[Cherry-pick] PRs #2022 #2026 #2032 #1981 #2010 #2043 #2042 #2038 #2050 #2041 #2031 #2061 #1983 #1628

[kevalmorabia97]

Cherry-picked PRs

• Update documentation guide for ONNX INT4 PTQ on Windows cuda13 host #2022
• tools/mcp: pin mcp<2 to fix unit CI collection #2026
• fix(export): [NVBug 6525534] preserve nested VLM namespaces #2032
• [5726458] [Experimental] Add NVFP4 projection-output-quantizer recipe and HF embedding ONNX export example #1981
• Add ModelOpt QAD skill for Slurm workflows #2010
• Skip zero-quota splits in get_dataset_samples #2043
• [skill] evaluation: correct the stale vLLM CUDA-13 image tag rule #2042
• [NVBug: 5987078] Fix unified HF export of compressed NVFP4 weights (--low_memory_mode) #2038
• Bump transformers dependency to >=4.57,<5.15 #2050
• [NVBug: 6524370] use sequential device_map for DiffusionGemma #2041
• Add Megatron-Bridge prune & quantize launcher pipelines #2031
• Fix saving pruned Nemotron-3-Nano hybrid_override_pattern with MTP or Pipe symbols #2061
• [6425069][ONNX][Autocast] Fix autocast metadata propagation #1983
• [6058841] Fix inconsistent tensor types on control-flow (If/Loop/Scan) subgraphs during FP16/BF16 conversion #1628

Summary by CodeRabbit

• New Features

  • Added recipe-driven ONNX quantization and export for embedding and reranking models.
  • Added launcher support for inline commands, package requirements, and configurable Docker users.
  • Added QAD workflows, Nemotron data recipes, pruning, quantization, and evaluation examples.
  • Added evaluation accuracy and pruning score thresholds.

• Bug Fixes

  • Improved ONNX control-flow conversion, compressed NVFP4 export, dataset loading, and multi-GPU DiffusionGemma handling.

• Documentation

  • Updated CUDA 12/13 installation guidance and vLLM Blackwell deployment instructions.

[coderabbitai]

📝 Walkthrough

Walkthrough

The PR adds QAD and launcher workflows, recipe-driven ONNX quantization, expanded ONNX autocast handling, deployment guidance updates, evaluation gates, export fixes, dataset-loading changes, and Transformers compatibility updates.

Changes

Deployment guidance

Layer / File(s)	Summary
vLLM CUDA tagging guidance .agents/skills/deployment/*, .agents/skills/evaluation/*	Updates vLLM CUDA-13 and CUDA-12 image-tag guidance for Blackwell NVFP4 deployments and adds manifest verification instructions.

QAD and Megatron workflows

Layer / File(s)	Summary
QAD execution policy .agents/skills/qad/*, examples/megatron_bridge/README.md, examples/megatron_bridge/data/*	Adds authorized QAD workflow, training defaults, dataset preparation, monitoring, recovery, and reporting rules.
Distillation controls and validation examples/megatron_bridge/distill.py, examples/megatron_bridge/prune_minitron.py, tests/examples/megatron_bridge/*	Adds interval validation, independent checkpoint cadence, exit controls, pruning score gates, and hybrid-pattern normalization.

Launcher execution workflows

Layer / File(s)	Summary
Launcher task contract and configuration tools/launcher/core.py, tools/launcher/slurm_config.py, tools/launcher/docs/*	Adds inline commands, pip requirements, environment overrides, and Docker user configuration.
Requirement installation and task execution tools/launcher/core.py, tools/launcher/examples/*	Installs requirements with Slurm synchronization and adds Nemotron pruning and quantization workflows.
Launcher validation and regression tests tools/launcher/tests/*	Tests inline-task validation, requirement installation, Docker users, environment isolation, and variable substitution.

ONNX quantization and autocast

Layer / File(s)	Summary
Recipe-driven ONNX examples examples/torch_onnx/*, modelopt_recipes/*, tests/examples/torch_onnx/*	Adds recipe-based quantization configuration and Hugging Face embedding/reranking export with NVFP4 and FP8 output quantization.
Autocast metadata and control-flow conversion modelopt/onnx/autocast/*, modelopt/onnx/utils.py, tests/unit/onnx/autocast/*	Preserves public I/O metadata, improves shape inference and constant metadata synchronization, and handles nested control-flow precision and casts.

ModelOpt runtime and export updates

Layer / File(s)	Summary
Evaluation gates and device mapping examples/llm_eval/*, examples/hf_ptq/*, tests/examples/*	Adds accuracy thresholds and DiffusionGemma detection with sequential multi-GPU mapping.
Export state and namespace handling modelopt/torch/export/*, modelopt/torch/prune/*, tests/gpu*/*, tests/unit/torch/export/*	Preserves compressed NVFP4 scales, filters shadowed reverse renames, and persists the best pruning candidate.
Compatibility and utilities modelopt/torch/*, noxfile.py, pyproject.toml, tests/_test_utils/*, tests/gpu/torch/puzzletron/*	Updates Transformers support, quantizer metadata, dataset split loading, Qwen3-VL imports, and related fixtures.

Estimated code review effort: 5 (Critical) | ~120 minutes

Possibly related PRs

• NVIDIA/Model-Optimizer#2031: Covers overlapping launcher, evaluation, pruning, and Megatron-Bridge workflow changes.
• NVIDIA/Model-Optimizer#2038: Covers compressed NVFP4 export handling and regression tests.
• NVIDIA/Model-Optimizer#2042: Covers the vLLM CUDA-13 image-tag guidance updates.

Suggested labels: cherry-pick-0.46.0

Suggested reviewers: cjluo-nv, edwardf0t1, meenchen

🚥 Pre-merge checks | ✅ 4 | ❌ 2

❌ Failed checks (1 warning, 1 inconclusive)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 46.81% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.
Title check	❓ Inconclusive	The title identifies the cherry-picked PRs but does not summarize the broad release changes included in this pull request.	Replace the PR number list with a concise summary of the main release integration changes, such as updates to quantization, ONNX, launcher, and deployment support.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
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.
Security Anti-Patterns	✅ Passed	No security anti-patterns detected: no unsafe torch.load, numpy.load, hardcoded trust_remote_code, eval/exec on external input, new nosec comments, or non-permissive dependencies in new PR code.

✨ Finishing Touches 💡 1

📝 Generate docstrings 💡

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch cherry-picks/release-0.46.0

---

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

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-08-04 16:58 UTC

[codecov]

Codecov Report

❌ Patch coverage is 83.75635% with 64 lines in your changes missing coverage. Please review.
✅ Project coverage is 77.99%. Comparing base (33d05b0) to head (5bbc2d2).
⚠️ Report is 1 commits behind head on release/0.46.0.

Files with missing lines	Patch %	Lines
modelopt/onnx/autocast/precisionconverter.py	81.26%	59 Missing ⚠️
modelopt/onnx/utils.py	77.77%	4 Missing ⚠️
modelopt/torch/export/quant_aware_conversion.py	95.45%	1 Missing ⚠️

Additional details and impacted files

@@                 Coverage Diff                 @@
##           release/0.46.0    #2062       +/-   ##
===================================================
+ Coverage           66.83%   77.99%   +11.15%     
===================================================
  Files                 519      519               
  Lines               58916    59278      +362     
===================================================
+ Hits                39376    46233     +6857     
+ Misses              19540    13045     -6495     

Flag	Coverage Δ
examples	43.19% <28.68%> (-0.20%)	⬇️
gpu	58.88% <23.85%> (+37.77%)	⬆️
regression	14.98% <1.77%> (-0.02%)	⬇️
unit	55.17% <80.96%> (+0.27%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[coderabbitai]

Warning

CodeRabbit couldn't request changes on this pull request because it doesn't have sufficient GitHub permissions.

Please grant CodeRabbit Pull requests: Read and write permission and re-run the review.

👉 Steps to fix this

Actionable comments posted: 12

🧹 Nitpick comments (5)

tests/examples/megatron_bridge/test_prune_minitron.py (2)

43-51: 🗄️ Data Integrity & Integration | 🔵 Trivial | ⚡ Quick win

Assert the MTP and hybrid-pattern output contract.

The fixture enables num_nextn_predict_layers=1 and mtp_hybrid_override_pattern, but the assertions only check output existence and parameter counts. They do not verify that MTP fields are removed or that the normalized hybrid pattern is persisted.

Add direct assertions against the generated configuration or checkpoint metadata. Use the appropriate inspection path for Megatron and Hugging Face outputs.

As per path instructions, tests must exercise the behavior they claim to validate.

🤖 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/examples/megatron_bridge/test_prune_minitron.py` around lines 43 - 51,
Extend the test case using num_nextn_predict_layers=1 and
mtp_hybrid_override_pattern to inspect both generated output formats: assert the
Megatron configuration or checkpoint metadata has MTP fields removed, and assert
the Hugging Face metadata persists the normalized hybrid pattern. Use each
format’s existing inspection/loading helpers rather than relying only on output
existence or parameter counts.

Source: Path instructions

---

79-79: 🎯 Functional Correctness | 🔵 Trivial | ⚡ Quick win

Exercise the failing score-gate branch.

Both tests use score_lower_bound=0.0, so they cover only the passing path. Add a case with a bound above the measured score and assert that the pruning command exits non-zero.

As per path instructions, focused tests must protect the changed behavior and use the real implementation where applicable.

Also applies to: 148-148

🤖 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/examples/megatron_bridge/test_prune_minitron.py` at line 79, Add a
focused test case in the existing pruning tests around the score_lower_bound
configuration, using a bound higher than the measured score and the real pruning
command implementation. Assert that this failing score-gate path exits non-zero,
while preserving the existing passing-path coverage.

Source: Path instructions

tests/examples/megatron_bridge/test_qad.py (1)

113-124: 🎯 Functional Correctness | 🔵 Trivial | ⚡ Quick win

Test duration-based exit independently.

This test always exits at iteration 2 through exit_interval. The ten-minute exit_duration_in_mins value cannot trigger. A broken duration-exit mapping could still pass this test.

Add a focused configuration test or a separate integration case with exit_interval unset and a controlled duration-based exit.

As per path instructions, tests must exercise the behavior they claim to validate.

🤖 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/examples/megatron_bridge/test_qad.py` around lines 113 - 124, Update
the test around the distillation command and tracker assertions to validate
duration-based exit independently: add a focused configuration test or separate
integration case with exit_interval unset, use a controlled
exit_duration_in_mins value, and retain assertions confirming the run exits at
the expected iteration/checkpoint. Ensure the existing interval-based test
remains scoped to interval exit rather than claiming duration coverage.

Source: Path instructions

modelopt/onnx/autocast/precisionconverter.py (1)

1620-1665: 🚀 Performance & Scalability | 🔵 Trivial | 💤 Low value

Reuse the existing maps instead of scanning the graph containers.

