[NVFP4] NVFP4 MOE emulation fallback for H100/MI300/MI350, standardize TritonExperts usage for OCP MX emulation (#35737)

Signed-off-by: Felix Marty <Felix.Marty@amd.com>
Signed-off-by: fxmarty-amd <felmarty@amd.com>
Co-authored-by: Kyle Sayers <kylesayrs@gmail.com>
This commit is contained in:
fxmarty-amd
2026-04-22 08:58:54 -07:00
committed by GitHub
co-authored by Kyle Sayers
parent 5f76b3fb30
commit d622e27d2b
12 changed files with 601 additions and 121 deletions
+5 -1
View File
@@ -115,6 +115,7 @@ MoEBackend = Literal[
"flashinfer_cutedsl",
"marlin",
"aiter",
"emulation",
]
@@ -142,7 +143,10 @@ class KernelConfig:
- "flashinfer_cutlass": Use FlashInfer with CUTLASS kernels
- "flashinfer_cutedsl": Use FlashInfer with CuteDSL kernels (FP4 only)
- "marlin": Use Marlin kernels (weight-only quantization)
- "aiter": Use AMD AITer kernels (ROCm only)"""
- "aiter": Use AMD AITer kernels (ROCm only)
- "emulation": use BF16/FP16 GEMM, dequantizing weights and
running QDQ on activations.
"""
@field_validator("moe_backend", mode="before")
@classmethod