"""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()