Skip to content

[CI]enable chunked prefill by default #4569

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
wangxiyuan merged 1 commit into from
Dec 2, 2025
Merged

[CI]enable chunked prefill by default #4569

Show file tree
Hide file tree
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
2 changes: 1 addition & 1 deletion tests/e2e/conftest.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -280,7 +280,7 @@ def __init__(
disable_log_stats: bool = True,
tensor_parallel_size: int = 1,
block_size: int = 16,
enable_chunked_prefill: bool = False,
enable_chunked_prefill: bool = True,
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist Bot Nov 29, 2025

Choose a reason for hiding this comment

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

high

Enabling enable_chunked_prefill by default will affect MLA (Multi-Layer Attention) models in tests. The logic in vllm_ascend/worker/model_runner_v1.py's _build_attn_state function will now set the attention state to ChunkedPrefill for all models, including MLA.

However, I found a configuration chunked_prefill_for_mla in vllm_ascend/ascend_config.py which seems intended to control this behavior for MLA models, but it is currently not used. This is confusing and could lead to unexpected behavior for developers trying to configure this feature.

If chunked prefill is now fully supported for MLA models and should be on by default, please consider removing the unused chunked_prefill_for_mla configuration to avoid confusion.

If chunked prefill for MLA is experimental or should be opt-in, the logic in _build_attn_state should be updated to respect this configuration. For example:

# In vllm_ascend/worker/model_runner_v1.py _build_attn_state
...
        elif self.scheduler_config.enable_chunked_prefill:
            if self.vllm_config.model_config.use_mla and not self.ascend_config.chunked_prefill_for_mla:
                 attn_state = AscendAttentionState.PrefillCacheHit
            else:
                 attn_state = AscendAttentionState.ChunkedPrefill
...

Given this PR changes the default for all tests, clarifying the intended behavior for MLA models is important for maintainability.

swap_space: int = 4,
enforce_eager: Optional[bool] = False,
quantization: Optional[str] = None,
Expand Down
1 change: 0 additions & 1 deletion tests/e2e/multicard/test_prefix_caching.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -58,7 +58,6 @@
]


@pytest.mark.skip(reason="Fix me, the accuracy is not correct")
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("max_tokens", [50])
def test_prefix_cache_with_v1_scheduler(model: str, max_tokens: int) -> None:
Expand Down
1 change: 0 additions & 1 deletion tests/e2e/singlecard/spec_decode_v1/test_v1_spec_decode.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -118,7 +118,6 @@ def test_eagle_correctness(
spec_model_name = eagle3_model_name() if use_eagle3 else eagle_model_name()
with VllmRunner(
model_name,
enable_chunked_prefill=True,
max_num_seqs=1,
max_num_batched_tokens=2048,
gpu_memory_utilization=0.6,
Expand Down
Loading