+if old_unet and new_unet:
+    code = code.replace(old_unet, new_unet)
+    patched = True
+    print("\n[OK] Patched unet_offload_device")
+
+if old_vae and new_vae:
+    code = code.replace(old_vae, new_vae)
+    patched = True
+    print("[OK] Patched vae_offload_device")
+
+if patched:
+    # Backup and write
+    sh('cp /home/fabian/ComfyUI/comfy/model_management.py /home/fabian/ComfyUI/comfy/model_management.py.bak2')
+    with sftp.open('/home/fabian/ComfyUI/comfy/model_management.py', 'w') as f:
+        f.write(code)
+    print("[OK] Written to disk")
+else:
+    print("[INFO] No patches needed (already applied or code changed)")
+
+# Verify
+with sftp.open('/home/fabian/ComfyUI/comfy/model_management.py', 'r') as f:
+    verify = f.read().decode()
+for i, line in enumerate(verify.split('\n')):
+    if 'def unet_offload_device' in line or 'def vae_offload_device' in line:
+        print(f"\n--- VERIFY {line.strip()} ---")
+        for j in range(i, min(i+8, len(verify.split(chr(10))))):
+            print(f"  {j+1}: {verify.split(chr(10))[j]}")
+
+# ============ RESTART ============
+print("\n=== RESTARTING ===")
+sh('rm -f /tmp/comfyui.log; rm -f ~/ComfyUI/output/ZImageTurbo_GPU*.png')
+sh('nohup /tmp/run_comfyui.sh > /tmp/comfyui.log 2>&1 &')
+time.sleep(3)
+pid = sh('pgrep -f "python3.*main.py"')
+print(f"PID: {pid}")
+
+# Wait for ready
+print("Waiting for HTTP", end='', flush=True)
+for i in range(90):
+    r = sh('curl -s -o /dev/null -w "%{http_code}" http://127.0.0.1:8188/ 2>/dev/null', timeout=5)
+    if '200' in r:
+        print(f" OK ({i*2}s)")
+        break
+    print('.', end='', flush=True)
+    time.sleep(2)
+else:
+    print(" TIMEOUT")
+
+# Check SHARED mode active
+with sftp.open('/tmp/comfyui.log', 'r') as f:
+    log = f.read().decode(errors='replace')
+for line in log.split('\n'):
+    s = line.strip()
+    if any(x in s.lower() for x in ['vram state', 'shared', 'device:', 'total vram']):
+        print(f"  {s}")
+
+# Submit workflow
+print("\nSubmitting workflow...")
+wf = {"prompt": {
+    "1": {"class_type": "UnetLoaderGGUF", "inputs": {"unet_name": "z_image_turbo-Q5_K_S.gguf"}},
+    "2": {"class_type": "CLIPLoaderGGUF", "inputs": {"clip_name": "Qwen3-4B.i1-Q5_K_S.gguf", "type": "qwen_image"}},
+    "3": {"class_type": "VAELoader", "inputs": {"vae_name": "ae.safetensors"}},
+    "4": {"class_type": "CLIPTextEncode", "inputs": {"text": "A red fox in a snowy forest, photorealistic", "clip": ["2", 0]}},
+    "5": {"class_type": "EmptyLatentImage", "inputs": {"width": 512, "height": 512, "batch_size": 1}},
+    "6": {"class_type": "KSampler", "inputs": {"model": ["1", 0], "positive": ["4", 0], "negative": ["4", 0],
+        "latent_image": ["5", 0], "seed": 99999, "steps": 8, "cfg": 1.0,
+        "sampler_name": "euler", "scheduler": "simple", "denoise": 1.0}},
+    "7": {"class_type": "VAEDecode", "inputs": {"samples": ["6", 0], "vae": ["3", 0]}},
+    "8": {"class_type": "SaveImage", "inputs": {"images": ["7", 0], "filename_prefix": "ZImageTurbo_GPU"}}
+}}
+with sftp.open('/tmp/wf.json', 'w') as f:
+    f.write(json.dumps(wf))
+resp = sh('curl -s -X POST http://127.0.0.1:8188/prompt -H "Content-Type: application/json" -d @/tmp/wf.json')
+print(f"  {resp[:150]}")
+
+# Monitor - watch for UNet+VAE both on GPU, fast VAE
+print("\nMonitoring...")
+t0 = time.time()
+for i in range(120):
+    el = int(time.time() - t0)
+    
+    try:
+        with sftp.open('/tmp/comfyui.log', 'r') as f:
+            log = f.read().decode(errors='replace')
+    except: log = ''
+    
+    # GPU usage
+    gpu = sh('cat /sys/class/drm/card0/device/gpu_busy_percent 2>/dev/null || echo ?', timeout=5)
+    temp = sh('cat /sys/class/drm/card0/device/hwmon/hwmon*/temp1_input 2>/dev/null', timeout=5)
+    tc = int(temp)//1000 if temp.isdigit() else '?'
+    
+    # Get latest progress line
+    last_progress = ''
+    last_line = ''
+    for line in log.split('\n'):
+        s = line.strip()
+        if '/8' in s and ('it/s' in s or 's/it' in s): last_progress = s
+        if 'loaded' in s.lower() or 'VAE' in s or 'Requested' in s or 'Prompt executed' in s:
+            last_line = s
+        if s and 'FETCH' not in s: last_line = s
+    
+    status = last_progress or last_line
+    print(f"  [{el:>3}s] GPU:{gpu}% {tc}C | {status[-100:]}")
+    
+    # Check for output
+    imgs = sh('ls ~/ComfyUI/output/ZImageTurbo_GPU*.png 2>/dev/null', timeout=5)
+    if imgs:
+        print(f"\n  *** IMAGE DONE! *** {imgs}")
+        print(f"  Wall time: {el}s")
+        # Print key log lines
+        for line in log.split('\n'):
+            s = line.strip()
+            if any(x in s for x in ['loaded', 'load device', 'offload device', 'Prompt executed', '/8', 'Requested', 'VAE']):
+                if 'FETCH' not in s:
+                    print(f"    {s}")
+        break
+    
+    # Check queue empty
+    if el > 30:
+        q = sh('curl -s http://127.0.0.1:8188/queue 2>/dev/null', timeout=5)
+        try:
+            qd = json.loads(q)
+            if not qd.get('queue_running') and not qd.get('queue_pending'):
+                time.sleep(3)
+                imgs = sh('ls ~/ComfyUI/output/ZImageTurbo_GPU*.png 2>/dev/null', timeout=5)
+                if imgs:
+                    print(f"\n  *** DONE: {imgs} ***")
+                else:
+                    print(f"\n  Queue empty, no image. Error?")
+                    for line in log.split('\n')[-20:]:
+                        if line.strip() and 'FETCH' not in line: print(f"    {line.strip()}")
+                break
+        except: pass
+    
+    # Check alive
+    if sh('pgrep -f "python3.*main.py" >/dev/null && echo Y || echo N', timeout=5) == 'N':
+        print("\n  CRASHED!")
+        for line in log.split('\n')[-20:]:
+            if line.strip(): print(f"    {line.strip()}")
+        break
+    
+    time.sleep(5)
+
+sftp.close()
+c.close()
+print("\nDone.")
@@ -0,0 +1,31 @@
+import paramiko
+k = paramiko.Ed25519Key.from_private_key_file(r'C:\Users\fabia\.ssh\id_ed25519')
+c = paramiko.SSHClient()
+c.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+c.connect('192.168.178.150', username='fabian', pkey=k, timeout=10)
+sftp = c.open_sftp()
+
+# Full log
+with sftp.open('/tmp/comfyui.log', 'r') as f:
+    log = f.read().decode(errors='replace')
+for line in log.split('\n'):
+    s = line.strip()
+    if not s or 'FETCH' in s or 'startup tasks' in s or 'DEPRECATION' in s: continue
+    print(s)
+
+# Process + GPU
+chan = c.get_transport().open_session()
+chan.settimeout(10)
+chan.exec_command('/bin/bash -c "echo === PROC ===; ps aux | grep main.py | grep -v grep; echo === GPU ===; rocm-smi 2>/dev/null | head -12; echo === VRAM ===; rocm-smi --showmeminfo vram 2>/dev/null"')
+o = b""
+while True:
+    try:
+        ch = chan.recv(65536)
+        if not ch: break
+        o += ch
+    except: break
+chan.close()
+print(o.decode(errors='replace'))
+
+sftp.close()
+c.close()
@@ -0,0 +1,23 @@
+#!/usr/bin/env python3
+"""Quick status check."""
+import paramiko
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+def run(cmd):
+    _, so, se = ssh.exec_command(cmd, timeout=15)
+    return so.read().decode()
+
+print("=== PROCESS ===")
+print(run("bash -c 'PID=$(pgrep -f \"python3 main.py\" | head -1); ps -p $PID -o pid,%cpu,%mem,nlwp --no-headers 2>/dev/null; echo LOAD: $(cat /proc/loadavg)'"))
+
+print("=== GPU ===")
+print(run("HSA_OVERRIDE_GFX_VERSION=10.1.0 rocm-smi 2>/dev/null | tail -8"))
+
+print("=== LOG (last 15) ===")
+print(run("tail -15 /home/fabian/comfyui.log 2>/dev/null"))
+
+print("=== MEMORY ===")
+print(run("free -h"))
+
+ssh.close()
@@ -0,0 +1,47 @@
+"""Read raw ComfyUI log - no filtering, no quoting issues."""
+import paramiko
+
+k = paramiko.Ed25519Key.from_private_key_file(r'C:\Users\fabia\.ssh\id_ed25519')
+c = paramiko.SSHClient()
+c.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+c.connect('192.168.178.150', username='fabian', pkey=k, timeout=10)
+
+# Read the ENTIRE log via SFTP - no shell, no grep, no quoting
+sftp = c.open_sftp()
+try:
+    with sftp.open('/tmp/comfyui.log', 'r') as f:
+        log = f.read().decode(errors='replace')
+    
+    lines = log.split('\n')
+    print(f"Total log lines: {len(lines)}")
+    print()
+    
+    # Print everything that's NOT ComfyUI-Manager registry spam
+    for line in lines:
+        if 'FETCH ComfyRegistry' in line:
+            continue
+        if 'All startup tasks' in line:
+            continue
+        if line.strip():
+            print(line)
+except Exception as e:
+    print(f"Error reading log: {e}")
+finally:
+    sftp.close()
+
+# Also check: is the process actually using GPU memory?
+chan = c.get_transport().open_session()
+chan.settimeout(10)
+chan.exec_command('/bin/bash -c "rocm-smi --showmeminfo vram 2>/dev/null"')
+out = b""
+while True:
+    try:
+        chunk = chan.recv(65536)
+        if not chunk: break
+        out += chunk
+    except: break
+chan.close()
+print("\n=== VRAM Info ===")
+print(out.decode(errors='replace'))
+
+c.close()
@@ -0,0 +1,64 @@
+"""Read more of model_management.py — find how --novram affects GPU compute."""
+import paramiko
+
+k = paramiko.Ed25519Key.from_private_key_file(r'C:\Users\fabia\.ssh\id_ed25519')
+c = paramiko.SSHClient()
+c.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+c.connect('192.168.178.150', username='fabian', pkey=k, timeout=10)
+sftp = c.open_sftp()
+
+with sftp.open('/home/fabian/ComfyUI/comfy/model_management.py', 'r') as f:
+    mgmt = f.read().decode()
+
+lines = mgmt.split('\n')
+
+# Find all functions related to device/offload
+print("=== KEY FUNCTIONS ===")
+for keyword in ['unet_offload_device', 'unet_device', 'NO_VRAM', 'should_use', 'get_torch_device', 
+                'def text_encoder_device', 'def text_encoder_offload', 'VRAMState']:
+    for i, line in enumerate(lines):
+        if keyword in line and ('def ' in line or 'class ' in line or '=' in line[:50]):
+            print(f"  L{i+1}: {line.strip()[:100]}")
+
+# Show VRAMState enum
+print("\n=== VRAMState ===")
+for i, line in enumerate(lines):
+    if 'class VRAMState' in line or (i > 0 and 'VRAMState' in lines[i-1] and 'class' in lines[i-1]):
+        for j in range(i, min(i+15, len(lines))):
+            print(f"  {j+1}: {lines[j]}")
+        break
+
+# Show unet_offload_device
+print("\n=== unet_offload_device ===")
+for i, line in enumerate(lines):
+    if 'def unet_offload_device' in line:
+        for j in range(max(0,i-2), min(i+15, len(lines))):
+            print(f"  {j+1}: {lines[j]}")
+
+# Show text_encoder functions
+print("\n=== text_encoder_device ===")
+for i, line in enumerate(lines):
+    if 'def text_encoder_device' in line:
+        for j in range(max(0,i-2), min(i+12, len(lines))):
+            print(f"  {j+1}: {lines[j]}")
+
+print("\n=== text_encoder_offload_device ===")
+for i, line in enumerate(lines):
+    if 'def text_encoder_offload_device' in line:
+        for j in range(max(0,i-2), min(i+12, len(lines))):
+            print(f"  {j+1}: {lines[j]}")
+
+# Show how NO_VRAM is used in loading logic
+print("\n=== NO_VRAM usage in model loading ===")
+for i, line in enumerate(lines):
+    if 'NO_VRAM' in line:
+        print(f"  L{i+1}: {line.strip()[:120]}")
+
+# Show the VRAM state setting logic
+print("\n=== vram_state assignment ===")
+for i, line in enumerate(lines):
+    if 'vram_state' in line and ('=' in line) and 'VRAMState' in line:
+        print(f"  L{i+1}: {line.strip()[:120]}")
+
+sftp.close()
+c.close()
@@ -0,0 +1,43 @@
+"""Read the SHARED vram state logic and CLI args."""
+import paramiko
+
+k = paramiko.Ed25519Key.from_private_key_file(r'C:\Users\fabia\.ssh\id_ed25519')
+c = paramiko.SSHClient()
+c.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+c.connect('192.168.178.150', username='fabian', pkey=k, timeout=10)
+sftp = c.open_sftp()
+
+# Read model_management.py around line 440-470 (where SHARED is set)
+with sftp.open('/home/fabian/ComfyUI/comfy/model_management.py', 'r') as f:
+    mgmt = f.read().decode()
+lines = mgmt.split('\n')
+
+print("=== L430-475: VRAM state setting ===")
+for i in range(429, min(475, len(lines))):
+    print(f"  {i+1}: {lines[i]}")
+
+print("\n=== L750-790: NO_VRAM model loading ===")
+for i in range(749, min(790, len(lines))):
+    print(f"  {i+1}: {lines[i]}")
+
+print("\n=== L850-870: Smart memory / offload ===")
+for i in range(849, min(870, len(lines))):
+    print(f"  {i+1}: {lines[i]}")
+
+# Check CLI args for shared memory
+with sftp.open('/home/fabian/ComfyUI/comfy/cli_args.py', 'r') as f:
+    cli = f.read().decode()
+
+print("\n=== CLI args with 'shared' or 'SHARED' ===")
+for i, line in enumerate(cli.split('\n')):
+    if 'shared' in line.lower():
+        print(f"  L{i+1}: {line.strip()}")
+
+# Check what --gpu-only does
+print("\n=== CLI args with 'gpu_only' ===")
+for i, line in enumerate(cli.split('\n')):
+    if 'gpu_only' in line.lower() or 'gpu-only' in line.lower():
+        print(f"  L{i+1}: {line.strip()}")
+
+sftp.close()
+c.close()
@@ -0,0 +1,55 @@
+"""Reconnect, check patches, fix missing ones, restart."""
+import paramiko, time, json
+
+k = paramiko.Ed25519Key.from_private_key_file(r'C:\Users\fabia\.ssh\id_ed25519')
+c = paramiko.SSHClient()
+c.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+c.connect('192.168.178.150', username='fabian', pkey=k, timeout=10)
+sftp = c.open_sftp()
+
+def sh(cmd, t=30):
+    ch = c.get_transport().open_session()
+    ch.settimeout(t)
+    ch.exec_command('/bin/bash -l -c ' + "'" + cmd.replace("'", "'\\''") + "'")
+    o = b""
+    while True:
+        try:
+            d = ch.recv(65536)
+            if not d: break
+            o += d
+        except: break
+    ch.close()
+    return o.decode(errors='replace').strip()
+
+# Read current state
+with sftp.open('/home/fabian/ComfyUI/comfy/model_management.py', 'r') as f:
+    code = f.read().decode()
+lines = code.split('\n')
+
+# Show ALL offload functions
+for fname in ['unet_offload_device', 'vae_offload_device', 'text_encoder_offload_device']:
+    for i, line in enumerate(lines):
+        if f'def {fname}' in line:
+            print(f"\n=== {fname} (L{i+1}) ===")
+            for j in range(i, min(i+10, len(lines))):
+                print(f"  {j+1}: {lines[j]}")
+            break
+
+# Also show unet_inital_load_device  
+for i, line in enumerate(lines):
+    if 'def unet_inital_load_device' in line:
+        print(f"\n=== unet_inital_load_device (L{i+1}) ===")
+        for j in range(i, min(i+10, len(lines))):
+            print(f"  {j+1}: {lines[j]}")
+        break
+
+# Show SHARED patch
+for i, line in enumerate(lines):
+    if 'COMFYUI_SHARED_MEMORY' in line:
+        print(f"\n=== SHARED patch (L{i+1}) ===")
+        for j in range(max(0,i-2), min(i+5, len(lines))):
+            print(f"  {j+1}: {lines[j]}")
+
+sftp.close()
+c.close()
+print("\nDone reading.")
@@ -0,0 +1,203 @@
+"""Fix: Remove --novram so ComfyUI actually uses the GPU for compute."""
+import paramiko, time, json, textwrap
+
+k = paramiko.Ed25519Key.from_private_key_file(r'C:\Users\fabia\.ssh\id_ed25519')
+c = paramiko.SSHClient()
+c.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+c.connect('192.168.178.150', username='fabian', pkey=k, timeout=15)
+
+sftp = c.open_sftp()
+
+def sh(cmd, timeout=60):
+    chan = c.get_transport().open_session()
+    chan.settimeout(timeout)
+    chan.exec_command('/bin/bash -l -c ' + "'" + cmd.replace("'", "'\\''") + "'")
+    out = b""
+    while True:
+        try:
+            chunk = chan.recv(65536)
+            if not chunk: break
+            out += chunk
+        except: break
+    chan.close()
+    return out.decode(errors='replace').strip()
+
+# STEP 1: Kill
+print("STEP 1: Kill")
+sh('pkill -9 -f "python3.*main.py" 2>/dev/null; sleep 2')
+print("  Done")
+
+# STEP 2: Write new launcher - NO --novram, NO --lowvram
+# ComfyUI sees 7602MB VRAM → will use NORMAL_VRAM mode → GPU compute
+print("\nSTEP 2: New launcher (NO memory flags = auto GPU)")
+launcher = textwrap.dedent("""\
+    #!/bin/bash
+    # GPU
+    export HSA_OVERRIDE_GFX_VERSION=10.1.0
+    export HIP_VISIBLE_DEVICES=0
+    export HSA_ENABLE_SDMA=0
+    export HSA_TOOLS_LIB=""
+    export HSA_TOOLS_REPORT_LOAD_FAILURE=0
+    export PYTORCH_HIP_ALLOC_CONF=expandable_segments:False
+    # Threading
+    export OMP_NUM_THREADS=12
+    export MKL_NUM_THREADS=12
+    export OPENBLAS_NUM_THREADS=12
+    # MIOpen - use fast mode, persistent cache
+    export MIOPEN_FIND_MODE=3
+    export MIOPEN_LOG_LEVEL=3
+    
+    cd ~/ComfyUI
+    source ~/comfyui-env/bin/activate
+    
+    # NO --novram, NO --lowvram = ComfyUI auto-detects 7602MB VRAM = GPU compute
+    # --force-fp16 = half precision saves memory
+    exec python3 main.py \\
+        --listen 0.0.0.0 --port 8188 \\
+        --force-fp16
+""")
+with sftp.open('/tmp/run_comfyui.sh', 'w') as f:
+    f.write(launcher)
+sh('chmod +x /tmp/run_comfyui.sh')
+print("  Flags: --force-fp16 ONLY (auto VRAM mode)")
+
+# STEP 3: Start
+print("\nSTEP 3: Start ComfyUI")
+sh('rm -f /tmp/comfyui.log; touch /tmp/comfyui.log')
+sh('nohup /tmp/run_comfyui.sh > /tmp/comfyui.log 2>&1 &')
+time.sleep(3)
+pid = sh('pgrep -f "python3.*main.py"')
+print(f"  PID: {pid}")
+
+# STEP 4: Wait for ready
+print("\nSTEP 4: Wait for HTTP ready", end='', flush=True)
+for i in range(120):
+    code = sh('curl -s -o /dev/null -w "%{http_code}" http://127.0.0.1:8188/ 2>/dev/null', timeout=5)
+    if '200' in code:
+        print(f" READY ({i*2}s)")
+        break
+    if i % 10 == 0 and i > 0:
+        try:
+            with sftp.open('/tmp/comfyui.log', 'r') as f:
+                log = f.read().decode(errors='replace')
+            lines = [l.strip() for l in log.split('\n') if l.strip() and 'FETCH' not in l and 'DEPRECATION' not in l]
+            print(f"\n  [{i*2}s] {lines[-1][:80] if lines else '...'}", end='', flush=True)
+        except: pass
+    else:
+        print('.', end='', flush=True)
+    time.sleep(2)
+else:
+    print("\n  TIMEOUT!")
+    exit(1)
+
+# Show VRAM state
+with sftp.open('/tmp/comfyui.log', 'r') as f:
+    log = f.read().decode(errors='replace')
+for line in log.split('\n'):
+    s = line.strip()
+    if any(k in s for k in ['vram state', 'Device:', 'Total VRAM', 'VRAM', 'pytorch version']):
+        print(f"  {s}")
+
+# STEP 5: Submit workflow
+print("\nSTEP 5: Submit workflow")
+workflow = {
+    "prompt": {
+        "1": {"class_type": "UnetLoaderGGUF", "inputs": {"unet_name": "z_image_turbo-Q5_K_S.gguf"}},
+        "2": {"class_type": "CLIPLoaderGGUF", "inputs": {"clip_name": "Qwen3-4B.i1-Q5_K_S.gguf", "type": "qwen_image"}},
+        "3": {"class_type": "VAELoader", "inputs": {"vae_name": "ae.safetensors"}},
+        "4": {"class_type": "CLIPTextEncode", "inputs": {"text": "A red fox in a snowy forest, photorealistic", "clip": ["2", 0]}},
+        "5": {"class_type": "EmptyLatentImage", "inputs": {"width": 512, "height": 512, "batch_size": 1}},
+        "6": {"class_type": "KSampler", "inputs": {
+            "model": ["1", 0], "positive": ["4", 0], "negative": ["4", 0],
+            "latent_image": ["5", 0], "seed": 42, "steps": 8, "cfg": 1.0,
+            "sampler_name": "euler", "scheduler": "simple", "denoise": 1.0
+        }},
+        "7": {"class_type": "VAEDecode", "inputs": {"samples": ["6", 0], "vae": ["3", 0]}},
+        "8": {"class_type": "SaveImage", "inputs": {"images": ["7", 0], "filename_prefix": "ZImageTurbo_GPU"}}
+    }
+}
+with sftp.open('/tmp/wf.json', 'w') as f:
+    f.write(json.dumps(workflow))
+resp = sh('curl -s -X POST http://127.0.0.1:8188/prompt -H "Content-Type: application/json" -d @/tmp/wf.json', timeout=10)
+print(f"  {resp[:150]}")
+if 'prompt_id' not in resp:
+    print("  FAILED!")
+    exit(1)
+
+# STEP 6: Monitor - focus on GPU usage and sampling speed
+print("\nSTEP 6: Monitor")
+print("  First step may be slow (MIOpen kernel compilation). Be patient.")
+t0 = time.time()
+for i in range(200):
+    elapsed = int(time.time() - t0)
+    
+    gpu_temp = sh('cat /sys/class/drm/card0/device/hwmon/hwmon*/temp1_input 2>/dev/null', timeout=5)
+    temp_c = int(gpu_temp) // 1000 if gpu_temp.isdigit() else '?'
+    
+    try:
+        with sftp.open('/tmp/comfyui.log', 'r') as f:
+            log = f.read().decode(errors='replace')
+    except: log = ''
+    
+    # Find sampling progress and last meaningful line
+    sampling = ''
+    last = ''
+    for line in log.split('\n'):
+        s = line.strip()
+        if '/8' in s and ('it/s' in s or 's/it' in s):
+            sampling = s
+        if s and 'FETCH' not in s and 'startup tasks' not in s and 'DEPRECATION' not in s:
+            last = s
+    
+    display = sampling if sampling else last[-100:]
+    print(f"  [{elapsed:>4}s] {temp_c}C | {display}")
+    
+    # Check output
+    imgs = sh('ls ~/ComfyUI/output/ZImageTurbo_GPU*.png 2>/dev/null', timeout=5)
+    if imgs:
+        # Get timing from log
+        exec_time = ''
+        for line in log.split('\n'):
+            if 'Prompt executed in' in line:
+                exec_time = line.strip()
+        print(f"\n  *** IMAGE GENERATED! ***")
+        print(f"  File: {imgs}")
+        print(f"  {exec_time}")
+        print(f"  Wall time: {elapsed}s")
+        
+        # Show vram state and model loading details
+        for line in log.split('\n'):
+            s = line.strip()
+            if any(k in s for k in ['loaded completely', 'loaded partially', 'vram state', '/8']):
+                print(f"  {s}")
+        break
+    
+    # Check queue empty
+    q = sh('curl -s http://127.0.0.1:8188/queue 2>/dev/null', timeout=5)
+    try:
+        qd = json.loads(q)
+        if not qd.get('queue_running') and not qd.get('queue_pending') and elapsed > 30:
+            time.sleep(3)
+            imgs = sh('ls ~/ComfyUI/output/ZImageTurbo_GPU*.png 2>/dev/null', timeout=5)
+            if imgs:
+                print(f"\n  *** IMAGE GENERATED: {imgs} ***")
+            else:
+                print(f"\n  Queue empty, no image. Log errors:")
+                for line in log.split('\n')[-25:]:
+                    if line.strip(): print(f"  {line.strip()}")
+            break
+    except: pass
+    
+    # Process alive?
+    alive = sh('pgrep -f "python3.*main.py" >/dev/null && echo Y || echo N', timeout=5)
+    if alive == 'N':
+        print(f"\n  *** CRASHED ***")
+        for line in log.split('\n')[-30:]:
+            if line.strip(): print(f"  {line.strip()}")
+        break
+    
+    time.sleep(15)
+
+sftp.close()
+c.close()
+print("\nDone.")
@@ -0,0 +1,133 @@
+"""Recreate launcher (lost on reboot) and start ComfyUI."""
+import paramiko, time, json, sys
+
+k = paramiko.Ed25519Key.from_private_key_file(r'C:\Users\fabia\.ssh\id_ed25519')
+c = paramiko.SSHClient()
+c.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+c.connect('192.168.178.150', username='fabian', pkey=k, timeout=10)
+sftp = c.open_sftp()
+
+def sh(cmd, t=30):
+    stdin, stdout, stderr = c.exec_command(f"bash -lc '{cmd}'", timeout=t)
+    return stdout.read().decode(errors='replace').strip()
+
+# Recreate launcher (wiped by reboot since /tmp)
+launcher = """#!/bin/bash
+export HSA_OVERRIDE_GFX_VERSION=10.1.0
+export HIP_VISIBLE_DEVICES=0
+export HSA_ENABLE_SDMA=0
+export HSA_TOOLS_LIB=""
+export HSA_TOOLS_REPORT_LOAD_FAILURE=0
+export PYTORCH_HIP_ALLOC_CONF=expandable_segments:False
+export OMP_NUM_THREADS=12
+export MKL_NUM_THREADS=12
+export OPENBLAS_NUM_THREADS=12
+export MIOPEN_FIND_MODE=3
+export COMFYUI_SHARED_MEMORY=1
+
+cd ~/ComfyUI
+source ~/comfyui-env/bin/activate
+
+exec python3 main.py \\
+    --listen 0.0.0.0 --port 8188 \\
+    --force-fp16 \\
+    --fp16-vae
+"""
+with sftp.open('/tmp/run_comfyui.sh', 'w') as f:
+    f.write(launcher)
+sh("chmod +x /tmp/run_comfyui.sh")
+print("Launcher recreated")
+
+# Verify patches
+p = sh("grep -c COMFYUI_SHARED_MEMORY ~/ComfyUI/comfy/model_management.py")
+print(f"SHARED patch refs: {p}")
+p2 = sh("grep 'def unet_offload_device' -A2 ~/ComfyUI/comfy/model_management.py | head -3")
+print(f"unet_offload: {p2}")
+p3 = sh("grep 'def vae_offload_device' -A2 ~/ComfyUI/comfy/model_management.py | head -3")
+print(f"vae_offload: {p3}")
+
+# Start
+sh("rm -f /tmp/comfyui.log")
+sh("nohup /tmp/run_comfyui.sh > /tmp/comfyui.log 2>&1 &")
+time.sleep(4)
+pid = sh("pgrep -f 'python3.*main.py'")
+print(f"PID: {pid}")
+
+if not pid:
+    print("FAILED! Log:")
+    print(sh("cat /tmp/comfyui.log 2>/dev/null"))
+    c.close()
+    sys.exit(1)
+
+# Wait for HTTP
+print("Waiting for HTTP...", end='', flush=True)
+for i in range(90):
+    code = sh('curl -s -o /dev/null -w "%{http_code}" http://127.0.0.1:8188/ 2>/dev/null')
+    if '200' in code:
+        print(f" ready ({i*2}s)")
+        break
+    print('.', end='', flush=True)
+    time.sleep(2)
+else:
+    print(" TIMEOUT")
+    print(sh("tail -30 /tmp/comfyui.log"))
+    c.close()
+    sys.exit(1)
+
+# Log state
+log = sh("cat /tmp/comfyui.log")
+for l in log.split('\n'):
+    s = l.strip()
+    if any(x in s for x in ['vram state', 'SHARED', 'Device:', 'Total VRAM']):
+        print(f"  {s}")
+
+# Submit
+wf = {"prompt": {
+    "1": {"class_type": "UnetLoaderGGUF", "inputs": {"unet_name": "z_image_turbo-Q5_K_S.gguf"}},
+    "2": {"class_type": "CLIPLoaderGGUF", "inputs": {"clip_name": "Qwen3-4B.i1-Q5_K_S.gguf", "type": "qwen_image"}},
+    "3": {"class_type": "VAELoader", "inputs": {"vae_name": "ae.safetensors"}},
+    "4": {"class_type": "CLIPTextEncode", "inputs": {"text": "A red fox in a snowy forest, photorealistic", "clip": ["2", 0]}},
+    "5": {"class_type": "EmptyLatentImage", "inputs": {"width": 512, "height": 512, "batch_size": 1}},
+    "6": {"class_type": "KSampler", "inputs": {"model": ["1", 0], "positive": ["4", 0], "negative": ["4", 0],
+        "latent_image": ["5", 0], "seed": 99999, "steps": 8, "cfg": 1.0,
+        "sampler_name": "euler", "scheduler": "simple", "denoise": 1.0}},
+    "7": {"class_type": "VAEDecode", "inputs": {"samples": ["6", 0], "vae": ["3", 0]}},
+    "8": {"class_type": "SaveImage", "inputs": {"images": ["7", 0], "filename_prefix": "ZImageTurbo_GPU"}}
+}}
+with sftp.open('/tmp/wf.json', 'w') as f:
+    f.write(json.dumps(wf))
+resp = sh('curl -s -X POST http://127.0.0.1:8188/prompt -H "Content-Type: application/json" -d @/tmp/wf.json')
+print(f"Submitted: {resp[:100]}")
+
+# Monitor
+t0 = time.time()
+shown = set()
+for _ in range(200):
+    el = int(time.time() - t0)
+    log = sh("cat /tmp/comfyui.log 2>/dev/null")
+
+    for l in log.split('\n'):
+        s = l.strip()
+        if s not in shown and any(x in s for x in ['/8', 'loaded completely', 'load device', 'offload device', 'Requested to load', 'Prompt executed', 'Error', 'OOM']):
+            if 'FETCH' not in s and 'audio' not in s and 'split attention' not in s:
+                gpu = sh("cat /sys/class/drm/card0/device/gpu_busy_percent 2>/dev/null")
+                print(f"  [{el:>3}s] GPU:{gpu}% | {s[-110:]}")
+                shown.add(s)
+
+    if 'Prompt executed' in log:
+        print(f"\n*** DONE in {el}s! ***")
+        imgs = sh("ls ~/ComfyUI/output/ZImageTurbo_GPU*.png 2>/dev/null")
+        print(f"  Images: {imgs}")
+        break
+
+    alive = sh("pgrep -c -f 'python3.*main.py' 2>/dev/null")
+    if alive == '0':
+        print(f"\nCRASHED at {el}s!")
