[CI] Solidify speculative decoding E2E coverage - #50713
Conversation
Signed-off-by: Andreas Karatzas <Andreas.Karatzas@amd.com>
|
Documentation preview: https://vllm--50713.org.readthedocs.build/en/50713/ |
mergify
Bot
added
documentation
Signed-off-by: Andreas Karatzas <Andreas.Karatzas@amd.com>
Signed-off-by: Andreas Karatzas <Andreas.Karatzas@amd.com>
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
AndreasKaratzas
added
the
ready
Signed-off-by: Andreas Karatzas <Andreas.Karatzas@amd.com>
| # GSM8K is text-only; avoid profiling unused multimodal towers. | ||
| limit_mm_per_prompt={"image": 0, "audio": 0, "video": 0}, |
There was a problem hiding this comment.
Nit: we can just use language_model_only=True
| @@ -43,10 +44,14 @@ def test_gemma4_dspark_correctness_and_acceptance_rate( | |||
| max_num_seqs=32, | |||
| gpu_memory_utilization=0.85, | |||
| enforce_eager=True, | |||
| enable_chunked_prefill=None, | |||
There was a problem hiding this comment.
Why do we need to set block_size and enable_chunked_prefill? These should be fine as their defaults
There was a problem hiding this comment.
Actually VllmRunner defaults these to 16 and False, unlike LLM construction. Removing them would change behavior and disable Gemma4's supported chunked prefill.
Signed-off-by: Andreas Karatzas <Andreas.Karatzas@amd.com>
|
/ci run |
|
✅ Triggered Buildkite CI #82443 for commit |
|
/ci run |
|
✅ Triggered Buildkite CI #82541 for commit |
|
This pull request has merge conflicts that must be resolved before it can be |
Retain the PR vllm_runner lifecycle and acceptance diagnostics while incorporating the async-scheduling fix from main. Co-authored-by: OpenAI Codex <codex@openai.com> Signed-off-by: Andreas Karatzas <Andreas.Karatzas@amd.com>
|
/ci run |
|
✅ Triggered Buildkite CI #82804 for commit |
Retain the PR's vllm_runner lifecycle, ROCm-safe dSpark setup, and diagnostic assertions while incorporating the Qwen3.6 dSpark acceptance coverage from main. Co-authored-by: OpenAI Codex <codex@openai.com> Signed-off-by: Andreas Karatzas <akaratza@amd.com>
|
/ci run |
|
✅ Triggered Buildkite CI #82930 for commit |
|
This pull request has merge conflicts that must be resolved before it can be |
Retain the PR's vllm_runner lifecycle and diagnostic assertions while incorporating hybrid Model Runner V2 MTP coverage from main. Co-authored-by: OpenAI Codex <codex@openai.com> Signed-off-by: Andreas Karatzas <akaratza@amd.com>
|
/ci run |
|
✅ Triggered Buildkite CI #83023 for commit |
2 participants
vllm_runnerfor consistent setup, defaults, and cleanup.This is a follow-up on #50330. It applies some NITS across the new spec-decode area while hardening deterministic setup and actionable failure reporting.