#!/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.")