Skip to content

[Bugfix] Remove NVFP4 scales assertions to fix load_format=dummy #18861

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
DarkLight1337 merged 4 commits into from
May 30, 2025
Merged

[Bugfix] Remove NVFP4 scales assertions to fix load_format=dummy #18861

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
dcef1ed
[Bugfix] Remove fp4 marlin scales assertion to fix load_format=dummy
mgoin May 28, 2025
d0f12bc
Change assert to warning
mgoin May 28, 2025
0e61be6
Change assert to warning
mgoin May 29, 2025
8d8e3e9
Merge branch 'main' into fix-fp4-marlin-dummy
mgoin May 29, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 5 additions & 3 deletions vllm/model_executor/layers/quantization/modelopt.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -585,9 +585,11 @@ def swizzle_blockscale(self, scale: torch.tensor):
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:

# GEMM 1
assert torch.allclose(
layer.w13_weight_scale_2[:, 0], layer.w13_weight_scale_2[:, 1]), (
"w1_weight_scale_2 must match w3_weight_scale_2")
if not torch.allclose(layer.w13_weight_scale_2[:, 0],
layer.w13_weight_scale_2[:, 1]):
logger.warning_once(
"w1_weight_scale_2 must match w3_weight_scale_2. "
"Accuracy may be affected.")

w13_weight_scale_2 = layer.w13_weight_scale_2[:, 0]
layer.w13_weight_scale_2 = Parameter(w13_weight_scale_2,
Expand Down
7 changes: 6 additions & 1 deletion vllm/model_executor/layers/quantization/utils/marlin_utils_fp4.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,7 +22,12 @@ def is_fp4_marlin_supported():


def fp4_marlin_process_scales(marlin_scales):
assert (marlin_scales >= 0).all()
Comment thread
mgoin marked this conversation as resolved.
if not (marlin_scales >= 0).all():
logger.warning_once(
"NVFP4 Marlin assumes the scales to be >=0, but has encountered "
"negative scales. Accuracy will likely be degraded. This is "
"because it changes the scales from FP8-S1E4M3 to a special "
"FP8-S0E5M3 format to speedup the dequantization.")

# convert to half first, we would convert to fp8 later
marlin_scales = marlin_scales.to(torch.half)
Expand Down