Commit Graph

  • eab0e74c82 Merge pull request #1 from alexpolo1/rig-local-tuning master origin/master origin/HEAD Alex 2026-08-28 14:47:25 +02:00
  • 79920ff5ee Local: correct base.yml profile header (drafter off, one Status) origin/rig-local-tuning Alex 2026-07-17 12:51:39 +02:00
  • 213d888c90 Local: Pi vs Hermes on a clean server = DEAD TIE (8/8 each) Alex 2026-07-15 06:20:38 +02:00
  • 6e30a10b29 Local: same-day quality A/B — MTP-off is quality-neutral; CORRECT the contamination claim Alex 2026-07-15 01:25:45 +02:00
  • 225f1b56b6 Local: ship MTP-off dual compose (fp8.yml) + systemd drop-in; :8010 flipped Alex 2026-07-14 23:55:30 +02:00
  • c09916adb1 Local: prove MTP n=3 corrupts chained-agent tool calls (2x2 A/B) -> club-3090 #710 Alex 2026-07-14 15:48:02 +02:00
  • fd56245672 Local: Hermes<->vLLM tool-call investigation + coldcase fixes writeup Alex 2026-07-14 10:39:55 +02:00
  • 06b1cfd58d Local: append save-confirmation result to session report Alex 2026-07-14 01:09:28 +02:00
  • 5b1cdf0393 Local: session report + L1 A/B artifacts + coldcase test notes Alex 2026-07-14 00:37:58 +02:00
  • b03a134925 Local: L1=30 vs L1=60 A/B — the RAM IS working (~1.6-2x faster warm reads) Alex 2026-07-13 21:44:15 +02:00
  • 8269e79e54 Local: L2=0 result — retention is L2 disk, not L1 RAM (0/8 vs 8/8) Alex 2026-07-13 17:37:45 +02:00
  • ffab073ef6 Local: preserve LMCache tuning artifacts + L2=0 watcher (branch-only) Alex 2026-07-13 07:14:49 +02:00
  • 99cbd81ac9 Add LMCache L1 retention test (proves what a bigger L1 buys) Alex 2026-07-13 06:59:12 +02:00
  • c88c188328 Local: raise LMCache shm_size 32→64 for LMCACHE_L1_GB=60 (124 GB rig) Alex 2026-07-12 23:15:37 +02:00
  • 5eff4eb61b docs: refresh expired Discord invite link (FAQ + README + issue template) noonghunna 2026-07-12 20:34:37 +00:00
  • 101077fdf2 docs: cross-link Cliff 3 ↔ LMCache — offload IS the mitigation, not a victim noonghunna 2026-07-12 20:11:27 +00:00
  • 8244b21920 QUALITY_TEST: pass@1 vs pass@N — the churn-harvest ceiling noonghunna 2026-07-12 18:09:19 +00:00
  • 11af94a113 Re-tier tess4-model-floor per Sean's b9967 x3 data + the A0 amendment noonghunna 2026-07-12 15:14:05 +00:00
  • 334b20cf01 Record Tess vLLM A0 baseline: 106 off / 113 on — fallback triggered, slug stays 🧪 noonghunna 2026-07-12 05:48:39 +00:00
  • 5621d9a129 Promote Tess dual llama.cpp compose to Production (refresh: OFF 116 / ON 117) noonghunna 2026-07-12 03:59:05 +00:00
  • f19139fbef nvfp4 composes: document the unsloth (compressed-tensors) provider variant noonghunna 2026-07-12 01:03:31 +00:00
  • be2ce99223 QUALITY_TEST: both-modes probe example + measured scenario-set runtimes noonghunna 2026-07-12 00:55:15 +00:00
  • 8035cb5f65 Wire benchlocal scenario selection + incremental/resume through the quality stack (#683) noonghunna 2026-07-12 05:30:52 +05:00
  • fed2ff75e6 BENCHMARKS: Tess dual b9967 re-bench + EAGLE3-vs-MTP external A/B (MTP wins) noonghunna 2026-07-11 23:09:21 +00:00
  • f9667e9e98 Docs: pin policy trigger (2) — stability pins; fix stale llama.cpp/vLLM pin rows noonghunna 2026-07-11 22:22:15 +00:00
