A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 6, 10:45 UTC
[Bugfix] Remove assertion for NVFP4 scale dynamic range (#37465)
Signed-off-by: Michael Goin <mgoin64@gmail.com>
Michael Goin committed
Mar 18, 2026 at 23:37 UTC
9482b0b085e044fe9db8926d0ba262fd70b56ca1
1 file changed
-3
vllm/model_executor/layers/quantization/utils/marlin_utils_fp4.py
-3
@@ -37,9 +37,6 @@ def _nvfp4_compute_scale_factor(marlin_scales: torch.Tensor) -> float:
37
min_val = ws_float[nonzero_mask].min()
38
if min_val < 2:
39
sf = (2 / min_val).log2().ceil().exp2()
40
- assert (ws_float[nonzero_mask] * sf <= 448 * (2**7)).all(), (
41
- "NVFP4 scale dynamic range too large for rescaling"
42
- )
40
return sf.item()
41
return 1.0
42