[https://nvbugs/6433376][fix] Update the Dense test to mirror the MoE sibling — assert bfloat16 under…

[trtllm-agent]

Summary

• Root cause: Dense test asserts old resolver semantics (fp32 cache from mamba_ssm_dtype) that PR [https://nvbugs/6422332][fix] Keep SSM cache in weights dtype when ma… #16065 narrowed away; sibling MoE test was updated but dense was missed.
• Fix: Update the Dense test to mirror the MoE sibling — assert bfloat16 under "auto", add explicit "float32" opt-in check; remove both waives.txt entries.
• 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/6433376

Summary by CodeRabbit

• Bug Fixes
  • Updated cache dtype handling for a specific model path so automatic resolution now uses the expected lower-precision setting.
  • Added coverage for explicitly selecting a standard precision setting, ensuring it is applied correctly.
• Tests
  • Removed outdated test waivers for the affected model checks.

[coderabbitai]

📝 Walkthrough

Walkthrough

This PR updates the expected dtype for Mamba SSM KV-cache resolution in a Qwen3.5 dense VL unit test, changing "auto" resolution from float32 to bfloat16 and adding coverage for explicit float32 opt-in. Two corresponding test waiver entries are removed from the waivers file.

Changes

Mamba SSM Cache Dtype Test Update

Layer / File(s)	Summary
Updated dtype expectations and removed waivers tests/unittest/_torch/modeling/test_modeling_qwen3_5_vl.py, tests/integration/test_lists/waives.txt	Test now expects bfloat16 for "auto" dtype resolution (previously float32) and adds a new assertion for explicit "float32" opt-in; the two related waiver entries are removed from the waivers list.

Estimated code review effort: 2 (Simple) | ~10 minutes

Possibly related PRs

• NVIDIA/TensorRT-LLM#15249: Adds the underlying Qwen3.5 VL dense config/dtype resolution and original test scaffolding modified here.
• NVIDIA/TensorRT-LLM#16065: Alters the "auto" dtype-resolution fallback for Mamba/SSM cache dtype that this PR's test now validates.
• NVIDIA/TensorRT-LLM#16186: Introduced the exact two waiver entries this PR removes from waives.txt.

Suggested reviewers: moraxu, tijyojwad, dongxuy04, laikhtewari

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title is specific, matches the bug fix, and includes the required ticket/type prefix.
Description check	✅ Passed	The description covers the bug, fix, tests, and bug link, but it omits the template's PR Checklist section.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
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.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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

[coderabbitai]

🧹 Nitpick comments (1)

tests/unittest/_torch/modeling/test_modeling_qwen3_5_vl.py (1)

160-163: 🎯 Functional Correctness | 🔵 Trivial | ⚡ Quick win

Opt-in coverage doesn't verify downstream propagation.

The "auto" case verifies both quant_config.mamba_ssm_cache_dtype and the downstream mamba_params.mamba_ssm_cache_dtype from extract_mamba_kv_cache_params. The new float32 opt-in case only checks quant_config, leaving the downstream propagation path (which reads quant_config.mamba_ssm_cache_dtype first, per config_utils.py's extract_mamba_kv_cache_params) unverified for the explicit opt-in.

🧪 Suggested addition to close the coverage gap

     opt_in_config = ModelConfig(pretrained_config=config)
     validate_and_set_mamba_ssm_cache_dtype(opt_in_config, "float32")
     assert opt_in_config.quant_config.mamba_ssm_cache_dtype is torch.float32
+
+    opt_in_mamba_params = extract_mamba_kv_cache_params(
+        config.text_config,
+        quant_config=opt_in_config.quant_config,
+    )
+    assert opt_in_mamba_params.mamba_ssm_cache_dtype is torch.float32

As per path instructions, coverage here is largely sufficient but this asymmetry between the auto/opt-in paths is a minor gap worth closing to mirror the auto-path assertions.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@tests/unittest/_torch/modeling/test_modeling_qwen3_5_vl.py` around lines 160
- 163, The explicit float32 opt-in test in the Qwen3_5 VL modeling suite only
checks `quant_config.mamba_ssm_cache_dtype` and misses the downstream
propagation verified in the auto path. Update the `test_modeling_qwen3_5_vl`
case around `validate_and_set_mamba_ssm_cache_dtype` to also exercise
`extract_mamba_kv_cache_params` on `opt_in_config` and assert the resulting
`mamba_params.mamba_ssm_cache_dtype` is `torch.float32`, matching the coverage
pattern used by the auto-case.

Source: Path instructions

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Nitpick comments:
In `@tests/unittest/_torch/modeling/test_modeling_qwen3_5_vl.py`:
- Around line 160-163: The explicit float32 opt-in test in the Qwen3_5 VL
modeling suite only checks `quant_config.mamba_ssm_cache_dtype` and misses the
downstream propagation verified in the auto path. Update the
`test_modeling_qwen3_5_vl` case around `validate_and_set_mamba_ssm_cache_dtype`
to also exercise `extract_mamba_kv_cache_params` on `opt_in_config` and assert
the resulting `mamba_params.mamba_ssm_cache_dtype` is `torch.float32`, matching
the coverage pattern used by the auto-case.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: c8fc2534-78f3-4195-bb49-bf432aefa2ca

📥 Commits

Reviewing files that changed from the base of the PR and between 8e7b98c and 89d8965.

📒 Files selected for processing (2)

• tests/integration/test_lists/waives.txt
• tests/unittest/_torch/modeling/test_modeling_qwen3_5_vl.py

💤 Files with no reviewable changes (1)

• tests/integration/test_lists/waives.txt

[VALLIS-NERIA]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #61250 [ run ] triggered by Bot. Commit: 7fabe92 Link to invocation

[tensorrt-cicd]

PR_Github #61250 [ run ] completed with state SUCCESS. Commit: 7fabe92
/LLM/main/L0_MergeRequest_PR pipeline #49487 completed with status: 'SUCCESS'

CI Report

Link to invocation

[VALLIS-NERIA]

/bot run

[tensorrt-cicd]

PR_Github #61800 [ run ] triggered by Bot. Commit: 5b87123 Link to invocation

[tensorrt-cicd]

PR_Github #61800 [ run ] completed with state SUCCESS. Commit: 5b87123
/LLM/main/L0_MergeRequest_PR pipeline #49999 completed with status: 'SUCCESS'

CI Report

Link to invocation