Merge branch 'main' into jennifchen/qat_slurm

Author: jenchen13

.github/CODEOWNERS

@@ -32,24 +32,24 @@ modelopt/torch/utils @NVIDIA/modelopt-torch-utils-codeowners
 # Examples
 /docker @NVIDIA/modelopt-docker-codeowners
 /README.md @NVIDIA/modelopt-examples-codeowners
-examples @NVIDIA/modelopt-examples-codeowners
-examples/chained_optimizations @NVIDIA/modelopt-torch-nas-prune-codeowners
-examples/cnn_qat @NVIDIA/modelopt-examples-cnn_qat-codeowners
-examples/deepseek @NVIDIA/modelopt-deploy-codeowners
-examples/diffusers @NVIDIA/modelopt-examples-diffusers-codeowners
-examples/gpt-oss @NVIDIA/modelopt-examples-gpt-oss-codeowners
-examples/llm_autodeploy @NVIDIA/modelopt-deploy-codeowners
-examples/llm_distill @NVIDIA/modelopt-torch-distill-codeowners
-examples/llm_eval @NVIDIA/modelopt-examples-llm_ptq-codeowners
-examples/llm_ptq @NVIDIA/modelopt-examples-llm_ptq-codeowners
-examples/llm_qat @NVIDIA/modelopt-examples-llm_qat-codeowners
-examples/llm_sparsity @NVIDIA/modelopt-torch-sparsity-codeowners
-examples/megatron-lm @NVIDIA/modelopt-examples-megatron-codeowners
-examples/model_hub @NVIDIA/modelopt-examples-model_hub-codeowners
-examples/nemo_run @NVIDIA/modelopt-examples-megatron-codeowners
-examples/onnx_ptq @NVIDIA/modelopt-onnx-codeowners
-examples/pruning @NVIDIA/modelopt-torch-nas-prune-codeowners
-examples/speculative_decoding @NVIDIA/modelopt-torch-speculative-codeowners
-examples/vlm_eval @NVIDIA/modelopt-examples-vlm-codeowners
-examples/vlm_ptq @NVIDIA/modelopt-examples-vlm-codeowners
-examples/windows @NVIDIA/modelopt-windows-codeowners
+/examples @NVIDIA/modelopt-examples-codeowners
+/examples/chained_optimizations @NVIDIA/modelopt-torch-nas-prune-codeowners
+/examples/cnn_qat @NVIDIA/modelopt-examples-cnn_qat-codeowners
+/examples/deepseek @NVIDIA/modelopt-deploy-codeowners
+/examples/diffusers @NVIDIA/modelopt-examples-diffusers-codeowners
+/examples/gpt-oss @NVIDIA/modelopt-examples-gpt-oss-codeowners
+/examples/llm_autodeploy @NVIDIA/modelopt-deploy-codeowners
+/examples/llm_distill @NVIDIA/modelopt-torch-distill-codeowners
+/examples/llm_eval @NVIDIA/modelopt-examples-llm_ptq-codeowners
+/examples/llm_ptq @NVIDIA/modelopt-examples-llm_ptq-codeowners
+/examples/llm_qat @NVIDIA/modelopt-examples-llm_qat-codeowners
+/examples/llm_sparsity @NVIDIA/modelopt-torch-sparsity-codeowners
+/examples/megatron-lm @NVIDIA/modelopt-examples-megatron-codeowners
+/examples/model_hub @NVIDIA/modelopt-examples-model_hub-codeowners
+/examples/nemo_run @NVIDIA/modelopt-examples-megatron-codeowners
+/examples/onnx_ptq @NVIDIA/modelopt-onnx-codeowners
+/examples/pruning @NVIDIA/modelopt-torch-nas-prune-codeowners
+/examples/speculative_decoding @NVIDIA/modelopt-torch-speculative-codeowners
+/examples/vlm_eval @NVIDIA/modelopt-examples-vlm-codeowners
+/examples/vlm_ptq @NVIDIA/modelopt-examples-vlm-codeowners
+/examples/windows @NVIDIA/modelopt-windows-codeowners

