[Kernel] TD operand loads for batched MoE GEMM (moe_mmk) on XPU (#46340)

Signed-off-by: oonyshch <xonyshch@gmail.com>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
This commit is contained in:
Lena Onyshchenko
2026-07-26 08:50:07 +08:00
committed by GitHub
co-authored by Kunshang Ji
parent d30b1ecd1b
commit 0111002323
9 changed files with 331 additions and 33 deletions
+3
View File
@@ -122,6 +122,7 @@ class IrOpPriorityConfig:
MoEBackend = Literal[
"auto",
"triton",
"batched_triton",
"deep_gemm",
"deep_gemm_mega_moe",
"cutlass",
@@ -192,6 +193,8 @@ class KernelConfig:
- "auto": Automatically select the best backend based on model and hardware
- "triton": Use Triton-based fused MoE kernels
- "batched_triton": Use batched Triton experts (moe_mmk) on the batched
activation format ([E_local, max_num_tokens, K])
- "deep_gemm": Use DeepGEMM kernels (FP8 block-quantized only)
- "deep_gemm_mega_moe": Use DeepGEMM mega MoE kernels
- "cutlass": Use vLLM CUTLASS kernels