Fix nvfp4 online scale with pcg (#32246)
This commit is contained in:
@@ -965,9 +965,15 @@ def fused_experts_none_to_flashinfer_trtllm_fp4(
|
|||||||
):
|
):
|
||||||
e4m3_max = 256.0
|
e4m3_max = 256.0
|
||||||
|
|
||||||
|
global_scale_inv = torch.full(
|
||||||
|
(1,),
|
||||||
|
1.0 / (e4m3_max * 6.0),
|
||||||
|
dtype=torch.float32,
|
||||||
|
device=hidden_states.device,
|
||||||
|
)
|
||||||
hs_fp4_bytes, hs_sf_bytes, per_token_scale = nvfp4_quantize(
|
hs_fp4_bytes, hs_sf_bytes, per_token_scale = nvfp4_quantize(
|
||||||
hidden_states,
|
hidden_states,
|
||||||
1.0 / (e4m3_max * 6.0),
|
global_scale_inv,
|
||||||
sfLayout=SfLayout.layout_linear,
|
sfLayout=SfLayout.layout_linear,
|
||||||
per_token_activation=True,
|
per_token_activation=True,
|
||||||
backend="cute-dsl",
|
backend="cute-dsl",
|
||||||
|
|||||||
Reference in New Issue
Block a user