Skip to content

[BugFix] Fix deepseek v3.2 mtp bug. #3900

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
merged 2 commits into from
Nov 4, 2025
Merged

[BugFix] Fix deepseek v3.2 mtp bug. #3900

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
d317577
fix deepseek 3.2 mtp bug
whx-sjtu Oct 30, 2025
926305f
add notes
whx-sjtu Oct 31, 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
11 changes: 11 additions & 0 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 @@ -12,6 +12,7 @@
from vllm.model_executor.model_loader.utils import \
process_weights_after_loading
from vllm.model_executor.models.deepseek_mtp import DeepSeekMTP
from vllm.model_executor.models.deepseek_v2 import DeepseekV32IndexerCache
from vllm.model_executor.models.llama_eagle3 import Eagle3LlamaForCausalLM
from vllm.v1.attention.backends.utils import (AttentionMetadataBuilder,
CommonAttentionMetadata)
Expand Down Expand Up @@ -129,6 +130,9 @@ def load_model(self, model) -> None:
target_attn_layer_names = set(
get_layers_from_vllm_config(self.vllm_config,
AttentionLayerBase).keys())
target_indexer_layer_names = set(
get_layers_from_vllm_config(self.vllm_config,
DeepseekV32IndexerCache).keys())
draft_model_config = \
self.vllm_config.speculative_config.draft_model_config
target_device = self.vllm_config.device_config.device
Expand All @@ -142,6 +146,13 @@ def load_model(self, model) -> None:
draft_attn_layer_names = (get_layers_from_vllm_config(
self.vllm_config, AttentionLayerBase).keys() -
target_attn_layer_names)
indexer_layers = get_layers_from_vllm_config(self.vllm_config,
DeepseekV32IndexerCache)
draft_indexer_layer_names = indexer_layers.keys(
) - target_indexer_layer_names
# NOTE: Currently we don't have specific attention backend and attention metadata
# for deepseek v3.2 indexer, so we just exclude the indexer layers here.
draft_attn_layer_names = draft_attn_layer_names - draft_indexer_layer_names

assert len(draft_attn_layer_names) == 1
self.attn_layer_name = list(draft_attn_layer_names)
Expand Down
Loading