[Bugfix][Kernel] Restore FP fusion for indexer Q quantization (#60716)

Signed-off-by: Warren Deng <warrdeng@meta.com>
This commit is contained in:
Warren Deng
2026-10-09 09:46:41 +00:00
committed by GitHub
parent 2186b89a9a
commit 59ffd73e85
@@ -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,
)