Commit Graph

4 Commits

Author SHA1 Message Date
ZacharyZcR 1d9f554715 cuda: cache layernorm weights on-device in pipe_layer_sparse (kill 152 sync H2D/token) + overlap-window profiling counters 2026-07-19 23:19:25 +08:00
ZacharyZcR ab55f4900c cuda: COLI_GROUP_ASYNC=1 — async expert-group issue/take with CPU/GPU overlap at decode (opt-in, +6-8% measured) 2026-07-19 23:19:25 +08:00
ZacharyZcR bc69a9a6d0 fix: remove duplicate argmax_v — use NaN-safe version from sample.h 2026-07-19 21:15:54 +08:00
ZacharyZcR f853ea8a0b refactor: split glm.c into colibri.c + 4 header modules
Rename glm.c → colibri.c and extract four self-contained modules
into header-only files (same pattern as st.h/tier.h/grammar.h):

  quant.h      (672 lines) — SIMD matmul kernels, quantization
  sample.h     (143 lines) — RNG, top-p sampling, stop-set
  kv_persist.h (121 lines) — .coli_kv disk persistence
  telemetry.h  (189 lines) — dashboard protocol, stats, usage

Main engine file shrinks from 6588 to 5396 lines (−18%).

Build system: primary target is now colibri$(EXE); `make glm`
remains as a phony alias for backward compat. CI, setup.sh,
coli CLI, and all 10 test files that include the engine are
updated. make check passes (C + Python, 73 tests, zero warnings).
2026-07-19 21:12:04 +08:00