Add assertion for no objects while hashing hf_config#16930
Add assertion for no objects while hashing hf_config#16930vllm-bot merged 1 commit intovllm-project:mainfrom
Conversation
|
👋 Hi! Thank you for contributing to the vLLM project. 💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels. Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging. To run CI, PR reviewers can either: Add 🚀 |
|
The language models test is failing consistently, PTAL |
We make the assumption that the hf_config is serializable. If there are any object ids in it, it will cause torch.compile to recompile every time the object id changes (which is always on warm start). Test Plan: - ran `python benchmark_latency.py` (smoke test) Signed-off-by: rzou <zou3519@gmail.com>
zou3519
force-pushed
the
hash_object_assertion
branch
from
7fa12db to
4a5d901
Compare
It ended up being a false positive, I tightened the regex some more. |
5 participants
We make the assumption that the hf_config is serializable. If there are any object ids in it, it will cause torch.compile to recompile every time the object id changes (which is always on warm start).
Test Plan:
python benchmark_latency.py(smoke test)