[Perf] Restore zero-init of swizzled NVFP4 scale buffer to recover Blackwell decode throughput (#45739)
Signed-off-by: Albert Cheng <albertching0112@gmail.com> Co-authored-by: Vadim Gimpelson <156319763+vadiklyutiy@users.noreply.github.com>
This commit is contained in:
co-authored by
Vadim Gimpelson
parent
ac521f6237
commit
92c7fac640
+7
-1
@@ -57,7 +57,13 @@ def create_fp4_scale_tensor(
|
|||||||
rounded_m = round_up(m, 128)
|
rounded_m = round_up(m, 128)
|
||||||
scale_n = n // block_size
|
scale_n = n // block_size
|
||||||
rounded_n = round_up(scale_n, 4)
|
rounded_n = round_up(scale_n, 4)
|
||||||
return torch.empty(
|
# Must be zero-initialized: the swizzled scale buffer is padded to
|
||||||
|
# (round_up(m, 128), round_up(scale_n, 4) // 4) but the NVFP4 quant
|
||||||
|
# kernel does not write every padded element that the downstream
|
||||||
|
# NVFP4 GEMM reads. torch.empty leaves those padded scale factors
|
||||||
|
# uninitialized, which corrupts dequantization and causes a severe
|
||||||
|
# Blackwell NVFP4 decode throughput/output-length regression.
|
||||||
|
return torch.zeros(
|
||||||
(rounded_m, rounded_n // 4), device=device, dtype=torch.int32
|
(rounded_m, rounded_n // 4), device=device, dtype=torch.int32
|
||||||
)
|
)
|
||||||
else:
|
else:
|
||||||
|
|||||||
Reference in New Issue
Block a user