+        for l in log.split('\n')[-20:]:
+            if l.strip(): print(f"  {l.strip()}")
+        break
+
+    time.sleep(5)
+
+sftp.close()
+c.close()
@@ -0,0 +1,43 @@
+#!/usr/bin/env python3
+"""Recon the BC-250 for PyTorch/ComfyUI installation."""
+import paramiko, sys
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+cmds = [
+    ("Python version", "python3 --version 2>&1"),
+    ("pip version", "pip --version 2>&1 || pip3 --version 2>&1"),
+    ("Disk space", "df -h / /home 2>&1"),
+    ("RAM", "free -h 2>&1"),
+    ("CPU cores", "nproc 2>&1"),
+    ("ROCm version", "cat /opt/rocm/.info/version 2>/dev/null || echo 'no version file'; ls /opt/rocm/lib/libamdhip64.so* 2>&1"),
+    ("hipcc", "which hipcc 2>&1 && hipcc --version 2>&1 | head -5"),
+    ("rocminfo GPU", "HSA_OVERRIDE_GFX_VERSION=10.1.0 rocminfo 2>&1 | grep -E 'Marketing|gfx|Name:' | head -10"),
+    ("Existing PyTorch", "python3 -c 'import torch; print(torch.__version__); print(torch.cuda.is_available()); print(torch.version.hip)' 2>&1"),
+    ("Existing venvs", "ls -la ~/venv* ~/env* ~/.local/lib/python*/site-packages/torch* 2>&1 | head -20"),
+    ("git version", "git --version 2>&1"),
+    ("cmake version", "cmake --version 2>&1 | head -1"),
+    ("ninja version", "ninja --version 2>&1"),
+    ("Available Python packages", "python3 -m venv --help >/dev/null 2>&1 && echo 'venv OK' || echo 'venv missing'"),
+    ("Swap", "swapon --show 2>&1"),
+    ("GPU device check", "ls -la /dev/kfd /dev/dri/render* 2>&1"),
+    ("Existing ComfyUI", "ls -la ~/ComfyUI 2>&1 || echo 'not found'"),
+    ("pacman cmake/ninja", "pacman -Q cmake ninja 2>&1"),
+]
+
+for label, cmd in cmds:
+    print(f"\n{'='*60}")
+    print(f"  {label}")
+    print(f"{'='*60}")
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=30)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    if out.strip():
+        print(out.strip())
+    if err.strip():
+        print(f"STDERR: {err.strip()}")
+
+ssh.close()
+print("\n\nDone.")
@@ -0,0 +1,34 @@
+#!/usr/bin/env python3
+"""Check Python versions and z-image-turbo info on BC-250."""
+import paramiko
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+cmds = [
+    ("Python 3.12 available?", "pacman -Ss python | grep -E 'python3\\.1[0-3]|python 3\\.' 2>&1 | head -20"),
+    ("All python packages", "pacman -Q | grep python 2>&1 | head -30"),
+    ("pip via python", "python3 -m pip --version 2>&1"),
+    ("pip package", "pacman -Q python-pip 2>&1"),
+    ("check pyenv", "which pyenv 2>&1; pacman -Q pyenv 2>&1"),
+    ("check python3.12", "which python3.12 2>&1; pacman -Q python312 2>&1; ls /usr/bin/python3.1* 2>&1"),
+    ("ninja available", "pacman -Ss '^ninja$' 2>&1 | head -5"),
+    ("check ccache", "which ccache 2>&1; pacman -Q ccache 2>&1"),
+    ("check z-image-turbo", "pacman -Ss z-image 2>&1; pip3 search z-image-turbo 2>&1 || true"),
+    ("check huggingface tools", "pacman -Q | grep -i hugging 2>&1; python3 -c 'import huggingface_hub' 2>&1 || true"),
+]
+
+for label, cmd in cmds:
+    print(f"\n{'='*60}")
+    print(f"  {label}")
+    print(f"{'='*60}")
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=30)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    if out.strip():
+        print(out.strip())
+    if err.strip():
+        print(f"STDERR: {err.strip()}")
+
+ssh.close()
@@ -0,0 +1,93 @@
+"""Post-reboot: start patched ComfyUI, submit, monitor. Single connection."""
+import paramiko, time, json, sys
+
+k = paramiko.Ed25519Key.from_private_key_file(r'C:\Users\fabia\.ssh\id_ed25519')
+c = paramiko.SSHClient()
+c.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+c.connect('192.168.178.150', username='fabian', pkey=k, timeout=10)
+
+def sh(cmd, t=30):
+    stdin, stdout, stderr = c.exec_command(f"bash -lc '{cmd}'", timeout=t)
+    return stdout.read().decode(errors='replace').strip()
+
+# Verify patches survived reboot
+p = sh("grep -c SHARED ~/ComfyUI/comfy/model_management.py")
+print(f"SHARED refs in code: {p}")
+
+# Start ComfyUI
+sh("rm -f /tmp/comfyui.log")
+sh("nohup /tmp/run_comfyui.sh > /tmp/comfyui.log 2>&1 &")
+time.sleep(4)
+pid = sh("pgrep -f main.py")
+print(f"PID: {pid}")
+
+# Wait for HTTP ready
+print("Waiting for HTTP...", end='', flush=True)
+for i in range(90):
+    code = sh('curl -s -o /dev/null -w "%{http_code}" http://127.0.0.1:8188/ 2>/dev/null')
+    if '200' in code:
+        print(f" ready ({i*2}s)")
+        break
+    print('.', end='', flush=True)
+    time.sleep(2)
+else:
+    print(" TIMEOUT")
+    print(sh("tail -30 /tmp/comfyui.log"))
+    sys.exit(1)
+
+# Show startup state
+log = sh("cat /tmp/comfyui.log")
+for l in log.split('\n'):
+    s = l.strip()
+    if any(x in s for x in ['vram state', 'SHARED', 'Device:', 'Total VRAM']):
+        print(f"  {s}")
+
+# Submit workflow
+wf = json.dumps({"prompt": {
+    "1": {"class_type": "UnetLoaderGGUF", "inputs": {"unet_name": "z_image_turbo-Q5_K_S.gguf"}},
+    "2": {"class_type": "CLIPLoaderGGUF", "inputs": {"clip_name": "Qwen3-4B.i1-Q5_K_S.gguf", "type": "qwen_image"}},
+    "3": {"class_type": "VAELoader", "inputs": {"vae_name": "ae.safetensors"}},
+    "4": {"class_type": "CLIPTextEncode", "inputs": {"text": "A red fox in a snowy forest, photorealistic", "clip": ["2", 0]}},
+    "5": {"class_type": "EmptyLatentImage", "inputs": {"width": 512, "height": 512, "batch_size": 1}},
+    "6": {"class_type": "KSampler", "inputs": {"model": ["1", 0], "positive": ["4", 0], "negative": ["4", 0],
+        "latent_image": ["5", 0], "seed": 99999, "steps": 8, "cfg": 1.0,
+        "sampler_name": "euler", "scheduler": "simple", "denoise": 1.0}},
+    "7": {"class_type": "VAEDecode", "inputs": {"samples": ["6", 0], "vae": ["3", 0]}},
+    "8": {"class_type": "SaveImage", "inputs": {"images": ["7", 0], "filename_prefix": "ZImageTurbo_GPU"}}
+}})
+# Write workflow and submit
+sh(f"echo '{wf}' > /tmp/wf.json")
+resp = sh('curl -s -X POST http://127.0.0.1:8188/prompt -H "Content-Type: application/json" -d @/tmp/wf.json')
+print(f"Submitted: {resp[:100]}")
+
+# Monitor
+t0 = time.time()
+shown = set()
+for _ in range(200):
+    el = int(time.time() - t0)
+    log = sh("cat /tmp/comfyui.log 2>/dev/null")
+    
+    for l in log.split('\n'):
+        s = l.strip()
+        if s not in shown and any(x in s for x in ['/8', 'loaded completely', 'load device', 'Requested to load', 'Prompt executed', 'Error', 'OOM']):
+            if 'FETCH' not in s and 'audio' not in s:
+                gpu = sh("cat /sys/class/drm/card0/device/gpu_busy_percent 2>/dev/null")
+                print(f"  [{el:>3}s] GPU:{gpu}% | {s[-110:]}")
+                shown.add(s)
+    
+    if 'Prompt executed' in log:
+        print(f"\n*** DONE in {el}s! ***")
+        imgs = sh("ls ~/ComfyUI/output/ZImageTurbo_GPU*.png 2>/dev/null")
+        print(f"  Images: {imgs}")
+        break
+    
+    alive = sh("pgrep -c -f main.py 2>/dev/null")
+    if alive == '0':
+        print(f"\nCRASHED at {el}s!")
+        for l in log.split('\n')[-15:]:
+            if l.strip(): print(f"  {l.strip()}")
+        break
+    
+    time.sleep(5)
+
+c.close()
@@ -0,0 +1,279 @@
+"""Patch ComfyUI for BC-250 APU: Use SHARED VRAM mode + force fp16 VAE.
+This is the correct mode for an APU where CPU and GPU share the same physical memory."""
+import paramiko, time, json, textwrap
+
+k = paramiko.Ed25519Key.from_private_key_file(r'C:\Users\fabia\.ssh\id_ed25519')
+c = paramiko.SSHClient()
+c.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+c.connect('192.168.178.150', username='fabian', pkey=k, timeout=15)
+sftp = c.open_sftp()
+
+def sh(cmd, timeout=60):
+    chan = c.get_transport().open_session()
+    chan.settimeout(timeout)
+    chan.exec_command('/bin/bash -l -c ' + "'" + cmd.replace("'", "'\\''") + "'")
+    out = b""
+    while True:
+        try:
+            chunk = chan.recv(65536)
+            if not chunk: break
+            out += chunk
+        except: break
+    chan.close()
+    return out.decode(errors='replace').strip()
+
+# ======================================
+# 1) Kill
+# ======================================
+print("1) Kill ComfyUI")
+sh('pkill -9 -f "python3.*main.py" 2>/dev/null; sleep 2')
+
+# ======================================
+# 2) Backup + Patch model_management.py
+# ======================================
+print("2) Patch model_management.py: SHARED mode for APU")
+
+with sftp.open('/home/fabian/ComfyUI/comfy/model_management.py', 'r') as f:
+    code = f.read().decode()
+
+# Backup
+with sftp.open('/home/fabian/ComfyUI/comfy/model_management.py.bak', 'w') as f:
+    f.write(code)
+print("   Backup saved")
+
+# PATCH 1: After MPS sets SHARED, also set SHARED for this AMD APU
+# Current code (L458-462):
+#   if cpu_state != CPUState.GPU:
+#       vram_state = VRAMState.DISABLED
+#   if cpu_state == CPUState.MPS:
+#       vram_state = VRAMState.SHARED
+#
+# We add: if the GPU has shared memory (small dedicated VRAM), set SHARED
+
+old_block = '''if cpu_state == CPUState.MPS:
+    vram_state = VRAMState.SHARED
+
+logging.info(f"Set vram state to: {vram_state.name}")'''
+
+new_block = '''if cpu_state == CPUState.MPS:
+    vram_state = VRAMState.SHARED
+
+# BC-250 APU: shared memory between CPU and GPU. Dedicated VRAM is tiny (512MB)
+# but the full system RAM is accessible to both. SHARED mode loads models
+# directly on GPU (zero-copy for shared memory APUs).
+if cpu_state == CPUState.GPU and vram_state not in (VRAMState.DISABLED, VRAMState.SHARED):
+    try:
+        import os
+        if os.environ.get("COMFYUI_SHARED_MEMORY") == "1":
+            vram_state = VRAMState.SHARED
+            logging.info("Forcing SHARED vram state (COMFYUI_SHARED_MEMORY=1)")
+    except:
+        pass
+
+logging.info(f"Set vram state to: {vram_state.name}")'''
+
+if old_block in code:
+    code = code.replace(old_block, new_block)
+    print("   PATCH 1 applied: COMFYUI_SHARED_MEMORY env var support")
+else:
+    print("   PATCH 1: Could not find exact block, trying alternate...")
+    # Try line by line
+    lines = code.split('\n')
+    for i, line in enumerate(lines):
+        if 'cpu_state == CPUState.MPS' in line and 'SHARED' in lines[i+1] if i+1 < len(lines) else '':
+            # Insert after the MPS block
+            insert_idx = i + 2  # After "vram_state = VRAMState.SHARED"
+            patch_lines = [
+                '',
+                '# BC-250 APU shared memory support',
+                'if cpu_state == CPUState.GPU and vram_state not in (VRAMState.DISABLED, VRAMState.SHARED):',
+                '    try:',
+                '        import os',
+                '        if os.environ.get("COMFYUI_SHARED_MEMORY") == "1":',
+                '            vram_state = VRAMState.SHARED',
+                '            logging.info("Forcing SHARED vram state (COMFYUI_SHARED_MEMORY=1)")',
+                '    except:',
+                '        pass',
+            ]
+            for j, pl in enumerate(patch_lines):
+                lines.insert(insert_idx + j, pl)
+            code = '\n'.join(lines)
+            print(f"   PATCH 1 applied (alternate) at line {insert_idx}")
+            break
+
+# Write patched file
+with sftp.open('/home/fabian/ComfyUI/comfy/model_management.py', 'w') as f:
+    f.write(code)
+print("   File written")
+
+# Verify patch
+with sftp.open('/home/fabian/ComfyUI/comfy/model_management.py', 'r') as f:
+    verify = f.read().decode()
+if 'COMFYUI_SHARED_MEMORY' in verify:
+    print("   Patch verified!")
+else:
+    print("   ERROR: Patch not found in file!")
+
+# ======================================
+# 3) Write launcher with SHARED mode
+# ======================================
+print("3) Write launcher with COMFYUI_SHARED_MEMORY=1")
+launcher = textwrap.dedent("""\
+    #!/bin/bash
+    # GPU
+    export HSA_OVERRIDE_GFX_VERSION=10.1.0
+    export HIP_VISIBLE_DEVICES=0
+    export HSA_ENABLE_SDMA=0
+    export HSA_TOOLS_LIB=""
+    export HSA_TOOLS_REPORT_LOAD_FAILURE=0
+    export PYTORCH_HIP_ALLOC_CONF=expandable_segments:False
+    # Threading
+    export OMP_NUM_THREADS=12
+    export MKL_NUM_THREADS=12
+    export OPENBLAS_NUM_THREADS=12
+    # MIOpen
+    export MIOPEN_FIND_MODE=3
+    # Shared memory APU mode: CPU and GPU share the same physical RAM
+    export COMFYUI_SHARED_MEMORY=1
+    
+    cd ~/ComfyUI
+    source ~/comfyui-env/bin/activate
+    
+    # --force-fp16: half precision (saves memory)
+    # --fp16-vae: VAE in fp16 (320MB instead of 640MB, fits in GPU memory)
+    # SHARED mode: models load directly on GPU, no offloading overhead
+    exec python3 main.py \\
+        --listen 0.0.0.0 --port 8188 \\
+        --force-fp16 \\
+        --fp16-vae
+""")
+with sftp.open('/tmp/run_comfyui.sh', 'w') as f:
+    f.write(launcher)
+sh('chmod +x /tmp/run_comfyui.sh')
+print("   Flags: --force-fp16 --fp16-vae + COMFYUI_SHARED_MEMORY=1")
+
+# ======================================
+# 4) Start
+# ======================================
+print("4) Start ComfyUI")
+sh('rm -f /tmp/comfyui.log; rm -f ~/ComfyUI/output/ZImageTurbo_GPU*.png')
+sh('nohup /tmp/run_comfyui.sh > /tmp/comfyui.log 2>&1 &')
+time.sleep(3)
+pid = sh('pgrep -f "python3.*main.py"')
+print(f"   PID: {pid}")
+
+# ======================================
+# 5) Wait ready
+# ======================================
+print("5) Wait HTTP", end='', flush=True)
+for i in range(90):
+    code = sh('curl -s -o /dev/null -w "%{http_code}" http://127.0.0.1:8188/ 2>/dev/null', timeout=5)
+    if '200' in code:
+        print(f" OK ({i*2}s)")
+        break
+    if i % 10 == 0 and i > 0:
+        try:
+            with sftp.open('/tmp/comfyui.log', 'r') as f:
+                log = f.read().decode(errors='replace')
+            ls = [l.strip() for l in log.split('\n') if l.strip() and 'FETCH' not in l and 'DEPRECATION' not in l]
+            print(f"\n   [{i*2}s] {ls[-1][:80] if ls else ''}", end='', flush=True)
+        except: pass
+    else:
+        print('.', end='', flush=True)
+    time.sleep(2)
+
+# Verify SHARED mode
+with sftp.open('/tmp/comfyui.log', 'r') as f:
+    log = f.read().decode(errors='replace')
+for line in log.split('\n'):
+    s = line.strip()
+    if any(k in s for k in ['vram state', 'SHARED', 'Device:', 'Total VRAM', 'pytorch version']):
+        print(f"   {s}")
+
+# ======================================
+# 6) Submit
+# ======================================
+print("6) Submit")
+wf = {"prompt": {
+    "1": {"class_type": "UnetLoaderGGUF", "inputs": {"unet_name": "z_image_turbo-Q5_K_S.gguf"}},
+    "2": {"class_type": "CLIPLoaderGGUF", "inputs": {"clip_name": "Qwen3-4B.i1-Q5_K_S.gguf", "type": "qwen_image"}},
+    "3": {"class_type": "VAELoader", "inputs": {"vae_name": "ae.safetensors"}},
+    "4": {"class_type": "CLIPTextEncode", "inputs": {"text": "A red fox in a snowy forest, photorealistic", "clip": ["2", 0]}},
+    "5": {"class_type": "EmptyLatentImage", "inputs": {"width": 512, "height": 512, "batch_size": 1}},
+    "6": {"class_type": "KSampler", "inputs": {"model": ["1", 0], "positive": ["4", 0], "negative": ["4", 0],
+        "latent_image": ["5", 0], "seed": 999, "steps": 8, "cfg": 1.0,
+        "sampler_name": "euler", "scheduler": "simple", "denoise": 1.0}},
+    "7": {"class_type": "VAEDecode", "inputs": {"samples": ["6", 0], "vae": ["3", 0]}},
+    "8": {"class_type": "SaveImage", "inputs": {"images": ["7", 0], "filename_prefix": "ZImageTurbo_GPU"}}
+}}
+with sftp.open('/tmp/wf.json', 'w') as f:
+    f.write(json.dumps(wf))
+resp = sh('curl -s -X POST http://127.0.0.1:8188/prompt -H "Content-Type: application/json" -d @/tmp/wf.json')
+print(f"   {resp[:150]}")
+
+# ======================================
+# 7) Monitor
+# ======================================
+print("7) Monitor (SHARED mode = everything on GPU)")
+t0 = time.time()
+for i in range(200):
+    el = int(time.time() - t0)
+    temp = sh('cat /sys/class/drm/card0/device/hwmon/hwmon*/temp1_input 2>/dev/null', timeout=5)
+    tc = int(temp)//1000 if temp.isdigit() else '?'
+    
+    try:
+        with sftp.open('/tmp/comfyui.log', 'r') as f:
+            log = f.read().decode(errors='replace')
+    except: log = ''
+    
+    samp = ''
+    last = ''
+    for line in log.split('\n'):
+        s = line.strip()
+        if '/8' in s and ('it/s' in s or 's/it' in s): samp = s
+        if s and 'FETCH' not in s and 'startup' not in s and 'DEPRECATION' not in s: last = s
+    
+    print(f"   [{el:>4}s] {tc}C | {(samp or last)[-90:]}")
+    
+    imgs = sh('ls ~/ComfyUI/output/ZImageTurbo_GPU*.png 2>/dev/null', timeout=5)
+    if imgs:
+        et = ''
+        for line in log.split('\n'):
+            if 'Prompt executed' in line: et = line.strip()
+        print(f"\n   *** DONE! ***")
+        print(f"   File: {imgs}")
+        print(f"   {et}")
+        print(f"   Wall: {el}s")
+        for line in log.split('\n'):
+            s = line.strip()
+            if any(k in s for k in ['loaded', '/8', 'Prompt executed', 'VAE load', 'Requested']):
+                if 'FETCH' not in s:
+                    print(f"   {s}")
+        break
+    
+    q = sh('curl -s http://127.0.0.1:8188/queue 2>/dev/null', timeout=5)
+    try:
+        qd = json.loads(q)
+        if not qd.get('queue_running') and not qd.get('queue_pending') and el > 30:
+            time.sleep(2)
+            imgs = sh('ls ~/ComfyUI/output/ZImageTurbo_GPU*.png 2>/dev/null', timeout=5)
+            if imgs:
+                print(f"\n   *** DONE: {imgs} ***")
+            else:
+                print(f"\n   Queue empty, no image. Log:")
+                for line in log.split('\n')[-20:]:
+                    if line.strip() and 'FETCH' not in line: print(f"   {line.strip()}")
+            break
+    except: pass
+    
+    if sh('pgrep -f "python3.*main.py" >/dev/null && echo Y || echo N', timeout=5) == 'N':
+        print("\n   CRASHED!")
+        for line in log.split('\n')[-30:]:
+            if line.strip(): print(f"   {line.strip()}")
+        break
+    
+    time.sleep(10)
+
+sftp.close()
+c.close()
+print("\nDone.")
@@ -0,0 +1,216 @@
+"""Quick: check current state, patch for SHARED mode, restart."""
+import paramiko, time, json, textwrap
+
+k = paramiko.Ed25519Key.from_private_key_file(r'C:\Users\fabia\.ssh\id_ed25519')
+c = paramiko.SSHClient()
+c.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+c.connect('192.168.178.150', username='fabian', pkey=k, timeout=10)
+sftp = c.open_sftp()
+
+def sh(cmd, timeout=60):
+    chan = c.get_transport().open_session()
+    chan.settimeout(timeout)
+    chan.exec_command('/bin/bash -l -c ' + "'" + cmd.replace("'", "'\\''") + "'")
+    out = b""
+    while True:
+        try:
+            chunk = chan.recv(65536)
+            if not chunk: break
+            out += chunk
+        except: break
+    chan.close()
+    return out.decode(errors='replace').strip()
+
+# Check current cmdline
+print("=== Current process ===")
+print(sh('ps aux | grep main.py | grep -v grep'))
+
+# Check if patch exists
+with sftp.open('/home/fabian/ComfyUI/comfy/model_management.py', 'r') as f:
+    code = f.read().decode()
+print(f"\n=== Patch status: {'APPLIED' if 'COMFYUI_SHARED_MEMORY' in code else 'NOT applied'} ===")
+
+# Check current vram state in log
+try:
+    with sftp.open('/tmp/comfyui.log', 'r') as f:
+        log = f.read().decode(errors='replace')
+    for line in log.split('\n'):
+        s = line.strip()
+        if 'vram state' in s or 'VAE load' in s or 'Device:' in s:
+            print(f"  {s}")
+except: pass
+
+# =============================
+# APPLY PATCH if not done
+# =============================
+if 'COMFYUI_SHARED_MEMORY' not in code:
+    print("\nApplying SHARED memory patch...")
+    # Backup
+    sh('cp /home/fabian/ComfyUI/comfy/model_management.py /home/fabian/ComfyUI/comfy/model_management.py.bak')
+    
+    old = 'if cpu_state == CPUState.MPS:\n    vram_state = VRAMState.SHARED'
+    new = '''if cpu_state == CPUState.MPS:
+    vram_state = VRAMState.SHARED
+
+# Shared memory APU: CPU+GPU share physical RAM (e.g. AMD BC-250)
+import os as _os
+if _os.environ.get("COMFYUI_SHARED_MEMORY") == "1" and cpu_state == CPUState.GPU:
+    vram_state = VRAMState.SHARED
+    logging.info("SHARED vram: APU shared memory mode enabled")'''
+    
+    if old in code:
+        code = code.replace(old, new)
+        with sftp.open('/home/fabian/ComfyUI/comfy/model_management.py', 'w') as f:
+            f.write(code)
+        print("  Patch applied!")
+    else:
+        print("  ERROR: Could not find patch target. Dumping area:")
+        for i, line in enumerate(code.split('\n')):
+            if 'MPS' in line and 'SHARED' in code.split('\n')[i+1] if i+1 < len(code.split('\n')) else '':
+                for j in range(max(0,i-3), min(i+5, len(code.split('\n')))):
+                    print(f"  {j+1}: {code.split(chr(10))[j]}")
+else:
+    print("  Patch already applied, good.")
+
+# =============================
+# KILL + RESTART with SHARED
+# =============================
+print("\nKilling ComfyUI...")
+sh('pkill -9 -f "python3.*main.py" 2>/dev/null; sleep 2')
+
+launcher = textwrap.dedent("""\
+    #!/bin/bash
+    export HSA_OVERRIDE_GFX_VERSION=10.1.0
+    export HIP_VISIBLE_DEVICES=0
+    export HSA_ENABLE_SDMA=0
+    export HSA_TOOLS_LIB=""
+    export HSA_TOOLS_REPORT_LOAD_FAILURE=0
+    export PYTORCH_HIP_ALLOC_CONF=expandable_segments:False
+    export OMP_NUM_THREADS=12
+    export MKL_NUM_THREADS=12
+    export OPENBLAS_NUM_THREADS=12
+    export MIOPEN_FIND_MODE=3
+    # APU shared memory mode: everything on GPU
+    export COMFYUI_SHARED_MEMORY=1
+    
+    cd ~/ComfyUI
+    source ~/comfyui-env/bin/activate
+    
+    # SHARED mode: models load on GPU directly (shared memory = zero copy)
+    # --force-fp16: half precision for models
+    # --fp16-vae: VAE in fp16 (160MB, fast on GPU)
+    exec python3 main.py \\
+        --listen 0.0.0.0 --port 8188 \\
+        --force-fp16 \\
+        --fp16-vae
+""")
+with sftp.open('/tmp/run_comfyui.sh', 'w') as f:
+    f.write(launcher)
+sh('chmod +x /tmp/run_comfyui.sh')
+
+sh('rm -f /tmp/comfyui.log; rm -f ~/ComfyUI/output/ZImageTurbo_GPU*.png')
+sh('nohup /tmp/run_comfyui.sh > /tmp/comfyui.log 2>&1 &')
+time.sleep(3)
+pid = sh('pgrep -f "python3.*main.py"')
+print(f"Started PID: {pid}")
+
+# Wait for ready
+print("Waiting for HTTP", end='', flush=True)
+for i in range(90):
+    r = sh('curl -s -o /dev/null -w "%{http_code}" http://127.0.0.1:8188/ 2>/dev/null', timeout=5)
+    if '200' in r:
+        print(f" OK ({i*2}s)")
+        break
+    print('.', end='', flush=True)
+    time.sleep(2)
+
+# Verify SHARED mode
+with sftp.open('/tmp/comfyui.log', 'r') as f:
+    log = f.read().decode(errors='replace')
+for line in log.split('\n'):
+    s = line.strip()
+    if any(k in s for k in ['vram state', 'SHARED', 'Device:', 'Total VRAM']):
+        print(f"  {s}")
+
+# Submit test
+print("\nSubmitting test workflow...")
+wf = {"prompt": {
+    "1": {"class_type": "UnetLoaderGGUF", "inputs": {"unet_name": "z_image_turbo-Q5_K_S.gguf"}},
+    "2": {"class_type": "CLIPLoaderGGUF", "inputs": {"clip_name": "Qwen3-4B.i1-Q5_K_S.gguf", "type": "qwen_image"}},
+    "3": {"class_type": "VAELoader", "inputs": {"vae_name": "ae.safetensors"}},
+    "4": {"class_type": "CLIPTextEncode", "inputs": {"text": "A red fox in a snowy forest, photorealistic", "clip": ["2", 0]}},
+    "5": {"class_type": "EmptyLatentImage", "inputs": {"width": 512, "height": 512, "batch_size": 1}},
+    "6": {"class_type": "KSampler", "inputs": {"model": ["1", 0], "positive": ["4", 0], "negative": ["4", 0],
+        "latent_image": ["5", 0], "seed": 12345, "steps": 8, "cfg": 1.0,
+        "sampler_name": "euler", "scheduler": "simple", "denoise": 1.0}},
+    "7": {"class_type": "VAEDecode", "inputs": {"samples": ["6", 0], "vae": ["3", 0]}},
+    "8": {"class_type": "SaveImage", "inputs": {"images": ["7", 0], "filename_prefix": "ZImageTurbo_GPU"}}
+}}
+with sftp.open('/tmp/wf.json', 'w') as f:
+    f.write(json.dumps(wf))
+resp = sh('curl -s -X POST http://127.0.0.1:8188/prompt -H "Content-Type: application/json" -d @/tmp/wf.json')
+print(f"  {resp[:120]}")
+
+# Monitor
+print("\nMonitoring (SHARED = VAE on GPU, everything on GPU)...")
+t0 = time.time()
+for i in range(200):
+    el = int(time.time() - t0)
+    temp = sh('cat /sys/class/drm/card0/device/hwmon/hwmon*/temp1_input 2>/dev/null', timeout=5)
+    tc = int(temp)//1000 if temp.isdigit() else '?'
