"""Remove --cpu-vae: Let VAE run on GPU (only 320MB, easily fits). Restart ComfyUI 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)

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()

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

# Kill old
print("Killing ComfyUI...")
sh('pkill -9 -f "python3.*main.py" 2>/dev/null; sleep 2')

# New launcher WITHOUT --cpu-vae
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 computes (correct for shared-memory APU)
    # --force-fp16: half precision
    # NO --cpu-vae: VAE is only 320MB, runs fine on GPU and much faster
    exec python3 main.py \\
        --listen 0.0.0.0 --port 8188 \\
        --novram \\
        --force-fp16 \\
        --disable-smart-memory
""")
sftp_write('/tmp/run_comfyui.sh', launcher)
sh('chmod +x /tmp/run_comfyui.sh')
print("Launcher updated: --novram --force-fp16 (NO --cpu-vae)")

# Start
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}")

# Wait for ready
print("Waiting for server...", 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:
        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)

# Submit workflow
print("\nSubmitting 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]}},
        "9": {"class_type": "CLIPTextEncode", "inputs": {"text": "", "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": ["9", 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_v2"}}
    }
}
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[:150]}")

# Monitor
print("\nMonitoring...")
t0 = time.time()
for i in range(120):
    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:
        log = sftp_read('/tmp/comfyui.log')
    except:
        log = ''
    
    # Find last meaningful line
    last = ''
    sampling = ''
    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_v2*.png 2>/dev/null', timeout=5)
    if imgs:
        print(f"\n*** IMAGE GENERATED! ***")
        print(f"File: {imgs}")
        print(f"Total time: {elapsed}s")
        # Show timing from log
        for line in log.split('\n')[-15:]:
            s = line.strip()
            if s and 'FETCH' not in s and 'startup' not in s:
                print(f"  {s}")
        break
    
    # Queue check
    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_v2*.png 2>/dev/null', timeout=5)
            if imgs:
                print(f"\n*** IMAGE GENERATED! ***")
                print(f"File: {imgs}")
                print(f"Total time: {elapsed}s")
            else:
                print(f"\nQueue 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("\n*** CRASHED ***")
        for line in log.split('\n')[-30:]:
            if line.strip():
                print(f"  {line.strip()}")
        break
    
    time.sleep(15)

c.close()
print("\nDone.")