[Bug]: LLM initialization time increases significantly with larger tensor parallel size and Ray

[piood]

Your current environment

vllm 0.5.2

The output of `python collect_env.py`

Collecting environment information...
PyTorch version: 2.3.1+cu121
Is debug build: False
CUDA used to build PyTorch: 12.1
ROCM used to build PyTorch: N/A

OS: Ubuntu 20.04.5 LTS (x86_64)
GCC version: (Ubuntu 9.4.0-1ubuntu1~20.04.1) 9.4.0
Clang version: Could not collect
CMake version: version 3.24.1
Libc version: glibc-2.31

Python version: 3.8.10 (default, Mar 13 2023, 10:26:41)  [GCC 9.4.0] (64-bit runtime)
Python platform: Linux-5.10.134-008.7.kangaroo.al8.x86_64-x86_64-with-glibc2.29
Is CUDA available: True
CUDA runtime version: 12.1.66
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: 
GPU 0: NVIDIA L20Z
GPU 1: NVIDIA L20Z
GPU 2: NVIDIA L20Z
GPU 3: NVIDIA L20Z
GPU 4: NVIDIA L20Z
GPU 5: NVIDIA L20Z
GPU 6: NVIDIA L20Z
GPU 7: NVIDIA L20Z

Nvidia driver version: 535.161.08
cuDNN version: Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.8.9.0
/usr/lib/x86_64-linux-gnu/libcudnn_adv_infer.so.8.9.0
/usr/lib/x86_64-linux-gnu/libcudnn_adv_train.so.8.9.0
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_infer.so.8.9.0
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_train.so.8.9.0
/usr/lib/x86_64-linux-gnu/libcudnn_ops_infer.so.8.9.0
/usr/lib/x86_64-linux-gnu/libcudnn_ops_train.so.8.9.0
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Architecture:                    x86_64
CPU op-mode(s):                  32-bit, 64-bit
Byte Order:                      Little Endian
Address sizes:                   52 bits physical, 57 bits virtual
CPU(s):                          100
On-line CPU(s) list:             0-99
Thread(s) per core:              1
Core(s) per socket:              100
Socket(s):                       1
NUMA node(s):                    1
Vendor ID:                       GenuineIntel
CPU family:                      6
Model:                           143
Model name:                      Intel(R) Xeon(R) Processor
Stepping:                        8
CPU MHz:                         2000.000
BogoMIPS:                        4000.00
Hypervisor vendor:               KVM
Virtualization type:             full
L1d cache:                       4.7 MiB
L1i cache:                       3.1 MiB
L2 cache:                        200 MiB
L3 cache:                        105 MiB
NUMA node0 CPU(s):               0-99
Vulnerability Itlb multihit:     Not affected
Vulnerability L1tf:              Not affected
Vulnerability Mds:               Not affected
Vulnerability Meltdown:          Not affected
Vulnerability Mmio stale data:   Not affected
Vulnerability Retbleed:          Not affected
Vulnerability Spec store bypass: Vulnerable
Vulnerability Spectre v1:        Vulnerable: __user pointer sanitization and usercopy barriers only; no swapgs barriers
Vulnerability Spectre v2:        Vulnerable, IBPB: disabled, STIBP: disabled, PBRSB-eIBRS: Vulnerable
Vulnerability Srbds:             Not affected
Vulnerability Tsx async abort:   Not affected
Flags:                           fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc rep_good nopl xtopology nonstop_tsc cpuid tsc_known_freq pni pclmulqdq ssse3 fma cx16 pdcm pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch cpuid_fault invpcid_single ssbd ibrs ibpb stibp ibrs_enhanced fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves avx_vnni avx512_bf16 wbnoinvd avx512vbmi umip pku waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq rdpid cldemote movdiri movdir64b fsrm md_clear serialize tsxldtrk avx512_fp16 arch_capabilities

Versions of relevant libraries:
[pip3] flake8==6.1.0
[pip3] mypy-extensions==1.0.0
[pip3] numpy==1.22.2
[pip3] onnx==1.13.1
[pip3] pytorch-quantization==2.1.2
[pip3] torch==2.3.1
[pip3] torch-tensorrt==1.4.0.dev0
[pip3] torchaudio==2.3.1
[pip3] torchtext==0.13.0a0+fae8e8c
[pip3] torchtyping==0.1.4
[pip3] torchvision==0.18.1
[pip3] triton==2.3.1
[conda] Could not collect

Model Input Dumps

just test the vllm init time

🐛 Describe the bug

Issue Description

We observed significant and unexpected increases in VLLM initialization time when scaling tensor parallelism (TP), especially with Ray enabled.

