From 59ffd73e8597d860730f4a8c4f778f27c75a0dfa Mon Sep 17 00:00:00 2001 From: Warren Deng <35615489+warrendeng@users.noreply.github.com> Date: Fri, 9 Oct 2026 02:46:41 -0700 Subject: [PATCH] [Bugfix][Kernel] Restore FP fusion for indexer Q quantization (#60716) Signed-off-by: Warren Deng --- vllm/models/deepseek_v4/common/ops/fused_indexer_q.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/vllm/models/deepseek_v4/common/ops/fused_indexer_q.py b/vllm/models/deepseek_v4/common/ops/fused_indexer_q.py index 669186edbb..54b1b9be7e 100644 --- a/vllm/models/deepseek_v4/common/ops/fused_indexer_q.py +++ b/vllm/models/deepseek_v4/common/ops/fused_indexer_q.py @@ -305,6 +305,7 @@ class FusedIndexerQRopeQuantTritonKernel( FP8_MAX=fp8_max, USE_FNUZ=use_fnuz, USE_EXPLICIT_FMA=current_platform.is_rocm(), + enable_fp_fusion=True, num_warps=1, ) @@ -540,6 +541,7 @@ class FusedIndexerQRopeMxFp4TritonKernel( MXFP4_BLOCK=MXFP4_BLOCK_SIZE, index_weights_stride=index_weights.stride(0), index_weights_out_stride=index_weights_out.stride(0), + enable_fp_fusion=True, num_warps=1, )