CI/Dev env bump

.github/codecov.yml

@@ -9,5 +9,5 @@ coverage:
     project:
       default:
         target: auto
-        threshold: 1% # Allow atmost 1% coverage drop from main branch.
+        threshold: 2% # Allow atmost 2% coverage drop from main branch.
     patch: false

.github/workflows/example_tests.yml

@@ -59,7 +59,7 @@ jobs:
     uses: ./.github/workflows/_example_tests_runner.yml
     secrets: inherit
     with:
-      docker_image: "nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc17"
+      docker_image: "nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc19"
       example: ${{ matrix.example }}
       pip_install_extras: "[hf,dev-test]"
       runner: linux-amd64-gpu-rtxpro6000-latest-1
@@ -73,7 +73,7 @@ jobs:
     uses: ./.github/workflows/_example_tests_runner.yml
     secrets: inherit
     with:
-      docker_image: "nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc17"
+      docker_image: "nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc19"
       example: ${{ matrix.example }}
       pip_install_extras: "[hf,dev-test]"
       runner: linux-amd64-gpu-rtxpro6000-latest-2
@@ -85,7 +85,7 @@ jobs:
     uses: ./.github/workflows/_example_tests_runner.yml
     secrets: inherit
     with:
-      docker_image: "nvcr.io/nvidia/nemo:26.04"
+      docker_image: "nvcr.io/nvidia/nemo:26.06"
       example: megatron_bridge
       timeout_minutes: 45
       pip_install_extras: "[hf,puzzletron,dev-test]"

.github/workflows/gpu_tests.yml

@@ -43,10 +43,10 @@ jobs:
             container_image: nvcr.io/nvidia/pytorch:26.05-py3
           - example: gpu_megatron
             timeout: 60
-            container_image: nvcr.io/nvidia/nemo:26.04
+            container_image: nvcr.io/nvidia/nemo:26.06
           - example: gpu_trtllm
             timeout: 15
-            container_image: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc17
+            container_image: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc19
           - example: gpu_vllm
             timeout: 15
             container_image: docker.io/vllm/vllm-openai:v0.20.0

.pre-commit-config.yaml

@@ -21,16 +21,14 @@ repos:
       - id: requirements-txt-fixer
 
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.12.11
+    rev: v0.15.20
     hooks:
       - id: ruff-check
         args: [--fix, --exit-non-zero-on-fix]
-        exclude: ^examples/specdec_bench/specdec_bench/datasets/speed\.py$
       - id: ruff-format
-        exclude: ^examples/specdec_bench/specdec_bench/datasets/speed\.py$
 
   - repo: https://github.com/pre-commit/mirrors-mypy
-    rev: v1.17.1
+    rev: v2.1.0
     hooks:
       - id: mypy
 

docs/source/conf.py

@@ -94,6 +94,11 @@
 exclude_patterns = []
 templates_path = ["_templates"]
 
+# Suppress ambiguous cross-reference warnings that arise because EMAConfig and
+# QuantizerAttributeConfig both define a field named `type`.  Renaming would be
+# an API break; silencing the warning here is the least-invasive fix.
+suppress_warnings = ["ref.python"]
+
 
 # -- Options for HTML output -------------------------------------------------
 

examples/dataset/add_nemotron_chat.py

@@ -157,7 +157,7 @@ def parse_nemotron_conversation(raw_conversations: list) -> list[dict] | None:
         if content:
             msgs.append({"role": role, "content": content})
 
-    return msgs if msgs else None
+    return msgs or None
 
 
 async def main(args: argparse.Namespace) -> None:

examples/diffusers/cache_diffusion/cache_diffusion/utils.py

@@ -24,8 +24,8 @@
 
 PIXART_DEFAULT_CONFIG = [
     {
-        "wildcard_or_filter_func": lambda name: not re.search(
-            r"transformer_blocks\.(2[1-7])\.", name
+        "wildcard_or_filter_func": lambda name: (
+            not re.search(r"transformer_blocks\.(2[1-7])\.", name)
         ),
         "select_cache_step_func": lambda step: (step % 3) != 0,
     }

examples/diffusers/quantization/diffusion_trt.py

@@ -194,7 +194,7 @@ def main():
     )
     args = parser.parse_args()
 
-    image_name = args.save_image_as if args.save_image_as else f"{args.model}.png"
+    image_name = args.save_image_as or f"{args.model}.png"
     model_dtype = DTYPE_MAP[args.model]
 
     pipe = PipelineManager.create_pipeline_from(

examples/diffusers/quantization/onnx_utils/export.py

@@ -417,9 +417,9 @@ def get_io_shapes(model_id, onnx_load_path, trt_dynamic_shapes):
     if onnx_load_path != "":
         if model_id in ["sdxl-1.0", "sdxl-turbo"]:
             output_name = "latent"
-        elif model_id in ["sd3-medium"]:
+        elif model_id == "sd3-medium":
             output_name = "sample"
-        elif model_id in ["sd3.5-medium"]:
+        elif model_id == "sd3.5-medium":
             output_name = "out_hidden_states"
         elif model_id in ["flux-dev", "flux-schnell"]:
             output_name = "output"
@@ -499,7 +499,7 @@ def modelopt_export_sd(backbone, onnx_dir, model_name, precision):
         if model_name == "flux-dev":
             input_names.append("guidance")
         output_names = ["latent"]
-    elif model_name in ["ltx-video-dev"]:
+    elif model_name == "ltx-video-dev":
         input_names = [
             "hidden_states",
             "encoder_hidden_states",
@@ -508,7 +508,7 @@ def modelopt_export_sd(backbone, onnx_dir, model_name, precision):
             "video_coords",
         ]
         output_names = ["latent"]
-    elif model_name in ["wan2.2-t2v-14b"]:
+    elif model_name == "wan2.2-t2v-14b":
         input_names = [
             "hidden_states",
             "timestep",

examples/gpt-oss/requirements.txt

@@ -2,4 +2,4 @@ kernels>=0.9.0,<0.13
 trackio<0.21
 # transformers>=5.3 avoids CVE-2026-4372 (RCE via the `kernels` Hub download path, which this example installs)
 transformers>=5.3
-trl>=0.21.0
+trl>=1.0

examples/gpt-oss/sft.py

@@ -70,7 +70,7 @@ def main(script_args, training_args, model_args, quant_args):
     # ------------------------
     model_kwargs = {
         "revision": model_args.model_revision,
-        "trust_remote_code": model_args.trust_remote_code,
+        "trust_remote_code": getattr(model_args, "trust_remote_code", False),
         "attn_implementation": model_args.attn_implementation,
         "dtype": getattr(model_args, "dtype", "bfloat16"),
         "use_cache": not training_args.gradient_checkpointing,

examples/llm_distill/README.md

@@ -25,7 +25,6 @@ This section focuses on demonstrating how to apply Model Optimizer to perform kn
 ### Docker
 
 For Hugging Face models, please use the PyTorch docker image (e.g., `nvcr.io/nvidia/pytorch:26.01-py3`).
-For Megatron-Bridge or Megatron-LM models, use the NeMo container (e.g., `nvcr.io/nvidia/nemo:26.02`) which has all the dependencies installed.
 Visit our [installation docs](https://nvidia.github.io/Model-Optimizer/getting_started/2_installation.html) for more information.
 
 Also follow the installation steps below to upgrade to the latest version of Model Optimizer and install example-specific dependencies.

examples/llm_distill/requirements.txt

@@ -1,3 +1,5 @@
 pyarrow
 torchao>=0.14.1
-trl>=0.23.0
+# TODO: trl>=1.7 stops materializing logits during eval, breaking KD eval loss in
+# modelopt/torch/distill/plugins/huggingface.py (_standard_kd_loss). Lift the cap once fixed.
+trl>=1.0,<1.7

examples/llm_ptq/README.md

@@ -28,7 +28,6 @@ This section focuses on Post-training quantization, a technique that reduces mod
 ### Docker
 
 For Hugging Face models, please use the TensorRT-LLM docker image (e.g., `nvcr.io/nvidia/tensorrt-llm/release:1.2.0`).
-For Megatron-Bridge or Megatron-LM models, use the NeMo container (e.g., `nvcr.io/nvidia/nemo:26.02`).
 Visit our [installation docs](https://nvidia.github.io/Model-Optimizer/getting_started/2_installation.html) for more information.
 
