[CI Bug] Fully solve accuracy issue for DSv3.2 + MTP + Sequence Parallel (#48036)

Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
Wentao Ye
2026-07-14 10:00:24 -04:00
committed by GitHub
parent af453e5647
commit 1ff9429655
3 changed files with 34 additions and 3 deletions
-1
View File
@@ -652,7 +652,6 @@ class ParallelConfig:
)
and self.enable_expert_parallel
and self.tensor_parallel_size > 1
and self.data_parallel_size > 1
)
@property
@@ -10,6 +10,7 @@ from transformers import PretrainedConfig
from vllm._aiter_ops import rocm_aiter_ops
from vllm.compilation.decorators import support_torch_compile
from vllm.config import VllmConfig
from vllm.distributed import tensor_model_parallel_all_gather
from vllm.logger import init_logger
from vllm.model_executor.layers.fused_moe import (
fused_moe_make_expert_params_mapping,
@@ -40,6 +41,24 @@ from .utils import get_pp_missing_layer_names, maybe_prefix
logger = init_logger(__name__)
def _restore_full_token_layout_if_needed(
hidden_states: torch.Tensor,
residual: torch.Tensor,
num_tokens: int,
) -> tuple[torch.Tensor, torch.Tensor]:
"""Restore full token rows for the MTP proposer after SP MoE layers."""
if hidden_states.shape[0] == num_tokens:
return hidden_states, residual
combined_states = torch.cat([hidden_states, residual], dim=-1)
combined_states = tensor_model_parallel_all_gather(combined_states, 0)
combined_states = combined_states[:num_tokens]
hidden_states, residual = combined_states.split(
[hidden_states.shape[-1], residual.shape[-1]], dim=-1
)
return hidden_states, residual
class SharedHead(nn.Module):
def __init__(
self,
@@ -119,6 +138,11 @@ class DeepSeekMultiTokenPredictorLayer(nn.Module):
hidden_states=hidden_states,
residual=None,
)
hidden_states, residual = _restore_full_token_layout_if_needed(
hidden_states,
residual,
positions.shape[0],
)
hidden_states = residual + hidden_states # pre-final-norm (logits hidden)
# Recycle the post-final-norm hidden into the next draft step.
# compute_logits applies shared_head (== final norm) to the pre-norm
+10 -2
View File
@@ -21,7 +21,10 @@ from vllm.model_executor.model_loader.weight_utils import (
default_weight_loader,
maybe_remap_kv_scale_name,
)
from vllm.model_executor.models.deepseek_mtp import SharedHead
from vllm.model_executor.models.deepseek_mtp import (
SharedHead,
_restore_full_token_layout_if_needed,
)
from vllm.model_executor.models.deepseek_v2 import (
DeepseekV2MixtureOfExperts,
DeepseekV2MoE,
@@ -89,7 +92,12 @@ class DeepseekV32MultiTokenPredictorLayer(nn.Module):
hidden_states, residual = self.mtp_block(
positions=positions, hidden_states=hidden_states, residual=None
)
# mtp_block's MoE output is left un-reduced (reduce_results=False); the
hidden_states, residual = _restore_full_token_layout_if_needed(
hidden_states,
residual,
positions.shape[0],
)
# mtp_block's MoE output is left un-reduced (skip_final_all_reduce); the
# 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)