+    
+    try:
+        with sftp.open('/tmp/comfyui.log', 'r') as f:
+            log = f.read().decode(errors='replace')
+    except: log = ''
+    
+    samp = ''
+    last = ''
+    for line in log.split('\n'):
+        s = line.strip()
+        if '/8' in s and ('it/s' in s or 's/it' in s): samp = s
+        if s and 'FETCH' not in s and 'startup' not in s and 'DEPRECATION' not in s: last = s
+    
+    print(f"  [{el:>4}s] {tc}C | {(samp or last)[-90:]}")
+    
+    imgs = sh('ls ~/ComfyUI/output/ZImageTurbo_GPU*.png 2>/dev/null', timeout=5)
+    if imgs:
+        et = ''
+        for line in log.split('\n'):
+            if 'Prompt executed' in line: et = line.strip()
+        print(f"\n  *** DONE! ***  {imgs}")
+        print(f"  {et}")
+        print(f"  Wall: {el}s")
+        for line in log.split('\n'):
+            s = line.strip()
+            if any(k in s for k in ['loaded', '/8', 'Prompt executed', 'VAE load', 'Requested']):
+                if 'FETCH' not in s: print(f"  {s}")
+        break
+    
+    q = sh('curl -s http://127.0.0.1:8188/queue 2>/dev/null', timeout=5)
+    try:
+        qd = json.loads(q)
+        if not qd.get('queue_running') and not qd.get('queue_pending') and el > 30:
+            time.sleep(2)
+            imgs = sh('ls ~/ComfyUI/output/ZImageTurbo_GPU*.png 2>/dev/null', timeout=5)
+            if imgs:
+                print(f"\n  *** DONE: {imgs} ***")
+            else:
+                print(f"\n  Queue empty, no image:")
+                for line in log.split('\n')[-20:]:
+                    s = line.strip()
+                    if s and 'FETCH' not in s: print(f"  {s}")
+            break
+    except: pass
+    
+    if sh('pgrep -f "python3.*main.py" >/dev/null && echo Y || echo N', timeout=5) == 'N':
+        print("\n  CRASHED!")
+        for line in log.split('\n')[-30:]:
+            if line.strip(): print(f"  {line.strip()}")
+        break
+    
+    time.sleep(10)
+
+sftp.close()
+c.close()
+print("\nDone.")
@@ -0,0 +1,118 @@
+"""Upload a self-contained bash script and run it in ONE SSH session."""
+import paramiko, time
+
+k = paramiko.Ed25519Key.from_private_key_file(r'C:\Users\fabia\.ssh\id_ed25519')
+c = paramiko.SSHClient()
+c.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+c.connect('192.168.178.150', username='fabian', pkey=k, timeout=10)
+sftp = c.open_sftp()
+
+# Write the entire fix+restart+test as ONE bash script
+script = r'''#!/bin/bash
+set -e
+
+# Kill any running ComfyUI
+pkill -9 -f "python3.*main.py" 2>/dev/null || true
+sleep 2
+
+# Clean
+rm -f /tmp/comfyui.log
+rm -f ~/ComfyUI/output/ZImageTurbo_GPU*.png
+
+# Start ComfyUI
+echo "Starting ComfyUI..."
+nohup /tmp/run_comfyui.sh > /tmp/comfyui.log 2>&1 &
+sleep 3
+PID=$(pgrep -f "python3.*main.py" | head -1)
+echo "PID: $PID"
+
+if [ -z "$PID" ]; then
+    echo "FAILED TO START!"
+    cat /tmp/comfyui.log 2>/dev/null | tail -20
+    exit 1
+fi
+
+# Wait for HTTP
+echo "Waiting for HTTP..."
+for i in $(seq 1 90); do
+    CODE=$(curl -s -o /dev/null -w "%{http_code}" http://127.0.0.1:8188/ 2>/dev/null || echo 000)
+    if [ "$CODE" = "200" ]; then
+        echo "Ready after ${i}s"
+        break
+    fi
+    sleep 2
+done
+
+# Show key startup info
+grep -E "vram state|SHARED|Device:|Total VRAM|load device|offload" /tmp/comfyui.log 2>/dev/null || true
+
+# Submit workflow
+echo ""
+echo "Submitting workflow..."
+cat > /tmp/wf.json << 'WFEOF'
+{"prompt":{"1":{"class_type":"UnetLoaderGGUF","inputs":{"unet_name":"z_image_turbo-Q5_K_S.gguf"}},"2":{"class_type":"CLIPLoaderGGUF","inputs":{"clip_name":"Qwen3-4B.i1-Q5_K_S.gguf","type":"qwen_image"}},"3":{"class_type":"VAELoader","inputs":{"vae_name":"ae.safetensors"}},"4":{"class_type":"CLIPTextEncode","inputs":{"text":"A red fox in a snowy forest, photorealistic","clip":["2",0]}},"5":{"class_type":"EmptyLatentImage","inputs":{"width":512,"height":512,"batch_size":1}},"6":{"class_type":"KSampler","inputs":{"model":["1",0],"positive":["4",0],"negative":["4",0],"latent_image":["5",0],"seed":99999,"steps":8,"cfg":1.0,"sampler_name":"euler","scheduler":"simple","denoise":1.0}},"7":{"class_type":"VAEDecode","inputs":{"samples":["6",0],"vae":["3",0]}},"8":{"class_type":"SaveImage","inputs":{"images":["7",0],"filename_prefix":"ZImageTurbo_GPU"}}}}
+WFEOF
+
+RESP=$(curl -s -X POST http://127.0.0.1:8188/prompt -H "Content-Type: application/json" -d @/tmp/wf.json)
+echo "Response: ${RESP:0:120}"
+
+# Monitor
+echo ""
+echo "Monitoring..."
+START=$(date +%s)
+LAST=""
+while true; do
+    NOW=$(date +%s)
+    ELAPSED=$((NOW - START))
+    
+    if [ $ELAPSED -gt 600 ]; then
+        echo "TIMEOUT after 600s"
+        tail -20 /tmp/comfyui.log
+        break
+    fi
+    
+    # Check for output image
+    if ls ~/ComfyUI/output/ZImageTurbo_GPU*.png 2>/dev/null; then
+        echo ""
+        echo "*** IMAGE DONE in ${ELAPSED}s! ***"
+        grep -E "loaded|load device|offload|Prompt executed|/8" /tmp/comfyui.log 2>/dev/null | grep -v FETCH || true
+        break
+    fi
+    
+    # Show progress
+    LINE=$(grep -E "/8|loaded|Requested|VAE|Prompt executed|Error|OOM" /tmp/comfyui.log 2>/dev/null | grep -v FETCH | grep -v audio_vae | grep -v "split attention" | tail -1)
+    GPU=$(cat /sys/class/drm/card0/device/gpu_busy_percent 2>/dev/null || echo "?")
+    
+    if [ "$LINE" != "$LAST" ] && [ -n "$LINE" ]; then
+        echo "[${ELAPSED}s] GPU:${GPU}% ${LINE:0:110}"
+        LAST="$LINE"
+    fi
+    
+    # Check alive
+    if ! pgrep -f "python3.*main.py" > /dev/null 2>&1; then
+        echo "CRASHED at ${ELAPSED}s!"
+        tail -20 /tmp/comfyui.log
+        break
+    fi
+    
+    sleep 3
+done
+'''
+
+with sftp.open('/tmp/fix_and_run.sh', 'w') as f:
+    f.write(script)
+sftp.close()
+
+# Execute in ONE session
+print("Running fix+restart+monitor on BC-250...")
+stdin, stdout, stderr = c.exec_command('bash /tmp/fix_and_run.sh', timeout=660)
+# Stream output
+for line in iter(stdout.readline, ''):
+    print(line.rstrip())
+err = stderr.read().decode(errors='replace').strip()
+if err:
+    for l in err.split('\n')[-10:]:
+        if l.strip(): print(f"STDERR: {l.strip()}")
+
+c.close()
+print("\nDone.")
@@ -0,0 +1,44 @@
+"""Quick status: is ComfyUI still running and what's the log say?"""
+import paramiko
+
+k = paramiko.Ed25519Key.from_private_key_file(r'C:\Users\fabia\.ssh\id_ed25519')
+c = paramiko.SSHClient()
+c.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+c.connect('192.168.178.150', username='fabian', pkey=k, timeout=10)
+
+sftp = c.open_sftp()
+
+# Read full log
+with sftp.open('/tmp/comfyui.log', 'r') as f:
+    log = f.read().decode(errors='replace')
+
+sftp.close()
+
+lines = log.split('\n')
+print(f"Total lines: {len(lines)}")
+print()
+
+# Show only meaningful lines
+for line in lines:
+    s = line.strip()
+    if not s:
+        continue
+    if 'FETCH ComfyRegistry' in s or 'All startup tasks' in s or 'FETCH DATA' in s:
+        continue
+    print(s)
+
+# Check process + GPU
+chan = c.get_transport().open_session()
+chan.settimeout(10)
+chan.exec_command('/bin/bash -c "echo; echo === PROCESS ===; ps aux | grep python3 | grep -v grep; echo; echo === GPU ===; rocm-smi 2>/dev/null | head -12; echo; echo === OUTPUT ===; ls -la ~/ComfyUI/output/ 2>/dev/null; echo; echo === QUEUE ===; curl -s http://127.0.0.1:8188/queue 2>/dev/null"')
+out = b""
+while True:
+    try:
+        chunk = chan.recv(65536)
+        if not chunk: break
+        out += chunk
+    except: break
+chan.close()
+print(out.decode(errors='replace'))
+
+c.close()
@@ -0,0 +1,185 @@
+#!/usr/bin/env python3
+"""Start PyTorch build on BC-250 properly using SFTP for the script."""
+import paramiko
+import time
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=120, desc=""):
+    if desc:
+        print(f"\n{'='*60}")
+        print(f"  {desc}")
+        print(f"{'='*60}")
+    print(f"$ {cmd}")
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    rc = stdout.channel.recv_exit_status()
+    if out.strip():
+        lines = out.strip().split('\n')
+        if len(lines) > 60:
+            print(f"  ... ({len(lines)} lines, showing last 60)")
+            print('\n'.join(lines[-60:]))
+        else:
+            print(out.strip())
+    if err.strip():
+        lines = err.strip().split('\n')
+        show = lines[-20:] if len(lines) > 20 else lines
+        print(f"STDERR: {chr(10).join(show)}")
+    print(f"  Exit code: {rc}")
+    return rc, out, err
+
+# Upload the build script via SFTP
+build_script = '''#!/bin/bash
+set -euo pipefail
+
+LOG="/home/fabian/pytorch_build.log"
+exec > >(tee -a "$LOG") 2>&1
+
+echo "=========================================="
+echo " PyTorch Build for ROCm gfx1010 (BC-250)"
+echo " Started: $(date)"
+echo "=========================================="
+
+# Activate venv
+source /home/fabian/comfyui-env/bin/activate
+
+cd /home/fabian/pytorch
+
+# ROCm build configuration
+export USE_ROCM=1
+export USE_CUDA=0
+export PYTORCH_ROCM_ARCH="gfx1010"
+export HIP_VISIBLE_DEVICES=0
+export HSA_OVERRIDE_GFX_VERSION=10.1.0
+export HSA_ENABLE_SDMA=0
+export ROCM_PATH=/opt/rocm
+export HIP_PATH=/opt/rocm
+export CMAKE_PREFIX_PATH="/opt/rocm;$(python3 -c 'import sys; print(sys.prefix)')"
+export PATH=/opt/rocm/bin:$PATH
+
+# Build settings
+export USE_NINJA=1
+export CMAKE_GENERATOR=Ninja
+export MAX_JOBS=6
+export USE_CCACHE=1
+export CCACHE_DIR=/home/fabian/.ccache
+
+# Disable unnecessary components for faster build
+export USE_FBGEMM=0
+export USE_KINETO=0
+export USE_CUPTI_SO=0
+export USE_NCCL=0
+export USE_DISTRIBUTED=0
+export USE_TENSORPIPE=0
+export USE_GLOO=0
+export USE_MPI=0
+export USE_OPENMP=1
+export USE_MKLDNN=1
+export BUILD_TEST=0
+export USE_CUDNN=0
+
+echo ""
+echo "Build config:"
+echo "  PYTORCH_ROCM_ARCH=$PYTORCH_ROCM_ARCH"
+echo "  USE_ROCM=$USE_ROCM"
+echo "  MAX_JOBS=$MAX_JOBS"
+echo "  Python: $(python3 --version)"
+echo "  hipcc: $(hipcc --version 2>&1 | head -1)"
+echo "  ROCm: $(cat /opt/rocm/.info/version)"
+echo ""
+
+# Install requirements
+echo "Installing PyTorch requirements..."
+pip install -r requirements.txt 2>&1 | tail -10
+echo ""
+
+# Clean any partial build
+echo "Cleaning previous build artifacts..."
+python3 setup.py clean 2>&1 || true
+echo ""
+
+# Build the wheel
+echo "Starting PyTorch build..."
+echo "=========================================="
+python3 setup.py bdist_wheel 2>&1
+
+BUILD_RC=$?
+echo ""
+echo "=========================================="
+echo " Build exit code: $BUILD_RC"
+echo " Finished: $(date)"
+echo "=========================================="
+
+if [ $BUILD_RC -eq 0 ]; then
+    echo ""
+    echo "Wheel files:"
+    ls -lh dist/*.whl 2>/dev/null
+
+    echo ""
+    echo "Installing wheel..."
+    pip install dist/*.whl 2>&1
+
+    echo ""
+    echo "=== VERIFICATION ==="
+    python3 -c "
+import torch
+print(f'PyTorch version: {torch.__version__}')
+print(f'HIP version: {torch.version.hip}')
+print(f'CUDA available (HIP): {torch.cuda.is_available()}')
+if torch.cuda.is_available():
+    print(f'Device name: {torch.cuda.get_device_name(0)}')
+    print(f'Device count: {torch.cuda.device_count()}')
+    t = torch.randn(4, 4, device=\"cuda\")
+    print(f'Tensor on GPU: {t.device}')
+    print(f'Tensor sum: {t.sum().item():.4f}')
+    print('GPU COMPUTE: WORKING')
+else:
+    print('WARNING: CUDA/HIP not available')
+" 2>&1
+fi
+
+echo ""
+echo "BUILD_COMPLETE_RC=$BUILD_RC"
+'''
+
+print("Uploading build script via SFTP...")
+sftp = ssh.open_sftp()
+with sftp.open('/home/fabian/build_pytorch.sh', 'w') as f:
+    f.write(build_script)
+sftp.close()
+
+run("chmod +x /home/fabian/build_pytorch.sh", desc="Make executable")
+
+# Remove old log if it exists
+run("rm -f /home/fabian/pytorch_build.log", desc="Clean old log")
+
+# Start the build using nohup inside bash (not fish)
+# Using bash explicitly to avoid fish issues with nohup
+run("bash -c 'nohup bash /home/fabian/build_pytorch.sh </dev/null >/dev/null 2>&1 & echo PID=$!'",
+    desc="Start build in background")
+
+# Wait for it to actually start
+time.sleep(10)
+
+# Verify it's running
+run("pgrep -fa 'build_pytorch\\|setup.py' | head -10",
+    desc="Verify build is running")
+
+# Check initial log
+time.sleep(5)
+run("cat /home/fabian/pytorch_build.log 2>/dev/null | head -30 || echo 'Log not yet available'",
+    desc="Initial build log")
+
+# Monitor for first compile steps
+time.sleep(30)
+run("tail -30 /home/fabian/pytorch_build.log 2>/dev/null || echo 'Waiting for log...'",
+    desc="Build progress after 30 seconds")
+
+ssh.close()
+print("\n" + "="*60)
+print("  PyTorch build running on BC-250!")
+print("  Monitor: tail -f ~/pytorch_build.log")
+print("="*60)
@@ -0,0 +1,45 @@
+#!/usr/bin/env python3
+"""Quick single-connection status check."""