 Also follow the installation steps below to upgrade to the latest version of Model Optimizer and install example-specific dependencies.

examples/llm_qat/dataset_utils.py

@@ -539,7 +539,7 @@ def _build_cache_path(
     cache_dir: str,
 ) -> str:
     """Build a deterministic cache path for the blend config."""
-    base = cache_dir if cache_dir else tempfile.gettempdir()
+    base = cache_dir or tempfile.gettempdir()
 
     tok_name, tok_fp = _tokenizer_fingerprint(tokenizer)
     splits_str = ",".join(f"{k}:{v}" for k, v in sorted(config.splits.items()))

examples/llm_qat/notebooks/requirements.txt

@@ -1,3 +1,3 @@
 ipywidgets
 nvidia-modelopt[all]
-trl
+trl>=1.0

examples/megatron_bridge/README.md

@@ -20,7 +20,7 @@ This directory contains examples of using Model Optimizer with [NeMo Megatron-Br
 
 ## Pre-Requisites
 
-Running these examples requires many additional dependencies to be installed (e.g., Megatron-Bridge, Megatron-core, etc.), hence we strongly recommend directly using the NeMo container (e.g., `nvcr.io/nvidia/nemo:26.04`) which has all the dependencies installed.
+Running these examples requires many additional dependencies to be installed (e.g., Megatron-Bridge, Megatron-core, etc.), hence we strongly recommend directly using the NeMo container (e.g., `nvcr.io/nvidia/nemo:26.06`) which has all the dependencies installed.
 
 To get the ModelOpt examples scripts, mount your Model-Optimizer repo to the container as follows:
 
@@ -30,7 +30,7 @@ if [ ! -d "${MODELOPT_DIR}" ]; then
   git clone https://github.com/NVIDIA/Model-Optimizer.git ${MODELOPT_DIR}
 fi
 
-export DOCKER_IMAGE=nvcr.io/nvidia/nemo:26.04
+export DOCKER_IMAGE=nvcr.io/nvidia/nemo:26.06
 docker run \
   --gpus all \
   --shm-size=16GB \

examples/megatron_bridge/prune_minitron.py

@@ -319,11 +319,9 @@ def main(args: argparse.Namespace):
         # NAS-based pruning: restrict search space to a smaller set of candidates.
         # Allow more choices for MoE FFN as they are generally smaller.
         # NOTE: Reduce divisors and increase config['top_k'] to potentially find a better model.
-        hidden_size_divisor = args.ss_channel_divisor if args.ss_channel_divisor else 256
-        ffn_hidden_size_divisor = (
-            args.ss_channel_divisor
-            if args.ss_channel_divisor
-            else (256 if (provider.num_moe_experts or 0) > 0 else 512)
+        hidden_size_divisor = args.ss_channel_divisor or 256
+        ffn_hidden_size_divisor = args.ss_channel_divisor or (
+            256 if (provider.num_moe_experts or 0) > 0 else 512
         )
         ss_config = mtp.mcore_minitron.get_mcore_minitron_config(
             hidden_size_divisor=hidden_size_divisor,

examples/pruning/README.md

@@ -26,7 +26,7 @@ This section focuses on applying Model Optimizer's state-of-the-art complementar
 
 ## Pre-Requisites
 
-For Minitron pruning for Megatron-Bridge / Megatron-LM models, use the NeMo container (e.g., `nvcr.io/nvidia/nemo:26.04`) which has all the dependencies installed.
+For Minitron pruning for Megatron-Bridge / Megatron-LM models, use the NeMo container (e.g., `nvcr.io/nvidia/nemo:26.06`) which has all the dependencies installed.
 
 For FastNAS pruning for PyTorch Computer Vision models, no additional dependencies are required.
 

examples/specdec_bench/specdec_bench/datasets/speed.py

@@ -14,6 +14,8 @@
 # limitations under the License.
 
