vh
74f596b1d3
Re-quantizes the fleet `gen` seat from weight-only NVFP4A16 to a mixed-precision build: NVFP4 W4A4 for layers 0-55 MLPs, FP8 W8A8 for the attention projections / linear_attn / lm_head / layers 56-63 MLPs, FP8 KV cache. Replicates the scheme of unsloth/Qwen3.8-27B-NVFP4 on the abliterated weights. The queued task named this "W4A8" (NVFP4 weights + FP8 activations). That checkpoint cannot be served: vLLM 0.24's compressed-tensors dispatcher (compressed_tensors.py:704-713) accepts NVFP4 weights with either no input quantization (W4A16, which forces the Marlin kernel) or NVFP4 input quantization (W4A4) -- anything else, FP8 included, raises ValueError at load. CompressedTensorsW4A8Fp8 is INT4 weights gated on an exact-sm90 check, so it is closed on Blackwell twice over. The ~20% intuition was correct; the scheme name was not. Getting FP8 into the mix has to be done per-layer-group. Established the gain before spending GPU time: unsloth's build was already on-box, so serving it as a probe measured +19.1% over our seat at identical MTP acceptance -- a kernel-level result, no requant needed to learn it. Measured, cache-busted, bs=1: decode 80.12 -> 94.53 tok/s (+18.0%) MTP acceptance 47.8% -> 47.7% (unchanged) perplexity (n=6) 6.941 -> 7.059 (+1.7%) abliteration 4/4 -> 4/4 (preserved) weights on disk 27.7 -> 22.5 GB (-19%) Surface test green on the live seat: plain chat, vision, tool calling, thinking split, 36K-token needle retrieval, streaming. All 7 LiteLLM aliases verified routing. GEN_GPU_MEM_UTIL 0.45 -> 0.43: the new weights are 5.2 GB smaller, and at 0.45 the seat absorbed that slack as KV, leaving meromero-charrp 0.18 GiB short of its budget on the shared GPU0 -- it crash-looped. Handing the space back leaves gen 422K tokens of KV (1.6x its 262K context) and both seats co-resident at 89.8/97.9 GB. Also records two measured negatives so they are not re-chased: GEN_SPEC_TOKENS is already optimal at 3 (swept 2/3/4/5 -> 77.1/80.1/78.7/ 75.9 tok/s), and vLLM's prompt_logprobs are ~uniform while speculative decoding is on, so perplexity must be measured with spec off. Pipeline, acceptance harness and raw measurements land in services/gen-seat-mixed-quant/. Rollback is one .env line; the previous build is untouched at /tank/aimodels/qwen38-27b-uncensored-nvfp4.
1.5 KiB
1.5 KiB
gen-seat — the fleet gen seat (ana-ml2 GPU0, :8015)
Serves qwen3.8-27b-uncensored (JonathanColetti/Qwen3.8-27B-Uncensored,
Heretic-abliterated Qwen3.8-27B, vision-intact, 262K context) plus the
-thinking served-name for the reasoning split.
Backs 7 LiteLLM aliases: gen, gen-reasoning, summarizer,
summarizer-large, classifier, image-judge, qwen-image-bench. Treat any
change here as fleet-wide.
- Quant: in-house mixed-precision — NVFP4 W4A4 for layers 0-55 MLPs,
FP8 W8A8 for attention /
linear_attn/lm_head/ layers 56-63 MLPs, FP8 KV. ~94.5 tok/s decode at bs=1, MTP n=3 @ ~48% acceptance. - Speculative decoding:
qwen3_5_mtp,num_speculative_tokens=3(measured optimum — see the quant service README). - Tunables:
.envon the host. ⚠ it is mode 0600 / lkraven-owned, so everydocker composecall against this stack needssudo— without it compose cannot read.env, fails withpermission denied, and leaves the old container running while appearing to have succeeded.
Pipeline, acceptance gate, and rollback
All of it — the recipe, the re:^mtp.* foot-gun, the benchmark harness, the GPU0
budget interaction with meromero-charrp, and the rollback command — lives in
services/gen-seat-mixed-quant/README.md.
Deploy
scripts/deploy-stack.sh ana-ml2 gen-seat # diffs vs live, prompts y/N
# on host:
ssh infra-ops@10.250.50.54
cd /opt/docker/compose/gen-seat && sudo docker compose up -d vllm-gen