Merge branch 'main' into akaufman/usd-2608-allow

Author: mzamoramora-nvidia

.github/workflows/aws_gpu_benchmarks.yml

@@ -149,18 +149,13 @@ jobs:
           # Keep resolution 128 in the shared benchmark for nightly ASV collection.
           NEWTON_ASV_PR_GATE: "1"
         run: |
-          # Gate this PR job on directly measured runtimes only.
-          uvx --with virtualenv asv continuous \
-            --launch-method spawn \
-            --interleave-rounds \
-            --append-samples \
-            --no-only-changed \
-            -e \
-            -b 'Fast\w*\.time_' \
-            -b 'Fast(?!NewtonOverhead)\w*\.track_simulate(\(|$)' \
-            -b 'Fast\w*\.track_(mean|p95)_\w*(time|ms)(\(|$)' \
-            ${{ inputs.base_ref }} \
-            ${{ inputs.ref }}
+          # The PR environment intentionally omits PyTorch: fast DR Legs metrics
+          # disable policy inference, and no selected benchmark imports it.
+          #
+          # Keep representative camera output modes and teleop tail latency;
+          # full ASV collection retains every metric. Interleave rounds across
+          # revisions to control cold-start and ordering bias.
+          uv run --no-project asv/run_pr_benchmarks.py ${{ inputs.base_ref }} ${{ inputs.ref }}
         continue-on-error: true
         id: benchmark
 

asv.conf.json

@@ -15,8 +15,8 @@
   "default_benchmark_timeout": 600,
   "build_command": ["python -m build --wheel -o {build_cache_dir} {build_dir}"],
   "install_command": [
-    "in-dir={env_dir} python -m pip install {wheel_file}[dev] --extra-index-url=https://pypi.nvidia.com/",
     "in-dir={env_dir} python -m pip install warp-lang==1.17.0.dev20260807 mujoco==3.11.0 mujoco-warp==3.11.0 --extra-index-url=https://pypi.nvidia.com/",
+    "in-dir={env_dir} python -m pip install {wheel_file}[examples] --extra-index-url=https://pypi.nvidia.com/",
     "python -m pip install torch==2.10.0+cu130 --index-url https://download.pytorch.org/whl/cu130",
     "python -m pip install 'asv_runner<0.3.0'"
   ]

asv/benchmarks/benchmark_config.py

@@ -0,0 +1,11 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 The Newton Developers
+# SPDX-License-Identifier: Apache-2.0
+
+import os
+
+
+def pr_gate_repeat(default: int, pr_repeat: int = 3) -> int:
+    """Return a bounded PR sample count without changing full ASV runs."""
+    if os.environ.get("NEWTON_ASV_PR_GATE"):
+        return min(default, pr_repeat)
+    return default

asv/benchmarks/benchmark_kamino.py

@@ -508,16 +508,28 @@ def create_model_builder(robot, world_count):
         return builder
 
     @staticmethod
-    def create_solver(model, sim_dt):
-        # Reuse the Kamino RL example's solver settings to mirror the deployed workload.
-        from newton._src.solvers.kamino.examples.rl.simulation import RigidBodySim  # noqa: PLC0415
-
-        settings = RigidBodySim.default_settings(sim_dt)
-        settings.solver.collision_detector = settings.collision_detector
-        # Pin the linear solver so a change to default_settings cannot
-        # silently switch what this benchmark measures.
-        settings.solver.dynamics.linear_solver_type = "LLTBRCM"
-        return newton.solvers.SolverKamino(model, config=settings.solver)
+    def create_solver(model, _sim_dt):
+        # Pin the deployed RL settings without importing its PyTorch-dependent module.
+        config = newton.solvers.SolverKamino.Config(
+            sparse_jacobian=True,
+            use_collision_detector=True,
+            integrator="moreau",
+        )
+        config.collision_detector.pipeline = "unified"
+        config.collision_detector.max_contacts_per_pair = 8
+        config.constraints.alpha = 0.1
+        config.padmm.primal_tolerance = 1e-4
+        config.padmm.dual_tolerance = 1e-4
+        config.padmm.compl_tolerance = 1e-4
+        config.padmm.max_iterations = 200
+        config.padmm.eta = 1e-5
+        config.padmm.rho_0 = 0.05
+        config.padmm.use_acceleration = True
+        config.padmm.warmstart_mode = "containers"
+        config.padmm.contact_warmstart_method = "geom_pair_net_force"
+        config.padmm.use_graph_conditionals = False
+        config.dynamics.linear_solver_type = "LLTBRCM"
+        return newton.solvers.SolverKamino(model, config=config)
 
 
 if __name__ == "__main__":

asv/benchmarks/simulation/bench_cloth.py

@@ -6,6 +6,14 @@
 
 wp.config.log_level = wp.LOG_WARNING
 
+import os
+import sys
+
+parent_dir = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
+sys.path.append(parent_dir)
+
+from benchmark_config import pr_gate_repeat
+
 import newton.examples
 from newton.examples.cloth.example_cloth_franka import Example as ExampleClothManipulation
 from newton.examples.cloth.example_cloth_twist import Example as ExampleClothTwist
@@ -33,7 +41,7 @@ def time_simulate(self):
 
 
 class FastExampleClothTwist:
-    repeat = 5
+    repeat = pr_gate_repeat(5)
     number = 1
 
     def setup(self):

asv/benchmarks/simulation/bench_contacts.py

@@ -8,10 +8,17 @@
 wp.config.log_level = wp.LOG_WARNING
 
 import importlib
+import os
+import sys
 from typing import ClassVar
 
 import numpy as np
 
+parent_dir = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
+sys.path.append(parent_dir)
+
+from benchmark_config import pr_gate_repeat
+
 import newton.examples
 from newton.viewer import ViewerNull
 
@@ -267,7 +274,7 @@ class FastConvexCollision:
 
     params = (CONVEX_COLLISION_CASES,)
     param_names: ClassVar[list[str]] = ["case"]
-    repeat = 5
+    repeat = pr_gate_repeat(5)
     number = 1
 
     def setup(self, case):

asv/benchmarks/simulation/bench_ik.py

@@ -16,6 +16,7 @@
 parent_dir = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
 sys.path.append(parent_dir)
 
+from benchmark_config import pr_gate_repeat
 from benchmark_ik import build_ik_solver, create_franka_model, eval_success, fk_targets, random_solutions
 
 
@@ -89,7 +90,7 @@ def teardown(self, batch_size):
 
 class FastIKSolve(_IKBenchmark):
     params = ([512],)
-    repeat = 6
+    repeat = pr_gate_repeat(6)
 
 
 if __name__ == "__main__":

asv/benchmarks/simulation/bench_inverse_dynamics.py

@@ -18,6 +18,7 @@
 parent_dir = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
 sys.path.append(parent_dir)
 
+from benchmark_config import pr_gate_repeat
 from benchmark_inverse_dynamics import create_franka_model, set_default_pose
 
 
@@ -127,7 +128,7 @@ class FastInverseDynamics(_InverseDynamicsBenchmark):
     ``eval_inverse_dynamics_force`` on a model replicating the Franka arm
     across ``WORLD_COUNT`` worlds (default 1024)."""
 
-    repeat = 6
+    repeat = pr_gate_repeat(6)
 
 
 if __name__ == "__main__":

asv/benchmarks/simulation/bench_mujoco.py

@@ -329,6 +329,9 @@ class FastAllegro(_KpiBenchmark):
     environment = "None"
 
     def setup_cache(self):
+        if os.environ.get("NEWTON_ASV_PR_GATE"):
+            self.num_frames = 200
+            self.samples = 1
         return self._collect_metrics()
 
 

asv/benchmarks/simulation/bench_quadruped_xpbd.py

@@ -13,6 +13,7 @@
 parent_dir = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
 sys.path.append(parent_dir)
 
+from benchmark_config import pr_gate_repeat
 from benchmark_metric_tracks import _SimulationMetricTracksUnparameterized
 from benchmark_metrics import (
     collect_simulation_metrics,
@@ -33,7 +34,7 @@ def _create_example(num_frames, world_count):
 
 
 class FastExampleQuadrupedXPBD:
-    repeat = 10
+    repeat = pr_gate_repeat(10)
     number = 1
 
     def setup(self):