Observed Behavior

• TP=1: ~7 seconds initialization time
• TP=4: ~14 seconds initialization time
• TP=4 with Ray: ~24 seconds initialization time

Expected Behavior

Initialization time should remain relatively constant or have minimal increase when scaling tensor parallelism and use ray.

Environment

• VLLM version: 0.5.2
• Model: Qwen2-7B
• GPU: NVIDIA L20Z
• Number of GPUs: 8

Additional Context

The initialization time increase appears disproportionate to the tensor parallel size, suggesting a potential bottleneck in the initialization process, particularly when Ray is involved.

Reproducible Steps

1. Run VLLM with TP=1
2. Run VLLM with TP=4
3. Run VLLM with TP=4 and Ray enabled
4. Compare initialization times

vllm start time

def run_vllm(
    requests: List[Tuple[str, int, int]],
    model: str,
    tokenizer: str,
    quantization: Optional[str],
    tensor_parallel_size: int,
    seed: int,
    n: int,
    use_beam_search: bool,
    trust_remote_code: bool,
    dtype: str,
    max_model_len: Optional[int],
    enforce_eager: bool,
    kv_cache_dtype: str,
    quantization_param_path: Optional[str],
    device: str,
    enable_prefix_caching: bool,
    enable_chunked_prefill: bool,
    max_num_batched_tokens: int,
    distributed_executor_backend: Optional[str],
    gpu_memory_utilization: float = 0.9,
    num_scheduler_steps: int = 1,
    use_v2_block_manager: bool = False,
    download_dir: Optional[str] = None,
    load_format: str = EngineArgs.load_format,
    disable_async_output_proc: bool = False,
) -> float:
    # 导入必要的库
    from vllm import LLM, SamplingParams

    print(f"Start initializing LLM at {time.strftime('%Y-%m-%d %H:%M:%S')}")
    start = time.perf_counter()
    llm = LLM(
        model=model,
        tokenizer=tokenizer,
        quantization=quantization,
        tensor_parallel_size=tensor_parallel_size,
        seed=seed,
        trust_remote_code=trust_remote_code,
        dtype=dtype,
        max_model_len=max_model_len,
        gpu_memory_utilization=gpu_memory_utilization,
        enforce_eager=enforce_eager,
        kv_cache_dtype=kv_cache_dtype,
        quantization_param_path=quantization_param_path,
        device=device,
        enable_prefix_caching=enable_prefix_caching,
        download_dir=download_dir,
        enable_chunked_prefill=enable_chunked_prefill,
        max_num_batched_tokens=max_num_batched_tokens,
        distributed_executor_backend=distributed_executor_backend,
        load_format=load_format,
        # num_scheduler_steps=num_scheduler_steps,
        # use_v2_block_manager=use_v2_block_manager,
        # disable_async_output_proc=disable_async_output_proc,
    )
    end = time.perf_counter()
    print(f"Finish initializing LLM at {time.strftime('%Y-%m-%d %H:%M:%S')}")
    print(f"vllm init time: {end - start}")

