Skip to content

[eagle3][pcp] fix acceptance rate for eagle3 and pcp enabled #7549

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
yiz-liu merged 4 commits into from
Mar 25, 2026
Merged

[eagle3][pcp] fix acceptance rate for eagle3 and pcp enabled #7549

Changes from 2 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
ed8d127
[eagle3][pcp] fix acceptance rate for eagle3 and pcp enabled
lilinsiman Mar 23, 2026
cf3bb78
add note
lilinsiman Mar 23, 2026
d1afb97
add fix me
lilinsiman Mar 24, 2026
fb9b52a
Merge branch 'main' into eagle_pcp_acc_bugfix3
lilinsiman Mar 24, 2026
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
13 changes: 12 additions & 1 deletion vllm_ascend/spec_decode/eagle_proposer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -577,6 +577,11 @@ def _propose(

# Copy the old attn_metadata and update
attn_metadata_i = per_layer_attn_metadata[self.attn_layer_names[0]]

# Clone the data so that when calculating the data at position 2 and position 3
# in the merged graph, it does not affect position 1
common_attn_metadata.block_table_tensor = common_attn_metadata.block_table_tensor.clone()

if self.pcp_size * self.dcp_size > 1:
if self.num_speculative_tokens > 1 and not attn_metadata_i.num_prefills:
# For pcp/dcp, tokens are split across different cp ranks,
Expand Down Expand Up @@ -1144,9 +1149,15 @@ def attn_update_stack_num_spec_norm(
common_attn_metadata.num_input_tokens = input_batch_size

# The loop part

used_update_positions += 1

# Clone the data so that when calculating the data at position 2 and position 3
# in the merged graph, it does not affect position 1
common_attn_metadata.seq_lens = common_attn_metadata.seq_lens.clone()
common_attn_metadata.seq_lens_cpu = common_attn_metadata.seq_lens_cpu.clone()
common_attn_metadata.num_computed_tokens_cpu = common_attn_metadata.num_computed_tokens_cpu.clone()
common_attn_metadata.positions = common_attn_metadata.positions.clone()

# NOTE(woosuk): We should handle the case where the draft model
# generates tokens beyond the max model length. Since it is complex
# to remove such requests from the batch, we keep them in the batch
Expand Down
Loading