[3/N][warmup][DSv4] Migrate FA4 MLA and shared CuTeDSL kernels (#53565)
Signed-off-by: LopezCastroRoberto <rocastro@redhat.com> Co-authored-by: Codex <codex@openai.com> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
co-authored by
Codex
mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent
1713b9866a
commit
42801b3a6b
@@ -1,53 +0,0 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""Warm up FA4 CuTeDSL kernels."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from typing import TYPE_CHECKING
|
||||
|
||||
from vllm.v1.attention.backends.mla.prefill import get_mla_prefill_backend
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from vllm.v1.worker.gpu_worker import Worker
|
||||
|
||||
|
||||
def _warm_fa4_mla_prefill(worker: Worker) -> None:
|
||||
runner = worker.model_runner
|
||||
if runner.is_pooling_model:
|
||||
return
|
||||
|
||||
vllm_config = runner.vllm_config
|
||||
if not vllm_config.model_config.use_mla:
|
||||
return
|
||||
|
||||
try:
|
||||
backend_cls = get_mla_prefill_backend(vllm_config)
|
||||
except ValueError:
|
||||
# fall back to top-k MQA prefill path.
|
||||
return
|
||||
if backend_cls.get_name() != "FLASH_ATTN":
|
||||
return
|
||||
|
||||
from vllm.v1.attention.backends.mla.prefill import flash_attn
|
||||
|
||||
flash_attn.FA4_MLA_PREFILL_KERNEL.warmup(vllm_config)
|
||||
|
||||
|
||||
def _warm_inkling_fa4_rel_attention(worker: Worker) -> None:
|
||||
from vllm.models.inkling.configs import InklingMMConfig, InklingModelConfig
|
||||
from vllm.models.inkling.nvidia.ops.fa4_rel_attention import (
|
||||
INKLING_FA4_REL_ATTENTION_KERNEL,
|
||||
)
|
||||
|
||||
vllm_config = worker.vllm_config
|
||||
hf_config = vllm_config.model_config.hf_config
|
||||
if not isinstance(hf_config, (InklingMMConfig, InklingModelConfig)):
|
||||
return
|
||||
|
||||
INKLING_FA4_REL_ATTENTION_KERNEL.warmup(vllm_config)
|
||||
|
||||
|
||||
def fa4_cutedsl_warmup(worker: Worker) -> None:
|
||||
_warm_fa4_mla_prefill(worker)
|
||||
_warm_inkling_fa4_rel_attention(worker)
|
||||
@@ -20,9 +20,6 @@ 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.fa4_cutedsl_warmup import (
|
||||
fa4_cutedsl_warmup,
|
||||
)
|
||||
from vllm.model_executor.warmup.flashinfer_autotune_cache import (
|
||||
resolve_flashinfer_autotune_file,
|
||||
write_flashinfer_autotune_cache,
|
||||
@@ -169,7 +166,6 @@ def kernel_warmup(worker: "Worker", *, process_local_only: bool = False):
|
||||
# Run next so input-prep kernels JIT against pristine runner state.
|
||||
if worker.vllm_config.kernel_config.enable_jit_warmup:
|
||||
kimi_k3_triton_warmup(worker)
|
||||
fa4_cutedsl_warmup(worker)
|
||||
spec_decode_rejection_warmup(worker)
|
||||
qwen4_exp_qsa_triton_warmup(worker)
|
||||
|
||||
|
||||
Reference in New Issue
Block a user