.github/workflows/gpu_tests.yml

@@ -22,20 +22,31 @@ jobs:
       any_changed: ${{ steps.changed-tests.outputs.any_changed }}
     steps:
       - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
       - id: get-pr-info
         uses: nv-gha-runners/get-pr-info@main
+      # Get commit from main branch that is present in the PR to use as base for changed files
+      - id: calculate-merge-base
+        env:
+          PR_SHA: ${{ fromJSON(steps.get-pr-info.outputs.pr-info).head.sha }}
+          BASE_SHA: ${{ fromJSON(steps.get-pr-info.outputs.pr-info).base.sha }}
+        run: |
+          (echo -n "merge-base="; git merge-base "$BASE_SHA" "$PR_SHA") | tee --append "${GITHUB_OUTPUT}"
       - name: Check for changes in test-relevant directories
         id: changed-tests
         uses: step-security/[email protected]
         with:
+          base_sha: ${{ steps.calculate-merge-base.outputs.merge-base }}
+          sha: ${{ fromJSON(steps.get-pr-info.outputs.pr-info).head.sha }}
           files: |
             .github/workflows/gpu_tests.yml
             modelopt/**
             tests/gpu/**
             tox.ini
             pyproject.toml
             setup.py
-          base_sha: ${{ fromJSON(steps.get-pr-info.outputs.pr-info).base.ref }}
+          fail_on_initial_diff_error: true
   wait-checks:
     needs: [check-file-changes]
     if: needs.check-file-changes.outputs.any_changed == 'true'
@@ -70,3 +81,12 @@ jobs:
     timeout-minutes: 90
     container: *gpu_container
     steps: *gpu_steps
+  gpu-pr-required-check:
+    # Run even if gpu-tests-pr is skipped
+    if: ${{ startsWith(github.ref, 'refs/heads/pull-request/') && always() }}
+    needs: [check-file-changes, gpu-tests-pr]
+    runs-on: ubuntu-latest
+    steps:
+      - name: Required GPU tests did not succeed
+        if: ${{ needs.check-file-changes.result != 'success' || (needs.check-file-changes.outputs.any_changed == 'true' && needs.gpu-tests-pr.result != 'success') }}
+        run: exit 1

.github/workflows/unit_tests.yml

@@ -4,13 +4,6 @@ name: Unit tests
 on:
   pull_request:
     branches: [main, release/*]
-    paths:
-      - ".github/workflows/unit_tests.yml"
-      - "modelopt/**"
-      - "tests/unit/**"
-      - "pyproject.toml"
-      - "setup.py"
-      - "tox.ini"
   push:
     branches: [main, release/*]
     paths:
@@ -126,3 +119,9 @@ jobs:
           python-version: "3.12"
       - name: Run unit tests
         run: pip install tox && tox -e py312-partial-unit-${{ matrix.test-env }}
+  unit-pr-required-check:
+    if: github.event_name == 'pull_request'
+    needs: [linux, windows, multi-py, multi-torch, multi-transformers, partial-install]
+    runs-on: ubuntu-latest
+    steps:
+      - run: echo "All PR unit test jobs completed"

CHANGELOG.rst

@@ -5,10 +5,12 @@ Model Optimizer Changelog (Linux)
 ^^^^^^^^^^^^^^^^^
 
 **Deprecations**
+- Deprecated ``quantize_mode`` argument in ``examples/onnx_ptq/evaluate.py`` to support strongly typing. Use ``engine_precision`` instead.
 
 **Bug Fixes**
 
 **New Features**
+- ``high_precision_dtype`` default to fp16 in ONNX quantization, i.e. quantized output model weights are now FP16 by default.
 
 0.35 (2025-09-04)
 ^^^^^^^^^^^^^^^^^

examples/llm_ptq/hf_ptq.py

@@ -742,7 +742,7 @@ def output_decode(generated_ids, input_shape):
     )
     parser.add_argument(
         "--verbose",
-        help="Print verbose output (e.g. quantization summary). Disable by --no_verbose.",
+        help="Print verbose output (e.g. quantization summary). Disable by --no-verbose.",
         default=True,
         action=argparse.BooleanOptionalAction,
     )

examples/onnx_ptq/README.md

@@ -120,7 +120,7 @@ The following evaluation requires the `val` directory of the [ImageNet dataset](
 python evaluate.py \
     --onnx_path=<path to classification model> \
     --imagenet_path=<path to the ImageNet dataset> \
-    --quantize_mode=<fp8|int8|int4> \
+    --engine_precision=stronglyTyped \
     --model_name=vit_base_patch16_224
 ```
 
@@ -165,7 +165,7 @@ If the input model is of type image classification, use the following script to
 python evaluate.py \
     --onnx_path=<path to the exported ONNX model> \
     --imagenet_path=<path to the ImageNet dataset> \
-    --quantize_mode=stronglyTyped \
+    --engine_precision=stronglyTyped \
     --model_name=vit_base_patch16_224
 ```
 