  • c02c28e2d3 Tess dual compose: date the streaming+thinking caveat as non-repro on b9967 noonghunna 2026-07-11 22:14:19 +00:00
  • 6da2456b4d BENCHMARKS: add EAGLE3-GGUF draft provenance to the #674 single-4090 row noonghunna 2026-07-11 21:59:54 +00:00
  • f2345fc4f1 BENCHMARKS: fix stale Tess intro pin (b9246 → b9967, missed by #680) noonghunna 2026-07-11 21:11:35 +00:00
  • d7b5edad42 BENCHMARKS: add single-4090 Tess-4 + Eagle3 8-pack quality row (#674) seanyourhighness 2026-07-11 14:10:40 -07:00
  • fcd8507323 Add engine-pin-bump.sh — mechanical half of an engine image-pin bump (#681) noonghunna 2026-07-12 01:45:07 +05:00
  • ee0624dd1b Bump llama.cpp pin to server-cuda-b9967 across all composes (+4 think-ON) (#680) noonghunna 2026-07-12 01:00:29 +05:00
  • 0b970ab0a7 Add vllm/tess-dual-nvfp4: first vLLM Tess slug (fastest Tess on 2x24GB) (#679) noonghunna 2026-07-12 01:00:25 +05:00
  • 36295beeae Add rerun-failed-packs.sh: re-test only failing packs + flake verdict (#678) noonghunna 2026-07-11 22:26:30 +05:00
  • e8f5c3d689 tess-dual-mtp: record first-party REASONING_BUDGET A/B result in header noonghunna 2026-07-11 15:57:36 +00:00
  • e3f6f02356 tess-dual-mtp: default REASONING_BUDGET=16384 (inert while thinking off) noonghunna 2026-07-11 14:43:10 +00:00
  • 2d7c1d06d7 AGENTS.md: log-file rule for long-running tests (no tail/grep pipes) noonghunna 2026-07-11 14:30:15 +00:00
  • 525f5a00cb c3: render caveats status as an orange checkmark instead of noonghunna 2026-07-11 13:19:46 +00:00
  • 3c84fe5a1f Promote vllm/qwen-35b-a3b-dual-nvfp4-fast: experimental -> caveats (#676) noonghunna 2026-07-11 17:55:41 +05:00
  • 86fb581569 KV scales settled: unsloth checkpoint scales LOAD on the hybrid (and tie) noonghunna 2026-07-11 12:37:58 +00:00
  • 48fc67d692 Add vllm/qwen-35b-a3b-dual-nvfp4-fast: Ampere-validated unsloth NVFP4 MoE slug (#675) noonghunna 2026-07-11 17:10:34 +05:00
  • f0c773e447 Add BENCHMARKS row: unsloth 27B NVFP4 provider A/B — tie with nvidia export noonghunna 2026-07-11 12:00:17 +00:00
  • 87627fdf54 Docs pass: lightweight-path checklist + retire nightly/Genesis examples noonghunna 2026-07-11 11:48:54 +00:00
  • 280a4e8475 Add BENCHMARKS row: unsloth 35B-A3B NVFP4-Fast ties AutoRound tier on 2x3090 noonghunna 2026-07-11 11:06:15 +00:00
  • 8b68541af0 Correct Tess-4-27B arch description: hybrid attention, not uniform GQA noonghunna 2026-07-11 09:36:05 +00:00
  • 7d81184647 Add spec-sweep.sh: standardized draft-depth n-sweep for spec-decode drafters (#673) noonghunna 2026-07-11 12:54:29 +05:00
  • fee2b40340 PROVENANCE: record froggeric v21.3 rejection (hermes-timeout gate failure) noonghunna 2026-07-11 05:04:12 +00:00
  • 76dbc009a9 BENCHMARKS: NVFP4-on-Ampere row (Marlin W4A16 fallback, 2x3090) noonghunna 2026-07-11 01:11:23 +00:00
  • 80f75c0df2 NVFP4: unhide on fallback-capable hardware (fallback_sm + hw_fallback badge) (#670) noonghunna 2026-07-11 05:48:00 +05:00
  • 11c979df9c AGENTS.md: add 'Running a full eval — two non-overlapping passes' to Tests noonghunna 2026-07-11 00:30:08 +00:00
  • eabd8d18b3 concurrency-probe: report aggregate tok/s (verdict + RESULT + sweep) (#669) noonghunna 2026-07-11 02:51:02 +05:00