+import paramiko, json
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian',
+            key_filename=r'C:\Users\fabia\.ssh\id_ed25519', timeout=10)
+
+def run(cmd):
+    _, so, se = ssh.exec_command(cmd, timeout=15)
+    return so.read().decode()
+
+try:
+    # Is ComfyUI running?
+    print("=== PROCESS ===")
+    print(run("bash -c 'PID=$(pgrep -f \"python3 main.py\" | head -1); "
+        "if [ -n \"$PID\" ]; then "
+        "  ps -p $PID -o pid,%cpu,%mem,nlwp,etime --no-headers; "
+        "else echo NOT_RUNNING; fi'").strip())
+
+    # Log
+    print("\n=== LOG (last 20) ===")
+    print(run("tail -20 /home/fabian/comfyui.log 2>/dev/null").strip())
+
+    # GPU
+    print("\n=== GPU ===")
+    gpu = run("HSA_OVERRIDE_GFX_VERSION=10.1.0 rocm-smi 2>/dev/null | grep -E '0x|GPU%'")
+    print(gpu.strip() if gpu.strip() else "no output")
+
+    # Output files
+    print("\n=== OUTPUT ===")
+    print(run("ls -lah /home/fabian/ComfyUI/output/ 2>/dev/null").strip())
+
+    # Queue
+    print("\n=== QUEUE ===")
+    q = run("curl -s http://localhost:8188/queue 2>/dev/null")
+    if q.strip():
+        qj = json.loads(q)
+        print(f"Running: {len(qj.get('queue_running',[]))}, Pending: {len(qj.get('queue_pending',[]))}")
+    else:
+        print("Server not responding")
+finally:
+    ssh.close()
+    print("\nSSH closed.")
@@ -0,0 +1,55 @@
+#!/usr/bin/env python3
+"""Step 1: Install build dependencies on BC-250 for PyTorch build."""
+import paramiko
+import sys
+import time
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=300, desc=""):
+    if desc:
+        print(f"\n{'='*60}")
+        print(f"  {desc}")
+        print(f"{'='*60}")
+    print(f"$ {cmd}")
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    rc = stdout.channel.recv_exit_status()
+    if out.strip():
+        print(out.strip()[-2000:])  # Last 2000 chars
+    if err.strip():
+        # Filter out common noise
+        lines = [l for l in err.strip().split('\n') if not l.startswith('warning:')]
+        if lines:
+            print(f"STDERR: {chr(10).join(lines[-20:])}")
+    print(f"  Exit code: {rc}")
+    return rc, out, err
+
+# Install build tools
+run("sudo pacman -S --needed --noconfirm python-pip ninja ccache",
+    desc="Install pip, ninja, ccache")
+
+# Install PyTorch build dependencies
+run("sudo pacman -S --needed --noconfirm cmake blas lapack openblas "
+    "python-numpy python-pyyaml python-typing_extensions "
+    "intel-oneapi-mkl 2>/dev/null; echo done",
+    desc="Install build dependencies (cmake, blas, numpy, etc.)")
+
+# Install additional deps that PyTorch needs
+run("sudo pacman -S --needed --noconfirm python-cffi python-setuptools "
+    "python-wheel python-filelock python-sympy python-networkx",
+    desc="Install Python dependencies")
+
+# Verify installs
+run("pip --version && ninja --version && ccache --version | head -1 && cmake --version | head -1",
+    desc="Verify installations")
+
+# Check pip can install packages  
+run("pip install --user --upgrade pip setuptools wheel 2>&1 | tail -5",
+    desc="Upgrade pip/setuptools")
+
+ssh.close()
+print("\n\nDone — build dependencies installed.")
@@ -0,0 +1,24 @@
+#!/usr/bin/env python3
+"""Step 1: Kill stuck ComfyUI and check flags."""
+import paramiko
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=60):
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    print(out.strip() if out.strip() else "")
+    if err.strip():
+        for l in err.strip().split('\n')[-5:]:
+            print(f"STDERR: {l}")
+
+print("=== Kill stuck ===")
+run("pkill -f 'python3 main.py' 2>/dev/null; sleep 2; pkill -9 -f 'python3 main.py' 2>/dev/null; sleep 1; echo killed")
+
+print("\n=== Check flags ===")
+run("bash -c 'source /home/fabian/comfyui-env/bin/activate && cd /home/fabian/ComfyUI && python3 main.py --help 2>&1 | grep -i -E \"vae|fp16|fp32|force|cpu|novram|lowvram\"'")
+
+ssh.close()
+print("\nDone.")
@@ -0,0 +1,75 @@
+#!/usr/bin/env python3
+"""Step 2: Create venv and clone PyTorch on BC-250."""
+import paramiko
+import sys
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=600, desc=""):
+    if desc:
+        print(f"\n{'='*60}")
+        print(f"  {desc}")
+        print(f"{'='*60}")
+    print(f"$ {cmd}")
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    rc = stdout.channel.recv_exit_status()
+    if out.strip():
+        # Print last portion for long outputs
+        lines = out.strip().split('\n')
+        if len(lines) > 50:
+            print(f"  ... ({len(lines)} lines, showing last 50)")
+            print('\n'.join(lines[-50:]))
+        else:
+            print(out.strip())
+    if err.strip():
+        lines = [l for l in err.strip().split('\n') if 'warning:' not in l.lower()]
+        if lines:
+            if len(lines) > 30:
+                print(f"STDERR ({len(lines)} lines, last 30):")
+                print('\n'.join(lines[-30:]))
+            else:
+                print(f"STDERR: {chr(10).join(lines)}")
+    print(f"  Exit code: {rc}")
+    return rc, out, err
+
+# Create the venv
+run("python3 -m venv ~/comfyui-env --system-site-packages",
+    desc="Create venv with system site-packages (for numpy, etc.)")
+
+# Activate and install basic build deps in venv
+run("bash -c 'source ~/comfyui-env/bin/activate && pip install --upgrade pip setuptools wheel'",
+    desc="Upgrade pip in venv")
+
+run("bash -c 'source ~/comfyui-env/bin/activate && pip install cmake ninja pyyaml typing-extensions cffi future six requests dataclasses filelock sympy networkx jinja2 numpy'",
+    desc="Install PyTorch build deps in venv", timeout=120)
+
+# Check if PyTorch source already exists
+rc, out, _ = run("test -d ~/pytorch && echo EXISTS || echo MISSING",
+    desc="Check for existing PyTorch source")
+
+if "EXISTS" in out:
+    print("\n  PyTorch source directory exists. Checking if it's a valid repo...")
+    run("cd ~/pytorch && git log --oneline -1 2>&1", desc="Check PyTorch repo")
+else:
+    # Clone PyTorch — this is the big download
+    print("\n  Cloning PyTorch (this will take a while)...")
+    run("git clone --depth 1 --recursive --shallow-submodules https://github.com/pytorch/pytorch.git ~/pytorch 2>&1 | tail -20",
+        desc="Clone PyTorch (shallow, with submodules)",
+        timeout=1200)  # 20 minutes timeout
+
+# Verify clone
+run("ls -la ~/pytorch/setup.py ~/pytorch/torch/ 2>&1 | head -5",
+    desc="Verify PyTorch source")
+
+run("cd ~/pytorch && git log --oneline -1",
+    desc="PyTorch version")
+
+run("du -sh ~/pytorch",
+    desc="PyTorch source size")
+
+ssh.close()
+print("\n\nDone — venv created and PyTorch cloned.")
@@ -0,0 +1,180 @@
+#!/usr/bin/env python3
+"""Step 3: Build PyTorch from source for ROCm gfx1010 on BC-250.
+
+This build will take a long time (1-3 hours on 12 cores).
+We run it non-interactively via nohup so it survives SSH disconnects.
+"""
+import paramiko
+import sys
+import time
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=600, desc=""):
+    if desc:
+        print(f"\n{'='*60}")
+        print(f"  {desc}")
+        print(f"{'='*60}")
+    print(f"$ {cmd}")
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    rc = stdout.channel.recv_exit_status()
+    if out.strip():
+        lines = out.strip().split('\n')
+        if len(lines) > 50:
+            print(f"  ... ({len(lines)} lines, showing last 50)")
+            print('\n'.join(lines[-50:]))
+        else:
+            print(out.strip())
+    if err.strip():
+        lines = [l for l in err.strip().split('\n')]
+        if lines:
+            show = lines[-30:] if len(lines) > 30 else lines
+            print(f"STDERR: {chr(10).join(show)}")
+    print(f"  Exit code: {rc}")
+    return rc, out, err
+
+# First, create the build script on the BC-250
+build_script = r'''#!/bin/bash
+set -euo pipefail
+
+LOG="/home/fabian/pytorch_build.log"
+exec > >(tee -a "$LOG") 2>&1
+
+echo "=========================================="
+echo " PyTorch Build for ROCm gfx1010 (BC-250)"
+echo " Started: $(date)"
+echo "=========================================="
+
+# Activate venv
+source /home/fabian/comfyui-env/bin/activate
+
+# Go to PyTorch source
+cd /home/fabian/pytorch
+
+# Set environment for ROCm build
+export USE_ROCM=1
+export USE_CUDA=0
+export PYTORCH_ROCM_ARCH="gfx1010"
+export HIP_VISIBLE_DEVICES=0
+export HSA_OVERRIDE_GFX_VERSION=10.1.0
+export HSA_ENABLE_SDMA=0
+export ROCM_PATH=/opt/rocm
+export HIP_PATH=/opt/rocm
+export CMAKE_PREFIX_PATH=/opt/rocm
+export PATH=/opt/rocm/bin:$PATH
+
+# Use ninja for faster builds
+export USE_NINJA=1
+export CMAKE_GENERATOR=Ninja
+
+# Limit parallel jobs to avoid OOM (14GB RAM + 14GB swap)
+# Each compilation unit can use ~1-2GB during link, so limit to 6 jobs
+export MAX_JOBS=6
+
+# Use ccache to speed up rebuilds
+export USE_CCACHE=1
+export CCACHE_DIR=/home/fabian/.ccache
+
+# Disable unnecessary components to speed up build
+export USE_FBGEMM=0
+export USE_KINETO=0
+export USE_CUPTI_SO=0
+export USE_NCCL=0
+export USE_DISTRIBUTED=0
+export USE_TENSORPIPE=0
+export USE_GLOO=0
+export USE_MPI=0
+export USE_OPENMP=1
+export USE_MKLDNN=1
+export BUILD_TEST=0
+
+# Disable CUDA-specific stuff
+export USE_CUDNN=0
+
+echo ""
+echo "Build config:"
+echo "  PYTORCH_ROCM_ARCH=$PYTORCH_ROCM_ARCH"
+echo "  USE_ROCM=$USE_ROCM"
+echo "  MAX_JOBS=$MAX_JOBS"
+echo "  USE_CCACHE=$USE_CCACHE"
+echo "  Python: $(python3 --version)"
+echo "  hipcc: $(hipcc --version 2>&1 | head -1)"
+echo ""
+
+# Install requirements
+echo "Installing PyTorch requirements..."
+pip install -r requirements.txt 2>&1 | tail -5
+
+# Run the build
+echo ""
+echo "Starting PyTorch build... (this will take 1-3 hours)"
+echo "=========================================="
+python3 setup.py bdist_wheel 2>&1
+
+BUILD_RC=$?
+echo ""
+echo "=========================================="
+echo " Build finished with exit code: $BUILD_RC"
+echo " Time: $(date)"
+echo "=========================================="
+
+if [ $BUILD_RC -eq 0 ]; then
+    echo "Wheel file:"
+    ls -lh dist/*.whl 2>/dev/null || echo "No wheel found, trying develop install..."
+    
+    # Install the wheel
+    echo "Installing PyTorch wheel..."
+    pip install dist/*.whl 2>&1 | tail -5
+    
+    # Verify
+    echo ""
+    echo "Verification:"
+    python3 -c "import torch; print(f'PyTorch {torch.__version__}'); print(f'ROCm: {torch.version.hip}'); print(f'CUDA available: {torch.cuda.is_available()}'); print(f'Device: {torch.cuda.get_device_name(0) if torch.cuda.is_available() else \"N/A\"}')"
+fi
+
+echo "BUILD_COMPLETE_RC=$BUILD_RC" >> "$LOG"
+'''
+
+# Write the build script to BC-250
+sftp = ssh.open_sftp()
+with sftp.open('/home/fabian/build_pytorch.sh', 'w') as f:
+    f.write(build_script)
+sftp.close()
+
+run("chmod +x /home/fabian/build_pytorch.sh", desc="Make build script executable")
+
+# Check if a build is already running
+rc, out, _ = run("pgrep -f 'setup.py bdist_wheel' || echo 'NOT RUNNING'",
+    desc="Check if build is already running")
+
+if 'NOT RUNNING' not in out:
+    print("\n  BUILD IS ALREADY RUNNING — not starting a new one.")
+    print("  Monitor with: tail -f ~/pytorch_build.log")
+else:
+    # Start the build in background using nohup
+    # This way it survives SSH disconnects
+    run("nohup bash /home/fabian/build_pytorch.sh > /dev/null 2>&1 &",
+        desc="Starting PyTorch build in background (nohup)")
+    
+    # Give it a moment to start
+    time.sleep(5)
+    
+    # Verify it started
+    run("pgrep -fa 'build_pytorch.sh' || pgrep -fa 'setup.py' || echo 'WARNING: Build may have failed to start'",
+        desc="Verify build process started")
+    
+    # Check initial log output
+    time.sleep(10)
+    run("tail -30 /home/fabian/pytorch_build.log 2>/dev/null || echo 'Log not yet created'",
+        desc="Initial build log output")
+
+ssh.close()
+print("\n" + "="*60)
+print("  PyTorch build started in background on BC-250!")
+print("  Monitor: ssh fabian@BC-250 'tail -f ~/pytorch_build.log'")
+print("  Check status: ssh fabian@BC-250 'pgrep -fa setup.py'")
+print("="*60)
@@ -0,0 +1,81 @@
+#!/usr/bin/env python3
+"""Install missing ROCm math libraries for PyTorch build on BC-250."""
+import paramiko
+import time
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=300, desc=""):
+    if desc:
+        print(f"\n{'='*60}")
+        print(f"  {desc}")
+        print(f"{'='*60}")
+    print(f"$ {cmd}")
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    rc = stdout.channel.recv_exit_status()
+    if out.strip():
+        lines = out.strip().split('\n')
+        if len(lines) > 60:
+            print(f"  ... ({len(lines)} lines, showing last 60)")
+            print('\n'.join(lines[-60:]))
+        else:
+            print(out.strip())
+    if err.strip():
+        lines = err.strip().split('\n')
+        show = lines[-20:] if len(lines) > 20 else lines
+        print(f"STDERR: {chr(10).join(show)}")
+    print(f"  Exit code: {rc}")
+    return rc, out, err
+
+# Find all available ROCm packages
+run("pacman -Ss rocm | grep -E '^cachyos|^extra|^core' | head -40",
+    desc="Available ROCm packages")
+
+# Install all ROCm math/compute libraries needed by PyTorch
+run("sudo pacman -S --needed --noconfirm "
+    "hiprand rocrand "
+    "hipblas rocblas "
+    "hipfft rocfft "
+    "hipsparse rocsparse "
+    "hipsolver rocsolver "
+    "miopen-hip "
+    "rocprim hipcub "
+    "rocthrust "
+    "rccl "
+    "hipblaslt "
+    "roctracer "
+    "2>&1 | tail -40",
+    desc="Install ROCm math libraries",
+    timeout=600)
+
+# Verify hiprand is now available
+run("find /opt/rocm -name 'hiprandConfig.cmake' -o -name 'hiprand-config.cmake' 2>/dev/null | head -5",
+    desc="Verify hiprand cmake config")
+
+# Check all libraries
+run("ls /opt/rocm/lib/libhiprand.so /opt/rocm/lib/librocblas.so /opt/rocm/lib/libhipblas.so /opt/rocm/lib/librocfft.so /opt/rocm/lib/libMIOpen.so 2>&1",
+    desc="Verify key libraries exist")
+
+# Clean the failed build and restart
+run("bash -c 'source ~/comfyui-env/bin/activate && cd ~/pytorch && python3 setup.py clean 2>&1 | tail -5'",
+    desc="Clean failed build")
+
+# Restart build
+run("rm -f /home/fabian/pytorch_build.log", desc="Clean old log")
+run("bash -c 'nohup bash /home/fabian/build_pytorch.sh </dev/null >/dev/null 2>&1 & echo PID=$!'",
+    desc="Restart PyTorch build")
+
+time.sleep(15)
+run("pgrep -fa 'setup.py\\|cmake\\|ninja' | head -10",
+    desc="Verify build restarted")
+
+time.sleep(45)
+run("tail -40 /home/fabian/pytorch_build.log 2>/dev/null",
+    desc="Build progress")
+
+ssh.close()
+print("\nDone — ROCm libs installed and build restarted.")
