[Bugfix] Remove assertion for NVFP4 scale dynamic range (#37465)
Signed-off-by: Michael Goin <mgoin64@gmail.com>
This commit is contained in:
@@ -37,9 +37,6 @@ def _nvfp4_compute_scale_factor(marlin_scales: torch.Tensor) -> float:
|
|||||||
min_val = ws_float[nonzero_mask].min()
|
min_val = ws_float[nonzero_mask].min()
|
||||||
if min_val < 2:
|
if min_val < 2:
|
||||||
sf = (2 / min_val).log2().ceil().exp2()
|
sf = (2 / min_val).log2().ceil().exp2()
|
||||||
assert (ws_float[nonzero_mask] * sf <= 448 * (2**7)).all(), (
|
|
||||||
"NVFP4 scale dynamic range too large for rescaling"
|
|
||||||
)
|
|
||||||
return sf.item()
|
return sf.item()
|
||||||
return 1.0
|
return 1.0
|
||||||
|
|
||||||
|
|||||||
Reference in New Issue
Block a user