Skip to content

[BugFix] Fix torchair+mtp bug after deleting deepseek_mtp. #3590

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
wangxiyuan merged 2 commits into from
Oct 21, 2025
Merged

[BugFix] Fix torchair+mtp bug after deleting deepseek_mtp. #3590

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
5bf8bf9
fix torchair mtp bug
whx-sjtu Oct 21, 2025
77d1e3d
fix ut
whx-sjtu Oct 21, 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
3 changes: 1 addition & 2 deletions tests/ut/torchair/models/test_torchair_deepseek_mtp.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -145,8 +145,7 @@ def test_compute_logits(self, mocker: MockerFixture, setup_predictor):
return_value=None)
predictor.logits_processor.return_value = torch.tensor([1.0, 2.0, 3.0])

result_logits = predictor.compute_logits(hidden_states=hidden_states,
sampling_metadata=None)
result_logits = predictor.compute_logits(hidden_states=hidden_states)
predictor.logits_processor.assert_called_once()
assert torch.allclose(result_logits, torch.tensor([1.0, 2.0, 3.0]))

Expand Down
4 changes: 2 additions & 2 deletions vllm_ascend/spec_decode/mtp_proposer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -175,7 +175,7 @@ def dummy_run(self,
torchair_compiled_model(
input_ids=input_ids,
positions=positions,
previous_hidden_states=previous_hidden_states,
hidden_states=previous_hidden_states,
inputs_embeds=None,
intermediate_tensors=None,
attn_metadata=attn_metadata,
Expand Down Expand Up @@ -460,7 +460,7 @@ def _propose(
hidden_states = torchair_compiled_model(
input_ids=self.input_ids[:num_input_tokens],
positions=self.positions[:num_input_tokens],
previous_hidden_states=self.
hidden_states=self.
hidden_states[:num_input_tokens],
inputs_embeds=None,
intermediate_tensors=None,
Expand Down
10 changes: 4 additions & 6 deletions vllm_ascend/torchair/models/torchair_deepseek_mtp.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -176,14 +176,12 @@ def forward(
def compute_logits(
self,
hidden_states: torch.Tensor,
sampling_metadata=None, # type: ignore
spec_step_idx: int = 0,
) -> torch.Tensor:
current_step_idx = (spec_step_idx % self.num_mtp_layers)
mtp_layer = self.layers_list[current_step_idx]
logits = self.logits_processor(mtp_layer.shared_head.head,
mtp_layer.shared_head(hidden_states),
sampling_metadata)
mtp_layer.shared_head(hidden_states))
return logits


Expand All @@ -209,12 +207,12 @@ def forward(
positions: torch.Tensor,
kv_caches: Optional[List[torch.Tensor]] = None,
attn_metadata: Optional[AttentionMetadata] = None,
previous_hidden_states: Optional[torch.Tensor] = None,
hidden_states: Optional[torch.Tensor] = None,
intermediate_tensors: Optional[IntermediateTensors] = None,
inputs_embeds: Optional[torch.Tensor] = None,
spec_step_idx: int = 0,
) -> torch.Tensor:
hidden_states = self.model(input_ids, positions, kv_caches,
attn_metadata, previous_hidden_states,
inputs_embeds, spec_step_idx)
attn_metadata, hidden_states, inputs_embeds,
spec_step_idx)
return hidden_states
Loading