 # mypy: disable-error-code="index"
+# Paper-derived prompt strings carry intentional whitespace/long lines; keep them verbatim.
+# ruff: noqa: E501, W291, W293, PLR1704
 import random
 import re
 from enum import Enum
@@ -74,13 +76,15 @@ class BenchmarkDataset(str, Enum):
     BenchmarkDataset.CNN_DAILYMAIL.value: lambda dataset_name, config_name: load_dataset(
         dataset_name, config_name, split="test"
     ),
-    BenchmarkDataset.HLE.value: lambda dataset_name, config_name: load_dataset(
-        dataset_name, split="test", revision="021a3d71f516a7ac28ceb8d284969902edf1edeb"
-    )
-    if config_name != "train_test_split"
-    else load_dataset(
-        dataset_name, split="test", revision="021a3d71f516a7ac28ceb8d284969902edf1edeb"
-    ).train_test_split(test_size=0.5, shuffle=True, seed=42),
+    BenchmarkDataset.HLE.value: lambda dataset_name, config_name: (
+        load_dataset(
+            dataset_name, split="test", revision="021a3d71f516a7ac28ceb8d284969902edf1edeb"
+        )
+        if config_name != "train_test_split"
+        else load_dataset(
+            dataset_name, split="test", revision="021a3d71f516a7ac28ceb8d284969902edf1edeb"
+        ).train_test_split(test_size=0.5, shuffle=True, seed=42)
+    ),
     BenchmarkDataset.LIVECODEBENCH.value: lambda dataset_name, config_name: load_dataset(
         "json",
         data_files={
@@ -243,7 +247,7 @@ def _generate_stackselect_prompt(
         answers_to_add = (
             answers[: answers_to_add_stop + 1]
             if answers_to_add_stop >= correct_answer_i
-            else [answers[correct_answer_i]] + answers[: answers_to_add_stop + 1]
+            else [answers[correct_answer_i], *answers[: answers_to_add_stop + 1]]
         )
         random.shuffle(answers_to_add)
         for i, answer in enumerate(answers_to_add):
@@ -368,7 +372,7 @@ def _generate_chatrag_bench_prompt(external_dataset: "Dataset") -> list[Any]:
             if message["role"] == "user"
         ]
 
-        return [prompt.format(context=context, question=questions[0])] + questions[1:]
+        return [prompt.format(context=context, question=questions[0]), *questions[1:]]
 
     @staticmethod
     def _generate_coser_prompt(external_dataset: "Dataset") -> str:
@@ -519,11 +523,9 @@ def _fetch_all_turns_data(
                 hle_train = hle_train.to_pandas()
                 hle_train = hle_train[hle_train["image"] == ""]
                 hle_train["demonstration"] = hle_train.apply(
-                    lambda e: "Question: "
-                    + e["question"]
-                    + "\n\nAnswer: "
-                    + e["rationale"]
-                    + "\n\n",
+                    lambda e: (
+                        "Question: " + e["question"] + "\n\nAnswer: " + e["rationale"] + "\n\n"
+                    ),
                     axis=1,
                 )
                 hle_train["tokens"] = hle_train["demonstration"].apply(

examples/windows/accuracy_benchmark/fvd_metrics/compute_fvd.py

@@ -342,9 +342,7 @@ def main():
     parser.add_argument("--output", default=None, help="Path to save JSON results")
     args = parser.parse_args()
 
-    device = torch.device(
-        args.device if args.device else ("cuda" if torch.cuda.is_available() else "cpu")
-    )
+    device = torch.device(args.device or ("cuda" if torch.cuda.is_available() else "cpu"))
     log.info(f"Device: {device}")
 
     weights_path = resolve_weights(args.weights)

modelopt/onnx/graph_surgery/encoder_cross_kv.py

@@ -154,10 +154,7 @@ def _add_cross_kv_outputs(
     # Update the graph output
     for output in list(graph.output):
         if output.name == hidden_state_output_name:
-            dims = [
-                d.dim_param if d.dim_param else d.dim_value
-                for d in output.type.tensor_type.shape.dim
-            ]
+            dims = [d.dim_param or d.dim_value for d in output.type.tensor_type.shape.dim]
             new_output = helper.make_tensor_value_info(
                 encoder_hidden_states_name,
                 output.type.tensor_type.elem_type,

modelopt/onnx/llm_export_utils/quantization_utils.py

@@ -123,7 +123,7 @@ def quantize(
         f"Only fp8(W8A8), int4_awq(W4A16), nvfp4(W4A4) is supported. You passed an unsupported precision: {precision}."
     )
 
-    assert lm_head_precision in ["fp16"], (
+    assert lm_head_precision == "fp16", (
         f"Only fp16(unquantized) is supported for lm_head. You passed an unsupported precision: {lm_head_precision}."
     )
 

modelopt/onnx/op_types.py

@@ -240,7 +240,7 @@ def is_control_flow_op(op_type: str):
 
 def is_multiclass_op(op_type: str):
     """Returns whether the given op type is of Multiclass category or not."""
-    return op_type in ["Einsum"]
+    return op_type == "Einsum"
 
 
 def is_recurrent_op(op_type: str):

modelopt/onnx/quantization/graph_utils.py

@@ -596,7 +596,7 @@ def build_non_residual_input_map(
     non_residual_inputs = {}
     no_quantize_inputs = []
     for node in graph.nodes:
-        if node.op in ["Add"]:
+        if node.op == "Add":
             # Add nodes with constant or graph input does not have non-residual input
             # Here, A = node.inputs[0], B = node.inputs[1] and A.inputs means producer nodes of A
             # TODO: make this check a util?

modelopt/onnx/quantization/qdq_utils.py

@@ -1490,7 +1490,7 @@ def quantize_weights_to_mxfp8(
 
     # set output type of DQ to FP16
     for node in graph.node:
-        if node.op_type in ["TRT_MXFP8DequantizeLinear"]:
+        if node.op_type == "TRT_MXFP8DequantizeLinear":
             for attr in node.attribute:
                 if attr.name == "output_dtype":
                     attr.i = onnx_dtype_map["Half"]
@@ -1568,7 +1568,7 @@ def _get_precision_dtype() -> str:
     logger.debug(f"Found {len(fp4_qdq_nodes)} FP4QDQ nodes to convert")
 
     for node in fp4_qdq_nodes:
-        idx1 = initializer_indices.get(node.input[0], None)
+        idx1 = initializer_indices.get(node.input[0])
         assert idx1 is not None, f"Initializer for weight '{node.input[0]}' not found."
         block_size_attr = next((attr for attr in node.attribute if attr.name == "block_size"), None)
         assert block_size_attr is not None, f"block_size attribute not found for {node.name}"

modelopt/onnx/trt_utils.py

@@ -154,7 +154,7 @@ def get_custom_layers(
 
         for i in range(layer.num_outputs):
             output_tensor = layer.get_output(i)
-            if output_tensor and not all_tensor_info.get(output_tensor.name, None):
+            if output_tensor and not all_tensor_info.get(output_tensor.name):
                 all_tensor_info[output_tensor.name] = {
                     "shape": ["unk" if (s == -1) else s for s in output_tensor.shape],
                     "dtype": output_tensor.dtype,

modelopt/torch/__init__.py

@@ -51,7 +51,7 @@
 
     if _Version(_transformers_version) < _Version("4.56") or _Version(
         _transformers_version
-    ) >= _Version("5.10"):
+    ) >= _Version("5.13"):
         _warnings.warn(
             f"transformers {_transformers_version} is not tested with current version of modelopt and may cause issues."
             " Please install recommended version with `pip install -U nvidia-modelopt[hf]` if working with HF models.",

modelopt/torch/_deploy/_runtime/tensorrt/tensorrt_utils.py

@@ -59,7 +59,7 @@ def __del__(self):
 
 def get_engine_bytes(engine: trt.tensorrt.ICudaEngine) -> bytes:
     """Return serialized TensorRT engine bytes."""
-    return bytearray(engine.serialize())  # type: ignore[return-value]
+    return bytes(engine.serialize())
 
 
 def load_engine(buffer: bytes, log_level: int = trt.Logger.ERROR) -> trt.tensorrt.ICudaEngine: