vLLM/vllm/model_executor/warmup/kernel_warmup.py

# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
Warmup kernels used during model execution.
This is useful specifically for JIT'ed kernels as we don't want JIT'ing to
happen during model execution.
"""

from typing import TYPE_CHECKING

import torch

import vllm.envs as envs
from vllm.logger import init_logger
from vllm.model_executor.warmup.deep_gemm_warmup import deep_gemm_warmup
from vllm.model_executor.warmup.deepseek_v4_mhc_warmup import (
    deepseek_v4_mhc_warmup,
)
from vllm.model_executor.warmup.flashinfer_autotune_cache import (
    resolve_flashinfer_autotune_file,
    write_flashinfer_autotune_cache,
)
from vllm.model_executor.warmup.flashinfer_sparse_mla_warmup import (
    deepseek_v4_sparse_mla_attention_warmup,
    flashinfer_sparse_mla_decode_autotune_warmup,
)
from vllm.model_executor.warmup.qwen_triton_warmup import qwen_triton_warmup
from vllm.platforms import current_platform
from vllm.utils.deep_gemm import is_deep_gemm_supported
from vllm.utils.flashinfer import has_flashinfer

if TYPE_CHECKING:
    from vllm.v1.worker.gpu_model_runner import GPUModelRunner
    from vllm.v1.worker.gpu_worker import Worker

logger = init_logger(__name__)


def kernel_warmup(worker: "Worker"):
    from vllm.model_executor.warmup.minimax_m3_msa_warmup import (
        minimax_m3_msa_warmup,
    )

    qwen_triton_warmup(worker.model_runner, worker.vllm_config.model_config)

    # DSv4 mHC TileLang kernels (hc_pre/hc_post/hc_head_op) run every decoder
    # layer per token; warm them across token sizes first so the first real
    # request doesn't pay JIT cost. No-op for non-DSv4 models (gated inside).
    deepseek_v4_mhc_warmup(
        worker.get_model(),
        max_tokens=worker.scheduler_config.max_num_batched_tokens,
        cudagraph_capture_sizes=(
            worker.vllm_config.compilation_config.cudagraph_capture_sizes or []
        ),
    )

    # Run next so input-prep kernels JIT against pristine runner state.
    flashinfer_sparse_mla_decode_autotune_warmup(worker)
    deepseek_v4_sparse_mla_attention_warmup(worker)

    # Deep GEMM warmup
    do_deep_gemm_warmup = (
        envs.VLLM_USE_DEEP_GEMM
        and is_deep_gemm_supported()
        and envs.VLLM_DEEP_GEMM_WARMUP != "skip"
    )
    if do_deep_gemm_warmup:
        model = worker.get_model()
        max_tokens = worker.scheduler_config.max_num_batched_tokens
        deep_gemm_warmup(model, max_tokens)

    minimax_m3_msa_warmup(worker)

    enable_flashinfer_autotune = (
        worker.vllm_config.kernel_config.enable_flashinfer_autotune
    )
    # FlashInfer autotune for Hopper (SM 9.0) and Blackwell (SM 10.0) GPUs
    if enable_flashinfer_autotune is False:
        logger.info("Skipping FlashInfer autotune because it is disabled.")
    elif has_flashinfer() and current_platform.has_device_capability(90):
        flashinfer_autotune(worker.model_runner)

    # FlashInfer attention warmup
    # Only warmup if the model has FlashInfer attention groups
    # and is not a pooling model
    def _is_flashinfer_backend(backend):
        try:
            return backend.get_name() == "FLASHINFER"
        except NotImplementedError:
            return False

    if (
        not worker.model_runner.is_pooling_model
        and worker.model_runner.attn_groups
        # NOTE: This should be `any` instead of `all` but other hybrid attention
        # backends don't support this dummy run. Once we remove
        # `build_for_cudagraph_capture`, we can change it to `any`.
        and all(
            _is_flashinfer_backend(group.backend)
            for groups in worker.model_runner.attn_groups
            for group in groups
        )
    ):
        logger.info("Warming up FlashInfer attention.")
        # Warmup with mixed batch containing both prefill and decode tokens
        # This is to warm up both prefill and decode attention kernels
        worker.model_runner._dummy_run(
            num_tokens=16,
            skip_eplb=True,
            is_profile=True,
            force_attention=True,
            create_mixed_batch=True,
        )


# TODO: remove once FlashInfer upstream fixes the persistent file cache
# to resolve collisions like `use_8x4_sf_layout=True/False`, which causes
# invalid tactics to be chosen
_FLASHINFER_USE_PERSISTENT_CACHE = False


def flashinfer_autotune(runner: "GPUModelRunner") -> None:
    """
    Autotune FlashInfer operations.
    FlashInfer have many implementations for the same operation,
    autotuning runs benchmarks for each implementation and stores
    the results. The results are cached transparently and
    future calls to FlashInfer will use the best implementation.
    Without autotuning, FlashInfer will rely on heuristics, which may
    be significantly slower.

    Tuning is performed only on rank 0. The resulting cache is broadcast
    to every rank so all ranks dispatch the same kernel tactic.
    """
    import vllm.utils.flashinfer as fi_utils
    from vllm.distributed.parallel_state import get_world_group

    if not _FLASHINFER_USE_PERSISTENT_CACHE:
        with torch.inference_mode(), fi_utils.autotune():
            runner._dummy_run(
                num_tokens=runner.scheduler_config.max_num_batched_tokens,
                skip_eplb=True,
                is_profile=True,
            )
        get_world_group().barrier()
        return

    world = get_world_group()
    is_leader = world.rank_in_group == 0

    cache_path = resolve_flashinfer_autotune_file(runner)
    if is_leader:
        logger.info("Using FlashInfer autotune cache file: %s", cache_path)

    # We skip EPLB here since we don't want to record dummy metrics.
    # When autotuning with number of tokens m, flashinfer will autotune
    # operations for all number of tokens up to m, so we only need to
    # run with the max number of tokens.
    dummy_run_kwargs = dict(
        num_tokens=runner.scheduler_config.max_num_batched_tokens,
        skip_eplb=True,
        is_profile=True,
    )

    with torch.inference_mode():
        if is_leader:
            with fi_utils.autotune(tune_mode=True, cache=str(cache_path)):
                runner._dummy_run(**dummy_run_kwargs)
        else:
            runner._dummy_run(**dummy_run_kwargs)

    # Broadcast autotune cache from rank 0 to all other ranks so every
    # rank loads the same set of chosen tactics.
    tune_results: bytes | None = None
    if is_leader and cache_path.exists():
        with open(cache_path, "rb") as f:
            tune_results = f.read()

    tune_results = world.broadcast_object(tune_results, src=0)

    if tune_results is None:
        logger.warning(
            "No FlashInfer autotune cache entries found."
            "Falling back to default tactics."
        )
    else:
        write_flashinfer_autotune_cache(cache_path, tune_results)
        world.barrier()
        from flashinfer.autotuner import AutoTuner

        AutoTuner.get().load_configs(str(cache_path))
        logger.info(
            "FlashInfer autotune cache loaded on rank %d from %s.",
            world.rank_in_group,
            cache_path,
        )