Repository navigation
Align cuTile Gaussian width derivatives with the analytic model - #88
Conversation
Form physical-width derivatives before combining image planes and applying the log-width chain rule. This preserves the analytic model's overflow and underflow behavior for extreme finite widths. Cover convergence and fit-step parity for both width axes in single and double precision. Signed-off-by: Trent Nelson <trentn@nvidia.com>
|
Important Review skippedAuto reviews are limited based on label configuration. 🏷️ Required labels (at least one) (1)
Please check the settings in the CodeRabbit UI or the ⚙️ Run configurationConfiguration used: Repository: NVIDIA/cuPhoton/.coderabbit.yaml Review profile: CHILL Plan: Enterprise Run ID: You can disable this status message by setting the Use the checkbox below for a quick retry:
Comment |
melo-gonzo
left a comment
melo-gonzo
left a comment
There was a problem hiding this comment.
LGTM. The derivative and chain-rule order match the analytic model. All 18 selected cuTile tests passed on RTX 6000 Ada, and all eight new regressions failed on the base revision. Focused CPU suites passed 74 tests with 31 GPU-dependent skips; lint and formatting passed.
2 participants
Why this is needed
At extreme positive Gaussian widths, cuTile can report a different convergence status or take a different fit step from the analytic CuPy path. Form the physical-width derivatives before combining image planes and applying the log-width chain rule, preserving the model's overflow and underflow behavior.
This extracts the existing arithmetic correction from #83 so it can be reviewed independently for 0.1.3. The observed discrepancy concerns backend consistency; it does not establish incorrect ordinary fits or falsely valid covariance.
Validation
make ci-lintpassed, including all-files hooks, type checking, and lock validation.