Skip to content

[Bugfix][Model][CPU] Fix Mamba2 quantized in_proj weight and scale loading for TP >1 #58083

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
bigPYJ1151 merged 2 commits into from
Sep 30, 2026
Merged

[Bugfix][Model][CPU] Fix Mamba2 quantized in_proj weight and scale loading for TP >1 #58083

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
39a0a66
[Bugfix][Model] Fix Mamba2 quantized in_proj weight and scale loading…
Sep 21, 2026
9b09711
Merge branch 'main' into fix/mamba-quant-tp-sharding
bigPYJ1151 Sep 30, 2026
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
22 changes: 12 additions & 10 deletions vllm/model_executor/layers/mamba/mamba_mixer2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -232,13 +232,14 @@ def loader(param: torch.Tensor, loaded_weight: torch.Tensor) -> None:
# - the ignore is for a mundane mypy error as it does not
# seem to handle slices well.
# https://github.com/python/mypy/issues/2410
target_slice = param.data[boundary : (boundary + take), ...]
param.data[
boundary : (boundary + take), ... # type: ignore[misc]
] = loaded_weight[
loaded_start_idx : (
loaded_start_idx + take
) # type: ignore[misc]
] # type: ignore[misc]
].view_as(target_slice)

# move indexing boundaries
boundary += shard_size
Expand Down Expand Up @@ -440,17 +441,18 @@ def __init__(
tp_rank,
)

# Apply the custom weight loader to in_proj.weight
# Apply the custom weight loader to in_proj.weight and its scales.
# Works for both non-quantized (Parameter) and quantized
# (ModelWeightParameter which extends BasevLLMParameter)
if isinstance(self.in_proj.weight, BasevLLMParameter):
# For BasevLLMParameter subclasses (quantized layers like FP8)
# These have a weight_loader property that can be directly set
self.in_proj.weight.weight_loader = mamba_loader
else:
# For standard Parameter (non-quantized layers)
delattr(self.in_proj.weight, "weight_loader")
set_weight_attrs(self.in_proj.weight, {"weight_loader": mamba_loader})
for attr_name in ("weight", "weight_scale", "input_scale"):
param = getattr(self.in_proj, attr_name, None)
if param is not None:
if isinstance(param, BasevLLMParameter):
param.weight_loader = mamba_loader
else:
if hasattr(param, "weight_loader"):
delattr(param, "weight_loader")
set_weight_attrs(param, {"weight_loader": mamba_loader})

# unsqueeze to fit conv1d weights shape into the linear weights shape.
# Can't do this in `weight_loader` since it already exists in
Expand Down
Loading