vllm ray start time

def run_ray_vllm(
    requests: List[Tuple[str, int, int]],
    model: str,
    tokenizer: str,
    quantization: Optional[str],
    tensor_parallel_size: int,
    seed: int,
    n: int,
    use_beam_search: bool,
    trust_remote_code: bool,
    dtype: str,
    max_model_len: Optional[int],
    enforce_eager: bool,
    kv_cache_dtype: str,
    quantization_param_path: Optional[str],
    device: str,
    enable_prefix_caching: bool,
    enable_chunked_prefill: bool,
    max_num_batched_tokens: int,
    distributed_executor_backend: Optional[str],
    gpu_memory_utilization: float = 0.9,
    num_scheduler_steps: int = 1,
    use_v2_block_manager: bool = False,
    download_dir: Optional[str] = None,
    load_format: str = EngineArgs.load_format,
    disable_async_output_proc: bool = False,
) -> float:
    # 导入必要的库
    from vllm import LLM, SamplingParams

    import ray

    @ray.remote
    class LLMWorker:
        def __init__(self, model, tokenizer, quantization, tensor_parallel_size, seed, trust_remote_code, dtype, max_model_len, gpu_memory_utilization, enforce_eager, kv_cache_dtype, quantization_param_path, device, enable_prefix_caching, download_dir, enable_chunked_prefill, max_num_batched_tokens, distributed_executor_backend, load_format, num_scheduler_steps, use_v2_block_manager, disable_async_output_proc):
            from vllm import LLM
            start = time.perf_counter()
            self.llm = LLM(
                model=model,
                tokenizer=tokenizer,
                quantization=quantization,
                tensor_parallel_size=tensor_parallel_size,
                seed=seed,
                trust_remote_code=trust_remote_code,
                dtype=dtype,
                max_model_len=max_model_len,
                gpu_memory_utilization=gpu_memory_utilization,
                enforce_eager=enforce_eager,
                kv_cache_dtype=kv_cache_dtype,
                quantization_param_path=quantization_param_path,
                device=device,
                enable_prefix_caching=enable_prefix_caching,
                download_dir=download_dir,
                enable_chunked_prefill=enable_chunked_prefill,
                max_num_batched_tokens=max_num_batched_tokens,
                distributed_executor_backend=distributed_executor_backend,
                load_format=load_format,
                # num_scheduler_steps=num_scheduler_steps,
                # use_v2_block_manager=use_v2_block_manager,
                # disable_async_output_proc=disable_async_output_proc,
            )
            end = time.perf_counter()
            print(f"Finish initializing LLM at {time.strftime('%Y-%m-%d %H:%M:%S')}")
            print(f"vllm init time: {end - start}")

        def generate(self, prompts, sampling_params):
            return self.llm.generate(prompts, sampling_params, use_tqdm=True)

    # Create LLM worker
    worker = LLMWorker.remote(
        model=model,
        tokenizer=tokenizer,
        quantization=quantization,
        tensor_parallel_size=tensor_parallel_size,
        seed=seed,
        trust_remote_code=trust_remote_code,
        dtype=dtype,
        max_model_len=max_model_len,
        gpu_memory_utilization=gpu_memory_utilization,
        enforce_eager=enforce_eager,
        kv_cache_dtype=kv_cache_dtype,
        quantization_param_path=quantization_param_path,
        device=device,
        enable_prefix_caching=enable_prefix_caching,
        download_dir=download_dir,
        enable_chunked_prefill=enable_chunked_prefill,
        max_num_batched_tokens=max_num_batched_tokens,
        distributed_executor_backend=distributed_executor_backend,
        load_format=load_format,
        num_scheduler_steps=num_scheduler_steps,
        use_v2_block_manager=use_v2_block_manager,
        disable_async_output_proc=disable_async_output_proc,
    )

    # Add the requests to the engine.
    prompts: List[str] = []
    sampling_params: List[SamplingParams] = []
    for prompt, _, output_len in requests:
        prompts.append(prompt)
        sampling_params.append(
            SamplingParams(
                n=n,
                temperature=0.0 if use_beam_search else 1.0,
                top_p=1.0,
                use_beam_search=use_beam_search,
                ignore_eos=True,
                max_tokens=output_len,
            )
        )

    start = time.perf_counter()
    ray.get(worker.generate.remote(prompts, sampling_params))
    end = time.perf_counter()
    return end - start

