amd-strix-halo-vllm-toolboxes/benchmarks/vllm_cluster_bench.py

#!/usr/bin/env python3
import subprocess, time, json, sys, os, requests, argparse, re
from pathlib import Path

# Import models immediately to access globals
try:
    import models
except ImportError:
    # If in /opt, this should work if path includes ., otherwise:
    sys.path.append(os.getcwd())
    try:
        import models
        # Also try parent/scripts for local dev if above failed?
    except ImportError:
        sys.path.append(str(Path(__file__).parent.parent / "scripts"))
        import models

# =========================
# ⚙️ GLOBAL SETTINGS
# =========================

# CLUSTER CONFIG: 2x Strix Halo (TP=2)
# User requested specifically to test with TP=2 on the cluster.
CLUSTER_TP = 2
GPU_UTIL = "0.90" 
FORCE_ETH = False
FORCE_DEBUG_NCCL = False

# THROUGHPUT CONFIG (Imported from models.py)
OFF_NUM_PROMPTS      = models.OFF_NUM_PROMPTS
OFF_FORCED_OUTPUT    = models.OFF_FORCED_OUTPUT
DEFAULT_BATCH_TOKENS = models.DEFAULT_BATCH_TOKENS

RESULTS_DIR = Path("benchmark_results")
RESULTS_DIR.mkdir(exist_ok=True)

# Reuse the model table from the main benchmark script
# We can just import it or copy it. Importing is cleaner but might rely on path.
# For standalone robustness, I will copy the minimal needed config or import if possible.
# Since this is a new file in root/benchmarks? No, likely scripts/ or same dir.
# Let's assume it's in the same dir as run_vllm_bench.py.


MODEL_TABLE = models.MODEL_TABLE
MODELS_TO_RUN = models.MODELS_TO_RUN


# =========================
# UTILS (Adapted for Cluster)
# =========================


# =========================
# CLUSTER MANAGER INTEGRATION
# =========================
try:
    import cluster_manager
except ImportError:
    sys.path.append(str(Path(__file__).parent.parent / "scripts"))
    import cluster_manager

# Defaults for Cluster
HEAD_IP = os.getenv("VLLM_HEAD_IP", "192.168.100.1")
WORKER_IP = os.getenv("VLLM_WORKER_IP", "192.168.100.2")

def log(msg): print(f"\n[CLUSTER-BENCH] {msg}")

def restart_cluster():
    log("Restarting Ray Cluster (Clean State)...")
    
    # Push config to env so cluster_manager picks it up for daemon injection
    os.environ["NCCL_IB_DISABLE"] = "1" if FORCE_ETH else "0"
    if FORCE_DEBUG_NCCL:
        os.environ["NCCL_DEBUG"] = "INFO"
        os.environ["NCCL_DEBUG_SUBSYS"] = "INIT,NET"
    else:
        os.environ.pop("NCCL_DEBUG", None)
        os.environ.pop("NCCL_DEBUG_SUBSYS", None)
        
    # 1. Stop Cluster (Best Effort)
    cluster_manager.stop_cluster()
    
    # 2. Start Head
    if not cluster_manager.setup_head_node(HEAD_IP):
        log("ERROR: Failed to start HEAD node.")
        sys.exit(1)
        
    # 3. Start Worker
    # Give head a moment
    time.sleep(5)
    if not cluster_manager.setup_worker_node(WORKER_IP, HEAD_IP):
        log("ERROR: Failed to start WORKER node.")
        sys.exit(1)
        
    # 4. Wait
    if not cluster_manager.wait_for_cluster():
        log("ERROR: Cluster failed to initialize.")
        sys.exit(1)
        
    log("Cluster Ready.")

