[Bugfix][Kernel] Restore FP fusion for indexer Q quantization (#60716)
Signed-off-by: Warren Deng <warrdeng@meta.com>
This commit is contained in:
@@ -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,
|
||||
)
|
||||
|
||||
|
||||
Reference in New Issue
Block a user