diff --git a/vllm/model_executor/models/deepseek_v4.py b/vllm/model_executor/models/deepseek_v4.py index e229549cb8..0332968663 100644 --- a/vllm/model_executor/models/deepseek_v4.py +++ b/vllm/model_executor/models/deepseek_v4.py @@ -1212,10 +1212,10 @@ class DeepseekV4DecoderLayer(nn.Module): x: torch.Tensor, positions: torch.Tensor, input_ids: torch.Tensor | None, - post_mix: torch.Tensor | None, - res_mix: torch.Tensor | None, - residual: torch.Tensor | None, - ) -> torch.Tensor: + post_mix: torch.Tensor | None = None, + res_mix: torch.Tensor | None = None, + residual: torch.Tensor | None = None, + ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: if residual is None: # Run standalone hc_pre on first layer residual = x diff --git a/vllm/model_executor/models/deepseek_v4_mtp.py b/vllm/model_executor/models/deepseek_v4_mtp.py index 8b7621aca7..9251605e6b 100644 --- a/vllm/model_executor/models/deepseek_v4_mtp.py +++ b/vllm/model_executor/models/deepseek_v4_mtp.py @@ -143,9 +143,12 @@ class DeepSeekV4MultiTokenPredictorLayer(nn.Module): hidden_states = self.h_proj(previous_hidden_states) + self.e_proj( inputs_embeds ).unsqueeze(-2) - hidden_states = self.mtp_block( + hidden_states, residual, post_mix, res_mix = self.mtp_block( positions=positions, x=hidden_states, input_ids=None ) + hidden_states = self.mtp_block.hc_post( + hidden_states, residual, post_mix, res_mix + ) # Return the flat pre-hc_head residual so it can be re-fed as the # next spec step's `previous_hidden_states` when # num_speculative_tokens > 1. hc_head is deferred to compute_logits.