[AaronWang04]

someone correct me if im wrong but the way the workers are initialized are done sequentially on the main process. which can be seen in the function I linked below

vllm/vllm/executor/ray_gpu_executor.py

Line 109 in bbd3e86

def _init_workers_ray(self, placement_group: "PlacementGroup",

ray add additional overhead because you have to send the whole worker configs through Ray which is a slower process

[piood]

someone correct me if im wrong but the way the workers are initialized are done sequentially on the main process. which can be seen in the function I linked below

vllm/vllm/executor/ray_gpu_executor.py

Line 109 in bbd3e86

def _init_workers_ray(self, placement_group: "PlacementGroup",

ray add additional overhead because you have to send the whole worker configs through Ray which is a slower process

Thank you for your answer! However, I still have some concerns about the initialization overhead:

For a 7B model:

• TP=1: 7s initialization time (baseline)
• TP=4: 14s initialization time (+7s overhead)
• TP=4 with Ray: 24s initialization time (+10s additional overhead)

The overhead seems disproportionately large considering:

1. The baseline initialization is only 7 seconds
2. Moving from TP=1 to TP=4 doubles the initialization time
3. Adding Ray introduces an additional 10s overhead, which is even larger than the TP scaling overhead

Is this level of overhead expected? It seems excessive for a 7B model, especially since:

• The TP=1 to TP=4 transition adds 100% overhead
• Ray integration adds another ~71% overhead on top of TP=4

Could there be potential optimization opportunities to reduce these initialization costs?

[AaronWang04]

I don't find that overhead too strange, and there definitely is room for optimizations (parallelizing the process) but engine startup time is not really an important metric that people worry about. (model reloading would probably be the solution more people are interested in that is currently not implemented?) is there a reason you're looking for faster initialization?

[Jack47]

I don't find that overhead too strange, and there definitely is room for optimizations (parallelizing the process) but engine startup time is not really an important metric that people worry about. (model reloading would probably be the solution more people are interested in that is currently not implemented?) is there a reason you're looking for faster initialization?

Great thanks for you relpy!

we want to improve the startup speed, IMHO, 34s is also too long to wait, especially when we are developing new features and what to run some tests to verify it.

[ruisearch42]

I did some benchmarking on Ray executor initialization:

Code:
#11272

Command:

python3 benchmarks/benchmark_latency.py --model meta-llama/Llama-3.1-8B-Instruct --tensor-parallel-size 4  --num-iters-warmup 1 --num-iters 2  --batch-size 8 --input-len 128 --output-len 256 --max-model-len 2048 --no-enable-prefix-caching --distributed-executor-backend ray 

TP = 2
INFO 12-17 22:57:58 llm_engine.py:507] time for initialize_ray_cluster: 2.265391
INFO 12-17 22:58:05 ray_gpu_executor.py:160] time for creating workers: 6.369353
INFO 12-17 22:58:05 ray_gpu_executor.py:252] time for _run_workers update_environment_variables: 0.002733
INFO 12-17 22:58:08 ray_gpu_executor.py:283] time for _run_workers init_device: 1.320933
INFO 12-17 22:58:12 ray_gpu_executor.py:290] time for _run_workers load_model: 3.812319
INFO 12-17 22:58:12 ray_gpu_executor.py:67] time for _init_workers_ray: 12.652736

TP = 4
INFO 12-17 23:00:31 llm_engine.py:507] time for initialize_ray_cluster: 2.435579
INFO 12-17 23:00:44 ray_gpu_executor.py:160] time for creating workers: 12.788581
INFO 12-17 23:00:44 ray_gpu_executor.py:252] time for _run_workers update_environment_variables: 0.003613
INFO 12-17 23:00:46 ray_gpu_executor.py:278] time for _run_workers init_worker: 1.149088
INFO 12-17 23:00:47 ray_gpu_executor.py:283] time for _run_workers init_device: 1.840026
INFO 12-17 23:00:50 ray_gpu_executor.py:290] time for _run_workers load_model: 2.787669
INFO 12-17 23:00:50 ray_gpu_executor.py:67] time for _init_workers_ray: 18.580796

The time to create workers increase proportionally to the number of workers, others are quite constant. I will look into potential ways to optimize.