Skip to content

Document TorchScript and allocator economics in the performance vignette - #66

Merged
TroyHernandez merged 1 commit into
mainfrom
allocator-vignette
Sep 8, 2026
Merged

TroyHernandez merged 1 commit into
mainfrom
allocator-vignette

Conversation

@TroyHernandez

Copy link
Copy Markdown
Contributor

Summary

  • New "Where the time goes" section in the Performance Levers vignette: compiling the hot step with jit_compile(), the two jit_trace() hazards, expandable_segments vs backend:native, the three allocator callback gates and why they must be pushed live, pool regrowth after cuda_empty_cache(), compact per-token conditioning, and verifying tuning inputs. This is the content of the former torchscript-and-allocator-economics.md note from the skill hub, rewritten for the vignette.
  • ltx23_tune_gc() roxygen cites the diffuseR and chatterbox vignettes instead of a path on one machine; man/ltx23_tune_gc.Rd regenerated.
  • Docs only, no version bump.

Move the LTX-2.3 tuning notes out of the personal skill hub and into
"Performance Levers": compiling the hot step, when tracing loses,
allocator cost models and the once-at-load callback gates, pool
regrowth between phases, compact per-token conditioning, and verifying
tuning inputs. ltx23_tune_gc() now cites the vignettes instead of a
path on Troy's machine.
@TroyHernandez
TroyHernandez merged commit 25d95f5 into main Sep 8, 2026
2 checks passed
@TroyHernandez
TroyHernandez deleted the allocator-vignette branch September 8, 2026 23:06
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant