[None][fix] Revert backend for Nemotron ViT to TRT-LLM

[yechank-nvidia]

Summary by CodeRabbit

• Chores
  • Updated the default attention backend for the vision model. Users relying on previous default behavior should explicitly configure their preferred backend in settings.

[yechank-nvidia]

/bot run

[github-actions]

👎 Promotion blocked, new vulnerability found

Vulnerability report

Component	Vulnerability	Description	Severity
encode/uvicorn	CVE-2020-7694	This affects all versions of package uvicorn. The request logger provided by the package is vulnerable to ASNI escape sequence injection. Whenever any HTTP request is received, the default behaviour of uvicorn is to log its details to either the console or a log file. When attackers request crafted URLs with percent-encoded escape sequences, the logging component will log the URL after it's been processed with urllib.parse.unquote, therefore converting any percent-encoded characters into their single-character equivalent, which can have special meaning in terminal emulators. By requesting URLs with crafted paths, attackers can: * Pollute uvicorn's access logs, therefore jeopardising the integrity of such files. * Use ANSI sequence codes to attempt to interact with the terminal emulator that's displaying the logs (either in real time or from a file).	HIGH
encode/uvicorn	CVE-2020-7695	Uvicorn before 0.11.7 is vulnerable to HTTP response splitting. CRLF sequences are not escaped in the value of HTTP headers. Attackers can exploit this to add arbitrary headers to HTTP responses, or even return an arbitrary response body, whenever crafted input is used to construct HTTP headers.	MEDIUM

[coderabbitai]

📝 Walkthrough

Walkthrough

The RADIOVisionModel.__init__ method's vision_attn_backend parameter default value is updated from "FLASHINFER" to "TRTLLM", changing the attention backend selected when callers do not explicitly provide this parameter.

Changes

Cohort / File(s)	Summary
Attention Backend Default tensorrt_llm/_torch/models/modeling_radio.py	Updated vision_attn_backend parameter default from "FLASHINFER" to "TRTLLM" in RADIOVisionModel.__init__.

Estimated code review effort

🎯 1 (Trivial) | ⏱️ ~2 minutes

🚥 Pre-merge checks | ✅ 2 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Description check	⚠️ Warning	The PR description is incomplete; it contains only the placeholder '@coderabbitai summary' without providing actual details about the change, rationale, or test coverage.	Replace the placeholder with a proper description explaining why the backend was reverted, what testing was done, and how this affects Nemotron ViT functionality.

✅ Passed checks (2 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly and specifically describes the main change: reverting the vision attention backend for Nemotron ViT from FLASHINFER to TRTLLM.
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (1)

tensorrt_llm/_torch/models/modeling_radio.py (1)

1013-1020: ⚠️ Potential issue | 🟡 Minor

Update stale constructor docstring default.

Line 1013 defaults vision_attn_backend to "TRTLLM", but Line 1019 still documents "FLASHINFER". Please sync docs with behavior.

Suggested patch

-            vision_attn_backend: Attention backend to use for the vision tower. Defaults to "FLASHINFER".
+            vision_attn_backend: Attention backend to use for the vision tower. Defaults to "TRTLLM".

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@tensorrt_llm/_torch/models/modeling_radio.py` around lines 1013 - 1020, The
constructor docstring for the RADIO model has a stale default: update the
docstring text that currently says "Defaults to \"FLASHINFER\"" to reflect the
actual parameter default "TRTLLM" for the vision_attn_backend argument (the
vision_attn_backend parameter in the model's __init__/constructor in
modeling_radio.py); keep the rest of the docstring unchanged.

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Outside diff comments:
In `@tensorrt_llm/_torch/models/modeling_radio.py`:
- Around line 1013-1020: The constructor docstring for the RADIO model has a
stale default: update the docstring text that currently says "Defaults to
\"FLASHINFER\"" to reflect the actual parameter default "TRTLLM" for the
vision_attn_backend argument (the vision_attn_backend parameter in the model's
__init__/constructor in modeling_radio.py); keep the rest of the docstring
unchanged.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: 5122d312-1b8d-45e3-a044-c216d5b863de

📥 Commits

Reviewing files that changed from the base of the PR and between 66431d8 and 76a2e36.

📒 Files selected for processing (1)

• tensorrt_llm/_torch/models/modeling_radio.py

[yechank-nvidia]

/bot run

[tensorrt-cicd]

PR_Github #44203 [ run ] triggered by Bot. Commit: d6aaa9c Link to invocation

[tensorrt-cicd]

PR_Github #44203 [ run ] completed with state SUCCESS. Commit: d6aaa9c
/LLM/main/L0_MergeRequest_PR pipeline #34629 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

• Please check the failed tests and fix your PR
• If you cannot view the failures, ask the CI triggerer to share details
• Once fixed, request an NVIDIA team member to trigger CI again

Link to invocation

[yechank-nvidia]

/bot run

[tensorrt-cicd]

PR_Github #44232 [ run ] triggered by Bot. Commit: d6aaa9c Link to invocation