Skip to content

v0.1.1 — hardened M0–M3 (review-clean)

Choose a tag to compare

View all tags
@TaewoooPark TaewoooPark released this 30 Jun 20:44
v0.1.1
b307545

Hardening release on top of v0.1.0, after a code review against Exo and other
split-inference references (Petals, llama.cpp RPC). Reached 2 consecutive all-OK
review rounds; round 3 was fully clean.

Changes since v0.1.0:

  • engine: pass only the kwargs the loaded decoder layer's forward() accepts
    (introspected, no hardcoded arg list); size the per-shard DynamicCache to the
    shard's local layer count; document the shared-model layer_idx reindex side-effect
  • orchestrator: narrow EOS to eos_token_id / generation_config (not all special ids)
  • parity_test: --selftest now covers n=1 and a 180-token sequence

Parity: --selftest 6/6 bitwise-equal to a clean reference (English/code/Korean;
n = 1..180). Scope unchanged: M4–M6 need the Windows node.

Assets 2
Loading