3 people 3c84fe5a1f Promote vllm/qwen-35b-a3b-dual-nvfp4-fast: experimental -> caveats (#676) ...

Full gate complete on the reference 2x3090: verify-stress 8/8 (NIAH
clean to 240,635 tok = 91% of 262K, ceiling margin 1,801 MB, zero drift
across the ladder) + soak-continuous PASS (0 errors, 0/25 silent-empty,
0 VRAM growth, p50 215.9, 100% retention), joining the existing bench
(179.5/179.4 decode) and 8-pack (103/150 tie) results.

Caveats (why not full Production): streaming tool-calls + thinking-on
hit finish=length (known family class, verify-full check 6); native-FP4
quality remains unvalidated (all numbers are the Ampere W4A16 bound).


Claude-Session: https://claude.ai/code/session_01EfF565T9eSLaqGzidyJ1Pm

Co-authored-by: noonghunna <[email protected]>
Co-authored-by: Claude Fable 5 <[email protected]>

2026-07-11 17:55:41 +05:00

..

fixtures

serve: neutral primary model name (qwen3.6-27b / gemma-4-31b), keep -autoround alias

2026-07-02 11:28:31 +00:00

lib

Promote vllm/qwen-35b-a3b-dual-nvfp4-fast: experimental -> caveats (#676)

2026-07-11 17:55:41 +05:00

maintenance

…

systemd

tools: add HTTP model-switch service (thin wrapper over switch.sh)

2026-07-03 17:44:17 -07:00

tests

Add vllm/qwen-35b-a3b-dual-nvfp4-fast: Ampere-validated unsloth NVFP4 MoE slug (#675)

2026-07-11 17:10:34 +05:00

arch-ab.sh

fp8w on Blackwell: auto-disable DeepGEMM + --force the arch-ab arm

2026-07-05 07:21:34 +00:00

beellama-pin-bump.sh

…

bench-agentic.sh

bench-agentic: don't let a truncated tool call poison the ramp (#665) (#667)

2026-07-11 01:27:22 +05:00

bench.sh

catalog-baselines slice 2c: canonical two-depth prefill/TTFT probe + anchor calibration

2026-07-04 16:22:48 +00:00

c3t

Add c3t — the club-3090 test-console TUI (built by Qwen Max from the spec) (#428)

2026-06-18 07:31:02 +05:00

catalog-baseline.sh

baselines: quality_env harness provenance on quality rows (friction #8)

2026-07-05 09:55:54 +00:00

concurrency-probe.sh

concurrency-probe: report aggregate tok/s (verdict + RESULT + sweep) (#669)

2026-07-11 02:51:02 +05:00

detect_nvlink.sh

detect_nvlink: auto-enable PCIe P2P when nvidia-smi reports it OK

2026-06-24 14:07:06 +00:00

diagnose-estate.sh

…

diagnose-profile.sh

…

generate-compose.sh

…

gpu-mode.sh

c3/gpu-mode: preview-clip fix + scene-boundary hardening (F2+F5, T1-lite tail)

2026-07-04 11:12:32 +00:00

health.sh

Phase 2b: add --json/CLI contracts to 7 stack scripts (additive)

2026-06-18 08:43:00 +00:00

launch.sh

fix(launcher): GGUF verify-glob (#634) + .env engine-image passthrough (#632) (#639)

2026-07-08 12:52:47 +05:00

pod.sh

refactor(pods): rename cluster → pod (#610) + heterogeneous-rig guidance

2026-07-07 02:10:17 +00:00

power-cap-sweep.sh

…

preflight.sh

preflight: warn on single-card GPU_MEMORY_UTILIZATION override above default (#641)

2026-07-08 23:18:33 +05:00

pull.sh

c3+pull: detect in-progress downloads + per-repo lock (#617) (#644)

2026-07-09 02:53:50 +05:00

quality-baseline.sh

…

quality-test.sh

fix(quality): precise benchlocal-cli source (sha/describe) — version is blind

2026-07-06 22:02:21 +00:00

rebench-full.sh

rebench-full: replace fragile --help with a self-contained usage() (#664)

2026-07-10 01:44:14 +05:00

rebench-report.py

…

rebench-runtime.sh

…

report.sh

fix(quality): precise benchlocal-cli source (sha/describe) — version is blind

2026-07-06 22:02:21 +00:00

setup-ai-studio.sh

setup-ai-studio.sh: make the "sign up then install the pipe" step unmissable (#510) (#527)

2026-06-30 00:08:52 +00:00

setup-image-studio.sh

studio: one-command setup-ai-studio.sh + auto-install the OWUI pipe

2026-06-24 03:47:18 +00:00

setup-video-studio.sh

studio: one-command setup-ai-studio.sh + auto-install the OWUI pipe

2026-06-24 03:47:18 +00:00

setup.sh

fix(launcher): GGUF verify-glob (#634) + .env engine-image passthrough (#632) (#639)

2026-07-08 12:52:47 +05:00

soak-helper.py

serve: neutral primary model name (qwen3.6-27b / gemma-4-31b), keep -autoround alias

2026-07-02 11:28:31 +00:00

soak-test.sh

…

spec-sweep.sh

Add spec-sweep.sh: standardized draft-depth n-sweep for spec-decode drafters (#673)

2026-07-11 12:54:29 +05:00

stream-toolcall-probe.py

serve: neutral primary model name (qwen3.6-27b / gemma-4-31b), keep -autoround alias

2026-07-02 11:28:31 +00:00

submit-bench.sh

…

switch.sh

preflight: warn on single-card GPU_MEMORY_UTILIZATION override above default (#641)

2026-07-08 23:18:33 +05:00

update.sh

…

verify-full.sh

serve: neutral primary model name (qwen3.6-27b / gemma-4-31b), keep -autoround alias

2026-07-02 11:28:31 +00:00

verify-stress.sh

STRESS_FAST mode + honest dual-rig VRAM margins (dogfood findings)

2026-07-04 22:22:54 +00:00

verify.sh

serve: neutral primary model name (qwen3.6-27b / gemma-4-31b), keep -autoround alias

2026-07-02 11:28:31 +00:00