frozenleaves/vllm
1
0
Fork 0
mirror of https://github.com/vllm-project/vllm.git synced 2025-10-20 14:53:52 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Fix DeepSeek MTP (#22934)

Browse Source 
Signed-off-by: Benjamin Chislett <benjamin.chislett@centml.ai>
This commit is contained in:
Benjamin Chislett
2025-08-15 21:25:06 -04:00
committed by GitHub
parent 070da660c1
commit fbd88728b3
3 changed files with 13 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
vllm/model_executor/models/deepseek_mtp.py
View File

@ -158,14 +158,13 @@ class DeepSeekMTP(nn.Module, SupportsPP):
self,
input_ids: torch.Tensor,
positions: torch.Tensor,
previous_hidden_states: torch.Tensor,
hidden_states: torch.Tensor,
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,
previous_hidden_states, inputs_embeds,
spec_step_idx)
hidden_states = self.model(input_ids, positions, hidden_states,
inputs_embeds, spec_step_idx)
return hidden_states
def compute_logits(
@ -213,13 +212,15 @@ class DeepSeekMTP(nn.Module, SupportsPP):
# for mlp.experts[0].gate_gate_up_proj, which breaks load.
if (("mlp.experts." in name) and name not in params_dict):
continue
name = name.replace(weight_name, param_name)
name_mapped = name.replace(weight_name, param_name)
# QKV fusion is optional, fall back to normal
# weight loading if it's not enabled
if ((param_name == "fused_qkv_a_proj")
and name not in params_dict):
and name_mapped not in params_dict):
continue
else:
name = name_mapped
# Skip loading extra bias for GPTQ models.
if name.endswith(".bias") and name not in params_dict:

7
vllm/model_executor/models/glm4_moe_mtp.py
View File

@ -180,14 +180,13 @@ class Glm4MoeMTP(nn.Module, SupportsPP):
self,
input_ids: torch.Tensor,
positions: torch.Tensor,
previous_hidden_states: torch.Tensor,
hidden_states: torch.Tensor,
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,
previous_hidden_states, inputs_embeds,
spec_step_idx)
hidden_states = self.model(input_ids, positions, hidden_states,
inputs_embeds, spec_step_idx)
return hidden_states
def compute_logits(

7
vllm/model_executor/models/mimo_mtp.py
View File

@ -164,15 +164,14 @@ class MiMoMTP(nn.Module):
self,
input_ids: torch.Tensor,
positions: torch.Tensor,
previous_hidden_states: torch.Tensor,
hidden_states: torch.Tensor,
intermediate_tensors: Optional[IntermediateTensors] = None,
inputs_embeds: Optional[torch.Tensor] = None,
spec_step_idx: int = 0,
) -> torch.Tensor:
assert spec_step_idx == 0, "mimo_mtp only support predict one token now"
hidden_states = self.model(input_ids, positions,
previous_hidden_states, inputs_embeds,
spec_step_idx)
hidden_states = self.model(input_ids, positions, hidden_states,
inputs_embeds, spec_step_idx)
return hidden_states
def compute_logits(