[https://nvbugs/6115039][fix] Override from_hf in Qwen3HybridConfig to pre-compute num_attention_layers

[tensorrt-cicd]

Summary

• Root cause: Qwen3HybridConfig.set_values_if_none validator calls load_pretrained_config(self.name) where self.name is the bench identifier ("qwen3.5_9b_hf"), not a valid model path, causing OSError during trtllm-bench model type resolution.
• Fix: Override from_hf in Qwen3HybridConfig to pre-compute num_attention_layers and num_linear_attention_layers from the already-loaded pretrained config, so the validator's reload path is never triggered.
• Automated fix generated by repair-bot

Test plan

• Verify fix on the same GPU type as the original failure
• Check for regressions in related tests

Links

• Bug: https://nvbugs/6115039

Summary by CodeRabbit

• New Features
  • Qwen3Hybrid configurations can now be constructed directly from Hugging Face model names, automatically loading pretrained settings and computing necessary model parameters.

[coderabbitai]

📝 Walkthrough

Walkthrough

A classmethod from_hf is added to Qwen3HybridConfig to construct config from HuggingFace models by loading pretrained config, computing parameter count, and counting attention layer types. Additionally, set_values_if_none refactors to use list.count() instead of generator-based sum().

Changes

Cohort / File(s)	Summary
Qwen3HybridConfig Enhancement tensorrt_llm/bench/build/dataclasses.py	Added from_hf classmethod to build config from HuggingFace models by loading pretrained config, computing param_count, counting full_attention and linear_attention layer types. Refactored set_values_if_none to use list.count() for deriving layer counts instead of generator-based sum().

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~10 minutes

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 50.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title is fully related to the main change: overriding from_hf in Qwen3HybridConfig to pre-compute layer counts, which directly addresses the root cause of the bug described in the PR.
Description check	✅ Passed	The PR description includes all critical required sections: a clear summary of the root cause and fix, test plan with verification steps, and a link to the bug. The description template structure is followed adequately despite minor formatting differences.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

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

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

_{Review rate limit: 9/10 reviews remaining, refill in 6 minutes.}

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

[coderabbitai]

🧹 Nitpick comments (1)

tensorrt_llm/bench/build/dataclasses.py (1)

275-291: ⚡ Quick win

Annotate and document the new public factory.

from_hf is a new public API, but it currently omits the required return type annotation and a Google-style docstring. Please add both so this entry point follows the repo’s Python guidelines.

Suggested patch

-    def from_hf(cls, model_hf_name, hf_model_path):
+    def from_hf(
+        cls,
+        model_hf_name: str,
+        hf_model_path: str | None,
+    ) -> "Qwen3HybridConfig":
+        """Build a Qwen3 hybrid config from Hugging Face metadata."""

🤖 Prompt for AI Agents

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

In `@tensorrt_llm/bench/build/dataclasses.py` around lines 275 - 291, The new
public factory method from_hf lacks a return type annotation and a Google-style
docstring; update the method signature to include the correct return type (the
dataclass type, e.g., -> "ModelDataclass" or use "cls" typing via
ClassVar/Type[...]/"Self" depending on project typing rules) and add a concise
Google-style docstring above from_hf describing parameters (model_hf_name,
hf_model_path), behavior (loads pretrained_config via load_pretrained_config,
derives layer types with get_qwen3_hybrid_layer_types, computes param_count via
cls.get_param_count) and the returned instance, and ensure any forward
references are quoted if necessary to satisfy static typing.

🤖 Prompt for all review comments with AI agents

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

Nitpick comments:
In `@tensorrt_llm/bench/build/dataclasses.py`:
- Around line 275-291: The new public factory method from_hf lacks a return type
annotation and a Google-style docstring; update the method signature to include
the correct return type (the dataclass type, e.g., -> "ModelDataclass" or use
"cls" typing via ClassVar/Type[...]/"Self" depending on project typing rules)
and add a concise Google-style docstring above from_hf describing parameters
(model_hf_name, hf_model_path), behavior (loads pretrained_config via
load_pretrained_config, derives layer types with get_qwen3_hybrid_layer_types,
computes param_count via cls.get_param_count) and the returned instance, and
ensure any forward references are quoted if necessary to satisfy static typing.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: b3846119-7d9f-45ff-ac9e-1d75915fd9a1

📥 Commits

Reviewing files that changed from the base of the PR and between b7aee0b and 51e70e7.

📒 Files selected for processing (1)

• tensorrt_llm/bench/build/dataclasses.py