[Elastic EP] Async preparation (#47288)

Signed-off-by: Itay Alroy <ialroy@nvidia.com>
This commit is contained in:
Itay Alroy
2026-07-28 05:19:13 -07:00
committed by GitHub
parent d2bfc6fe20
commit 35efdf6b34
21 changed files with 701 additions and 677 deletions
+17 -14
View File
@@ -93,7 +93,7 @@ def _warmup_ll_bf16_router_gemm(model: torch.nn.Module) -> None:
)
def kernel_warmup(worker: "Worker"):
def kernel_warmup(worker: "Worker", *, process_local_only: bool = False):
from vllm.model_executor.warmup.minimax_m3_msa_warmup import (
minimax_m3_msa_warmup,
)
@@ -118,6 +118,22 @@ def kernel_warmup(worker: "Worker"):
)
# Run next so input-prep kernels JIT against pristine runner state.
if worker.vllm_config.kernel_config.enable_jit_warmup:
fa4_cutedsl_warmup(worker)
sparse_mla_triton_warmup(worker)
if current_platform.has_device_capability(90):
_warmup_ll_bf16_router_gemm(worker.get_model())
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 process_local_only:
return
flashinfer_sparse_mla_decode_autotune_warmup(worker)
deepseek_v4_sparse_mla_attention_warmup(worker)
@@ -143,9 +159,6 @@ def kernel_warmup(worker: "Worker"):
elif has_flashinfer() and current_platform.has_device_capability(90):
flashinfer_autotune(worker.model_runner)
if current_platform.has_device_capability(90):
_warmup_ll_bf16_router_gemm(worker.get_model())
# FlashInfer attention warmup
# Only warmup if the model has FlashInfer attention groups
# and is not a pooling model
@@ -178,16 +191,6 @@ def kernel_warmup(worker: "Worker"):
create_mixed_batch=True,
)
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: