[Kimi-K3] Overlap low-M TP8 KDA projections (#54697)

Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
Yongye Zhu
2026-09-02 03:35:28 +00:00
committed by GitHub
co-authored by OpenAI Codex
parent 12b9573c98
commit 3ba9907a1d
7 changed files with 1092 additions and 12 deletions
@@ -114,6 +114,12 @@ def _warmup_kimi_k3_gemm_rs_ar() -> None:
logger.info_once("Warmed up %d Kimi-K3 GEMM-RS/AR variants.", compiled)
def _autotune_kimi_k3_kda_qkvg(model: torch.nn.Module) -> None:
module = sys.modules.get("vllm.models.kimi_k3.nvidia.low_latency_gemm")
if module is not None:
module.autotune_kda_qkvg(model)
def kernel_warmup(worker: "Worker", *, process_local_only: bool = False):
from vllm.model_executor.warmup.minimax_m3_msa_warmup import (
minimax_m3_msa_warmup,
@@ -343,6 +349,7 @@ def flashinfer_autotune(runner: "GPUModelRunner") -> None:
):
_run_flashinfer_autotune_dummy_runs(runner)
replayssm_autotune_warmup(runner)
_autotune_kimi_k3_kda_qkvg(runner.get_model())
finally:
set_autotune_process_group(None)