diff --git a/vllm/models/deepseek_v32/nvidia/mtp.py b/vllm/models/deepseek_v32/nvidia/mtp.py index 118f27459b..baeb8b25e0 100644 --- a/vllm/models/deepseek_v32/nvidia/mtp.py +++ b/vllm/models/deepseek_v32/nvidia/mtp.py @@ -93,10 +93,18 @@ class DeepseekV32MultiTokenPredictorLayer(nn.Module): # main model fuses that all-reduce into the next norm, but here the # recycle hidden is consumed directly, so reduce it now. hidden_states = tensor_model_parallel_all_reduce(hidden_states) - # Return the pre-final-norm recycle hidden (re-fed as the next spec - # step's previous_hidden_states); shared_head norm is applied in - # compute_logits. Matches the V2-runner / deepseek_v4 MTP contract. - return residual + hidden_states + # Recycle the POST-final-norm hidden into the next draft step. The + # residual-add is fused into the final RMSNorm so it is computed + # exactly once, and the result is returned for both tuple positions: + # the draft-logits hidden (compute_logits applies the LM head only) and + # the recycled previous_hidden_states. Recycling the pre-final-norm + # hidden mismatches the draft model's hnorm and lowers MTP acceptance; + # post-norm recycle matches deepseek_mtp.py (PR #45895). The tuple form + # is understood by both the V2 speculator (isinstance-tuple check) and + # the legacy proposer (model_returns_tuple is True for the + # DeepSeekMTPModel architecture). + hidden_states, _ = self.shared_head.norm(hidden_states, residual) + return hidden_states, hidden_states class DeepseekV32MultiTokenPredictor(nn.Module): @@ -168,9 +176,10 @@ class DeepseekV32MultiTokenPredictor(nn.Module): ) -> torch.Tensor: current_step_idx = spec_step_idx % self.num_mtp_layers mtp_layer = self.layers[str(self.mtp_start_layer_idx + current_step_idx)] - return self.logits_processor( - mtp_layer.shared_head.head, mtp_layer.shared_head(hidden_states) - ) + # hidden_states is already post-final-norm (produced in the layer + # forward and recycled as-is); apply the LM head only, without a + # second RMSNorm. + return self.logits_processor(mtp_layer.shared_head.head, hidden_states) class DeepseekV32MTP(nn.Module, DeepseekV2MixtureOfExperts):