Repository navigation

Production paged + tiered KV cache for the agent army (MAD-126) - #12

Merged

kmbandy merged 20 commits into

masterfrom

feat/MAD-126-army-goal

May 11, 2026

Merged

Production paged + tiered KV cache for the agent army (MAD-126)#12
kmbandy merged 20 commits into
masterfrom
feat/MAD-126-army-goal

Commits on May 10, 2026
- mt:: paged-attn — paged-block semantic prefetch infra (MAD-122, partial)
  
  kmbandy
  and
  claude
  committed
  b5f029b
- mt:: server — paged-block semantic fingerprinting + restore (MAD-122)
  
  kmbandy
  and
  claude
  committed
  d432230
- mt:: paged-attn — relocate semantic prefetch API to llama_kv_cache_paged (MAD-125 follow-up)
  
  kmbandy
  and
  claude
  committed
  e16916d
- mt:: ARCH-CLEANUP — remove legacy tiered paths; thin the wrapper (MAD-127)
  
  kmbandy
  and
  claude
  committed
  7bf8a5b
- mt:: paged-attn — block-aligned partial seq_rm (MAD-128, part 1)
  
  kmbandy
  and
  claude
  committed
  d23d327
- mt:: server — clearer paged ctx-shift fallback messaging (MAD-128, part 2)
  
  kmbandy
  and
  claude
  committed
  db75156
- mt:: paged-attn — seq_cp CoW via BlockPool refcounting (MAD-128, part 3)
  
  kmbandy
  and
  claude
  committed
  f3ec51c
- mt:: semantic prefetch — prefill-time write trigger on llama_kv_cache_paged (MAD-129)
  
  kmbandy
  and
  claude
  committed
  9299eab
- mt:: paged-attn — state persistence + cold-tier resume + fingerprint save/load (MAD-130)
  
  kmbandy
  and
  claude
  committed
  e2f9e8f
- mt:: multi-instance — per-instance cold subdir + lockfile + budget cap (MAD-131)
  
  kmbandy
  and
  claude
  committed
  4f667ab
- mt:: paged-attn — concurrency hardening (MAD-132)
  
  kmbandy
  and
  claude
  committed
  50e1ce6
- mt:: observability — tier counters + /metrics/tier + /slots tier extension + per-batch log (MAD-133)
  
  kmbandy
  and
  claude
  committed
  de644a5
- mt:: ergonomics — paged-default-on for tiered + bge warmup + config validation (MAD-134)
  
  kmbandy
  and
  claude
  committed
  93796b6
- mt:: paged-attn — F16 cold-tier int4 compression + round-trip tests (MAD-135)
  
  kmbandy
  and
  claude
  committed
  6358dcc
- mt:: build — multi-target HIP (gfx1201;gfx1030) + comment fix (MAD-136 partial)
  
  kmbandy
  and
  claude
  committed
  7d520e4
- mt:: tests — unit tests for tier primitives (MAD-137 partial)
  
  kmbandy
  and
  claude
  committed
  d788f7a
- mt:: tests — integration + stress + matrix runner + CI (MAD-137)
  
  kmbandy
  and
  claude
  committed
  b9ad370
- server: fail-fast on paged admission deadlock instead of GGML_ABORT (MAD-141)
  
  kmbandy
  and
  claude
  committed
  0d66d8a
- ci: strip self-hosted matrix from army workflow
  
  kmbandy
  and
  claude
  committed
  1871f4f
- docs: tiered KV cache user guide + operator runbook + architecture (MAD-138)
  
  kmbandy
  and
  claude
  committed
  9c16a2f

Commits

Commits on May 10, 2026