  • 28863ace42 bench-agentic: don't let a truncated tool call poison the ramp (#665) (#667) noonghunna 2026-07-11 01:27:22 +05:00
  • a81ff368ae Promote qwen-35b-a3b-single-nvfp4 → ⚠️ Production w/ caveats (#666) noonghunna 2026-07-10 04:58:15 +05:00
  • da063d7714 announcement template: add §7 'Run the evals' section noonghunna 2026-07-09 22:23:28 +00:00
  • 30e7cd2a51 rebench-full: replace fragile --help with a self-contained usage() (#664) noonghunna 2026-07-10 01:44:14 +05:00
  • 977d34fe8d Record Tess-4-27B MTP n-sweep: n=2 is the sweet spot noonghunna 2026-07-09 18:35:38 +00:00
  • 8eff77e2be Add Tess-4-27B to the catalog (llama.cpp dual, external MTP, 262K) (#661) noonghunna 2026-07-09 21:45:38 +05:00
  • cd568fc9d6 feat(c3): route-G [D] downloads GGUF files directly, not via pull.sh (#660) noonghunna 2026-07-09 18:01:11 +05:00
  • c2c42e3332 feat(c3): route-G supports embedded MTP, not just external drafters (#659) noonghunna 2026-07-09 17:28:57 +05:00
  • 42238f1452 fix(c3): put repo root on sys.path so ② Serve emit can import scripts (#657) noonghunna 2026-07-09 17:08:34 +05:00
  • b04c1dc3af fix(test): sync test-pull-swap.sh emit asserts with P2b env-gating (#656) noonghunna 2026-07-09 16:58:13 +05:00
  • 158f779292 fix(c3): route-C swap compose — absolute mounts + runtime-dir location (#655) noonghunna 2026-07-09 15:39:54 +05:00
  • 2d6557ca44 fix(c3): route-G GGUF serve — single /models mount + runtime-dir compose (#654) noonghunna 2026-07-09 15:30:17 +05:00
  • 3599fab24d docs(pull): document MODEL_DIR storage convention (curated vs BYO tiers) noonghunna 2026-07-09 10:01:37 +00:00
  • 176c1c3ccf fix(c3): honest "won't fit" verdict + reachable custom-slug hatch (#653) noonghunna 2026-07-09 14:42:44 +05:00
  • f242b6c607 fix(c3): honest ② Serve copy + GB disk-fit labels (post-#647/#649 nits) (#651) noonghunna 2026-07-09 07:42:41 +05:00
  • 6bbeba08a3 feat(c3): GGUF route-G fit + compose emit (#650) noonghunna 2026-07-09 07:35:38 +05:00
  • ed6c80e696 feat(c3): UI/UX phase 3 — download preflight, jobs chip, route language (#649) noonghunna 2026-07-09 07:00:55 +05:00
  • 7be8d6ab59 feat(c3): UI/UX phase 2 — B&V stage hierarchy (state → action → details) (#648) noonghunna 2026-07-09 07:00:10 +05:00
  • 6d7ce6c18c fix(c3): UI/UX phase 1 — serve honesty, footer labels, promote preview (#647) noonghunna 2026-07-09 06:52:35 +05:00
  • 3036766ae8 deriver: resolve HF_HOME from MODEL_DIR so bare pull.sh lands on the model disk (#646) noonghunna 2026-07-09 04:40:20 +05:00
  • a12bf4c9ca deriver: include a dedicated MTP head in the weight download set (#645) noonghunna 2026-07-09 03:05:15 +05:00
  • c25b9947ff c3+pull: detect in-progress downloads + per-repo lock (#617) (#644) noonghunna 2026-07-09 02:53:50 +05:00
  • 431fe983b4 c3: bring [D] download parity with the catalog path (#617) (#643) noonghunna 2026-07-09 00:05:24 +05:00
  • 8ae205bcce preflight: warn on single-card GPU_MEMORY_UTILIZATION override above default (#641) noonghunna 2026-07-08 23:18:33 +05:00
  • 44734bd42d feat(c3): ② Serve spec-dec — engine-driven drafter SELECTOR (not just on/off) (#638) noonghunna 2026-07-08 12:56:50 +05:00
