[Feat][Perf] Add new warmup infrastructure for JITs (#47451)

Signed-off-by: LopezCastroRoberto <rocastro@redhat.com>
Signed-off-by: Roberto L. Castro <38211239+LopezCastroRoberto@users.noreply.github.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Roberto L. Castro
2026-07-20 13:21:55 -07:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent fbfe58133d
commit 7ca017778f
14 changed files with 1855 additions and 693 deletions
+11 -2
View File
@@ -17,6 +17,9 @@ 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,
@@ -27,7 +30,7 @@ from vllm.model_executor.warmup.flashinfer_sparse_mla_warmup import (
)
from vllm.model_executor.warmup.qwen_triton_warmup import qwen_triton_warmup
from vllm.model_executor.warmup.sparse_mla_triton_warmup import (
sparse_mla_triton_warmup_if_needed,
sparse_mla_triton_warmup,
)
from vllm.model_executor.warmup.v1_block_table_warmup import (
warm_v1_block_table_kernels,
@@ -94,7 +97,6 @@ def kernel_warmup(worker: "Worker"):
)
# Run next so input-prep kernels JIT against pristine runner state.
sparse_mla_triton_warmup_if_needed(worker)
flashinfer_sparse_mla_decode_autotune_warmup(worker)
deepseek_v4_sparse_mla_attention_warmup(worker)
@@ -156,8 +158,15 @@ def kernel_warmup(worker: "Worker"):
)
if worker.vllm_config.kernel_config.enable_cutedsl_warmup:
# TODO(roberto): Remove after registered CuTeDSL warmups are migrated
# to the shared JIT warmup infrastructure.
# https://github.com/vllm-project/vllm/pull/47451
cutedsl_warmup()
if worker.vllm_config.kernel_config.enable_jit_warmup:
fa4_cutedsl_warmup(worker)
sparse_mla_triton_warmup(worker)
def _flashinfer_autotune_skip_ops(runner: "GPUModelRunner") -> set[str] | None:
if envs.VLLM_FLASHINFER_AUTOTUNE_SKIP_OPS is not None: