Skip to content

fixture_validation.md

Latest commit

 

History

History
53 lines (39 loc) · 2.22 KB

fixture_validation.md

File metadata and controls

53 lines (39 loc) · 2.22 KB
Raw

SparseDrive Fixture Validation

The recipe runs a deterministic single-card SparseDrive training step on one nuScenes scene. It validates model launch and native dispatch; it is not full-dataset training, an accuracy benchmark, or a performance benchmark.

Prerequisites

Prepare the pinned checkout, the dependencies, and the info files and anchors as described in README.md. Install muSPACE with the native deformable_aggregation symbols, and run inside an environment with an available MUSA device.

The pinned upstream imports flash_attn while the model package is imported, and its attention blocks call flash-attn's variable-length CUDA kernel on a device that reports itself as CUDA, so the launch check needs a flash_attn that can actually run there; see Known limitations.

One-command validation

model_examples/SparseDrive/scripts/validate_fixture.sh \
  --checkout ./SparseDrive \
  --data-root <PREPARED_NUSCENES_ROOT> \
  --infos ./SparseDrive/data/infos/nuscenes_infos_train.pkl

--scene-name, --seed, and --max-iters change the fixture. --device must be a MUSA device, because the check asserts native MUSA dispatch and has no reference path to select.

Acceptance criteria

Validation passes only when the result JSON reports:

  • a finite loss;
  • finite_gradients: true;
  • optimizer_step: true;
  • a dispatch report whose only operator is deformable_aggregation, with selected_backend: "musa", native_available: true, and no fallback.

The pinned upstream revision is checked before model construction. A mismatched checkout, a missing info file, an unavailable MUSA device, a non-finite update, or a non-native dispatch fails the command.

Troubleshooting

If the info or anchor files are missing, run prepare_dataset.sh and prepare_anchors.sh first. If the revision check fails, re-run prepare_checkout.sh.

If dependency imports fail, install the versions in requirements.lock alongside the matching platform PyTorch build, and make flash_attn importable. If strict dispatch fails, verify that the installed muSPACE wheel exports the native deformable_aggregation entry points; the recipe does not silently fall back to the PyTorch reference path.