biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Bugfix] Enable Triton MoE for FP8 per-tensor dynamic (#33300)

Browse Source 
Signed-off-by: mgoin <mgoin64@gmail.com>
(cherry picked from commit bfb9bdaf3f)
This commit is contained in:
Michael Goin
2026-01-29 15:15:17 -05:00
committed by khluu
parent 1ed963d43a
commit 6ff16b77f8
2 changed files with 3 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
vllm/model_executor/layers/fused_moe/fused_batched_moe.py
View File

@@ -927,6 +927,7 @@ class BatchedTritonExperts(mk.FusedMoEPermuteExpertsUnpermute):
SUPPORTED_W_A_FP8 = [ SUPPORTED_W_A_FP8 = [
(kFp8Static128BlockSym, kFp8Dynamic128Sym), (kFp8Static128BlockSym, kFp8Dynamic128Sym),
(kFp8StaticChannelSym, kFp8DynamicTokenSym), (kFp8StaticChannelSym, kFp8DynamicTokenSym),
(kFp8StaticTensorSym, kFp8DynamicTokenSym),
(kFp8StaticTensorSym, kFp8StaticTensorSym), (kFp8StaticTensorSym, kFp8StaticTensorSym),
(kFp8StaticTensorSym, kFp8DynamicTensorSym), (kFp8StaticTensorSym, kFp8DynamicTensorSym),
] ]

2
vllm/model_executor/layers/fused_moe/fused_moe.py
View File

@@ -45,6 +45,7 @@ from vllm.model_executor.layers.quantization.utils.ocp_mx_utils import OCP_MX_Sc
from vllm.model_executor.layers.quantization.utils.quant_utils import ( from vllm.model_executor.layers.quantization.utils.quant_utils import (
QuantKey, QuantKey,
kFp8Dynamic128Sym, kFp8Dynamic128Sym,
kFp8DynamicTensorSym,
kFp8DynamicTokenSym, kFp8DynamicTokenSym,
kFp8Static128BlockSym, kFp8Static128BlockSym,
kFp8StaticChannelSym, kFp8StaticChannelSym,
@@ -1942,6 +1943,7 @@ class TritonExperts(mk.FusedMoEPermuteExpertsUnpermute):
(kFp8StaticChannelSym, kFp8DynamicTokenSym), (kFp8StaticChannelSym, kFp8DynamicTokenSym),
(kFp8StaticTensorSym, kFp8DynamicTokenSym), (kFp8StaticTensorSym, kFp8DynamicTokenSym),
(kFp8StaticTensorSym, kFp8StaticTensorSym), (kFp8StaticTensorSym, kFp8StaticTensorSym),
(kFp8StaticTensorSym, kFp8DynamicTensorSym),
] ]
return (weight_key, activation_key) in SUPPORTED_W_A return (weight_key, activation_key) in SUPPORTED_W_A