def get_net_iface():
    return cluster_manager.get_net_iface()

def get_local_ip(iface):
    return cluster_manager.get_local_ip(iface)

def nuke_vllm_cache():
    # We use explicit IPs because ray status might return Hex IDs which we can't SSH to.
    cluster_manager.nuke_vllm_cache_cluster(nodes=[HEAD_IP, WORKER_IP])


def get_dataset():
    # Same as original
    data_path = Path("ShareGPT_V3_unfiltered_cleaned_split.json")
    if data_path.exists(): return str(data_path)
    
    log("Downloading ShareGPT dataset...")
    url = "https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json"
    try:
        r = requests.get(url, stream=True, timeout=15)
        r.raise_for_status()
        with open(data_path, 'wb') as f:
            for chunk in r.iter_content(chunk_size=8192): f.write(chunk)
        return str(data_path)
    except Exception as e:
        log(f"WARNING: ShareGPT download failed ({e}). using RANDOM.")
        return None

def get_cluster_env():
    # Detect Interface and IP
    rdma_iface = get_net_iface()
    host_ip = get_local_ip(rdma_iface)
    
    env = os.environ.copy()
    
    # Critical Cluster Envs (Match start_vllm_cluster.py)
    env["RAY_EXPERIMENTAL_NOSET_ROCR_VISIBLE_DEVICES"] = "1"
    env["VLLM_HOST_IP"] = host_ip
    env["NCCL_SOCKET_IFNAME"] = rdma_iface
    env["GLOO_SOCKET_IFNAME"] = rdma_iface
    # RCCL specific
    env["NCCL_IB_GID_INDEX"] = "1"
    env["NCCL_IB_DISABLE"] = "1" if FORCE_ETH else "0"
    env["NCCL_NET_GDR_LEVEL"] = "0"
    
    # Stability for RDMA (Fix for high-throughput models like Gemma 3)
    env["NCCL_IB_TIMEOUT"] = "23"  # ~32 seconds (default is 18/~1s)
    env["NCCL_IB_RETRY_CNT"] = "7" # Default is 3, increase for lossy networks
    
    if FORCE_DEBUG_NCCL:
        env["NCCL_DEBUG"] = "INFO"
        env["NCCL_DEBUG_SUBSYS"] = "INIT,NET"
    
    return env

def get_model_args(model):
    config = MODEL_TABLE.get(model, {"max_num_seqs": "32"})
    util = config.get("gpu_util", GPU_UTIL)

    cmd = [
        "--model", model,
        "--gpu-memory-utilization", util,
        "--dtype", "auto",
        "--tensor-parallel-size", str(CLUSTER_TP),
        "--max-num-seqs", config["max_num_seqs"],
        "--distributed-executor-backend", "ray"
    ]
    
    # Optional ctx
    if "ctx" in config:
        cmd.extend(["--max-model-len", config["ctx"]])
        
    if config.get("trust_remote"): cmd.append("--trust-remote-code")
    
    # Force eager mode for cluster stability
    cmd.append("--enforce-eager")
    
    return cmd

def get_benchmark_output_file(model, output_dir):
    model_safe = model.replace("/", "_")
    output_dir_path = Path(output_dir)
    eth_suffix = "_eth" if FORCE_ETH else ""
    return output_dir_path / f"{model_safe}_cluster_tp{CLUSTER_TP}{eth_suffix}_throughput.json"

def run_bench_set(model, backend_name, output_dir, extra_env=None):
    output_dir_path = Path(output_dir)
    output_dir_path.mkdir(parents=True, exist_ok=True)
    
    output_file = get_benchmark_output_file(model, output_dir)
    
    if output_file.exists():
        log(f"SKIP {model} [{backend_name}] (Result exists)")
        return

    dataset_path = get_dataset()
    dataset_args = ["--dataset-name", "sharegpt", "--dataset-path", dataset_path] if dataset_path else ["--input-len", "1024"]
    
    batch_tokens = MODEL_TABLE[model].get("max_tokens", DEFAULT_BATCH_TOKENS)

    log(f"START {model} [TP={CLUSTER_TP} | {backend_name}]...")
    
