Pull requests / #2015

#2015 Draft: immutable branching contexts and shared CUDA prefixes

open · draft · @CC-David-CC · 0 comentários · No GitHub

AMD / HIPNVIDIA / CUDAModels & quantsDocumentationLinux

Descrição

## Summary

Foundation for [#1938](https://github.com/Niko1221/Strata/issues/1938) and [#1962](https://github.com/Niko1221/Strata/issues/1962): branch from a consumed prefix while preserving its history and sharing immutable CUDA pages through copy-on-write.

## What changed

Local POSIX/Python history and controller API, movable branch references, optional checkpoints, direct ownership/work counters and a separate fresh-process oracle. New branch from upstream main at 7695859a; earlier server work is excluded.

## Validation

Linux RTX 4090, Q8_0 weights, FP16 KV:

- G0: 51/51; demo passed; four mutants rejected.
- CUDA build, nine native support executables and two digest tests passed.
- G3-004: eight exact full-vocabulary comparisons, maximum logit error 0.0; 28 retained-page audits passed.
- Six hot appends: one token computed each, zero prefix recomputation. All eight image/handle owners released.

[Usage, measurements and fingerprints](https://github.com/CC-David-CC/Strata-a5500/blob/6dd5f9df7e1aa7f00045cf72e997fd8595175847/experimental/branching_context/README.md)

## Extra notes

Draft foundation: 1 of 16 G3 cases executed; full certification is incomplete. Host-RAM deduplication and concurrent server/batch-slot integration remain. HIP/SYCL builds, default serving-path parity and batched-prefill equivalence are unverified. This does not close either issue.

No site

Links install, modelos, releases.