[Perf][2/N] Expand Triton kernel warmup coverage, Qwen (#46750)
Signed-off-by: LopezCastroRoberto <rocastro@redhat.com>
This commit is contained in:
@@ -24,6 +24,7 @@ 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
|
||||
@@ -40,6 +41,8 @@ def kernel_warmup(worker: "Worker"):
|
||||
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).
|
||||
|
||||
Reference in New Issue
Block a user