[Kernel][DSV4.1] Fuse MXFP8 wo_b GEMM with sequence-parallel reduce-scatter (#57428)

Signed-off-by: Thien Tran <gau.nernst@yahoo.com.sg>
Signed-off-by: Canlin <canlinguosdu@gmail.com>
Co-authored-by: Thien Tran <gau.nernst@yahoo.com.sg>
Co-authored-by: OpenAI Codex <codex@openai.com>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Co-authored-by: Jee Jee Li <pandaleefree@gmail.com>
This commit is contained in:
Canlin Guo
2026-09-22 01:07:17 -07:00
committed by GitHub
co-authored by Thien Tran OpenAI Codex Claude Fable 5.1 Jee Jee Li
parent a6c47fbbf4
commit f92b78f6ef
18 changed files with 1037 additions and 624 deletions
+8 -7
View File
@@ -136,16 +136,17 @@ def _warmup_bf16x3_router_gemm(
logger.info_once("Warmed up BF16x3 router GEMM configs: %s.", configs)
def _warmup_kimi_k3_gemm_rs_ar() -> None:
# Kimi-K3 model construction imports this module only when GEMM-RS/AR is
# enabled and initializes its singleton before kernel_warmup runs. Avoid
# importing it here so other models do not compile the RS/AR variants.
module = sys.modules.get("vllm.models.kimi_k3.nvidia.ops.cute_dsl.gemm_rs_ar")
def _warmup_gemm_rs_ar() -> None:
# Model construction (Kimi-K3, DeepSeek-V4.1) imports this module only
# when GEMM-RS/AR is enabled and initializes its singleton before
# kernel_warmup runs. Avoid importing it here so other models do not
# compile the RS/AR variants.
module = sys.modules.get("vllm.model_executor.kernels.linear.cute_dsl.gemm_rs_ar")
if module is None:
return
compiled = module.warmup_gemm_rs_ar()
if compiled:
logger.info_once("Warmed up %d Kimi-K3 GEMM-RS/AR variants.", compiled)
logger.info_once("Warmed up %d GEMM-RS/AR variants.", compiled)
def _autotune_kimi_k3_kda_qkvg(model: torch.nn.Module) -> None:
@@ -208,7 +209,7 @@ def kernel_warmup(worker: "Worker", *, process_local_only: bool = False):
if current_platform.has_device_capability(90):
_warmup_ll_bf16_router_gemm(worker.get_model())
_warmup_kimi_k3_gemm_rs_ar()
_warmup_gemm_rs_ar()
if worker.vllm_config.kernel_config.enable_cutedsl_warmup:
# TODO(roberto): Remove after registered CuTeDSL warmups are migrated