Skip to content

[Feat] MTP support DeepSeekV3.2 #4465

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 1 commit into from
Dec 3, 2025
+31 −27
Merged

[Feat] MTP support DeepSeekV3.2 #4465

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
58 changes: 31 additions & 27 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 @@ -47,6 +47,8 @@

_MTP_MODELS = {
"DeepseekV3ForCausalLM":
("vllm.model_executor.models.deepseek_mtp", "DeepSeekMTP"),
"DeepseekV32ForCausalLM":
("vllm.model_executor.models.deepseek_mtp", "DeepSeekMTP")
}

Expand Down Expand Up @@ -817,26 +819,28 @@ def _propose(
attn_metadata_i.slot_mapping.fill_(-1)
attn_metadata_i.query_start_loc = self.arange[:batch_size + 1]
last_token_indices = self.arange[:batch_size]
if attn_metadata_i.num_decode_tokens != 0:
if getattr(attn_metadata_i, "num_decode_tokens", 0):
attn_metadata_i.num_decode_tokens = batch_size

input_ids = draft_token_ids_list[-1].int()
positions += 1

decode_metadata = getattr(attn_metadata_i, "decode", None)
prefill_metadata = getattr(attn_metadata_i, "prefill", None)
# When disable_padded_drafter_batch=False, it should not to be updating these params, maybe.
if self.speculative_config.disable_padded_drafter_batch or \
aclgraph_runtime_mode != CUDAGraphMode.FULL:
attn_metadata_i.decode.actual_seq_lengths_q = attn_metadata_i.query_start_loc[
if decode_metadata is not None and (self.speculative_config.disable_padded_drafter_batch or \
aclgraph_runtime_mode != CUDAGraphMode.FULL):
decode_metadata.actual_seq_lengths_q = attn_metadata_i.query_start_loc[
1:batch_size + 1].tolist()
if aclgraph_runtime_mode == CUDAGraphMode.FULL:
attn_metadata_i.decode.actual_seq_lengths_q = \
decode_metadata.actual_seq_lengths_q = \
builder.pad_actual_seq_len_q_mtp_disable_pad(
graph_pad_size - batch_size,
batch_size,
attn_metadata_i.decode.actual_seq_lengths_q)
attn_metadata_i.decode.cos = builder.cos_cache[
decode_metadata.actual_seq_lengths_q)
decode_metadata.cos = builder.cos_cache[
positions[:batch_size]].unsqueeze(1).unsqueeze(2)
attn_metadata_i.decode.sin = builder.sin_cache[
decode_metadata.sin = builder.sin_cache[
positions[:batch_size]].unsqueeze(1).unsqueeze(2)
# NOTE(woosuk): We should handle the case where the draft model
# generates tokens beyond the max model length. Since it is complex
Expand Down Expand Up @@ -874,32 +878,32 @@ def _propose(
self.input_ids[batch_size:num_input_tokens] = 0
self.hidden_states[batch_size:num_input_tokens].fill_(0)

if attn_metadata_i.prefill is not None:
attn_metadata_i.prefill.seq_lens = attn_metadata_i.seq_lens
attn_metadata_i.prefill.seq_lens_list = attn_metadata_i.prefill.seq_lens.tolist(
if prefill_metadata is not None:
prefill_metadata.seq_lens = attn_metadata_i.seq_lens
prefill_metadata.seq_lens_list = prefill_metadata.seq_lens.tolist(
)
attn_metadata_i.prefill.context_lens = attn_metadata_i.seq_lens
attn_metadata_i.prefill.input_positions = self.positions[:
num_input_tokens]
attn_metadata_i.prefill.max_seq_lens += 1
attn_metadata_i.prefill.max_seq_lens = min(
attn_metadata_i.prefill.max_seq_lens,
prefill_metadata.context_lens = attn_metadata_i.seq_lens
prefill_metadata.input_positions = self.positions[:
num_input_tokens]
prefill_metadata.max_seq_lens += 1
prefill_metadata.max_seq_lens = min(
prefill_metadata.max_seq_lens,
self.runner.model_config.max_model_len)
if attn_metadata_i.decode is not None:
attn_metadata_i.decode.seq_lens = attn_metadata_i.seq_lens
attn_metadata_i.decode.seq_lens_list = attn_metadata_i.decode.seq_lens.tolist(
if decode_metadata is not None:
decode_metadata.seq_lens = attn_metadata_i.seq_lens
decode_metadata.seq_lens_list = decode_metadata.seq_lens.tolist(
)
decode_seq_lens_list = attn_metadata_i.decode.seq_lens_list
decode_seq_lens_list = decode_metadata.seq_lens_list
if aclgraph_runtime_mode == CUDAGraphMode.FULL and \
self.speculative_config.disable_padded_drafter_batch:
attn_metadata_i.decode.seq_lens_list = decode_seq_lens_list + [
decode_metadata.seq_lens_list = decode_seq_lens_list + [
0
] * (graph_pad_size - len(decode_seq_lens_list))
attn_metadata_i.decode.input_positions = self.positions[:
num_input_tokens]
attn_metadata_i.decode.max_seq_lens += 1
attn_metadata_i.decode.max_seq_lens = min(
attn_metadata_i.decode.max_seq_lens,
decode_metadata.input_positions = self.positions[:
num_input_tokens]
decode_metadata.max_seq_lens += 1
decode_metadata.max_seq_lens = min(
decode_metadata.max_seq_lens,
self.runner.model_config.max_model_len)

# mtp>1: [batch_size, k]
Expand Down
Loading