Skip to content

[Bugfix] We have fixed the bug that occurred when using FlashInfer as the backend in vLLM Speculative Decoding. #5412

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

Closed
wants to merge 6 commits into from
Closed

[Bugfix] We have fixed the bug that occurred when using FlashInfer as the backend in vLLM Speculative Decoding. #5412

Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
cf9edaf
flashinfer bug at _prepare_model_input function is fixed
bong-furiosa Jun 11, 2024
d9f14fc
fix comment
bong-furiosa Jun 13, 2024
12a00cf
Merge remote-tracking branch 'upstream/main'
bong-furiosa Jun 14, 2024
12469af
fix comment and applied new commits
bong-furiosa Jun 14, 2024
61df3e8
Merge remote-tracking branch 'upstream/main'
bong-furiosa Jun 15, 2024
170c942
merged with upstream and modified comment
bong-furiosa Jun 15, 2024
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
10 changes: 8 additions & 2 deletions vllm/worker/model_runner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -395,9 +395,15 @@ def _prepare_model_input(
block_table = block_table[
-curr_sliding_window_blocks:]
if self.attn_backend.get_name() == "flashinfer":
paged_kv_indices.extend(block_table)
# flashinfer wrong indexing bug fix
block_table_bound = seq_data.get_len(
) // self.block_size + 1 if seq_data.get_len(
) % self.block_size != 0 else seq_data.get_len(
) // self.block_size
paged_kv_indices.extend(
block_table[:block_table_bound])
paged_kv_indptr.append(paged_kv_indptr[-1] +
len(block_table))
block_table_bound)
last_page_len = seq_data.get_len(
) % self.block_size
if last_page_len == 0:
Expand Down