@@ -0,0 +1,95 @@
+#!/usr/bin/env python3
+"""Install ComfyUI and dependencies on BC-250."""
+import paramiko
+import time
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=300, desc=""):
+    if desc:
+        print(f"\n{'='*60}")
+        print(f"  {desc}")
+        print(f"{'='*60}")
+    print(f"$ {cmd}")
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    rc = stdout.channel.recv_exit_status()
+    if out.strip():
+        lines = out.strip().split('\n')
+        if len(lines) > 60:
+            print(f"  ... ({len(lines)} lines, showing last 60)")
+            print('\n'.join(lines[-60:]))
+        else:
+            print(out.strip())
+    if err.strip():
+        lines = err.strip().split('\n')
+        show = lines[-20:] if len(lines) > 20 else lines
+        print(f"STDERR: {chr(10).join(show)}")
+    print(f"  Exit code: {rc}")
+    return rc, out, err
+
+# Clone ComfyUI
+run("git clone --depth 1 https://github.com/comfyanonymous/ComfyUI.git ~/ComfyUI 2>&1 | tail -10",
+    desc="Clone ComfyUI")
+
+# Install ComfyUI requirements in venv
+run("bash -c 'source ~/comfyui-env/bin/activate && cd ~/ComfyUI && pip install -r requirements.txt 2>&1 | tail -20'",
+    desc="Install ComfyUI requirements",
+    timeout=300)
+
+# Install diffusers from source (needed for ZImagePipeline)
+run("bash -c 'source ~/comfyui-env/bin/activate && pip install git+https://github.com/huggingface/diffusers 2>&1 | tail -10'",
+    desc="Install diffusers from source (for ZImagePipeline)",
+    timeout=300)
+
+# Install additional deps that ComfyUI/Z-Image might need
+run("bash -c 'source ~/comfyui-env/bin/activate && pip install transformers accelerate safetensors sentencepiece huggingface_hub aiohttp einops torchvision 2>&1 | tail -15'",
+    desc="Install transformers, accelerate, etc.",
+    timeout=300)
+
+# Install ComfyUI-Manager (custom node manager)
+run("git clone --depth 1 https://github.com/Comfy-Org/ComfyUI-Manager.git ~/ComfyUI/custom_nodes/ComfyUI-Manager 2>&1 | tail -5",
+    desc="Install ComfyUI-Manager")
+
+# Install ComfyUI-GGUF (needed for GGUF checkpoint format)
+run("git clone --depth 1 https://github.com/city96/ComfyUI-GGUF.git ~/ComfyUI/custom_nodes/ComfyUI-GGUF 2>&1 | tail -5",
+    desc="Install ComfyUI-GGUF nodes")
+
+# Install GGUF dependencies
+run("bash -c 'source ~/comfyui-env/bin/activate && pip install gguf 2>&1 | tail -5'",
+    desc="Install gguf Python package")
+
+# Install Z-Image Power Nodes
+run("git clone --depth 1 https://github.com/martin-rizzo/ComfyUI-ZImagePowerNodes.git ~/ComfyUI/custom_nodes/ComfyUI-ZImagePowerNodes 2>&1 | tail -5",
+    desc="Install Z-Image Power Nodes")
+
+# Verify ComfyUI structure
+run("ls -la ~/ComfyUI/main.py ~/ComfyUI/custom_nodes/ 2>&1",
+    desc="Verify ComfyUI structure")
+
+run("ls ~/ComfyUI/custom_nodes/",
+    desc="Custom nodes installed")
+
+# Create model directories
+run("mkdir -p ~/ComfyUI/models/diffusion_models ~/ComfyUI/models/text_encoders ~/ComfyUI/models/vae ~/ComfyUI/models/checkpoints",
+    desc="Create model directories")
+
+# Quick test: can ComfyUI import?
+run("""bash -c 'source ~/comfyui-env/bin/activate && \
+    HSA_OVERRIDE_GFX_VERSION=10.1.0 \
+    HIP_VISIBLE_DEVICES=0 \
+    HSA_ENABLE_SDMA=0 \
+    cd ~/ComfyUI && python3 -c "
+import torch
+print(f\\"torch {torch.__version__} hip={torch.version.hip} cuda={torch.cuda.is_available()}\\")
+import comfy
+print(\\"ComfyUI import OK\\")
+" 2>&1'""",
+    desc="Test ComfyUI import",
+    timeout=60)
+
+ssh.close()
+print("\nDone — ComfyUI installed.")
@@ -0,0 +1,156 @@
+#!/usr/bin/env python3
+"""Download Z-Image-Turbo GGUF model checkpoints on BC-250.
+
+Files needed (GGUF format — memory-efficient for 14GB RAM):
+1. z_image_turbo-Q5_K_S.gguf (5.19 GB) → diffusion_models/
+2. Qwen3-4B.i1-Q5_K_S.gguf  (2.82 GB) → text_encoders/
+3. ae.safetensors             (335 MB)  → vae/
+Total: ~8.35 GB
+"""
+import paramiko
+import time
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=3600, desc=""):
+    if desc:
+        print(f"\n{'='*60}")
+        print(f"  {desc}")
+        print(f"{'='*60}")
+    print(f"$ {cmd}")
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    rc = stdout.channel.recv_exit_status()
+    if out.strip():
+        lines = out.strip().split('\n')
+        if len(lines) > 30:
+            print(f"  ... ({len(lines)} lines, showing last 30)")
+            print('\n'.join(lines[-30:]))
+        else:
+            print(out.strip())
+    if err.strip():
+        lines = err.strip().split('\n')
+        show = lines[-15:] if len(lines) > 15 else lines
+        print(f"STDERR: {chr(10).join(show)}")
+    print(f"  Exit code: {rc}")
+    return rc, out, err
+
+# Create download script for background execution
+dl_script = '''#!/bin/bash
+set -euo pipefail
+
+LOG="/home/fabian/model_download.log"
+exec > >(tee -a "$LOG") 2>&1
+
+source /home/fabian/comfyui-env/bin/activate
+
+COMFY="$HOME/ComfyUI"
+echo "=========================================="
+echo " Downloading Z-Image-Turbo GGUF Models"
+echo " Started: $(date)"
+echo "=========================================="
+
+# 1. Diffusion model (5.19 GB)
+echo ""
+echo "[1/3] Downloading z_image_turbo-Q5_K_S.gguf (5.19 GB)..."
+if [ -f "$COMFY/models/diffusion_models/z_image_turbo-Q5_K_S.gguf" ]; then
+    echo "  Already exists, skipping."
+else
+    HF_XET_HIGH_PERFORMANCE=1 huggingface-cli download \
+        jayn7/Z-Image-Turbo-GGUF \
+        z_image_turbo-Q5_K_S.gguf \
+        --local-dir "$COMFY/models/diffusion_models/" \
+        --local-dir-use-symlinks False
+    echo "  Done."
+fi
+
+# 2. Text encoder Qwen3-4B (2.82 GB)
+echo ""
+echo "[2/3] Downloading Qwen3-4B.i1-Q5_K_S.gguf (2.82 GB)..."
+if [ -f "$COMFY/models/text_encoders/Qwen3-4B.i1-Q5_K_S.gguf" ]; then
+    echo "  Already exists, skipping."
+else
+    HF_XET_HIGH_PERFORMANCE=1 huggingface-cli download \
+        mradermacher/Qwen3-4B-i1-GGUF \
+        Qwen3-4B.i1-Q5_K_S.gguf \
+        --local-dir "$COMFY/models/text_encoders/" \
+        --local-dir-use-symlinks False
+    echo "  Done."
+fi
+
+# 3. VAE (335 MB)
+echo ""
+echo "[3/3] Downloading ae.safetensors (VAE, 335 MB)..."
+if [ -f "$COMFY/models/vae/ae.safetensors" ]; then
+    echo "  Already exists, skipping."
+else
+    HF_XET_HIGH_PERFORMANCE=1 huggingface-cli download \
+        Comfy-Org/z_image_turbo \
+        split_files/vae/ae.safetensors \
+        --local-dir "$COMFY/models/vae/" \
+        --local-dir-use-symlinks False
+    # Move from subdirectory if needed
+    if [ -f "$COMFY/models/vae/split_files/vae/ae.safetensors" ]; then
+        mv "$COMFY/models/vae/split_files/vae/ae.safetensors" "$COMFY/models/vae/ae.safetensors"
+        rm -rf "$COMFY/models/vae/split_files"
+    fi
+    echo "  Done."
+fi
+
+echo ""
+echo "=========================================="
+echo " Model Download Summary"
+echo "=========================================="
+echo "Diffusion model:"
+ls -lh "$COMFY/models/diffusion_models/"*.gguf 2>/dev/null || echo "  NOT FOUND"
+echo "Text encoder:"
+ls -lh "$COMFY/models/text_encoders/"*.gguf 2>/dev/null || echo "  NOT FOUND"
+echo "VAE:"
+ls -lh "$COMFY/models/vae/"*.safetensors 2>/dev/null || echo "  NOT FOUND"
+echo ""
+echo "Total model size:"
+du -sh "$COMFY/models/"
+echo ""
+echo "DOWNLOAD_COMPLETE"
+echo "Finished: $(date)"
+'''
+
+# Upload download script
+sftp = ssh.open_sftp()
+with sftp.open('/home/fabian/download_models.sh', 'w') as f:
+    f.write(dl_script)
+sftp.close()
+
+run("chmod +x /home/fabian/download_models.sh", desc="Make download script executable")
+run("rm -f /home/fabian/model_download.log", desc="Clean old log")
+
+# Start download in background
+run("bash -c 'nohup bash /home/fabian/download_models.sh </dev/null >/dev/null 2>&1 & echo PID=$!'",
+    desc="Start model download in background")
+
+# Wait and check progress
+time.sleep(10)
+run("tail -20 /home/fabian/model_download.log 2>/dev/null || echo 'Waiting for log...'",
+    desc="Initial download progress")
+
+# Keep checking
+for i in range(6):
+    time.sleep(30)
+    rc, out, _ = run(f"tail -10 /home/fabian/model_download.log 2>/dev/null",
+        desc=f"Download progress check {i+1}")
+    if 'DOWNLOAD_COMPLETE' in out:
+        print("\n  ALL DOWNLOADS COMPLETE!")
+        break
+
+# Final check
+run("tail -20 /home/fabian/model_download.log 2>/dev/null",
+    desc="Final download status")
+
+run("du -sh ~/ComfyUI/models/diffusion_models/ ~/ComfyUI/models/text_encoders/ ~/ComfyUI/models/vae/ 2>/dev/null",
+    desc="Model directory sizes")
+
+ssh.close()
+print("\nDone.")
@@ -0,0 +1,147 @@
+#!/usr/bin/env python3
+"""Download Z-Image-Turbo GGUF models on BC-250 using wget."""
+import paramiko
+import time
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=7200, desc=""):
+    if desc:
+        print(f"\n{'='*60}")
+        print(f"  {desc}")
+        print(f"{'='*60}")
+    print(f"$ {cmd}")
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    rc = stdout.channel.recv_exit_status()
+    if out.strip():
+        lines = out.strip().split('\n')
+        if len(lines) > 30:
+            print(f"  ... ({len(lines)} lines, showing last 30)")
+            print('\n'.join(lines[-30:]))
+        else:
+            print(out.strip())
+    if err.strip():
+        lines = err.strip().split('\n')
+        show = lines[-15:] if len(lines) > 15 else lines
+        print(f"STDERR: {chr(10).join(show)}")
+    print(f"  Exit code: {rc}")
+    return rc, out, err
+
+dl_script = r'''#!/bin/bash
+set -euo pipefail
+
+LOG="/home/fabian/model_download.log"
+exec > >(tee -a "$LOG") 2>&1
+
+COMFY="$HOME/ComfyUI"
+
+echo "=========================================="
+echo " Downloading Z-Image-Turbo GGUF Models"
+echo " Started: $(date)"
+echo "=========================================="
+
+# 1. Diffusion model (5.19 GB)
+echo ""
+echo "[1/3] Downloading z_image_turbo-Q5_K_S.gguf (5.19 GB)..."
+DEST1="$COMFY/models/diffusion_models/z_image_turbo-Q5_K_S.gguf"
+if [ -f "$DEST1" ]; then
+    echo "  Already exists ($(du -h "$DEST1" | cut -f1)), skipping."
+else
+    wget -c -q --show-progress \
+        "https://huggingface.co/jayn7/Z-Image-Turbo-GGUF/resolve/main/z_image_turbo-Q5_K_S.gguf" \
+        -O "$DEST1.tmp"
+    mv "$DEST1.tmp" "$DEST1"
+    echo "  Done: $(du -h "$DEST1" | cut -f1)"
+fi
+
+# 2. Text encoder Qwen3-4B (2.82 GB)
+echo ""
+echo "[2/3] Downloading Qwen3-4B.i1-Q5_K_S.gguf (2.82 GB)..."
+DEST2="$COMFY/models/text_encoders/Qwen3-4B.i1-Q5_K_S.gguf"
+if [ -f "$DEST2" ]; then
+    echo "  Already exists ($(du -h "$DEST2" | cut -f1)), skipping."
+else
+    wget -c -q --show-progress \
+        "https://huggingface.co/mradermacher/Qwen3-4B-i1-GGUF/resolve/main/Qwen3-4B.i1-Q5_K_S.gguf" \
+        -O "$DEST2.tmp"
+    mv "$DEST2.tmp" "$DEST2"
+    echo "  Done: $(du -h "$DEST2" | cut -f1)"
+fi
+
+# 3. VAE (335 MB)
+echo ""
+echo "[3/3] Downloading ae.safetensors (VAE, 335 MB)..."
+DEST3="$COMFY/models/vae/ae.safetensors"
+if [ -f "$DEST3" ]; then
+    echo "  Already exists ($(du -h "$DEST3" | cut -f1)), skipping."
