[Kernel] Add b12x FP4 MoE backend (#52018)

Signed-off-by: Luke Alonso <lalonso@gmail.com>
Signed-off-by: mgoin <mgoin64@gmail.com>
Signed-off-by: Misha Goin <mgoin64@gmail.com>
Co-authored-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Luke Alonso
2026-08-21 08:04:36 -07:00
committed by GitHub
co-authored by mgoin
parent 5ee84d3c52
commit 2740c817ff
23 changed files with 2417 additions and 60 deletions
+2
View File
@@ -130,6 +130,7 @@ MoEBackend = Literal[
"flashinfer_cutlass",
"flashinfer_cutedsl",
"flashinfer_b12x",
"b12x",
"marlin",
"humming",
"triton_unfused",
@@ -204,6 +205,7 @@ class KernelConfig:
- "flashinfer_cutedsl": Use FlashInfer with CuteDSL kernels (FP4 only)
- "flashinfer_b12x": Use FlashInfer CuteDSL fused MoE for SM12x
(RTX Pro 6000 / DGX Spark)
- "b12x": Use b12x FP4 MoE kernels on SM12x
- "marlin": Use Marlin kernels (weight-only quantization)
- "humming": Use Humming Mixed Precision kernels
- "triton_unfused": Use Triton unfused MoE kernels