examples/pruning/cifar_resnet.ipynb (1)

25-33: Make the install cell idempotent and dev‑env friendly.

Avoid reinstalling over editable installs; gate on missing packages.

Apply:

-! pip install nvidia-modelopt torchvision
+import importlib.util, subprocess, sys
+missing = [m for m in ("modelopt", "torchvision") if importlib.util.find_spec(m) is None]
+if missing:
+    subprocess.check_call([sys.executable, "-m", "pip", "install", "-q", "nvidia-modelopt", "torchvision"])
+else:
+    print("Dependencies already present; skipping install")

README.md (1)

69-73: Provide HTTPS clone as default and keep SSH as optional.

SSH requires keys; HTTPS works for most users.

- git clone [email protected]:NVIDIA/TensorRT-Model-Optimizer.git
+ # HTTPS (recommended for most users)
+ git clone https://github.com/NVIDIA/TensorRT-Model-Optimizer.git
+ # Or SSH (requires GitHub SSH keys)
+ # git clone [email protected]:NVIDIA/TensorRT-Model-Optimizer.git

examples/onnx_ptq/docker/Dockerfile (2)

8-12: Drop SSH host scan unless you actually fetch via SSH.

It adds network dependency and build variability.

-    && mkdir -p -m 0600 ~/.ssh \
-    && ssh-keyscan github.com >> ~/.ssh/known_hosts
+    # SSH known_hosts not needed since we don't fetch via SSH

---

28-32: Tighten pip steps and validate env — TensorRT Python not present

Running import tensorrt raised ModuleNotFoundError — the Python TensorRT package isn't installed in the current environment, so duplication with the base image was not observed.

File: examples/onnx_ptq/docker/Dockerfile Lines: 28-32

-RUN pip install -r TensorRT-Model-Optimizer/examples/onnx_ptq/requirements.txt
+RUN pip install --no-cache-dir -r TensorRT-Model-Optimizer/examples/onnx_ptq/requirements.txt && pip check
@@
-RUN pip install -e "./TensorRT-Model-Optimizer[hf,onnx]"
+RUN pip install --no-cache-dir -e "./TensorRT-Model-Optimizer[hf,onnx]" && pip check

docs/source/getting_started/_installation_for_Linux.rst (1)

40-41: Minor grammar/clarity.

"examples's" → "examples'"; also link to a concrete file for discoverability.

-    dependencies pre-installed. You may need to install additional dependencies from the examples's `requirements.txt` file.
+    dependencies pre-installed. You may need to install additional dependencies from the examples' `requirements.txt` file (e.g., ``examples/llm_ptq/requirements.txt``).

.github/workflows/example_tests.yml (3)

63-65: Runner label is custom; configure actionlint or add a comment.

actionlint warns on unknown label. If you lint this repo, add the label to actionlint.yaml or gate actionlint for this workflow.

---

75-81: Install order OK; consider caching to speed up.

You already use a proxy cache. Optionally add “pip install -U pip wheel” first and “--upgrade” to ensure fresh extras.

-          pip install ".[all]"
+          python -m pip install -U pip wheel
+          pip install -U ".[all]"

---

82-91: YAML anchors trip actionlint; inline container/steps for non‑PR job.

GitHub Actions supports anchors, but actionlint’s diagnostics indicate it flags alias usage for mapping nodes here. To keep lint green, duplicate the mapping and steps.

   example-tests-non-pr:
     if: ${{ !startsWith(github.ref, 'refs/heads/pull-request/') }}
     # Runner list at https://github.com/nv-gha-runners/enterprise-runner-configuration/blob/main/docs/runner-groups.md
     runs-on: linux-amd64-gpu-h100-latest-1
     timeout-minutes: 90
     strategy:
       matrix:
         EXAMPLE: [llm_ptq]
-    container: *example_container
-    steps: *example_steps
+    container:
+      image: nvcr.io/nvidia/tensorrt-llm/release:1.1.0rc2.post2
+      env:
+        LD_LIBRARY_PATH: "/usr/lib/x86_64-linux-gnu:/usr/local/tensorrt/targets/x86_64-linux-gnu/lib:${LD_LIBRARY_PATH}"
+        PATH: "/usr/local/tensorrt/targets/x86_64-linux-gnu/bin:${PATH}"
+        PIP_CONSTRAINT: ""
+    steps:
+      - uses: actions/checkout@v4
+      - uses: nv-gha-runners/setup-proxy-cache@main
+      - name: Run example tests
+        run: |
+          python -m pip install -U pip wheel
+          pip install -U ".[all]"
+          if [ -f examples/$EXAMPLE/requirements.txt ]; then pip install -r examples/$EXAMPLE/requirements.txt; fi
+          pytest -s tests/examples/$EXAMPLE

Configuration used: CodeRabbit UI

Review profile: CHILL

Plan: Pro

docker/Dockerfile (2)

21-27: Don’t preinstall example deps in the base image; it contradicts PR goals.

This bakes llm_ptq deps globally and can clash with user torch/torchvision installs. Install at runtime in workflows/examples only.

 RUN pip install -e "./TensorRT-Model-Optimizer[all,dev-test]"
 RUN rm -rf TensorRT-Model-Optimizer/.git
 RUN python -c "import modelopt.torch.quantization.extensions as ext; ext.precompile()"
-RUN pip install -r TensorRT-Model-Optimizer/examples/llm_ptq/requirements.txt

---

7-7: Fix invalid TORCH_CUDA_ARCH_LIST entry (“12.0+PTX”) and stray “10.0”.

PyTorch expects SM versions; “12.0+PTX” is not a valid SM, and adding non‑supported SMs slows/ breaks builds. Keep only supported SMs and add +PTX to the highest.

-    TORCH_CUDA_ARCH_LIST="8.0 8.6 8.7 8.9 9.0 10.0 12.0+PTX"
+    TORCH_CUDA_ARCH_LIST="8.0 8.6 8.7 8.9 9.0+PTX"

If you truly target SM 10.0, confirm toolchain support and use 10.0+PTX as the highest instead of 12.0+PTX.

.github/workflows/code_quality.yml (1)

18-18: Timeout bump looks fine; consider pip caching to keep it under 30 mins.

Add setup-python’s pip cache to speed up tox/pre-commit runs and reduce flakiness.

 - uses: actions/setup-python@v5
   with:
     python-version: "3.12"
+    cache: 'pip'
+    cache-dependency-path: |
+      requirements*.txt
+      pyproject.toml
+      tox.ini

docker/Dockerfile (1)

3-6: Optional: enable pip no‑cache to reduce layer size.

You set PIP_NO_CACHE_DIR=off. Consider enabling for image builds.

-    PIP_NO_CACHE_DIR=off \
+    PIP_NO_CACHE_DIR=on \

.github/workflows/example_tests.yml (1)

71-74: Consider exporting PATH to TRT bin if tests need CLI tools.

If tests call trtexec or other TRT binaries, uncomment PATH; otherwise this is fine.

Configuration used: CodeRabbit UI

Review profile: CHILL

Plan: Pro

docker/Dockerfile (3)

7-7: Fix TORCH_CUDA_ARCH_LIST; "12.0+PTX" is not a valid SM and only the highest SM should have +PTX.
Use SMs (8.0/8.6/8.7/8.9/9.0). Add 10.0 only if your toolchain supports it.

-    TORCH_CUDA_ARCH_LIST="8.0 8.6 8.7 8.9 9.0 10.0 12.0+PTX"
+    TORCH_CUDA_ARCH_LIST="8.0 8.6 8.7 8.9 9.0+PTX"
+# If SM 10.0 is supported end-to-end, use:
+# TORCH_CUDA_ARCH_LIST="8.0 8.6 8.7 8.9 9.0 10.0+PTX"

---

27-27: Remove baked llm_ptq requirements; install at runtime in workflows/examples.
This contradicts “install example deps at runtime” and can clash with user torch builds.

-RUN pip install -r TensorRT-Model-Optimizer/examples/llm_ptq/requirements.txt
+# Example-specific dependencies are installed at runtime in CI or example entrypoints.

---

29-30: Avoid chmod 777; switch to non-root user with proper ownership.
777 is a security smell and breaks umask expectations.

-# Allow users to run without root
-RUN chmod -R 777 /workspace
+# Allow users to run without root (least privilege)
+RUN useradd -m -u 1000 appuser && chown -R appuser:appuser /workspace
+USER appuser

.github/workflows/example_tests.yml (1)

82-91: Inline anchors for container/steps in non‑PR job; actionlint flags alias/syntax errors.
Replace YAML aliases with concrete blocks.

   example-tests-non-pr:
     if: ${{ !startsWith(github.ref, 'refs/heads/pull-request/') }}
     # Runner list at https://github.com/nv-gha-runners/enterprise-runner-configuration/blob/main/docs/runner-groups.md
     runs-on: linux-amd64-gpu-h100-latest-1
     timeout-minutes: 90
     strategy:
       matrix:
         EXAMPLE: [llm_ptq]
-    container: *example_container
-    steps: *example_steps
+    container:
+      image: nvcr.io/nvidia/tensorrt-llm/release:1.1.0rc2.post2
+      env:
+        LD_LIBRARY_PATH: "/usr/lib/x86_64-linux-gnu:/usr/local/tensorrt/targets/x86_64-linux-gnu/lib"
+        PIP_CONSTRAINT: ""
+    steps:
+      - uses: actions/checkout@v4
+      - uses: nv-gha-runners/setup-proxy-cache@main
+      - name: Run example tests
+        run: |
+          pip install ".[all,dev-test]"
+          find examples/$EXAMPLE -name "requirements.txt" | while read req_file; do pip install -r "$req_file" || exit 1; done
+          pytest -s tests/examples/$EXAMPLE

.github/workflows/pages.yml (1)

26-26: Timeout bump acknowledged; add pip cache to reduce need for longer timeouts.
Caching pip deps will speed up tox builds and may let you keep the old timeout.

-      - uses: actions/setup-python@v5
-        with:
-          python-version: "3.12"
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+          cache: pip
+          cache-dependency-path: tox.ini

.github/workflows/example_tests.yml (1)

71-74: Env var expansion won’t happen here; LD_LIBRARY_PATH will literally contain “${LD_LIBRARY_PATH}”.
Set the full value explicitly or export in a run step.

-        LD_LIBRARY_PATH: "/usr/lib/x86_64-linux-gnu:/usr/local/tensorrt/targets/x86_64-linux-gnu/lib:${LD_LIBRARY_PATH}"
+        LD_LIBRARY_PATH: "/usr/lib/x86_64-linux-gnu:/usr/local/tensorrt/targets/x86_64-linux-gnu/lib"

Configuration used: CodeRabbit UI

Review profile: CHILL

Plan: Pro

.github/workflows/example_tests.yml (1)

82-91: Inline container/steps for non‑PR job to satisfy actionlint (anchors here trip parser).

actionlint flags “container alias mapping expected” and “steps must be sequence” for the aliases. Duplicate the blocks inline for example-tests-non-pr.

   example-tests-non-pr:
     if: ${{ !startsWith(github.ref, 'refs/heads/pull-request/') }}
     # Runner list at https://github.com/nv-gha-runners/enterprise-runner-configuration/blob/main/docs/runner-groups.md
-    runs-on: linux-amd64-gpu-h100-latest-1
+    runs-on: linux-amd64-gpu-h100-latest-1
     timeout-minutes: 120
     strategy:
       matrix:
         EXAMPLE: [llm_ptq]
-    container: *example_container
-    steps: *example_steps
+    container:
+      image: nvcr.io/nvidia/tensorrt-llm/release:1.1.0rc2.post2
+      env:
+        LD_LIBRARY_PATH: "/usr/lib/x86_64-linux-gnu:/usr/local/tensorrt/targets/x86_64-linux-gnu/lib:${LD_LIBRARY_PATH}"
+        # PATH: "/usr/local/tensorrt/targets/x86_64-linux-gnu/bin:${PATH}"
+        PIP_CONSTRAINT: ""
+    steps:
+      - uses: actions/checkout@v4
+      - uses: nv-gha-runners/setup-proxy-cache@main
+      - name: Run example tests
+        run: |
+          pip install ".[all,dev-test]"
+          find examples/${{ matrix.EXAMPLE }} -name "requirements.txt" | while read req_file; do pip install -r "$req_file" || exit 1; done
+          pytest -s tests/examples/${{ matrix.EXAMPLE }}

.github/workflows/example_tests.yml (4)

63-63: actionlint “unknown runner label”: add linter config or include self‑hosted labels.

Either:

• Add an actionlint config whitelisting the custom label; or
• Use a label array including self-hosted + linux to be explicit.

Option A (add config file at .github/actionlint.yaml):

runnerLabels:
  - self-hosted
  - linux
  - linux-amd64-gpu-h100-latest-1

Option B (explicit labels):

-    runs-on: linux-amd64-gpu-h100-latest-1
+    runs-on: [self-hosted, linux, linux-amd64-gpu-h100-latest-1]

Run actionlint with the config to confirm it’s clean.

Also applies to: 85-85

---

69-74: LD_LIBRARY_PATH/PATH expansion in container.env won’t interpolate shell vars.

In the container env block, ${LD_LIBRARY_PATH} and ${PATH} are treated as literals, not expanded. Either drop the suffix or export/merge in a step.

Minimal safe fix: set in a step before installs so you preserve existing values:

     container: &example_container
       image: nvcr.io/nvidia/tensorrt-llm/release:1.1.0rc2.post2
-      env:
-        LD_LIBRARY_PATH: "/usr/lib/x86_64-linux-gnu:/usr/local/tensorrt/targets/x86_64-linux-gnu/lib:${LD_LIBRARY_PATH}"
-        # PATH: "/usr/local/tensorrt/targets/x86_64-linux-gnu/bin:${PATH}"
-        PIP_CONSTRAINT: "" # Disable pip constraint for upgrading packages
+      env:
+        PIP_CONSTRAINT: "" # Disable pip constraint for upgrading packages
@@
     steps: &example_steps
       - uses: actions/checkout@v4
       - uses: nv-gha-runners/setup-proxy-cache@main
+      - name: Merge env paths
+        shell: bash
+        run: |
+          echo "LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu:/usr/local/tensorrt/targets/x86_64-linux-gnu/lib:${LD_LIBRARY_PATH}" >> "$GITHUB_ENV"
+          echo "PATH=/usr/local/tensorrt/targets/x86_64-linux-gnu/bin:${PATH}" >> "$GITHUB_ENV"

Apply the same change in the non‑PR job if you inline steps there.

---

36-48: Broaden change detection to catch packaging/test dependency edits.

Include pyproject.toml and example requirements variants so PR gating doesn’t miss impactful changes.

           files: |
             .github/workflows/example_tests.yml
             examples/llm_ptq/**
             modelopt/torch/**
             tests/examples/llm_ptq/**
             setup.py
+            pyproject.toml
+            examples/**/requirements*.txt

---

75-81: Harden steps: pin external actions, upgrade pip toolchain, fail fast.

• Pin nv-gha-runners action to a commit SHA (avoid @main drift).
• Upgrade pip/setuptools/wheel before installing.
• Prefer python -m pip and set -euo pipefail.

       - uses: actions/checkout@v4
-      - uses: nv-gha-runners/setup-proxy-cache@main
+      - uses: nv-gha-runners/setup-proxy-cache@<commit-sha>
       - name: Run example tests
-        run: |
-          pip install ".[all,dev-test]"
-          find examples/${{ matrix.EXAMPLE }} -name "requirements.txt" | while read req_file; do pip install -r "$req_file" || exit 1; done
-          pytest -s tests/examples/${{ matrix.EXAMPLE }}
+        shell: bash
+        run: |
+          set -euo pipefail
+          python -m pip install -U pip setuptools wheel
+          python -m pip install ".[all,dev-test]"
+          find examples/${{ matrix.EXAMPLE }} -name "requirements.txt" | while read req_file; do python -m pip install -r "$req_file" || exit 1; done
+          pytest -s tests/examples/${{ matrix.EXAMPLE }}

If you want, I can resolve the exact commit SHA and raise a follow-up PR to pin it.

Configuration used: CodeRabbit UI

Review profile: CHILL

Plan: Pro

docker/Dockerfile (2)

3-6: Invalid/overbroad TORCH_CUDA_ARCH_LIST; drop 12.0 and only +PTX on highest SM.

“12.0+PTX” is not recognized by many current PyTorch/CUDA combos and bloats build. Keep supported SMs; add +PTX only to the highest.

Apply:

-    TORCH_CUDA_ARCH_LIST="8.0 8.6 8.7 8.9 9.0 10.0 12.0+PTX"
+    TORCH_CUDA_ARCH_LIST="8.0 8.6 8.7 8.9 9.0+PTX"
+# If your toolchain truly supports SM 10.0, use:
+# TORCH_CUDA_ARCH_LIST="8.0 8.6 8.7 8.9 9.0 10.0+PTX"

---

27-27: Avoid world-writable perms; use non-root user with ownership.

chmod -R 777 /workspace is a security smell and breaks shared volumes’ perms.

Apply:

-# Allow users to run without root
-RUN chmod -R 777 /workspace
+# Allow users to run without root
+RUN useradd -m -u 1000 appuser && \
+    chown -R appuser:appuser /workspace
+USER appuser

.github/workflows/example_tests.yml (1)

82-91: Fix YAML anchor usage for container/steps — actionlint errors; inline blocks.

Anchors here trigger “container alias mapping expected” and “steps must be sequence” in actionlint; inline the container and steps for the non-PR job.

Apply:

   example-tests-non-pr:
     if: ${{ !startsWith(github.ref, 'refs/heads/pull-request/') }}
     # Runner list at https://github.com/nv-gha-runners/enterprise-runner-configuration/blob/main/docs/runner-groups.md
     runs-on: linux-amd64-gpu-h100-latest-1
     timeout-minutes: 90
     strategy:
       matrix:
         EXAMPLE: [llm_ptq]
-    container: *example_container
-    steps: *example_steps
+    container:
+      image: nvcr.io/nvidia/tensorrt-llm/release:1.1.0rc2.post2
+      env:
+        LD_LIBRARY_PATH: "/usr/lib/x86_64-linux-gnu:/usr/local/tensorrt/targets/x86_64-linux-gnu/lib"
+        PIP_CONSTRAINT: ""
+    steps:
+      - uses: actions/checkout@v4
+      - uses: nv-gha-runners/setup-proxy-cache@main
+      - name: Run example tests
+        run: |
+          pip install ".[all,dev-test]"
+          find examples/${{ matrix.EXAMPLE }} -name "requirements.txt" | while read req_file; do pip install -r "$req_file" || exit 1; done
+          pytest -s tests/examples/${{ matrix.EXAMPLE }}

examples/pruning/cifar_resnet.ipynb (4)

25-33: Avoid unpinned, unconditional installs; gate with import checks and use %pip.

Unconditionally installing nvidia-modelopt can override the editable/source install in your Docker/CI and create version skew. Prefer conditional installs and Jupyter-native %pip; only install torchvision explicitly here.

Apply:

-! pip install nvidia-modelopt torchvision
+%pip install -q torchvision
+import importlib
+if importlib.util.find_spec("modelopt") is None:
+    %pip install -q nvidia-modelopt

---

60-61: Graceful CPU fallback for device selection.

torch.device("cuda") will hard-fail on CPU-only envs (e.g., CI preview, Colab T4 quota exhaustion).

Apply:

-device = torch.device("cuda")
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

---

287-288: Loss averaging mismatch.

epoch_loss sums per-batch mean losses; dividing by number of samples underestimates the value. Divide by number of batches or accumulate sum of per-sample losses.

Apply:

-    epoch_loss /= len(train_loader.dataset)
+    epoch_loss /= len(train_loader)

---

662-666: Typo in checkpoint filename (“seaarch”).

This will create inconsistent artifacts and breaks resume-by-name tooling.

Apply:

-        "checkpoint": "modelopt_seaarch_checkpoint_fastnas.pth",
+        "checkpoint": "modelopt_search_checkpoint_fastnas.pth",

.github/workflows/example_tests.yml (1)

71-74: Env var expansion in container.env won’t interpolate ${LD_LIBRARY_PATH}.

GitHub Actions doesn’t expand env references inside the container env block; the literal string is passed. Either set a fixed value or export in a run step.

Apply:

-      env:
-        LD_LIBRARY_PATH: "/usr/lib/x86_64-linux-gnu:/usr/local/tensorrt/targets/x86_64-linux-gnu/lib:${LD_LIBRARY_PATH}"
-        # PATH: "/usr/local/tensorrt/targets/x86_64-linux-gnu/bin:${PATH}"
-        PIP_CONSTRAINT: "" # Disable pip constraint for upgrading packages
+      env:
+        LD_LIBRARY_PATH: "/usr/lib/x86_64-linux-gnu:/usr/local/tensorrt/targets/x86_64-linux-gnu/lib"
+        PIP_CONSTRAINT: "" # Disable pip constraint for upgrading packages

Optionally append in the step:

- name: Append LD_LIBRARY_PATH
  run: echo "LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/tensorrt/targets/x86_64-linux-gnu/lib" >> $GITHUB_ENV

Configuration used: CodeRabbit UI

Review profile: CHILL

Plan: Pro