Repository navigation
fix: bound the image fetch in the TRT-LLM preprocessing template - #129
Conversation
 |
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
| # Read one byte past the limit so that an oversized response is | ||
| # rejected rather than silently truncated. | ||
| image_data = response.raw.read(self.IMAGE_FETCH_MAX_BYTES + 1, | ||
| decode_content=True) | ||
| if len(image_data) > self.IMAGE_FETCH_MAX_BYTES: | ||
| raise ValueError( | ||
| f"[TensorRT-LLM][ERROR] Image at {img_url} exceeds the " | ||
| f"{self.IMAGE_FETCH_MAX_BYTES} byte limit.") |
There was a problem hiding this comment.
When one request in a dynamic batch receives an HTTP error, exceeds the size limit, or times out, these new exceptions escape the unguarded per-request loop and abort the entire execute invocation, causing otherwise valid requests in the same batch to fail. Convert fetch failures into an error response for only the offending request.
Knowledge Base Used: TensorRT-LLM model pipeline templates
|
Want your agent to iterate on Greptile's feedback? Try greploops. |
What does the PR do?
Bounds the remote image fetch in the TRT-LLM preprocessing template.
load_images_from_urlsfetched a request-supplied URL with no timeout, no size limit, and no status check. Adds an inactivity timeout, a total transfer deadline, and a max response size, and raises on HTTP errors instead of passing an error page to PIL. The inline base64 branch is unchanged.Checklist
<commit_type>: <Title>Commit Type:
Related PRs:
Test plan:
Ran this example locally after the updates