security: bump vllm, GitPython for CVE remediation - #2344
Closed
kajalj22 wants to merge 9 commits into
Closed
Conversation
- vllm 0.20.0 → 0.26.0 (security fix; also updates flashinfer-python 0.6.8.post1 → 0.6.14) - GitPython 3.1.50 → 3.1.57 (security fix; resolved to 3.1.58 by uv) - pyarrow already at 25.0.0 in lock (constraint >=23.0.1 unchanged, CVE addressed) - diskcache removed as transitive dep of old vllm (no standalone fix available) Signed-off-by: Kajal Jain <kajalj@nvidia.com>
vllm 0.26.0 imports NamespaceTool from openai.types.responses, which was added in openai 2.53.0. No breaking changes in the SDK between 2.7.2 and 2.53.0. Bump cap and lock to 2.53.0. Signed-off-by: Kajal Jain <kajalj@nvidia.com>
setuptools reads [tool.distutils.egg_info] egg_base = "cache" from pyproject.toml and fails if the directory doesn't exist. The hermetic stage only copies pyproject.toml + uv.lock, so cache/ is never present. This was masked by Docker layer caching until the lock file changed. Signed-off-by: Kajal Jain <kajalj@nvidia.com>
kajalj22
added
the
r0.5.0
Docker hermetic stage uses uv 0.11.29; regenerate lock to match and avoid --locked mismatch errors in the container build. Signed-off-by: Kajal Jain <kajalj@nvidia.com>
openai 2.53.0 made two breaking changes for Gym: 1. InputTokensDetails.cache_write_tokens is now a required field. Add cache_write_tokens=0 as a default in NeMoGymResponseInputTokensDetails so all existing callers (anthropic_converter, responses_converter, server tests) continue to work without individual changes. 2. Reasoning.context is now a valid optional field (was extra_forbidden before). Update test_prunes_nested_extra_fields to use a genuinely unknown field, and remove the stale reasoning.context example from the validate_streaming_responses_params docstring. Signed-off-by: Kajal Jain <kajalj@nvidia.com>
…nd validation openai 2.53.0 added three new Optional fields to FunctionTool (BaseModel) and FunctionToolParam (TypedDict): allowed_callers, defer_loading, output_schema. These fields were not present in 2.7.2 and cause two kinds of breakage: 1. NeMoGymResponse.model_dump() now includes them in response.tools, breaking server tests that compare serialized responses to expected dicts. 2. validate_streaming_responses_params fails when callers round-trip tool dicts via FunctionTool.model_dump() (e.g. single_step_tool_use_with_argument_comparison). Fix: override model_dump in NeMoGymResponse to strip the three new fields from each tool dict, and pre-strip them from input tools in validate_streaming_responses_params before validation. Signed-off-by: Kajal Jain <kajalj@nvidia.com>
…nd serialization NeMoGymResponseCreateParamsNonStreaming now strips allowed_callers, defer_loading, and output_schema (all None-valued) from tool dicts: - Before validation via model_validator(mode='before'), so that callers using FunctionTool.model_dump() don't fail when defer_loading=None - After serialization via model_dump() override, because Pydantic v2 adds absent TypedDict fields as None when dumping List[ToolParam]. Also fix ruff import-sort in responses_streaming.py. Signed-off-by: Kajal Jain <kajalj@nvidia.com>
1. cvdp/requirements.txt: update openai==2.7.2 -> >=2.53.0,<=2.53.0 2. NeMoGymResponseInputTokensDetails.cache_write_tokens: use Field(default=0, exclude=True) so the new required field has a default but is excluded from serialized output, preserving backward-compat format. 3. NeMoGymResponseCreateParamsNonStreaming.tools: change List[ToolParam] to List[Any] so Pydantic stops applying TypedDict schema normalization (which was adding defer_loading/allowed_callers/output_schema as None to every function tool, corrupting round-tripped params). 4. NeMoGymResponse: replace model_dump() override with @model_serializer (mode='wrap') so the tool-field stripping applies to BOTH Python and Rust/JSON serialization paths (FastAPI, model_dump_json). model_dump() overrides are bypassed by Pydantic's internal Rust serializer. Signed-off-by: Kajal Jain <kajalj@nvidia.com>
diskcache has unfixed CVEs and is no longer a transitive dep since vllm>=0.26.0. Adding it to exclude-dependencies prevents re-introduction via future transitive pulls, including in server venvs that install nemo-gym editably. Signed-off-by: Kajal Jain <kajalj@nvidia.com>
thomasdhc
approved these changes
Aug 5, 2026
Contributor
Author
|
Closing in favor of #2353 |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Test plan
🤖 Generated with Claude Code