_get_tensor_shape and _get_tensor_elem_type scan self.model.graph.initializer and the three value-info containers on every lookup. _refresh_gathernd_pre_cast_declarations already rebuilds self.initializer_map and self.value_info_map immediately before calling them, and both maps cover exactly these tensors. utils.setup_mappings populates value_info_map from value_info, input, and output.

Reading the maps keeps the behavior identical and removes the repeated linear scans over initializers, which matter on large models.

♻️ Proposed change to use the maintained maps

     def _get_tensor_shape(self, tensor_name: str) -> list[int | str | None] | None:
-        initializer = next(
-            (value for value in self.model.graph.initializer if value.name == tensor_name), None
-        )
+        initializer = self.initializer_map.get(tensor_name)
         if initializer is not None:
             return list(initializer.dims)
 
-        for value_info in (
-            *self.model.graph.input,
-            *self.model.graph.output,
-            *self.model.graph.value_info,
-        ):
-            if value_info.name != tensor_name:
-                continue
+        value_info = self.value_info_map.get(tensor_name)
+        if value_info is not None:
             tensor_type = value_info.type.tensor_type
-            if not tensor_type.HasField("shape"):
-                continue
-            shape = []
-            for dim in tensor_type.shape.dim:
-                if dim.HasField("dim_value"):
-                    shape.append(dim.dim_value)
-                elif dim.HasField("dim_param"):
-                    shape.append(dim.dim_param)
-                else:
-                    shape.append(None)
-            return shape
+            if tensor_type.HasField("shape"):
+                shape = []
+                for dim in tensor_type.shape.dim:
+                    if dim.HasField("dim_value"):
+                        shape.append(dim.dim_value)
+                    elif dim.HasField("dim_param"):
+                        shape.append(dim.dim_param)
+                    else:
+                        shape.append(None)
+                return shape

🤖 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 `@modelopt/onnx/autocast/precisionconverter.py` around lines 1620 - 1665,
Update _get_tensor_shape and _get_tensor_elem_type to resolve initializers
through self.initializer_map and value metadata through self.value_info_map
instead of scanning graph.initializer, input, output, and value_info. Preserve
the existing shape, element-type, Cast-producer fallback, and missing-value
behavior while relying on the maps rebuilt by
_refresh_gathernd_pre_cast_declarations.

docs/source/getting_started/windows/_installation_standalone.rst (1)

16-16: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

Fix inconsistent capitalization of "cuDNN".

Line 16 uses "CuDNN". Line 71 uses "cudnn". Lines 54 and 87 use "cuDNN". Use "cuDNN" consistently throughout the document.

Proposed fix

-      - CUDA Toolkit and matching CuDNN for using CUDA path during calibration (e.g. for calibration of ONNX models using `onnxruntime-gpu` or CUDA EP)
+      - CUDA Toolkit and matching cuDNN for using CUDA path during calibration (e.g. for calibration of ONNX models using `onnxruntime-gpu` or CUDA EP)

-The steps below assume a CUDA 13.x Toolkit, compatible cudnn, and a compatible driver are already installed on the host.
+The steps below assume a CUDA 13.x Toolkit, compatible cuDNN, and a compatible driver are already installed on the host.

Also applies to: 71-71, 87-87

🤖 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 `@docs/source/getting_started/windows/_installation_standalone.rst` at line 16,
Update the CUDA dependency references in the Windows standalone installation
document, including the entries around the CUDA Toolkit requirement and later
cuDNN mentions, to use the exact capitalization “cuDNN” consistently.

🤖 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.

Inline comments:
In @.agents/skills/qad/SKILL.md:
- Around line 22-24: Update the common skill references in the QAD skill
documentation to use the canonical repository-relative
`.agents/skills/common/...` paths, including `environment-setup.md`,
`workspace-management.md`, `slurm-setup.md`, and `remote-execution.md`; leave
the Megatron bridge paths unchanged.

In `@examples/llm_eval/lm_eval_hf.py`:
- Around line 298-308: Update the result-loading logic around the output_path
glob and scores extraction to support both directory paths and direct JSON file
paths. Load the selected JSON file directly when output_path is a file;
otherwise retain the existing latest-results-file selection. Extract the
requested task metrics from either the JSON's "results" or "groups" mapping,
preferring the available task entry before computing acc.

In `@examples/megatron_bridge/prune_minitron.py`:
- Around line 687-698: Move the accuracy-gate block in main immediately after
mtp.prune() completes and before either Megatron or Hugging Face output-saving
path writes artifacts. Preserve the existing score-bound evaluation and
sys.exit(1) behavior, ensuring failed pruning runs cannot leave outputs that
bypass the gate on rerun.
- Around line 242-247: Update the argparse definition for score_lower_bound in
prune_minitron.py to reject non-finite float values at the CLI boundary by
validating math.isfinite() after parsing. Preserve None as the unset value and
continue accepting finite thresholds.

In `@examples/torch_onnx/torch_quant_to_onnx.py`:
- Around line 125-136: Keep the recipe format and export mode consistent by
resolving a validated format profile once and using it throughout. In
examples/torch_onnx/torch_quant_to_onnx.py lines 125-136, derive Conv2d
overrides from the resolved recipe; document the compatibility contract in lines
458-468, reject incompatible recipe/mode combinations in lines 544-548, pass the
profile at lines 598-598, and use it for TensorRT guards at lines 613-625.
Update examples/torch_onnx/README.md lines 72-75 to document the enforced or
derived relationship, and add coverage for valid custom recipes and rejected
mismatches.