+else
+    wget -c -q --show-progress \
+        "https://huggingface.co/Comfy-Org/z_image_turbo/resolve/main/split_files/vae/ae.safetensors" \
+        -O "$DEST3.tmp"
+    mv "$DEST3.tmp" "$DEST3"
+    echo "  Done: $(du -h "$DEST3" | cut -f1)"
+fi
+
+echo ""
+echo "=========================================="
+echo " Model Download Summary"
+echo "=========================================="
+echo "Diffusion model:"
+ls -lh "$COMFY/models/diffusion_models/"*.gguf 2>/dev/null || echo "  NOT FOUND"
+echo "Text encoder:"
+ls -lh "$COMFY/models/text_encoders/"*.gguf 2>/dev/null || echo "  NOT FOUND"
+echo "VAE:"
+ls -lh "$COMFY/models/vae/"*.safetensors 2>/dev/null || echo "  NOT FOUND"
+echo ""
+echo "Total model size:"
+du -sh "$COMFY/models/"
+echo ""
+echo "DOWNLOAD_COMPLETE"
+echo "Finished: $(date)"
+'''
+
+# Upload download script
+sftp = ssh.open_sftp()
+with sftp.open('/home/fabian/download_models.sh', 'w') as f:
+    f.write(dl_script)
+sftp.close()
+
+run("chmod +x /home/fabian/download_models.sh", desc="Make script executable")
+run("rm -f /home/fabian/model_download.log", desc="Clean old log")
+run("bash -c 'which wget'", desc="Verify wget exists")
+
+# Start download in background via nohup
+run("bash -c 'nohup bash /home/fabian/download_models.sh </dev/null >/dev/null 2>&1 & echo PID=$!'",
+    desc="Start model download in background")
+
+time.sleep(15)
+run("tail -20 /home/fabian/model_download.log 2>/dev/null || echo 'Waiting for log...'",
+    desc="Initial download progress")
+
+# Monitor download progress - check every 60s for up to 30 minutes
+for i in range(30):
+    time.sleep(60)
+    rc, out, _ = run(f"tail -5 /home/fabian/model_download.log 2>/dev/null; echo '---'; "
+                     f"ls -lh ~/ComfyUI/models/diffusion_models/ ~/ComfyUI/models/text_encoders/ ~/ComfyUI/models/vae/ 2>/dev/null",
+        desc=f"Progress check {i+1}/30 ({(i+1)}min)")
+    if 'DOWNLOAD_COMPLETE' in out:
+        print("\n  ALL DOWNLOADS COMPLETE!")
+        break
+    # Check if background process still running
+    _, pout, _ = run("bash -c 'pgrep -f download_models.sh || echo NOPROCESS'")
+    if 'NOPROCESS' in pout and 'DOWNLOAD_COMPLETE' not in out:
+        print("\n  WARNING: Download process exited without completion!")
+        run("cat /home/fabian/model_download.log", desc="Full download log")
+        break
+
+# Final verification
+run("tail -25 /home/fabian/model_download.log 2>/dev/null", desc="Final download status")
+run("du -sh ~/ComfyUI/models/diffusion_models/ ~/ComfyUI/models/text_encoders/ ~/ComfyUI/models/vae/",
+    desc="Model directory sizes")
+
+ssh.close()
+print("\nDone.")
@@ -0,0 +1,177 @@
+#!/usr/bin/env python3
+"""Create ComfyUI startup script and launch it on BC-250."""
+import paramiko
+import time
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian', key_filename=r'C:\Users\fabia\.ssh\id_ed25519')
+
+def run(cmd, timeout=120, desc=""):
+    if desc:
+        print(f"\n{'='*60}")
+        print(f"  {desc}")
+        print(f"{'='*60}")
+    print(f"$ {cmd}")
+    _, stdout, stderr = ssh.exec_command(cmd, timeout=timeout)
+    out = stdout.read().decode()
+    err = stderr.read().decode()
+    rc = stdout.channel.recv_exit_status()
+    if out.strip():
+        lines = out.strip().split('\n')
+        if len(lines) > 40:
+            print(f"  ... ({len(lines)} lines, showing last 40)")
+            print('\n'.join(lines[-40:]))
+        else:
+            print(out.strip())
+    if err.strip():
+        lines = err.strip().split('\n')
+        show = lines[-20:] if len(lines) > 20 else lines
+        print(f"STDERR: {chr(10).join(show)}")
+    print(f"  Exit code: {rc}")
+    return rc, out, err
+
+# ──────────────────────────────────────────────────────────
+# 1. Create the startup script
+# ──────────────────────────────────────────────────────────
+startup_script = r'''#!/bin/bash
+# ComfyUI Startup Script for AsRock BC-250 (AMD Cyan Skillfish / ROCm 7.2)
+# Usage: ~/start_comfyui.sh [--listen] [--port PORT]
+
+set -euo pipefail
+
+# ═══════════════════════════════════════════════════════════
+#  BC-250 AMD GPU Environment Variables
+# ═══════════════════════════════════════════════════════════
+# Override gfx1013 → gfx1010 (RDNA 1.5 → RDNA 1 compat)
+export HSA_OVERRIDE_GFX_VERSION=10.1.0
+
+# Use device 0
+export HIP_VISIBLE_DEVICES=0
+
+# Disable SDMA (avoids queue errors on Cyan Skillfish)
+export HSA_ENABLE_SDMA=0
+
+# Suppress tool library warnings
+export HSA_TOOLS_LIB=""
+export HSA_TOOLS_REPORT_LOAD_FAILURE=0
+
+# PyTorch / ROCm tuning
+export PYTORCH_HIP_ALLOC_CONF="expandable_segments:True"
+export TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1
+
+# Avoid OOM on 14GB shared VRAM — force float16 where possible
+export COMFY_PRECISION=16
+
+# ═══════════════════════════════════════════════════════════
+#  Activate Virtual Environment
+# ═══════════════════════════════════════════════════════════
+source "$HOME/comfyui-env/bin/activate"
+
+# ═══════════════════════════════════════════════════════════
+#  Launch ComfyUI
+# ═══════════════════════════════════════════════════════════
+cd "$HOME/ComfyUI"
+
+echo "=========================================="
+echo " ComfyUI on BC-250 (ROCm 7.2)"
+echo "=========================================="
+echo " GPU:    AMD Cyan Skillfish (gfx1013→gfx1010)"
+echo " PyTorch: $(python -c 'import torch; print(torch.__version__)')"
+echo " HIP:    $(python -c 'import torch; print(torch.version.hip)')"
+echo " CUDA:   $(python -c 'import torch; print(torch.cuda.is_available())')"
+echo " Device: $(python -c 'import torch; print(torch.cuda.get_device_name(0) if torch.cuda.is_available() else "N/A")')"
+echo "=========================================="
+
+# Default: listen on all interfaces for remote access
+LISTEN_ARGS="--listen 0.0.0.0 --port 8188"
+
+# Parse arguments (override defaults if provided)
+if [ $# -gt 0 ]; then
+    LISTEN_ARGS="$@"
+fi
+
+echo ""
+echo "Starting ComfyUI with: $LISTEN_ARGS"
+echo "Access at: http://$(hostname -I | awk '{print $1}'):8188"
+echo ""
+
+exec python main.py $LISTEN_ARGS
+'''
+
+# ──────────────────────────────────────────────────────────
+# 2. Create fish shell wrapper too
+# ──────────────────────────────────────────────────────────
+fish_script = r'''#!/usr/bin/env fish
+# ComfyUI launcher for fish shell on BC-250
+
+# BC-250 GPU env vars
+set -gx HSA_OVERRIDE_GFX_VERSION 10.1.0
+set -gx HIP_VISIBLE_DEVICES 0
+set -gx HSA_ENABLE_SDMA 0
+set -gx HSA_TOOLS_LIB ""
+set -gx HSA_TOOLS_REPORT_LOAD_FAILURE 0
+set -gx PYTORCH_HIP_ALLOC_CONF "expandable_segments:True"
+set -gx TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL 1
+
+# Activate venv
+source $HOME/comfyui-env/bin/activate.fish
+
+# Launch
+cd $HOME/ComfyUI
+echo "Starting ComfyUI on BC-250..."
+python main.py --listen 0.0.0.0 --port 8188 $argv
+'''
+
+# Upload scripts
+sftp = ssh.open_sftp()
+with sftp.open('/home/fabian/start_comfyui.sh', 'w') as f:
+    f.write(startup_script)
+with sftp.open('/home/fabian/start_comfyui.fish', 'w') as f:
+    f.write(fish_script)
+sftp.close()
+
+run("chmod +x /home/fabian/start_comfyui.sh /home/fabian/start_comfyui.fish",
+    desc="Make startup scripts executable")
+
+# ──────────────────────────────────────────────────────────
+# 3. Quick pre-flight check
+# ──────────────────────────────────────────────────────────
+run("bash -c 'source /home/fabian/comfyui-env/bin/activate && "
+    "export HSA_OVERRIDE_GFX_VERSION=10.1.0 && "
+    "export HIP_VISIBLE_DEVICES=0 && "
+    "export HSA_ENABLE_SDMA=0 && "
+    "cd /home/fabian/ComfyUI && "
+    "python -c \""
+    "import torch; "
+    "print(f\\\"PyTorch {torch.__version__}, HIP {torch.version.hip}, CUDA {torch.cuda.is_available()}\\\"); "
+    "print(f\\\"Device: {torch.cuda.get_device_name(0)}\\\"); "
+    "import comfy.model_management; "
+    "print(f\\\"ComfyUI model_management imported OK\\\"); "
+    "\"'",
+    desc="Pre-flight: PyTorch + ComfyUI import check")
+
+# ──────────────────────────────────────────────────────────
+# 4. Launch ComfyUI in background
+# ──────────────────────────────────────────────────────────
+run("bash -c 'pkill -f \"python main.py\" 2>/dev/null; echo killed || echo no_existing'",
+    desc="Kill any existing ComfyUI process")
+
+time.sleep(2)
+
+run("bash -c 'nohup bash /home/fabian/start_comfyui.sh > /home/fabian/comfyui.log 2>&1 & echo PID=$!'",
+    desc="Launch ComfyUI in background")
+
+# Wait for startup
+time.sleep(15)
+run("tail -30 /home/fabian/comfyui.log 2>/dev/null", desc="ComfyUI startup log")
+
+# Check if port is listening
+time.sleep(10)
+run("bash -c 'ss -tlnp | grep 8188 || echo PORT_NOT_LISTENING'",
+    desc="Check if port 8188 is listening")
+
+run("tail -50 /home/fabian/comfyui.log 2>/dev/null", desc="Full startup log")
+
+ssh.close()
+print("\nDone.")
@@ -0,0 +1,127 @@
+#!/usr/bin/env python3
+"""Fix workflow submission: write JSON to file, curl from file, monitor."""
+import paramiko, json, time
+
+ssh = paramiko.SSHClient()
+ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+ssh.connect('192.168.178.150', username='fabian',
+            key_filename=r'C:\Users\fabia\.ssh\id_ed25519', timeout=10)
+
+def run(cmd, timeout=60):
+    _, so, se = ssh.exec_command(cmd, timeout=timeout)
+    return so.read().decode(), se.read().decode()
+
+try:
+    # Step 1: Write workflow JSON via SFTP (reliable, no shell escaping)
+    workflow = {
+        "prompt": {
+            "1": {"class_type": "UnetLoaderGGUF",
+                  "inputs": {"unet_name": "z_image_turbo-Q5_K_S.gguf"}},
+            "2": {"class_type": "CLIPLoaderGGUF",
+                  "inputs": {"clip_name": "Qwen3-4B.i1-Q5_K_S.gguf", "type": "qwen_image"}},
+            "3": {"class_type": "VAELoader",
+                  "inputs": {"vae_name": "ae.safetensors"}},
+            "4": {"class_type": "CLIPTextEncode",
+                  "inputs": {"text": "A red fox in a snowy forest, photorealistic", "clip": ["2", 0]}},
+            "5": {"class_type": "CLIPTextEncode",
+                  "inputs": {"text": "", "clip": ["2", 0]}},
+            "6": {"class_type": "EmptyLatentImage",
+                  "inputs": {"width": 512, "height": 512, "batch_size": 1}},
+            "7": {"class_type": "KSampler",
+                  "inputs": {"model": ["1", 0], "seed": 42, "steps": 8, "cfg": 1.0,
+                             "sampler_name": "euler", "scheduler": "simple",
+                             "positive": ["4", 0], "negative": ["5", 0],
+                             "latent_image": ["6", 0], "denoise": 1.0}},
+            "8": {"class_type": "VAEDecode",
+                  "inputs": {"samples": ["7", 0], "vae": ["3", 0]}},
+            "9": {"class_type": "SaveImage",
+                  "inputs": {"images": ["8", 0], "filename_prefix": "ZImageTurbo_test"}}
+        }
+    }
+
+    sftp = ssh.open_sftp()
+    with sftp.open('/tmp/wf.json', 'w') as f:
+        f.write(json.dumps(workflow))
+    sftp.close()
+    print("Workflow JSON written via SFTP.")
+
+    # Verify JSON is valid
+    out, _ = run("python3 -c \"import json; d=json.load(open('/tmp/wf.json')); print('nodes:', sorted(d['prompt'].keys()))\"")
+    print(f"Verify: {out.strip()}")
+
+    # Step 2: Check startup script has --cpu-vae
+    out, _ = run("cat /home/fabian/start_comfyui.sh")
+    has_cpu_vae = '--cpu-vae' in out
+    print(f"Startup has --cpu-vae: {has_cpu_vae}")
+    print(f"Startup has --novram: {'--novram' in out}")
+    print(f"Startup has --force-fp16: {'--force-fp16' in out}")
+
+    # Step 3: Submit
+    out, err = run("curl -s -X POST http://localhost:8188/prompt -H 'Content-Type: application/json' -d @/tmp/wf.json")
+    print(f"\nSubmit response: {out.strip()[:500]}")
+
+    try:
+        resp = json.loads(out.strip())
+    except:
+        print(f"Failed to parse response!")
+        raise SystemExit(1)
+
+    if 'error' in resp:
+        print(f"\nAPI ERROR: {resp['error']}")
+        print(f"Details: {resp.get('details','')}")
+        print(f"Node errors: {resp.get('node_errors',{})}")
+        raise SystemExit(1)
+
+    prompt_id = resp.get('prompt_id', 'unknown')
+    print(f"Prompt ID: {prompt_id}")
+
+    # Step 4: Monitor (15s intervals, up to 30 min)
+    print("\nMonitoring generation (GPU sampling + CPU VAE)...")
+    last_log = ""
+    for i in range(120):
+        time.sleep(15)
+
+        stats, _ = run("bash -c 'PID=$(pgrep -f \"python3 main.py\" | head -1); "
+            "if [ -n \"$PID\" ]; then "
+            "  CPU=$(ps -p $PID -o %cpu --no-headers); "
+            "  MEM=$(ps -p $PID -o rss --no-headers); "
+            "  echo \"CPU:${CPU}% RSS:$((MEM/1024))M LOAD:$(cut -d\" \" -f1-3 /proc/loadavg)\"; "
+            "else echo DEAD; fi'")
+
+        log, _ = run("tail -8 /home/fabian/comfyui.log 2>/dev/null")
+
+        m, s = divmod((i+1)*15, 60)
+        print(f"  [{m}m{s:02d}s] {stats.strip()}")
+
+        if log.strip() != last_log:
+            for line in reversed(log.strip().split('\n')):
+                l = line.strip()
+                if l and not l.startswith('FETCH') and not l.startswith('[DEPRECATION') and not l.startswith('[ComfyUI-Manager]'):
+                    print(f"    LOG: {l[:120]}")
+                    break
+            last_log = log.strip()
+
+        if 'DEAD' in stats:
+            print("\nPROCESS DIED!")
+            out, _ = run("tail -50 /home/fabian/comfyui.log")
+            print(out)
+            break
+        if 'Prompt executed in' in log:
+            print("\nSUCCESS! Image generated!")
+            out, _ = run("tail -20 /home/fabian/comfyui.log")
+            print(out)
+            break
+        if 'Traceback' in log or 'RuntimeError' in log:
+            print("\nERROR detected!")
+            out, _ = run("tail -50 /home/fabian/comfyui.log")
+            print(out)
+            break
+
+    # Output files
+    print("\n=== Output files ===")
+    out, _ = run("ls -lah /home/fabian/ComfyUI/output/")
+    print(out.strip())
+
+finally:
+    ssh.close()
+    print("\nSSH closed.")
--- a/Show More
+++ b/Show More