    nuke_vllm_cache()

    cmd = ["vllm", "bench", "throughput"] + get_model_args(model)
    cmd.extend([
        "--num-prompts", str(OFF_NUM_PROMPTS),
        "--max-num-batched-tokens", batch_tokens,
        "--output-len", OFF_FORCED_OUTPUT,
        "--output-json", str(output_file),
        "--disable-log-stats"
    ])
    cmd.extend(dataset_args)

    if backend_name == "ROCm-Attn":
        cmd.extend(["--attention-backend", "ROCM_ATTN"])

    env = get_cluster_env()
    
    # Model specific envs
    model_env = MODEL_TABLE[model].get("env", {})
    env.update(model_env)
    
    # Run specific envs (e.g. ROCm attention)
    if extra_env:
        env.update(extra_env)

    try: 
        log(f"Command: {' '.join(cmd)}")
        subprocess.run(cmd, check=True, env=env)
    except subprocess.CalledProcessError as e:
        log(f"ERROR: Failed {model} [{backend_name}] (Exit {e.returncode})")
    except Exception as e:
        log(f"ERROR: System error: {e}")

def run_cluster_throughput(model):
    # 1. Default Run (Triton)
    if get_benchmark_output_file(model, RESULTS_DIR).exists():
        log(f"SKIP {model} [Default] (Result exists)")
    else:
        restart_cluster()
        run_bench_set(
            model, 
            "Default", 
            RESULTS_DIR
        )
    
    # 2. ROCm Attention Run
    if get_benchmark_output_file(model, "benchmark_results_rocm").exists():
        log(f"SKIP {model} [ROCm-Attn] (Result exists)")
    else:
        restart_cluster()
        run_bench_set(
            model,
            "ROCm-Attn",
            "benchmark_results_rocm",
            extra_env={}
        )


def print_summary():
    eth_suffix = "_eth" if FORCE_ETH else ""
    title_suffix = " (Ethernet ONLY)" if FORCE_ETH else ""
    print(f"\n{f'MODEL (TP=2){title_suffix}':<50} | {'Triton':<8} | {'ROCm':<8}")
    print("-" * 75)
    
    for m in MODELS_TO_RUN:
        msafe = m.replace("/", "_")
        
        # Default
        try: 
            p1 = RESULTS_DIR / f"{msafe}_cluster_tp{CLUSTER_TP}{eth_suffix}_throughput.json"
            d1 = json.loads(p1.read_text())
            val1 = f"{d1.get('tokens_per_second', 0):.1f}"
        except: val1 = "N/A"
        
        # ROCm
        try:
            p2 = Path("benchmark_results_rocm") / f"{msafe}_cluster_tp{CLUSTER_TP}{eth_suffix}_throughput.json"
            d2 = json.loads(p2.read_text())
            val2 = f"{d2.get('tokens_per_second', 0):.1f}"
        except: val2 = "N/A"

        name_cell = m.split('/')[-1]
        print(f"{name_cell:<50} | {val1:<8} | {val2:<8}")
    print("-" * 75)

if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="VLLM Cluster Benchmark")
    parser.add_argument("--eth-only", action="store_true", help="Run benchmark using only Ethernet (disable RDMA/RoCE)")
    parser.add_argument("--debug-nccl", action="store_true", help="Enable NCCL Debug logging (INFO level for Transport tracking)")
    args = parser.parse_args()
    
    FORCE_ETH = args.eth_only
    FORCE_DEBUG_NCCL = args.debug_nccl

    log("Ray Cluster Detected. Starting Benchmarks (Dual Backend)...")
    if FORCE_ETH:
        log("Note: Ethernet ONLY mode enabled. RDMA/RoCE disabled.")
    if FORCE_DEBUG_NCCL:
        log("Note: NCCL Debug mode enabled (Transport Logging).")
    log("Note: Eager Mode (--enforce-eager) is ENABLED for cluster stability.")
    
    for m in MODELS_TO_RUN:
        run_cluster_throughput(m)
        
    print_summary()