In `@modelopt/onnx/autocast/convert.py`:
- Around line 49-57: Update _capture_network_io_metadata and the corresponding
PrecisionConverter._restore_original_io_metadata flow so keep_io_types=True
remains valid after sanitization: preserve or reconcile renamed/disconnected
outputs and FP64 input/output element types before restoration, preventing
RuntimeError while retaining the original public I/O contract. Add regression
tests covering disconnected output names and FP64 I/O types.

In `@modelopt/onnx/autocast/precisionconverter.py`:
- Around line 307-420: Update _get_const_values to support Constant nodes
encoded with value_int, value_ints, or sparse_value, extracting their values for
shape inference without assuming attrs["value"] exists. When an encoding cannot
be safely converted to constant values, return None so _infer_unsqueeze_op_shape
exits safely instead of raising KeyError.

In `@tests/examples/llm_eval/test_llm_eval.py`:
- Around line 39-41: Add a focused test for _enforce_accuracy_gate that provides
a results JSON with a score below accuracy_lower_bound and asserts the function
raises SystemExit with code 1. Keep the existing passing-bound test unchanged so
both successful and failing gate behavior are covered.

In `@tests/unit/torch/export/test_quant_aware_conversion.py`:
- Around line 269-270: Add brief reason comments before each conditional local
Transformers import at tests/unit/torch/export/test_quant_aware_conversion.py
lines 180, 183, 237, 269-270, 297-298, and 358, explaining that transformers is
optional and conversion_mapping is version-dependent. Keep the existing
pytest.importorskip guards and imports unchanged.

In `@tools/launcher/core.py`:
- Around line 114-126: Preserve positional construction compatibility by moving
the new inline, reqs, and reqs_file fields in SandboxTask after all existing
fields, or making them keyword-only. In tools/launcher/slurm_config.py at lines
51-53 and 77-99, likewise append docker_user after the existing SlurmConfig
fields and slurm_factory parameters, or make it keyword-only.

In
`@tools/launcher/examples/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/mbridge_prune.yaml`:
- Around line 23-25: Remove the hardcoded --trust_remote_code flag from all five
command invocations:
tools/launcher/examples/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/mbridge_prune.yaml
lines 23-25 and 49-52, and mbridge_quantize.yaml lines 20-22, 31-34, and 49-52.
Keep these workflows disabled for remote code execution by default; if
compatibility requires the flag, replace it with an explicit opt-in parameter or
environment variable defaulting to false, with a pinned model revision and
security documentation.

In `@tools/launcher/tests/test_examples_resolve.py`:
- Around line 79-91: Update the task validation around the inline check in the
examples resolver to reject any non-empty args list when inline is set, matching
run_jobs() behavior. Use the existing task, inline, and args values and include
the task path/name in the assertion message; preserve validation for script
tasks and empty or absent args.

---

Nitpick comments:
In `@docs/source/getting_started/windows/_installation_standalone.rst`:
- Line 16: Update the CUDA dependency references in the Windows standalone
installation document, including the entries around the CUDA Toolkit requirement
and later cuDNN mentions, to use the exact capitalization “cuDNN” consistently.

In `@modelopt/onnx/autocast/precisionconverter.py`:
- Around line 1620-1665: Update _get_tensor_shape and _get_tensor_elem_type to
resolve initializers through self.initializer_map and value metadata through
self.value_info_map instead of scanning graph.initializer, input, output, and
value_info. Preserve the existing shape, element-type, Cast-producer fallback,
and missing-value behavior while relying on the maps rebuilt by
_refresh_gathernd_pre_cast_declarations.

In `@tests/examples/megatron_bridge/test_prune_minitron.py`:
- Around line 43-51: Extend the test case using num_nextn_predict_layers=1 and
mtp_hybrid_override_pattern to inspect both generated output formats: assert the
Megatron configuration or checkpoint metadata has MTP fields removed, and assert
the Hugging Face metadata persists the normalized hybrid pattern. Use each
format’s existing inspection/loading helpers rather than relying only on output
existence or parameter counts.
- Line 79: Add a focused test case in the existing pruning tests around the
score_lower_bound configuration, using a bound higher than the measured score
and the real pruning command implementation. Assert that this failing score-gate
path exits non-zero, while preserving the existing passing-path coverage.

In `@tests/examples/megatron_bridge/test_qad.py`:
- Around line 113-124: Update the test around the distillation command and
tracker assertions to validate duration-based exit independently: add a focused
configuration test or separate integration case with exit_interval unset, use a
controlled exit_duration_in_mins value, and retain assertions confirming the run
exits at the expected iteration/checkpoint. Ensure the existing interval-based
test remains scoped to interval exit rather than claiming duration coverage.

🪄 Autofix

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 29d0031c-cd36-487c-9744-6092bdc07c37

📥 Commits

Reviewing files that changed from the base of the PR and between 33d05b0 and 5bbc2d2.

📒 Files selected for processing (62)

• .agents/skills/deployment/SKILL.md
• .agents/skills/deployment/references/benchmarking.md
• .agents/skills/deployment/references/support-matrix.md
• .agents/skills/evaluation/SKILL.md
• .agents/skills/evaluation/recipes/examples/example_eval.yaml
• .agents/skills/evaluation/recipes/examples/example_eval_next.yaml
• .agents/skills/evaluation/references/nel-next.md
• .agents/skills/qad/SKILL.md
• .claude/skills/qad
• CHANGELOG.rst
• docs/source/getting_started/windows/_installation_standalone.rst
• examples/hf_ptq/example_utils.py
• examples/llm_eval/lm_eval_hf.py
• examples/megatron_bridge/README.md
• examples/megatron_bridge/data/nemotron-cascade-2-blend.yaml
• examples/megatron_bridge/distill.py
• examples/megatron_bridge/prune_minitron.py
• examples/torch_onnx/README.md
• examples/torch_onnx/hf_embedding_quant_to_onnx.py
• examples/torch_onnx/torch_quant_to_onnx.py
• examples/windows/onnx_ptq/genai_llm/README.md
• examples/windows/onnx_ptq/genai_llm/quantize.py
• modelopt/onnx/autocast/convert.py
• modelopt/onnx/autocast/precisionconverter.py
• modelopt/onnx/utils.py
• modelopt/torch/__init__.py
• modelopt/torch/export/quant_aware_conversion.py
• modelopt/torch/export/unified_export_hf.py
• modelopt/torch/prune/plugins/mcore_minitron.py
• modelopt/torch/puzzletron/anymodel/models/__init__.py
• modelopt/torch/quantization/export_onnx.py
• modelopt/torch/quantization/nn/modules/quant_linear.py
• modelopt/torch/utils/dataset_utils.py
• modelopt_recipes/huggingface/nemotron_llama/ptq/fp8_output_quant_proj.yaml
• modelopt_recipes/huggingface/nemotron_llama/ptq/nvfp4_output_quant_proj.yaml
• modelopt_recipes/ptq.md
• noxfile.py
• pyproject.toml
• tests/_test_utils/torch/transformers_models.py
• tests/examples/hf_ptq/test_example_utils.py
• tests/examples/llm_eval/test_llm_eval.py
• tests/examples/megatron_bridge/test_prune_minitron.py
• tests/examples/megatron_bridge/test_qad.py
• tests/examples/torch_onnx/test_hf_embedding_quant_to_onnx.py
• tests/examples/torch_onnx/test_torch_quant_to_onnx.py
• tests/gpu/torch/export/test_export_weight_gpu.py
• tests/gpu/torch/puzzletron/test_puzzletron.py
• tests/gpu_trtllm/torch/export/test_export_compressed_nvfp4.py
• tests/unit/onnx/autocast/test_precisionconverter.py
• tests/unit/torch/export/test_quant_aware_conversion.py
• tests/unit/torch/utils/test_dataset_utils.py
• tools/launcher/core.py
• tools/launcher/docs/configuration.md
• tools/launcher/examples/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/mbridge_prune.yaml
• tools/launcher/examples/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/mbridge_quantize.yaml
• tools/launcher/slurm_config.py
• tools/launcher/tests/test_core.py
• tools/launcher/tests/test_core_extended.py
• tools/launcher/tests/test_docker_execution.py
• tools/launcher/tests/test_examples_resolve.py
• tools/launcher/tests/test_yaml_formats.py
• tools/mcp/pyproject.toml

💤 Files with no reviewable changes (1)

• tests/gpu/torch/puzzletron/test_puzzletron.py

[coderabbitai]

🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

Use the canonical paths for common skills.

Lines 23-24 reference skills/common/.... The canonical skill tree is .agents/skills/common/.... The current paths can prevent the QAD skill from loading the required environment and Slurm instructions.

Proposed fix

-  `skills/common/{environment-setup,workspace-management,slurm-setup}.md`; also
-  `skills/common/remote-execution.md` for remote Slurm
+  `.agents/skills/common/{environment-setup,workspace-management,slurm-setup}.md`; also
+  `.agents/skills/common/remote-execution.md` for remote Slurm

As per path instructions, use relative paths from the repository root and keep .agents/ as the canonical source.

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	- `examples/megatron_bridge/{quantize.py,distill.py}` via `--help`
	- `skills/common/{environment-setup,workspace-management,slurm-setup}.md`; also
	`skills/common/remote-execution.md` for remote Slurm
	- `.agents/skills/common/{environment-setup,workspace-management,slurm-setup}.md`; also
	`.agents/skills/common/remote-execution.md` for remote Slurm

🤖 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 @.agents/skills/qad/SKILL.md around lines 22 - 24, Update the common skill
references in the QAD skill documentation to use the canonical
repository-relative `.agents/skills/common/...` paths, including
`environment-setup.md`, `workspace-management.md`, `slurm-setup.md`, and
`remote-execution.md`; leave the Megatron bridge paths unchanged.

Source: Path instructions

[coderabbitai]

🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

🧩 Analysis chain

🏁 Script executed:

#!/bin/bash
set -euo pipefail

get_file() {
  gh api -X GET "repos/EleutherAI/lm-evaluation-harness/contents/$1?ref=v0.4.10" \
    --jq '.content' | tr -d '\n' | base64 --decode
}

# Verify that v0.4.10 writes group results separately from task results.
get_file lm_eval/evaluator.py | rg -n -C 3 'groups|results|_to_eval_results'

# Verify the documented output-path behavior for the pinned release.
get_file docs/interface.md | rg -n -C 2 'output_path|JSON file'

Repository: NVIDIA/Model-Optimizer

Length of output: 5852

---

🏁 Script executed:

#!/bin/bash
set -euo pipefail

echo '--- CONTRIBUTING coding standards ---'
rg -n -A80 -B5 'coding standards|pre-commit hooks' CONTRIBUTING.md | head -n 180

echo '--- target implementation ---'
cat -n examples/llm_eval/lm_eval_hf.py | sed -n '220,325p'

echo '--- related tests and call sites ---'
rg -n -C4 'accuracy_threshold|results\*\.json|output_path|mmlu' examples tests 2>/dev/null | head -n 240

Repository: NVIDIA/Model-Optimizer

Length of output: 30609

---

🏁 Script executed:

#!/bin/bash
set -euo pipefail

python3 - <<'PY'
import json
from pathlib import Path

path = Path("examples/llm_eval/lm_eval_hf.py")
text = path.read_text()
for needle in ("glob.glob", '["results"]', '["groups"]', "accuracy_threshold", 'tasks="mmlu"', "--output_path"):
    print(f"{needle!r}: {text.count(needle)}")

# Model the exact current lookup against representative lm-eval reports.
reports = [
    {"results": {"mmlu": {"acc,none": 0.5}}},
    {"groups": {"mmlu": {"acc,none": 0.5}}},
    {"results": {"mmlu": {"acc": 0.5}}, "groups": {"mmlu": {"acc": 0.4}}},
]
for report in reports:
    scores = report.get("results", {}).get("mmlu", {})
    try:
        acc = next(
            float(v) for k, v in scores.items()
            if k == "acc" or k.startswith("acc,")
        )
    except StopIteration:
        acc = None
    print(json.dumps(report), "-> current lookup:", acc)
PY

Repository: NVIDIA/Model-Optimizer

Length of output: 490

---

Support JSON output files and group metrics.

When --output_path names a JSON file, line 298 treats it as a directory. For group tasks such as mmlu, the aggregate acc can be stored under groups, but line 304 reads only results. Read the task from results or groups, and support direct JSON file paths.

🧰 Tools

🪛 ast-grep (0.45.0)

[warning] 302-302: File path is request-/variable-derived; validate and normalize to prevent path traversal.
Context: open(max(files, key=os.path.getmtime))
Note: [CWE-22] Improper Limitation of a Pathname to a Restricted Directory ('Path Traversal').

(open-filename-from-request)

🤖 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 `@examples/llm_eval/lm_eval_hf.py` around lines 298 - 308, Update the
result-loading logic around the output_path glob and scores extraction to
support both directory paths and direct JSON file paths. Load the selected JSON
file directly when output_path is a file; otherwise retain the existing
latest-results-file selection. Extract the requested task metrics from either
the JSON's "results" or "groups" mapping, preferring the available task entry
before computing acc.

[coderabbitai]

🎯 Functional Correctness | 🟡 Minor | ⚡ Quick win

Reject non-finite score bounds.

type=float accepts nan, inf, and -inf. nan makes every finite score fail. -inf makes every finite score pass. Validate math.isfinite() at the argument boundary.

Suggested validation

+def _finite_float(value: str) -> float:
+    parsed = float(value)
+    if not math.isfinite(parsed):
+        raise argparse.ArgumentTypeError("must be finite")
+    return parsed
+
...
-        type=float,
+        type=_finite_float,

Based on the new CLI boundary, reject values that cannot represent a valid score threshold.

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	parser.add_argument(
	"--score_lower_bound",
	type=float,
	default=None,
	help="If set, fail the job when the NAS-based pruned model's score is below this bound.",
	)
	def _finite_float(value: str) -> float:
	parsed = float(value)
	if not math.isfinite(parsed):
	raise argparse.ArgumentTypeError("must be finite")
	return parsed
	parser.add_argument(
	"--score_lower_bound",
	type=_finite_float,
	default=None,
	help="If set, fail the job when the NAS-based pruned model's score is below this bound.",
	)

🤖 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 `@examples/megatron_bridge/prune_minitron.py` around lines 242 - 247, Update
the argparse definition for score_lower_bound in prune_minitron.py to reject
non-finite float values at the CLI boundary by validating math.isfinite() after
parsing. Preserve None as the unset value and continue accepting finite
thresholds.

[coderabbitai]

🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

Run the accuracy gate before saving output artifacts.

The gate runs after the Megatron and Hugging Face output paths have written artifacts. If the score is below the bound, sys.exit(1) leaves a model and configuration behind. On a rerun, the earlier output-existence check returns successfully before evaluating the score. A below-bound artifact can therefore bypass the gate.

Move the gate immediately after mtp.prune() and before saving, or remove failed artifacts and make existing outputs re-enter the gate.

Based on the existing output-existence guard in main, the current ordering can turn a failed gate into a successful rerun.

🤖 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 `@examples/megatron_bridge/prune_minitron.py` around lines 687 - 698, Move the
accuracy-gate block in main immediately after mtp.prune() completes and before
either Megatron or Hugging Face output-saving path writes artifacts. Preserve
the existing score-bound evaluation and sys.exit(1) behavior, ensuring failed
pruning runs cannot leave outputs that bypass the gate on rerun.

[coderabbitai]

🎯 Functional Correctness | 🟠 Major | 🏗️ Heavy lift

Keep the recipe format and export mode consistent.

--recipe can replace the loaded quantization format, but Conv2d overrides and TensorRT guards still use --quantize_mode. For example, --quantize_mode=fp8 --recipe=nvfp4 loads NVFP4 settings but skips the NVFP4 high-rank dynamic-quantizer guard. Derive format-specific behavior from validated recipe metadata, or reject incompatible pairs before quantization. Add coverage for both allowed custom recipes and rejected mismatches.

• examples/torch_onnx/torch_quant_to_onnx.py#L125-L136: select Conv2d overrides from the resolved recipe format.
• examples/torch_onnx/torch_quant_to_onnx.py#L458-L468: define the compatibility contract in the CLI help.
• examples/torch_onnx/torch_quant_to_onnx.py#L544-L548: reject incompatible recipe and mode combinations.
• examples/torch_onnx/torch_quant_to_onnx.py#L598-L598: pass a resolved format profile with the recipe configuration.
• examples/torch_onnx/torch_quant_to_onnx.py#L613-L625: use the resolved format profile for TensorRT export guards.
• examples/torch_onnx/README.md#L72-L75: document the enforced or derived recipe/mode relationship.

📍 Affects 2 files

• examples/torch_onnx/torch_quant_to_onnx.py#L125-L136 (this comment)
• examples/torch_onnx/torch_quant_to_onnx.py#L458-L468
• examples/torch_onnx/torch_quant_to_onnx.py#L544-L548
• examples/torch_onnx/torch_quant_to_onnx.py#L598-L598
• examples/torch_onnx/torch_quant_to_onnx.py#L613-L625
• examples/torch_onnx/README.md#L72-L75

🤖 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 `@examples/torch_onnx/torch_quant_to_onnx.py` around lines 125 - 136, Keep the
recipe format and export mode consistent by resolving a validated format profile
once and using it throughout. In examples/torch_onnx/torch_quant_to_onnx.py
lines 125-136, derive Conv2d overrides from the resolved recipe; document the
compatibility contract in lines 458-468, reject incompatible recipe/mode
combinations in lines 544-548, pass the profile at lines 598-598, and use it for
TensorRT guards at lines 613-625. Update examples/torch_onnx/README.md lines
72-75 to document the enforced or derived relationship, and add coverage for
valid custom recipes and rejected mismatches.

[coderabbitai]

🎯 Functional Correctness | 🟡 Minor | ⚡ Quick win

Cover the failing accuracy-gate path.

This test only uses a passing bound. A regression that removes the non-zero exit for a failing score will still pass. Add a focused test with a result JSON below the bound and assert SystemExit(1) from _enforce_accuracy_gate.

As per path instructions, “Tests must exercise the behavior they claim to validate.”

🤖 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/examples/llm_eval/test_llm_eval.py` around lines 39 - 41, Add a focused
test for _enforce_accuracy_gate that provides a results JSON with a score below
accuracy_lower_bound and asserts the function raises SystemExit with code 1.
Keep the existing passing-bound test unchanged so both successful and failing
gate behavior are covered.

Source: Path instructions

[coderabbitai]

📐 Maintainability & Code Quality | 🟠 Major | ⚡ Quick win

🧩 Analysis chain

🏁 Script executed:

#!/bin/bash
set -eu
printf '%s\n' '--- target file ---'
sed -n '1,340p' tests/unit/torch/export/test_quant_aware_conversion.py
printf '%s\n' '--- relevant contribution guidance ---'
rg -n -A8 -B4 'coding standards|imports|optional depend|unit tests|writing and running tests' CONTRIBUTING.md

Repository: NVIDIA/Model-Optimizer

Length of output: 20597

---

🏁 Script executed:

#!/bin/bash
set -eu
python3 - <<'PY'
from pathlib import Path
import ast

path = Path("tests/unit/torch/export/test_quant_aware_conversion.py")
source = path.read_text()
tree = ast.parse(source)
lines = source.splitlines()

print("Optional-import sites and preceding source lines:")
for node in ast.walk(tree):
    if isinstance(node, ast.ImportFrom) and node.module == "transformers.core_model_loading":
        start = max(0, node.lineno - 4)
        print(f"\nimport line {node.lineno}: {lines[node.lineno - 1].strip()}")
        for number in range(start + 1, node.lineno):
            print(f"{number}: {lines[number - 1].strip()}")

print("\nAll pytest.importorskip calls:")
for node in ast.walk(tree):
    if isinstance(node, ast.Call) and isinstance(node.func, ast.Attribute):
        if isinstance(node.func.value, ast.Name) and node.func.value.id == "pytest" and node.func.attr == "importorskip":
            argument = ast.get_source_segment(source, node.args[0]) if node.args else ""
            print(f"line {node.lineno}: pytest.importorskip({argument})")
PY

Repository: NVIDIA/Model-Optimizer

Length of output: 1364

---

🏁 Script executed:

#!/bin/bash
set -eu
printf '%s\n' '--- all optional-import contexts ---'
sed -n '228,305p' tests/unit/torch/export/test_quant_aware_conversion.py
sed -n '348,365p' tests/unit/torch/export/test_quant_aware_conversion.py

Repository: NVIDIA/Model-Optimizer

Length of output: 4591

---

🏁 Script executed:

#!/bin/bash
set -eu
python3 - <<'PY'
from pathlib import Path
import ast

path = Path("tests/unit/torch/export/test_quant_aware_conversion.py")
source = path.read_text()
lines = source.splitlines()
tree = ast.parse(source)

parents = {}
for parent in ast.walk(tree):
    for child in ast.iter_child_nodes(parent):
        parents[id(child)] = parent

print("All imports nested inside test functions:")
for node in ast.walk(tree):
    if isinstance(node, (ast.Import, ast.ImportFrom)):
        parent = parents.get(id(node))
        while parent is not None and not isinstance(parent, (ast.FunctionDef, ast.AsyncFunctionDef)):
            parent = parents.get(id(parent))
        if parent is not None:
            text = ast.get_source_segment(source, node).replace("\n", " ")
            preceding = lines[node.lineno - 2].strip() if node.lineno >= 2 else ""
            print(f"line {node.lineno}, function {parent.name}: {text}")
            print(f"  preceding line: {preceding!r}")
PY

Repository: NVIDIA/Model-Optimizer

Length of output: 1421

---

Document conditional Transformers imports.

Add a brief reason comment before each local Transformers import at lines 180, 183, 237, 270, 298, and 358. Keep the imports conditional because transformers is optional and conversion_mapping is version-dependent.

📍 Affects 1 file

• tests/unit/torch/export/test_quant_aware_conversion.py#L269-L270 (this comment)
• tests/unit/torch/export/test_quant_aware_conversion.py#L297-L298

🤖 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/unit/torch/export/test_quant_aware_conversion.py` around lines 269 -
270, Add brief reason comments before each conditional local Transformers import
at tests/unit/torch/export/test_quant_aware_conversion.py lines 180, 183, 237,
269-270, 297-298, and 358, explaining that transformers is optional and
conversion_mapping is version-dependent. Keep the existing pytest.importorskip
guards and imports unchanged.

Sources: Coding guidelines, Path instructions

[coderabbitai]

🎯 Functional Correctness | 🟡 Minor | ⚡ Quick win

🧩 Analysis chain

🏁 Script executed:

#!/bin/bash
set -euo pipefail

ast-grep outline tools/launcher/core.py --items all
ast-grep outline tools/launcher/slurm_config.py --items all

rg -n -C 3 --glob '*.py' \
  '\b(SandboxTask[0-4]?|SlurmConfig|slurm_factory)\s*\(' .

Repository: NVIDIA/Model-Optimizer

Length of output: 28282

---

🏁 Script executed:

#!/bin/bash
set -euo pipefail

printf '%s\n' '--- CONTRIBUTING coding standards ---'
rg -n -A80 -B5 -- 'coding standards|pre-commit hooks|submitting your code' CONTRIBUTING.md

printf '%s\n' '--- SandboxTask definition ---'
cat -n tools/launcher/core.py | sed -n '105,135p'

printf '%s\n' '--- SlurmConfig and slurm_factory definitions ---'
cat -n tools/launcher/slurm_config.py | sed -n '25,115p'

printf '%s\n' '--- focused diff ---'
git diff -- tools/launcher/core.py tools/launcher/slurm_config.py

Repository: NVIDIA/Model-Optimizer

Length of output: 12348

---

Preserve positional configuration construction.

The new fields change positional argument mapping:

• Append inline, reqs, and reqs_file after the existing SandboxTask fields, or make them keyword-only.
• Append docker_user after the existing SlurmConfig fields and slurm_factory() parameters, or make it keyword-only.

📍 Affects 2 files

• tools/launcher/core.py#L114-L126 (this comment)
• tools/launcher/slurm_config.py#L51-L53
• tools/launcher/slurm_config.py#L77-L99

🤖 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 `@tools/launcher/core.py` around lines 114 - 126, Preserve positional
construction compatibility by moving the new inline, reqs, and reqs_file fields
in SandboxTask after all existing fields, or making them keyword-only. In
tools/launcher/slurm_config.py at lines 51-53 and 77-99, likewise append
docker_user after the existing SlurmConfig fields and slurm_factory parameters,
or make it keyword-only.

[coderabbitai]

🎯 Functional Correctness | 🟡 Minor | ⚡ Quick win

Reject non-empty args for inline tasks.

run_jobs() rejects non-empty args when inline is set. This validator only checks that args is a list. Validate this combination here so invalid example YAML fails before job submission.

Proposed fix

         inline = task.get("inline")
+        args = task.get("args")
         # A task runs either a `script:` wrapper or an `inline:` command — exactly one.
         assert (isinstance(script, str) and script.strip()) or (
             isinstance(inline, str) and inline.strip()
         ), f"{path}:{name}: task needs a `script` or `inline`"
         assert not (script and inline), f"{path}:{name}: set only one of `script`/`inline`"
+        assert not (inline and args), f"{path}:{name}: `args` is only valid with `script`"

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	inline = task.get("inline")
	# A task runs either a `script:` wrapper or an `inline:` command — exactly one.
	assert (isinstance(script, str) and script.strip()) or (
	isinstance(inline, str) and inline.strip()
	), f"{path}:{name}: task needs a `script` or `inline`"
	assert not (script and inline), f"{path}:{name}: set only one of `script`/`inline`"
	# nemo-run's --yaml CLI layer rejects multi-line override values, so an
	# inline command must stay single-line (use a folded scalar `>-`).
	if inline:
	assert "\n" not in inline.strip(), (
	f"{path}:{name}: `inline` must be single-line (use folded `>-`, chain with `&&`)"
	)
	inline = task.get("inline")
	args = task.get("args")
	# A task runs either a `script:` wrapper or an `inline:` command — exactly one.
	assert (isinstance(script, str) and script.strip()) or (
	isinstance(inline, str) and inline.strip()
	), f"{path}:{name}: task needs a `script` or `inline`"
	assert not (script and inline), f"{path}:{name}: set only one of `script`/`inline`"
	assert not (inline and args), f"{path}:{name}: `args` is only valid with `script`"
	# nemo-run's --yaml CLI layer rejects multi-line override values, so an
	# inline command must stay single-line (use a folded scalar `>-`).
	if inline:
	assert "\n" not in inline.strip(), (
	f"{path}:{name}: `inline` must be single-line (use folded `>-`, chain with `&&`)"
	)

🤖 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 `@tools/launcher/tests/test_examples_resolve.py` around lines 79 - 91, Update
the task validation around the inline check in the examples resolver to reject
any non-empty args list when inline is set, matching run_jobs() behavior. Use
the existing task, inline, and args values and include the task path/name in the
assertion message; preserve validation for script tasks and empty or absent
args.