[Feature] Non-contiguous Support for FP8 Quantization (#21961)

Signed-off-by: yewentao256 <zhyanwentao@126.com>
Co-authored-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Wentao Ye
2025-08-05 02:36:43 -07:00
committed by GitHub
co-authored by mgoin
parent 05fae02175
commit 4771df7b2b
4 changed files with 207 additions and 201 deletions
+4 -5
View File
@@ -1279,14 +1279,13 @@ def scaled_fp8_quant(
device=input.device,
dtype=torch.float32)
torch.ops._C.dynamic_per_token_scaled_fp8_quant(
output, input.contiguous(), scale, scale_ub)
output, input, scale, scale_ub)
else:
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
torch.ops._C.dynamic_scaled_fp8_quant(output, input.contiguous(),
scale)
scale = torch.empty(1, device=input.device, dtype=torch.float32)
torch.ops._C.dynamic_scaled_fp8_quant(output, input, scale)
else:
assert scale.numel() == 1, f"{scale.shape}"
torch.ops._C.static_scaled_fp8_quant(output, input.contiguous(), scale)
torch.ops._C.static_scaled_fp8_quant(output, input, scale)
return output, scale