Skip to content

[CI][SpecDecode] Fix spec decode tests, use flash attention backend for spec decode CI tests. #8975

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
Oct 1, 2024
Merged

[CI][SpecDecode] Fix spec decode tests, use flash attention backend for spec decode CI tests. #8975

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
f90f27d
fix spec dec test
LiuXiaoxuanPKU Sep 30, 2024
9b1ef6b
reverse
LiuXiaoxuanPKU Sep 30, 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
2 changes: 0 additions & 2 deletions .buildkite/test-pipeline.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -199,8 +199,6 @@ steps:
- vllm/spec_decode
- tests/spec_decode
commands:
# See https://github.com/vllm-project/vllm/issues/5152
- export VLLM_ATTENTION_BACKEND=XFORMERS
- pytest -v -s spec_decode/e2e/test_multistep_correctness.py
- pytest -v -s spec_decode --ignore=spec_decode/e2e/test_multistep_correctness.py

Expand Down
6 changes: 4 additions & 2 deletions tests/spec_decode/test_multi_step_worker.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -173,7 +173,6 @@ def test_same_output_for_multi_step():
block_size,
num_gpu_blocks,
seed,
model_runner_cls=TP1DraftModelRunner,
Copy link
Collaborator

@comaniac comaniac Sep 30, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Wait I just found this. Why we remove this line? I suppose this test is for TP1 draft model runner?

Copy link
Collaborator Author

@LiuXiaoxuanPKU LiuXiaoxuanPKU Sep 30, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Actually the current main brach fails this test as well.
This is because
(1) the test did not configure MultiStepWorker correctly. It did not explicitly persist GPU tensors. When used with SpecDecodeWorker, it configures this sampler parameter as shown here.
(2) not sure which PR modifies the sampler output, the current test does not output correct CPU sampler output. Therefore, even if (1) is fixed, this test will fail when extracting the final output because it cannot get the output tokens for comparison.

We need to either change this test, or maybe just use the default model runner to pass this test.

Copy link
Collaborator Author

@LiuXiaoxuanPKU LiuXiaoxuanPKU Sep 30, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

reverse this change for now, will fix it in future prs.

)

worker = create_worker(
Expand Down Expand Up @@ -673,7 +672,10 @@ def test_use_draft_model_runner_advance_step():
worker.model_runner._gpu_advance_step.side_effect = ValueError(
exception_secret)

seq_group_metadata_list, _, _ = create_batch(batch_size, k)
seq_group_metadata_list, _, _ = create_batch(batch_size,
k,
block_size=block_size,
num_gpu_blocks=num_gpu_blocks)

# Fallback (should not call) when num_steps=1.
execute_model_req = ExecuteModelRequest(
Expand Down