[Kernel] Enable fp8 support for pplx and BatchedTritonExperts. (#18864)

Signed-off-by: Bill Nell <bnell@redhat.com>
This commit is contained in:
bnellnm
2025-07-03 14:55:40 -07:00
committed by GitHub
parent 2f2fcb31b8
commit 78fe77534b
25 changed files with 1277 additions and 663 deletions
+34 -3
View File
@@ -99,9 +99,20 @@ def _fp8_perm(m: torch.Tensor, idx: torch.Tensor) -> torch.Tensor:
return m[idx, ...]
# TODO(bnell): better name
def maybe_fix_scales(scales: Optional[torch.Tensor],
num_experts: int) -> Optional[torch.Tensor]:
def normalize_scales_shape(
scales: Optional[torch.Tensor]) -> Optional[torch.Tensor]:
if scales is not None:
if scales.numel() == 1:
scales = scales.view(1, 1)
else:
scales = scales.view(-1, scales.size(-1))
return scales
def normalize_batched_scales_shape(
scales: Optional[torch.Tensor],
num_experts: int,
) -> Optional[torch.Tensor]:
if scales is not None and scales.ndim < 3:
if scales.numel() == 1:
scales = scales.view(1)
@@ -111,3 +122,23 @@ def maybe_fix_scales(scales: Optional[torch.Tensor],
scales = scales.view(num_experts, -1, scales.size(-1))
return scales
def _validate_scale_shape(
a: torch.Tensor,
a_scale: Optional[torch.Tensor],
per_act_token_quant: bool,
block_shape: Optional[list[int]],
) -> None:
if a_scale is None:
return
if not per_act_token_quant and block_shape is None:
assert a_scale.numel() == 1, f"{a_scale.shape}"
elif per_act_token_quant:
assert a_scale.shape[0] == a.shape[0] and a_scale.shape[1] == 1, (
f"{a_scale.shape[0]} == {a.shape[0]} and {a_scale.shape[1]} == 1")
else:
assert block_shape is not None
expected = (a.shape[0], cdiv(a.shape[1], block_shape[1]))
assert a_scale.shape == expected, f"{a_scale.shape} == {expected}"