  • 662fde03d3 fix(ik-llama): driver-aware cu13/cu12 image select + preflight hint + docs (#633) (#640) noonghunna 2026-07-08 12:52:52 +05:00
  • c4c72018ab fix(launcher): GGUF verify-glob (#634) + .env engine-image passthrough (#632) (#639) noonghunna 2026-07-08 12:52:47 +05:00
  • 7cffc84d14 feat(c3): ② Serve editor — custom values, engine KV, real spec, preview (#637) noonghunna 2026-07-08 09:41:03 +05:00
  • dbfbd598e7 feat(c3): ② Serve override editor — retune served-name/ctx/KV/spec/util (#636) noonghunna 2026-07-08 08:49:51 +05:00
  • 9cddcf6228 fix(c3): ② Serve honest for Route-C + [s] advance key (kill the dead-end) (#635) noonghunna 2026-07-08 08:23:01 +05:00
  • 3f1456d016 fix(c3): ⏎ on ① Bring advances to ② Serve when weights present (#631) noonghunna 2026-07-08 07:27:18 +05:00
  • 97d3523a9d fix(c3): hide [D] when brought weights on disk; ② Serve emits swap (#630) noonghunna 2026-07-08 07:12:18 +05:00
  • bc815b3272 feat(byo): apply-swap — download + serve a Route-C fine-tune (#628) noonghunna 2026-07-08 05:50:16 +05:00
  • 7d16e2763f Merge pull request #629 from noonghunna/fix/nvfp4-27b-mtp-65k-derate noonghunna 2026-07-08 05:36:22 +05:00
  • f4cbd5a5a2 Re-derate 27B nvfp4 single default: SPEC=on + 65K (keep MTP's 2x) noonghunna 2026-07-08 00:35:04 +00:00
  • e3e114f7fe Merge pull request #627 from noonghunna/fix/byo-swap-detect-mtp-head noonghunna 2026-07-08 03:38:43 +05:00
  • fdbd68bc3a Detect MTP head in BYO Route-C swap instead of blanket spec-drop noonghunna 2026-07-07 22:38:19 +00:00
  • fbf6ab9022 fix(nvfp4): disable deepgemm on consumer cards noonghunna 2026-07-07 22:05:09 +00:00
  • a5778484fe Merge pull request #624 from noonghunna/feat/c3-spec-dec-column noonghunna 2026-07-08 03:01:12 +05:00
  • 800642ec6b Merge pull request #626 from noonghunna/feat/qwen-multi-max-production noonghunna 2026-07-08 03:01:08 +05:00
  • 10f5b55adb Promote vllm/qwen-27b-multi-max to Production (v0.24.0 4-card gate) noonghunna 2026-07-07 21:56:10 +00:00
  • 23fc84f079 docs(qwen): sync multi-max fp8 kv header noonghunna 2026-07-07 21:10:19 +00:00
  • a62784c132 feat(c3): add Spec Dec column to Catalog + fold into funnel labels noonghunna 2026-07-07 17:16:08 +00:00
  • 233e6c0bdd Merge pull request #623 from noonghunna/feat/nvfp4-27b-spec-toggle noonghunna 2026-07-07 17:24:07 +05:00
  • f50db81ee8 feat(nvfp4): SPEC=off toggle for the 27B nvfp4 composes (#617) noonghunna 2026-07-07 12:24:00 +00:00
  • 5e61d74e09 Merge pull request #622 from noonghunna/data/nvfp4-35b-corroboration noonghunna 2026-07-07 17:03:48 +05:00
  • ef51f7778f data(nvfp4): 35B-A3B 5090 baseline — corroborate #619 with #612 (120K NIAH) noonghunna 2026-07-07 12:03:41 +00:00
  • c903d48278 Merge pull request #621 from noonghunna/data/nvfp4-moe-single-first-validation noonghunna 2026-07-07 16:57:17 +05:00
  • c3f32f8fd6 data(nvfp4): first NVFP4 validation — 35B-A3B single on RTX 5090 (#619) noonghunna 2026-07-07 11:57:11 +00:00
  • 1cf3da1208 Merge pull request #620 from noonghunna/feat/rename-cluster-to-pod noonghunna 2026-07-07 07:10:23 +05:00