[Perf][Feat] Add generic cuteDSL LL BF16 router (GEMM) (#42562)

Signed-off-by: LopezCastroRoberto <rocastro@redhat.com>
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Signed-off-by: Roberto L. Castro <38211239+LopezCastroRoberto@users.noreply.github.com>
Co-authored-by: Lucas Wilkinson <lwilkins@redhat.com>
This commit is contained in:
Roberto L. Castro
2026-07-13 23:01:21 -07:00
committed by GitHub
co-authored by Lucas Wilkinson
parent 31be872f55
commit 0762f2afeb
8 changed files with 1812 additions and 8 deletions
@@ -42,6 +42,31 @@ if TYPE_CHECKING:
logger = init_logger(__name__)
_LL_BF16_WARMUP_MODEL_SHAPES: tuple[tuple[int, int], ...] = (
(7168, 256), # DSV3
(7168, 384), # DSV4-Pro
(14400, 256), # DSV4-Flash
)
_LL_BF16_WARMUP_M_RANGE = range(1, 17)
def _warmup_ll_bf16_router_gemm() -> None:
from vllm.model_executor.kernels.linear.cute_dsl.ll_bf16 import (
is_available as is_ll_bf16_gemm_available,
)
from vllm.model_executor.kernels.linear.cute_dsl.ll_bf16 import (
ll_bf16_gemm_kernel,
)
if not is_ll_bf16_gemm_available():
return
logger.info("Warming up ll_bf16 router GEMM kernels.")
ll_bf16_gemm_kernel.warmup(
shapes=_LL_BF16_WARMUP_MODEL_SHAPES,
m_values=_LL_BF16_WARMUP_M_RANGE,
)
def kernel_warmup(worker: "Worker"):
from vllm.model_executor.warmup.minimax_m3_msa_warmup import (
@@ -94,6 +119,9 @@ 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()
# FlashInfer attention warmup
# Only warmup if the model has FlashInfer attention groups
# and is not a pooling model