Repository navigation
Validate checkpoint contracts before constructing xScan models - #105
Conversation
|
@coderabbitai review |
80065a0 to
0564170
Compare
|
Important Review skippedReview was skipped as selected files did not have any reviewable changes. ⚙️ Run configuration
You can disable this status message by setting the Use the checkbox below for a quick retry:
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info
📝 Walkthrough
Priority: ⬇️ Low Merge Risk: ⚪ Minimal · up to
|
|
@coderabbitai review |
✅ Action performedReview finished.
|
|
melo-gonzo
left a comment
There was a problem hiding this comment.
LGTM. Checkpoint contracts now fail before model construction while preserving existing loader behavior. Focused CPU tests, rejection probes, and both selected GPU pipeline tests passed. Deserialization still precedes validation.
Let the checkpoint loader run an optional validator after deserialization and before model construction. Bind the worker and benchmark contract checks there so incompatible dimensions or feature identities fail early. Signed-off-by: Trent Nelson <trentn@nvidia.com>
0564170 to
5c0421d
Compare
The xScan device pipeline and stage benchmark checked checkpoint dimensions and feature identity after constructing, transferring, and optionally compiling the model. An incompatible checkpoint could therefore incur expensive setup before reporting the mismatch.
Let the loader run the existing contract validator immediately after checkpoint deserialization, and use it in both consumers. Existing policy, model, device, and file-identity checks remain in place. This change does not limit checkpoint deserialization or valid model sizes.
Validation: 106 focused tests passed; 50 optional GPU tests skipped. The regression exercises the real loader for both consumers and rejects incompatible image size and feature identity before model construction. All repository pre-commit checks passed.