examples/onnx_ptq/docker/Dockerfile

@@ -12,10 +12,11 @@ RUN python -m pip install --upgrade pip \
 
 WORKDIR /workspace
 
-RUN pip install tensorrt==10.13.2.6 && \
-    export TRT_PATH=$(python -c "import tensorrt; import os; print(os.path.dirname(tensorrt.__file__))") && \
-    export LD_LIBRARY_PATH="$TRT_PATH/lib:/usr/include:${LD_LIBRARY_PATH}" && \
-    export PATH="$TRT_PATH/bin:${PATH}"
+RUN pip install tensorrt==10.13.2.6
+ENV TRT_PATH=/usr/local/lib/python3.12/dist-packages/tensorrt
+ENV CUDNN_LIB_DIR=/usr/local/lib/python3.12/dist-packages/nvidia/cudnn/lib
+ENV LD_LIBRARY_PATH="${CUDNN_LIB_DIR}:${TRT_PATH}/lib:/usr/include:${LD_LIBRARY_PATH}"
+ENV PATH="${TRT_PATH}/bin:${PATH}"
 
 # Copy application code and install requirements
 COPY modelopt modelopt/modelopt

examples/onnx_ptq/evaluate.py

@@ -48,29 +48,22 @@ def main():
     parser.add_argument(
         "--eval_data_size", type=int, default=None, help="Number of examples to evaluate"
     )
-    # By default, TensorRT autotunes tensor types to generate the fastest engine. When you specify
-    # to TensorRT that a network is strongly typed, it infers a type for each intermediate and
-    # output tensor using the rules in the operator type specification. For networks quantized in
-    # INT4 or FP8 mode, stronglyTyped as the mode is recommended for TensorRT deployment. Though
-    # INT8 networks are generally compiled with int8 mode, certain INT8 ViT networks compiled with
-    # stronglyTyped precision have shown better performance.
     parser.add_argument(
-        "--quantize_mode",
+        "--engine_precision",
         type=str,
         default="stronglyTyped",
-        choices=["fp8", "fp16", "fp32", "int4", "int8", "int8_iq", "bf16", "best", "stronglyTyped"],
-        help="Quantization mode for the TensorRT engine. \
-            Supported options: fp8, fp16, fp32, int8, int8_iq(implicit quantization), bf16, best, stronglyTyped",
+        choices=["best", "fp16", "stronglyTyped"],
+        help="Precision mode for the TensorRT engine. \
+            stronglyTyped is recommended, all other modes have been deprecated in TensorRT",
     )
     parser.add_argument(
         "--results_path", type=str, default=None, help="Save the results to the specified path"
     )
 
     args = parser.parse_args()
-
     deployment = {
         "runtime": "TRT",
-        "precision": args.quantize_mode,
+        "precision": args.engine_precision,
     }
 
     # Create an ONNX bytes object with the specified path

examples/onnx_ptq/evaluation.py

@@ -29,7 +29,7 @@
 deployment = {
     "runtime": "TRT",
     "accelerator": "GPU",
-    "precision": "fp32",
+    "precision": "stronglyTyped",
     "onnx_opset": "21",
 }
 

examples/onnx_ptq/torch_quant_to_onnx.py

@@ -83,12 +83,12 @@ def forward_loop(model):
     return quantized_model
 
 
-def get_model_input_shape(model_name):
+def get_model_input_shape(model_name, batch_size):
     """Get the input shape from timm model configuration."""
     model = timm.create_model(model_name, pretrained=True, num_classes=1000)
     data_config = timm.data.resolve_model_data_config(model)
     input_size = data_config["input_size"]
-    return (1, *tuple(input_size))  # Add batch dimension
+    return (batch_size, *tuple(input_size))  # Add batch dimension
 
 
 def main():
@@ -119,11 +119,17 @@ def main():
         default=512,
         help="Number of images to use in calibration [1-512]",
     )
+    parser.add_argument(
+        "--batch_size",
+        type=int,
+        default=1,
+        help="Batch size for calibration and ONNX model export.",
+    )
 
     args = parser.parse_args()
 
     # Get input shape from model config
-    input_shape = get_model_input_shape(args.timm_model_name)
+    input_shape = get_model_input_shape(args.timm_model_name, args.batch_size)
 
     # Create model and move to appropriate device
     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")