Uploaded sanitized BC250/ROCm Repository.

Claude Test Scripts/arch_check.py

@@ -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')

Claude Test Scripts/build_gfx1010.sh

@@ -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!"

Claude Test Scripts/build_notriton.sh

@@ -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!"

Claude Test Scripts/build_pytorch.sh

@@ -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!"

Claude Test Scripts/build_pytorch_gfx1010.sh

@@ -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"

Claude Test Scripts/build_torchvision.sh

@@ -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__)"

Claude Test Scripts/build_v3.sh

@@ -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!"

Claude Test Scripts/build_v4.sh

@@ -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

Claude Test Scripts/check_aotriton.sh

@@ -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

Claude Test Scripts/check_build.sh

@@ -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"

Claude Test Scripts/check_build_state.sh

@@ -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 ==="

Claude Test Scripts/check_clip.sh

@@ -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/

Claude Test Scripts/check_errors.sh

@@ -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"

Claude Test Scripts/check_nodes.sh

@@ -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

Claude Test Scripts/check_progress.sh

@@ -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

Claude Test Scripts/check_state.sh

@@ -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 ==="

Claude Test Scripts/download_models.sh

@@ -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 ==="

Claude Test Scripts/find_tv.sh

@@ -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

Claude Test Scripts/gpu_monitor.sh

@@ -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

Claude Test Scripts/gpu_test.py

@@ -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")

Claude Test Scripts/gpu_test2.py

@@ -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}")

Claude Test Scripts/launch.sh

@@ -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

Claude Test Scripts/launch2.sh

@@ -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

Claude Test Scripts/progress.sh

@@ -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

Claude Test Scripts/start-comfyui.sh

@@ -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:$PATH"
+export LD_LIBRARY_PATH="/opt/rocm/lib"
+
+# --- Performance: DO NOT set these ---
+unset GPU_MAX_HW_QUEUES 2>/dev/null || true
+unset HIP_LAUNCH_BLOCKING 2>/dev/null || true
+unset GGML_CUDA_ENABLE_UNIFIED_MEMORY 2>/dev/null || true
+unset GGML_HIP_HOST_ALLOC 2>/dev/null || true
+unset GGML_CUDA_NO_PINNED 2>/dev/null || true
+unset GGML_HIP_NO_COARSE_GRAIN 2>/dev/null || true
+unset HSA_DISABLE_FRAGMENT_ALLOCATOR 2>/dev/null || true
+
+# --- PyTorch ROCm tuning ---
+export PYTORCH_HIP_ALLOC_CONF="expandable_segments:False"
+
+echo "[OK] ROCm environment configured (gfx1030 spoof)"
+
+# --- Activate venv ---
+cd ~/ComfyUI
+source venv/bin/activate
+
+# --- Launch ComfyUI ---
+echo "[START] Launching ComfyUI on http://0.0.0.0:8188"
+echo "=========================================="
+python main.py \
+    --listen 0.0.0.0 \
+    --port 8188 \
+    --force-fp32 \
+    --lowvram \
+    "$@"

Claude Test Scripts/start_comfyui.sh

@@ -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: $OMP_NUM_THREADS"
+
+# --- Activate venv ---
+cd ~/ComfyUI
+source venv/bin/activate
+
+# --- Launch ComfyUI ---
+echo "[START] Launching ComfyUI on http://0.0.0.0:8188"
+echo "=========================================="
+python main.py \
+    --listen 0.0.0.0 \
+    --port 8188 \
+    --force-fp32 \
+    --lowvram \
+    "$@"

Claude Test Scripts/test_arch.sh

@@ -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

Claude Test Scripts/test_gen2.sh

@@ -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

Claude Test Scripts/test_gen3.sh

@@ -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"

Claude Test Scripts/test_generate.sh

@@ -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

Claude Test Scripts/test_gpu.sh

@@ -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')
+"

Claude Test Scripts/test_runtime.sh

@@ -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

Claude Test Scripts/test_runtime2.sh

@@ -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

Claude Test Scripts/test_tv.sh

@@ -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')
+"