noonghunnaandClaude Opus 4.8 3a5ece7bb5 gemma-26ba4b-single: promote INT8-PTH single → ⚠️ Production-w/-caveats (gate PASS) ...

rebench-full (gemma-26ba4b-int8r, 1x 3090 @370W) PASSED every gate:
  verify-full ✓ · bench 168 narr / 217 code TPS (MTP AL 3.06-3.79) ·
  verify-stress NIAH-clean → 161K (91% of 176K) ·
  quality 109/150 think-ON (98/150 off) — on par with gemma-4-31B's 107/150 ·
  soak 20x5 PASS, 0 MiB growth.

Flip Status 🧪 Experimental → ⚠️ Production-w/-caveats in the int8.yml header
(+ Quality line) and compose_registry (status experimental→caveats + gate-result
note). Add the BENCHMARKS.md row. Caveats: needs the #40391 overlay; 176K @
mem_util 0.94 (262K only without the MTP drafter — drafter weights + cudagraph
capture cost ~86K tok, 0.96 OOMs the capture tail); think-OFF agentic/extraction
softer (cli-40 30% / DataExtract 60%, recover to 52%/73% with thinking). Gate 42/42.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

2026-06-06 14:17:50 +00:00

..

calibration

Wire gemma-4-26b-a4b AWQ on stock vLLM v0.22.0; retire AutoRound (#326)

2026-06-06 00:46:13 +00:00

drafters

Wire gemma-4-12b into curated catalog (vLLM gemma4-unified, bf16 + MTP)

2026-06-04 01:17:19 +00:00

engines

Repoint vllm/gemma-26ba4b-single to INT8-PTH long-ctx (#465)

2026-06-06 11:30:27 +00:00

hardware

feat(beellama): add beellama.cpp DFlash as first-class compose engine (#268)

2026-05-30 22:47:19 +05:00

models

kv-calc: per-sequence KV-pool floor (capped at 1 GB), fix KV-light false-FAIL

2026-06-06 01:41:10 +00:00

workloads

feat(profiles): ship v0.7.0 data layer

2026-05-14 14:46:42 +00:00

__init__.py

feat(profiles): ship v0.7.0 data layer

2026-05-14 14:46:42 +00:00

arch_patches.yml

Wire gemma-4-26b-a4b AWQ on stock vLLM v0.22.0; retire AutoRound (#326)

2026-06-06 00:46:13 +00:00

booter.py

v0.8.0 [F] F3: G6-A 3-part additive [E] touch (pt1.predicted_b_breakdown, pt3.failure_log_excerpt+actual, container-log capture) + §6.1 Tier-1 (CONTRACT-2)

2026-05-17 04:19:53 +00:00

calibration_seed.yml

Add v0.8 Phase A patch attribution data

2026-05-16 05:13:00 +00:00

canonical_scenarios.py

feat(compat): add profile validator and estate self-test

2026-05-14 15:25:28 +00:00

capture.py

feat(pull): v0.8.2 STEP V1 — capture-on-hard-block pt1-gate emitter + BaseCaptureBundle protocol lift

2026-05-18 16:43:28 +00:00

classifier.py

feat(pull): v0.8.2 STEP V1 — capture-on-hard-block pt1-gate emitter + BaseCaptureBundle protocol lift

2026-05-18 16:43:28 +00:00

compat.py

gemma-4-12b: add single-card vLLM INT8 compose (full 256K on one 3090)

2026-06-04 17:16:00 +00:00

compose_registry.py

gemma-26ba4b-single: promote INT8-PTH single → ⚠️ Production-w/-caveats (gate PASS)

2026-06-06 14:17:50 +00:00

dedup.py

feat(pull): v0.8.2 STEP V1 — capture-on-hard-block pt1-gate emitter + BaseCaptureBundle protocol lift

2026-05-18 16:43:28 +00:00

deriver.py

fix(deriver): correct stale "GGUF not supported until v0.8.1" message — now misleading post-v0.8.1-ship

2026-05-18 11:38:12 +00:00

diagnose_profile_cli.py

gemma-4-12b: drop vendored p-RoPE overlay — upstream config fix supersedes it

2026-06-04 16:04:24 +00:00

downloader.py

v0.8.0 [E] E2-fix-2: verify *.safetensors against HF API lfs.sha256 (not Xet-redirect-fragile HEAD x-linked-etag) — on-rig E5 caught false no-etag

2026-05-17 01:28:22 +00:00

einput.py

v0.8.0 [E] E1: generate_from_profile + derived-vllm template + EInput + CONTRACT-5 gate

2026-05-17 00:12:26 +00:00

estate_cli.py

feat(estate): add parallel boot mode

2026-05-15 22:18:09 +05:00

failure_fingerprints.yml

feat(pull): v0.8.2 STEP V1 — capture-on-hard-block pt1-gate emitter + BaseCaptureBundle protocol lift

2026-05-18 16:43:28 +00:00

gates.py

llama.cpp single: thinking-off policy alignment + MTP profile family

2026-05-20 01:01:26 +00:00

hwdetect.py

feat(pull): v0.8.2 STEP V4 — optional whichllm hw-detect subprocess (CONTRACT-3, hw-detect-only)

2026-05-18 19:29:43 +00:00

launch_compat.py

feat(beellama): release v0.3.0 Q8_K_XL dual composes (experimental) + centralize image pin (#296)

2026-06-02 16:35:48 +05:00

loop_input.py

feat(pull): v0.8.2 STEP V1 — capture-on-hard-block pt1-gate emitter + BaseCaptureBundle protocol lift

2026-05-18 16:43:28 +00:00

measurement_record.py

Fail loud on hollow measured records in measurement_record (#251)

2026-05-29 03:31:25 +05:00

patch_attribution.py

feat(pull): v0.8.2 STEP V3 — arch-registry expansion + chat-template attribution/drift_guard

2026-05-18 18:29:58 +00:00

patches.yml

Repoint vllm/gemma-26ba4b-single to INT8-PTH long-ctx (#465)

2026-06-06 11:30:27 +00:00

profile_runtime.yml

Repoint vllm/gemma-26ba4b-single to INT8-PTH long-ctx (#465)

2026-06-06 11:30:27 +00:00

pull.py

fix(pull): v0.8.2 STEP V5 — recommend must not label a fits-clean model "DOES NOT FIT"

2026-05-18 21:05:12 +00:00

submit_pull.py

fix(pull): v0.8.2 STEP V2 — gh-less issue body must not carry the absolute capture path

2026-05-18 17:49:00 +00:00

trust_pipeline.py

v0.8.0 [review] pre-tag fixes: scrub internal-path leaks from shipped source + make .pull-captures-corpus tests CI-safe (skip-when-absent)

2026-05-17 12:01:25 +00:00

weights.py

refactor(compose): insert <quant> layer + make the registry the single source of truth

2026-05-26 15:58:08 +00:00