#!/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.")