From 6ff16b77f8c0d15735829e5e85b1c50972536c59 Mon Sep 17 00:00:00 2001 From: Michael Goin Date: Thu, 29 Jan 2026 15:15:17 -0500 Subject: [PATCH] [Bugfix] Enable Triton MoE for FP8 per-tensor dynamic (#33300) Signed-off-by: mgoin (cherry picked from commit bfb9bdaf3f44fdcd4c6ff66b5e8c84834b092534) --- vllm/model_executor/layers/fused_moe/fused_batched_moe.py | 1 + vllm/model_executor/layers/fused_moe/fused_moe.py | 2 ++ 2 files changed, 3 insertions(+) diff --git a/vllm/model_executor/layers/fused_moe/fused_batched_moe.py b/vllm/model_executor/layers/fused_moe/fused_batched_moe.py index 509bacfbc..fd6c365fe 100644 --- a/vllm/model_executor/layers/fused_moe/fused_batched_moe.py +++ b/vllm/model_executor/layers/fused_moe/fused_batched_moe.py @@ -927,6 +927,7 @@ class BatchedTritonExperts(mk.FusedMoEPermuteExpertsUnpermute): SUPPORTED_W_A_FP8 = [ (kFp8Static128BlockSym, kFp8Dynamic128Sym), (kFp8StaticChannelSym, kFp8DynamicTokenSym), + (kFp8StaticTensorSym, kFp8DynamicTokenSym), (kFp8StaticTensorSym, kFp8StaticTensorSym), (kFp8StaticTensorSym, kFp8DynamicTensorSym), ] diff --git a/vllm/model_executor/layers/fused_moe/fused_moe.py b/vllm/model_executor/layers/fused_moe/fused_moe.py index 669a6e74b..0335339b7 100644 --- a/vllm/model_executor/layers/fused_moe/fused_moe.py +++ b/vllm/model_executor/layers/fused_moe/fused_moe.py @@ -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 ( QuantKey, kFp8Dynamic128Sym, + kFp8DynamicTensorSym, kFp8DynamicTokenSym, kFp8Static128BlockSym, kFp8StaticChannelSym, @@ -1942,6 +1943,7 @@ class TritonExperts(mk.FusedMoEPermuteExpertsUnpermute): (kFp8StaticChannelSym, kFp8DynamicTokenSym), (kFp8StaticTensorSym, kFp8DynamicTokenSym), (kFp8StaticTensorSym, kFp8StaticTensorSym), + (kFp8StaticTensorSym, kFp8DynamicTensorSym), ] return (weight_key, activation_key) in SUPPORTED_W_A