[Perf] Improve BF16x3 router GEMM accuracy and make it default on sm100 (#55899)

Signed-off-by: Thien Tran <gau.nernst@yahoo.com.sg>
Co-authored-by: OpenAI Codex <noreply@openai.com>
Co-authored-by: Jee Jee Li <pandaleefree@gmail.com>
This commit is contained in:
Thien Tran
2026-09-09 07:33:12 -07:00
committed by GitHub
co-authored by OpenAI Codex Jee Jee Li
parent 3116c5d06b
commit cc4210f671
6 changed files with 314 additions and 88 deletions
-3
View File
@@ -230,9 +230,6 @@ class KernelConfig:
enable_jit_warmup: bool = True
"""If True, run JIT compile warmup during kernel warmup."""
enable_bf16x3_router_gemm: bool = False
"""If True, use the experimental SM100 BF16x3 CuteDSL router GEMM."""
moe_backend: MoEBackend = "auto"
"""Backend for MoE expert computation kernels. Available options: