[6078291][OMNIML-3716] Add ViT FP8 + Torch-TRT example, wire softmax_quantizer in _QuantAttention

[ajrasane]

What does this PR do?

Type of change: new feature + bug fix

Adds a Torch-TensorRT deployment path for HuggingFace ViT and closes the
modelopt-side gap that prevented *softmax_quantizer from being applied
on the standard attention forward path.

• New ViT PTQ recipes under modelopt_recipes/huggingface/vit/ptq/:

  • fp8.yaml — W8A8 per-tensor FP8 E4M3 on encoder Linear weights/inputs;
    attention Q/K/V BMMs + softmax output at FP8; per-block LayerNorm output
    at FP8 (one shared Q/DQ feeds Q/K/V + MLP); patch-embed nn.Conv2d,
    classifier, and the final vit.layernorm left FP16. Uses max
    calibration.
  • The recipe is self-contained (no $import of shared snippets) and
    use the "specific-enable" style: narrow parent_class + path scoping
    on every enable rule, so no enable: false carve-outs are needed.

• New example under examples/torch_trt/:

  • torch_tensorrt_ptq.py — single-model pipeline (load HF model,
    calibrate from zh-plus/tiny-imagenet, mtq.quantize,
    torch_tensorrt.compile, verify the compiled-model argmax matches the
    fake-quant argmax). Defaults to google/vit-large-patch16-224; pass
    --model_id and --recipe to target any model + recipe combination.
    --no_pretrained + --model_kwargs shrink the model for fast tests.
  • README.md documenting the flow, the shipped recipes, hardware
    requirements, and CLI usage.
  • requirements.txt.

• Bug fix in modelopt/torch/quantization/plugins/huggingface.py — inside
  _QuantAttention._quantized_attention, the non-kitchen branch now
  temporarily replaces torch.nn.functional.softmax (via the existing
  replace_function context manager) with a wrapper that pipes the softmax
  output through self.softmax_quantizer. Previously the slot was created
  on every registered attention class but only consumed by the optional
  Kitchen MXFP8 flash-attention path, so FP8 / NVFP4 recipes that enabled
  *softmax_quantizer saw it stay uncalibrated (amax=None) and emitted
  no Q/DQ around the softmax output during ONNX / Torch-TRT export. With
  this fix the softmax_quantizer is calibrated alongside the rest of
  the model, and both the modelopt ONNX exporter and torch_tensorrt.compile
  pick up the Q/DQ pair. The patch short-circuits to the unwrapped call
  when the quantizer is disabled (zero-overhead) and has no effect on SDPA
  paths that fuse softmax inside a C++ kernel.

• New e2e integration test at
  tests/examples/torch_trt/test_torch_tensorrt_ptq.py — mirrors the
  torch_onnx test pattern: invokes the example through
  run_example_command, parametrizes over the two precision modes (fp8,
  nvfp4), uses a 1-layer ViT config (--no_pretrained + --model_kwargs)
  so each parametrized case completes in under a minute. importorskip on
  torch_tensorrt so the test is automatically skipped on hosts without
  the package.

Usage

# FP8 (Hopper / Ada) — default model is google/vit-large-patch16-224
python examples/torch_trt/torch_tensorrt_ptq.py \
    --precision fp8 \
    --calib_samples 128 \
    --batch_size 1

# Custom model + custom recipe
python examples/torch_trt/torch_tensorrt_ptq.py \
    --model_id <huggingface/model-id> \
    --recipe <recipe-path-relative-to-modelopt_recipes-or-absolute-yaml>

Testing

• Recipes load via modelopt.recipe.load_recipe() and pass
  QuantizeConfig schema validation.

• Run pytest tests/examples/torch_trt/test_torch_tensorrt_ptq.py →
  1 parametrized case passes on RTX 6000 Ada (fp8).

• End-to-end on google/vit-base-patch16-224: mtq.quantize with the new
  FP8 recipe followed by torch_tensorrt.compile(ir="dynamo") produces a
  TRT engine whose argmax matches the FP16 baseline.

• ONNX exported from the torch path now contains Q/DQ on 12 / 12
  softmax outputs (was 0 / 12 before this PR's _QuantAttention fix),
  matching the ONNX-CLI output's quantization layout.

  Both FP8 paths land within 0.13 pp Top-1 of the FP16 baseline; Top-5 is
  within 0.02 pp across all three.

• ImageNet-1k validation accuracy via the new torch_tensorrt_accuracy.py
  (full 50000 samples, batch=1, every model Torch-TensorRT-compiled —
  including the baseline — so the comparison is apples-to-apples) for the
  example's default google/vit-large-patch16-224:

  Model (Torch-TRT)	Top-1	Top-5	Δ Top-1 vs baseline
  Baseline (FP16)	81.99%	96.01%	—
  FP8	82.01%	96.05%	+0.02 pp

  FP8 is within noise of the FP16 TRT baseline and NVFP4 W4A4 costs only
  −0.13 pp Top-1 / −0.05 pp Top-5. Absolute Top-1 sits below the model card's
  ~85.5% because evaluation uses the HF AutoImageProcessor default
  preprocessing (direct 224×224 resize, no resize-then-center-crop), applied
  identically to all three models — so the deltas are the comparison signal.

Before your PR is "Ready for review"

Make sure you read and follow Contributor guidelines and your commits are signed (git commit -s -S).

Make sure you read and follow the Security Best Practices (e.g. avoiding hardcoded trust_remote_code=True, torch.load(..., weights_only=False), pickle, etc.).

• Is this change backward compatible?: ✅
• If you copied code from any other sources or added a new PIP dependency, did you follow guidance in CONTRIBUTING.md: N/A
• Did you write any new necessary tests?: ✅ — new e2e integration test under tests/examples/torch_trt/.
• Did you update Changelog?: ✅

Summary by CodeRabbit

• New Features

  • Added Torch‑TensorRT FP8/NVFP4 deployment examples and end‑to‑end scripts for HuggingFace ViT, plus ViT-specific PTQ recipes and ImageNet-1k vs FP16 accuracy reporting.

• Bug Fixes

  • Fixed softmax quantization and export/compilation edge cases (softmax calibration during export, IO casting for empty tensors, routed expert weight syncing, importer key handling).

• Documentation

  • Added comprehensive example README with setup, usage, recipes, evaluation, and hardware guidance.

• Requirements

  • Pinned minimum versions for example dependencies.

• Tests

  • Added tests validating the Torch‑TensorRT quantization examples for fp8.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[coderabbitai]

Note

Reviews paused

It looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the reviews.auto_review.auto_pause_after_reviewed_commits setting.

Use the following commands to manage reviews:

• @coderabbitai resume to resume automatic reviews.
• @coderabbitai review to trigger a single review.

Use the checkboxes below for quick actions:

• ▶️ Resume reviews
• 🔍 Trigger review

📝 Walkthrough

Walkthrough

Adds an end-to-end ModelOpt -> Torch-TensorRT PTQ deployment example for HuggingFace ViT (FP8/NVFP4), two PTQ recipes, an accuracy evaluation script, documentation and requirements, an integration test, CI/ownership updates, and a softmax quantization fix for HuggingFace attention.

Changes

ViT PTQ Torch-TensorRT Deployment

Layer / File(s)	Summary
Softmax quantization fix for HuggingFace attention modelopt/torch/quantization/plugins/huggingface.py	Adds conditional softmax quantization in the non-kitchen attention path by monkey-patching torch.nn.functional.softmax when self.softmax_quantizer.is_enabled, and limits KitchenFlashAttention to enabled MXFP8(8) softmax quantizer.
PTQ recipes for ViT FP8 and NVFP4 modelopt_recipes/huggingface/vit/ptq/fp8.yaml, modelopt_recipes/huggingface/vit/ptq/nvfp4.yaml	Two YAML recipe files: FP8 (metadata.recipe_type: ptq, quantize.algorithm: max) and NVFP4 (metadata.recipe_type: ptq, quantize.algorithm: awq_lite) with imported unit configs and explicit module-level FP8 overrides.
Torch-TensorRT PTQ example orchestration examples/torch_trt/torch_tensorrt_ptq.py	Main example script with helpers for model & processor loading, calibration loader from tiny-imagenet, recipe validation and ModelOpt PTQ execution, ViT logits wrapper, Torch-TensorRT Dynamo compilation with execution workarounds, and argmax-based comparison across baseline, fake-quant, and compiled models.
Example documentation and dependencies examples/torch_trt/requirements.txt, examples/torch_trt/README.md	Requirements pin minimum versions for datasets, torch-tensorrt, and transformers; README documents setup, CLI usage, workflow steps, shipped ViT recipes, hardware requirements, checkpoint resume, and custom recipe guidance.
Torch-TensorRT accuracy evaluation example examples/torch_trt/torch_tensorrt_accuracy.py	Evaluation script that adapts inputs/outputs for the onnx_ptq evaluation harness, supports baseline/unquantized scoring, quantizes with chosen recipe, optionally compiles for TRT (static batch), runs ImageNet top-1/top-5 evaluation, and optionally writes CSV results.
Integration test for PTQ example tests/examples/torch_trt/test_torch_tensorrt_ptq.py	Parametrized pytest over fp8 and nvfp4 that runs the example with a tiny ViT config, disables pretrained weights, and verifies the torch_trt backend executes.
CI matrix and CODEOWNERS .github/workflows/example_tests.yml, .github/CODEOWNERS	Adds torch_trt to the ONNX/TensorRT workflow matrix and assigns /examples/torch_trt to the modelopt-onnx CODEOWNERS.
Changelog entries CHANGELOG.rst	Documents the new Torch-TensorRT ViT PTQ example and multiple bug fixes including the attention softmax quantization fix.

---

🎯 3 (Moderate) | ⏱️ ~25 minutes

Suggested reviewers:

• ChenhanYu
• cjluo-nv
• meenchen
• realAsma
• jenchen13

🚥 Pre-merge checks | ✅ 5 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 38.89% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (5 passed)

Check name	Status	Explanation
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 found. All new dependencies use permissive licenses (Apache 2.0/BSD). No hardcoded trust_remote_code=True, unsafe eval/exec, pickle, or # nosec comments detected.
Title check	✅ Passed	The title is specific and matches the main changes: a ViT Torch-TRT example plus the _QuantAttention softmax quantizer fix.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch ajrasane/nvbug_6078291-vit-torch-trt-fp8

---

_{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.
2026-07-01 20:29 UTC

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 76.98%. Comparing base (1b03381) to head (3b0a302).
⚠️ Report is 2 commits behind head on main.

Additional details and impacted files

@@             Coverage Diff             @@
##             main    #1569       +/-   ##
===========================================
+ Coverage   61.17%   76.98%   +15.81%     
===========================================
  Files         515      515               
  Lines       57207    57216        +9     
===========================================
+ Hits        34994    44046     +9052     
+ Misses      22213    13170     -9043     

Flag	Coverage Δ
examples	42.95% <100.00%> (+10.11%)	⬆️
gpu	57.83% <100.00%> (+37.29%)	⬆️
regression	14.83% <11.11%> (+0.06%)	⬆️
unit	54.90% <11.11%> (-0.01%)	⬇️

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]

🧹 Nitpick comments (1)

examples/torch_trt/torch_tensorrt_ptq.py (1)

277-294: ⚡ Quick win

Argmax "match" results are informational only — the example never fails on a mismatch.

fq_match/trt_match are computed and printed but never used to exit non-zero. The e2e test docstring (test_torch_tensorrt_ptq.py, Lines 44-46) claims the CLI "exits non-zero ... if the compiled-model argmax doesn't match the fake-quant argmax", but as written the test only validates that the pipeline runs to completion. Either enforce the check here or correct the test docstring.

Note: also worth deciding intent — the comparison is against baseline_pred (BF16), while the docstring/README phrase it as matching the fake-quant argmax. If you do enforce, be cautious: a tiny --no_pretrained model under NVFP4 can legitimately flip argmax on random input, so a hard gate may be flaky for that path.

♻️ One option: gate the run on mismatch

     trt_match = (trt_pred == baseline_pred).all().item()
     print(f"TRT argmax class: {trt_pred.tolist()} (matches baseline: {trt_match})")
+    if not trt_match:
+        raise SystemExit("Torch-TensorRT argmax did not match the BF16 baseline.")

🤖 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_trt/torch_tensorrt_ptq.py` around lines 277 - 294, The script
currently computes fq_match and trt_match but doesn't fail on mismatch; update
the run to enforce non-zero exit when mismatches occur: after computing
fq_match/trt_match, if fq_match is False or trt_match is False, log an error
with context (include fq_pred, trt_pred, baseline_pred) and call sys.exit(1).
Locate the checks around fq_pred/baseline_pred and trt_pred (symbols: fq_pred,
fq_match, trt_pred, trt_match, baseline_pred, ViTLogitsWrapper,
compile_with_torch_tensorrt) and add the exit-on-mismatch logic (or, if you
prefer the other approach, instead update the test/docstring to accurately state
that mismatches are informational rather than failing).

🤖 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 `@examples/torch_trt/torch_tensorrt_ptq.py`:
- Around line 277-294: The script currently computes fq_match and trt_match but
doesn't fail on mismatch; update the run to enforce non-zero exit when
mismatches occur: after computing fq_match/trt_match, if fq_match is False or
trt_match is False, log an error with context (include fq_pred, trt_pred,
baseline_pred) and call sys.exit(1). Locate the checks around
fq_pred/baseline_pred and trt_pred (symbols: fq_pred, fq_match, trt_pred,
trt_match, baseline_pred, ViTLogitsWrapper, compile_with_torch_tensorrt) and add
the exit-on-mismatch logic (or, if you prefer the other approach, instead update
the test/docstring to accurately state that mismatches are informational rather
than failing).

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: aea38a8d-a5f8-44cf-8b02-9e7a700116ba

📥 Commits

Reviewing files that changed from the base of the PR and between ed0a4b1 and 7db9fad.

📒 Files selected for processing (8)

• CHANGELOG.rst
• examples/torch_trt/README.md
• examples/torch_trt/requirements.txt
• examples/torch_trt/torch_tensorrt_ptq.py
• modelopt/torch/quantization/plugins/huggingface.py
• modelopt_recipes/huggingface/vit/ptq/fp8.yaml
• modelopt_recipes/huggingface/vit/ptq/nvfp4.yaml
• tests/examples/torch_trt/test_torch_tensorrt_ptq.py

[cjluo-nv]

Bot review — DM the bot to share feedback.

Bug-fix portion has a regression risk on hosts that have the optional kitchen package installed: _init_kitchen_attn_fn runs before the new non-kitchen FP8 softmax path and raises NotImplementedError for any non-MXFP8 softmax quantizer. The new ViT FP8/NVFP4 recipes both enable plain E4M3 FP8 on *softmax_quantizer, so on a kitchen-equipped GPU they will fail before the new patched-softmax code path is reached. The new test uses importorskip("torch_tensorrt") but does not skip when kitchen is also installed, so this gap isn't caught in CI. Otherwise the design is reasonable (this is "new example + tuned recipes + 11-line plugin fix" rather than a new subsystem; the deterministic complexity gate fired only because of directory count) and the recipes/example look correct.

[kevalmorabia97]

Suggested change

	pip install -U "nvidia-modelopt[torch]"
	pip install -U "nvidia-modelopt"

[cjluo-nv]

Bot review — DM the bot to share feedback.

Re-review — several previously-flagged issues still unresolved in the latest diff.

Critical, still present:

• Kitchen-vs-FP8 ordering bug in _QuantAttention._quantized_attention (cjluo-nv's earlier review). On hosts with the kitchen package installed, _init_kitchen_attn_fn runs before the new non-kitchen softmax path and raises NotImplementedError whenever softmax_quantizer.is_enabled is True and the quantizer is not MXFP8. Both new ViT recipes enable plain E4M3 FP8 on *softmax_quantizer, so calibrating either recipe on a kitchen-equipped GPU will fail at mtq.quantize time before the new wrapper-softmax branch is reachable. The new branch's if not self.use_kitchen guard is dead code in exactly the configuration this PR ships. Fix in _init_kitchen_attn_fn (don't set use_kitchen=True and don't raise for non-MXFP8 — just fall through to the new wrapper path), or skip kitchen init entirely when the quantizer isn't MXFP8.
• CI gap reinforces the above: the new test only importorskips torch_tensorrt, not kitchen, so the regression mode isn't exercised in isolation either. Either fix the underlying ordering bug or add a if kitchen is not None: pytest.skip(...) guard.
• tests/examples/torch_trt/ is not wired into .github/workflows/example_tests.yml (kevalmorabia97). I confirmed the workflow file still has the matrix [llm_distill, llm_qat, llm_sparsity, diffusers_sparsity] with no torch_trt entry, so the new e2e test won't run in CI even on PRs that touch it. The "tested in CI" claim in the PR body is currently not true on main.

Minor, still unaddressed (verified against current diff):

• examples/torch_trt/README.md:19 still has pip install -U "nvidia-modelopt[torch]"; reviewer asked for "nvidia-modelopt".
• examples/torch_trt/requirements.txt:3 still pins transformers>=4.40; reviewer asked for >=4.56.
• examples/torch_trt/ is missing from .github/CODEOWNERS — confirmed not added in this PR.
• The "argmax must match" claim in the test docstring (lines 44-46) is still informational only — fq_match / trt_match in torch_tensorrt_ptq.py (lines 277-294) are printed but never gate exit status, so the test passes whenever the pipeline runs to completion regardless of correctness. Either enforce the check (carefully — small --no_pretrained NVFP4 runs can flip argmax legitimately) or correct the docstring.

Design note: The deterministic complexity gate fired on directory count (5 dirs), but cjluo-nv correctly characterized this as "new example + tuned recipes + 11-line plugin fix" rather than a new subsystem. Recipes reuse the existing YAML recipe loader (modelopt.recipe.load_recipe / ModelOptPTQRecipe) and don't introduce a new abstraction. Design is fine; the blockers above are correctness/CI-coverage, not architecture.

---

Additional comments (outside the PR diff):

• .github/workflows/example_tests.yml:38 — > Bot comment.

Still missing torch_trt from the example matrix (kevalmorabia97 raised this previously). Without an entry here, tests/examples/torch_trt/test_torch_tensorrt_ptq.py is not run in CI on PRs that touch examples/torch_trt/ or modelopt/torch/quantization/plugins/huggingface.py, so the PR-body claim of 2 parametrized cases passing is hand-verified-only. Add torch_trt to the appropriate matrix (looks like the onnx group, since both run under nvcr.io/nvidia/tensorrt:26.02-py3).

• .github/CODEOWNERS:49 — > Bot comment.

Still missing /examples/torch_trt (kevalmorabia97). Add an entry alongside the other example owners so PRs touching this directory get the right reviewers.

[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: 1

🤖 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 `@examples/torch_trt/torch_tensorrt_accuracy.py`:
- Around line 145-148: The help text for the "--baseline" CLI flag says the
baseline is BF16 but the code sets dtype = torch.float16; update them to match:
either change the argparse help for "--baseline" to say "FP16" if you intend
dtype=torch.float16, or change the dtype selection logic so that when the
baseline flag is used it sets dtype = torch.bfloat16; ensure the variable name
dtype and the "--baseline" flag (and any reporting code that prints the baseline
precision) are consistent.

🪄 Autofix (Beta)

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: c5b7ad31-84f2-45e3-bf4e-31e4d002ac8d

📥 Commits

Reviewing files that changed from the base of the PR and between 33366a8 and bd6da9c.

📒 Files selected for processing (3)

• examples/torch_trt/torch_tensorrt_accuracy.py
• modelopt_recipes/huggingface/vit/ptq/fp8.yaml
• modelopt_recipes/huggingface/vit/ptq/nvfp4.yaml

[cjluo-nv]

Bot review — DM the bot to share feedback.

Re-review: all previously-flagged critical issues are resolved.

• Kitchen-vs-FP8 ordering bug fixed in _init_kitchen_attn_fn: the MXFP8 check is now folded into the early-return guard, so non-MXFP8 (e.g. plain FP8 E4M3) softmax quantizers fall through with use_kitchen=False and the new wrapper-softmax path in _quantized_attention actually runs. The new if not self.use_kitchen branch is now reachable on kitchen-equipped hosts, which is exactly the regression mode the previous review flagged.
• CI wiring: torch_trt is now in the onnx matrix in .github/workflows/example_tests.yml, and /examples/torch_trt is in .github/CODEOWNERS under the onnx codeowners team.
• README/requirements nits: install line is now pip install -U "nvidia-modelopt" and transformers>=4.56.
• Test docstring: now accurately describes the argmax comparison as informational (a tiny random-weight ViT under NVFP4 can legitimately flip argmax), so the test–behavior contract is consistent.
• torch_tensorrt_accuracy.py baseline label: the BF16/FP16 mismatch is gone — --baseline help text is generic, and the printed tag derives from str(dtype) so it always matches.

Design is appropriate: this is "new example + tuned recipes + 11-line plugin fix" reusing the existing recipe loader (modelopt.recipe.load_recipe / ModelOptPTQRecipe), not a new subsystem. The deterministic complexity gate fired only on directory count.

Complex PR: spans 7 directories (≥ 5). Looping in a human for approval.

[cjluo-nv]

Bot review — DM the bot to share feedback.

Re-review: all previously-flagged critical issues are resolved in the latest diff.

• 💬 Kitchen-vs-FP8 ordering bug — fixed in _init_kitchen_attn_fn (modelopt/torch/quantization/plugins/huggingface.py:85-89): the early-return guard now folds in not self.softmax_quantizer.is_mxfp(8), so plain FP8 E4M3 (the configuration both new ViT recipes ship) falls through with use_kitchen=False and the new wrapper-softmax path in _quantized_attention actually runs on kitchen-equipped hosts. The dead-code concern from the prior review no longer applies.
• 💬 CI wiring — torch_trt is now in the onnx matrix in .github/workflows/example_tests.yml:101, and /examples/torch_trt is in .github/CODEOWNERS:56 under the onnx codeowners team. PR-body claim of "tested in CI" is now true.
• 💬 README install line + transformers pin updated to pip install -U "nvidia-modelopt" and transformers>=4.56.
• 💬 Test docstring (tests/examples/torch_trt/test_torch_tensorrt_ptq.py:43-46) now correctly states the baseline/fake-quant/TRT argmax comparison is informational — consistent with the script not gating on it.
• 💬 torch_tensorrt_accuracy.py baseline label now derives from str(dtype), so the BF16/FP16 mismatch is gone.

Design is appropriate: this is "new example + tuned recipes + 11-line plugin fix" reusing the existing recipe loader (modelopt.recipe.load_recipe / ModelOptPTQRecipe), not a new subsystem. The deterministic complexity gate fired only on directory count.

Nudging for a human reviewer to sign off — the PR spans 7 directories and the design-review gate fired, so per the re-review guidance the architectural-scope check warrants a human.

[kevalmorabia97]

changelog entries would go to 0.46 section