Skip to content

fix: MambaLMHeadModel and create_block robustness #1008

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

Open
Chessing234 wants to merge 3 commits into
base: main
Choose a base branch
from
+14 −5
Open

fix: MambaLMHeadModel and create_block robustness #1008

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
1bdc7e6
fix: accept legacy Mamba alias and clarify ssm_layer errors
Chessing234 Jul 31, 2026
5d1b107
fix: expose strict flag in MambaLMHeadModel.from_pretrained
Chessing234 Jul 31, 2026
6489c23
fix: write save_pretrained artifacts atomically
Chessing234 Jul 31, 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
19 changes: 14 additions & 5 deletions mamba_ssm/models/mixer_seq_simple.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -54,11 +54,15 @@ def create_block(
ssm_layer = ssm_cfg.pop("layer", "Mamba1")
ssm_layer_map = {
"Mamba1": Mamba,
"Mamba": Mamba,
"Mamba2": Mamba2,
"Mamba3": Mamba3,
}
if ssm_layer not in ssm_layer_map:
raise ValueError(f"Invalid ssm_layer: {ssm_layer}, only support Mamba1, Mamba2, and Mamba3")
supported = ", ".join(sorted(ssm_layer_map))
raise ValueError(
f"Invalid ssm_layer: {ssm_layer!r}. Supported values: {supported}."
)
mixer_cls = partial(
ssm_layer_map[ssm_layer],
layer_idx=layer_idx,
Expand Down Expand Up @@ -290,11 +294,12 @@ def forward(self, input_ids, position_ids=None, inference_params=None, num_last_
return CausalLMOutput(logits=lm_logits)

@classmethod
def from_pretrained(cls, pretrained_model_name, device=None, dtype=None, **kwargs):
def from_pretrained(cls, pretrained_model_name, device=None, dtype=None, strict=True, **kwargs):
config_data = load_config_hf(pretrained_model_name)
config = MambaConfig(**config_data)
model = cls(config, device=device, dtype=dtype, **kwargs)
model.load_state_dict(load_state_dict_hf(pretrained_model_name, device=device, dtype=dtype))
state_dict = load_state_dict_hf(pretrained_model_name, device=device, dtype=dtype)
model.load_state_dict(state_dict, strict=strict)
return model

def save_pretrained(self, save_directory):
Expand All @@ -307,9 +312,13 @@ def save_pretrained(self, save_directory):

# Save the model's state_dict
model_path = os.path.join(save_directory, 'pytorch_model.bin')
torch.save(self.state_dict(), model_path)
tmp_model_path = model_path + ".tmp"
torch.save(self.state_dict(), tmp_model_path)
os.replace(tmp_model_path, model_path)

# Save the configuration of the model
config_path = os.path.join(save_directory, 'config.json')
with open(config_path, 'w') as f:
tmp_config_path = config_path + ".tmp"
with open(tmp_config_path, 'w', encoding='utf-8') as f:
json.dump(self.config.__dict__, f, indent=4)
os.replace(tmp_config_path, config_path)