license: apache-2.0 library_name: llama.cpp pipeline_tag: image-text-to-text base_model:

• Qwen/Qwen3.8-27B
• JonathanColetti/Qwen3.8-27B-Uncensored
• mradermacher/Qwen3.8-27B-Uncensored-i1-GGUF
• incoai/Qwen3.8-27B-DFlash2 language:
• en
• zh tags:
• gguf
• qwen3.8
• qwen35
• uncensored
• abliterated
• speculative-decoding
• dflash2
• loopspec
• pascal
• sm_60
• gp100
• q4_1
• q4_1_g64
• vision

Qwen3.8-27B Uncensored — Q4_1_G64 4.5 bpw for GP100

Q4_1_G64 · 4.5 bpw · 15.0 GiB · +10% no-spec vs Q4_1 · 128k wall −4.8%

GGUF pack of the JonathanColetti abliteration for 2× Tesla P100 (sm_60).

The served target is type-43 Q4_1_G64: mradermacher i1 Q4_1 body, output head left Q8_0, other Q4_1 weights refit to one (d,m) per 64 weights (4.5 bpw) with the local imatrix. Same HFMA2 decode as Q4_1; fewer DRAM sectors (36 B / 64 w vs 20 B / 32 w). The Q4_1 parent stays in the repo so stock llama.cpp still has a file it can load.

Why this type, and how a decode round actually spends time: DESIGN.md. It walks GP100 vs GP104 (Pascal Tuning Guide + Arafa), the lab DECODE_TIMELINE.md trust tags, why Q4_K is a 2.3× tax at the same 4.5 bpw, why G64 is a storage grouping not a new inner product, and which levers are closed.

The point of G64 is speed, not extra context. On the realistic vibe bar — a 128K session that grows, then compacts — G64 finishes first:

G64 wins by 76.86 s (−4.8%). Q4_1 on --graph-slots 8 OOMed CUDA1 (~436 MiB) at the ~110k regrow; the Q4_1 number above is the run that finished.

Do not quote 100 tok/s exact-reuse, or a short HumanEval wall, as the headline. Those are below.

Two paths

Path A — stock llama.cpp. Load the Q4_1 target + Q4_1 DFlash + F16 mmproj. G64 is type 43; stock llama.cpp refuses it cleanly. Path A does not include LoopSpec (ngram-mod,draft-dflash) or the speed table.

Path B — this fork, latest rebase. thefallentree/llama.cpp-gp100 branch cursor/gemma4-mtp-host-6006 @ 958bce0bb (onto ggml-org master 930e2fa59, 2026-09-16). LoopSpec, sm_60 HFMA2, slot-release, empty-src sampling, GDN checkpoint eviction. Do not freeze on a8c9f3a7b / gp100-qwen38-g64slot (2026-09-05 archive). Do not track pr-sm60 HEAD — commit 31 (19d05ddb7) aborts Qwen LoopSpec in cpy.cu.

sm_60 only. P40 / GTX 10-series are sm_61; the HFMA2 path is gated off there.

Files

Why / how: DESIGN.md. Convert any model + serve recipe: CONVERT.md. Header dump: TENSORS.md. Bench protocol + raw rows: BENCH.md.

The G64 DFlash draft is held. Production uses the Q4_1 draft.

Speed (Path B, 2× Tesla P100)

Same coding fixture, 512 gen, ignore-eos, GPU already warm. Token budgets from POST /tokenize. Prefill and decode are separate. Restart between sizes so novel ngram is clean.

Headline: 128K session + compact

tools/bench_vibe_128k_compact.py: warmup, grow 1k → pin-keep → 16k → 64k → 128k, compact (keep 4k + tail 16k), regrow ~110k, compact again. Score is the sum of novel walls. Prefix cache on (token-ID prefixes, --slot-prompt-similarity 0.05, --ctx-checkpoints 8 --checkpoint-min-step 0). mmproj is loaded (disables ctx_shift).

G64 1532.79 s vs Q4_1 1609.65 s (Q4_1 needed --graph-slots 1).

Exact-reuse decode (G64 constant-w8 vs Q4_1 cap 7)

Same fixture, same g64slot binary, restart between sizes. G64 reuse is faster at 16k and 64k.

No-spec llama-bench (same g64slot binary)

llama-bench -p 512 -n 128, tensor split, FA on, Q8 KV. G64 is faster at every depth:

Honest short-prompt loss

G64 constant-w8 (DFlash-7 / ngram 7) loses the short novel-wall bar. Publish it:

1k accept 0.286 vs Q4_1 ~0.53 — DFlash-7 rejects most drafts on a short prompt. Do not serve cap 3 / DFlash-2 to hide this; those arms do not beat Q4_1 on the 128k bar.

Wiki PPL (shipped files)

Same g64slot llama-perplexity, eval/wiki.test.raw, 40 chunks, tensor split:

Under the +2% publish gate. Older G64 emulation (different GGUF) was +1.0% (6.113 vs 6.051).

Run (Path B)

git clone https://github.com/thefallentree/llama.cpp-gp100
cd llama.cpp-gp100 && git checkout cursor/gemma4-mtp-host-6006   # 958bce0bb; not g64slot
# build llama-server with CUDA (sm_60)

./build/bin/llama-server \
  --model Qwen3.8-27B-Uncensored.i1-Q4_1_G64-Q8_0-head.gguf \
  --model-draft Qwen3.8-27B-Uncensored-DFlash2-Q4_1-private-q4-head-exact-embd.gguf \
  --mmproj mmproj-Qwen3.8-27B-Uncensored-F16.gguf \
  --jinja \
  --spec-type ngram-mod,draft-dflash \
  --spec-draft-n-max 7 --spec-draft-n-min 0 \
  --spec-ngram-mod-n-min 7 --spec-ngram-mod-n-max 7 \
  --backend-sampling --graph-slots 8 \
  --ctx-checkpoints 8 --checkpoint-min-step 0 \
  --slot-prompt-similarity 0.05 \
  --device CUDA0,CUDA1 --split-mode tensor --tensor-split 1,1 \
  --device-draft CUDA1 \
  --cache-type-k q8_0 --cache-type-v q8_0 \
  --cache-type-k-draft q4_0 --cache-type-v-draft q4_0 \
  --ctx-size 131072 --flash-attn on

Path A: same command with the Q4_1 target, --spec-draft-n-max 3, --spec-ngram-mod-n-min 3 --spec-ngram-mod-n-max 7, and --graph-slots 1 if you want the 128k+compact path to finish on 2×16 GB.

Context

Default is 131072 + Q8_0 K/V. Native qwen35.context_length is 262144.

G64 saves ~2 GiB of weights. That does not buy a larger Q8 window for multi-turn chat. 163840 and 196608 Q8 decode the first high-n_kv turn and OOM CUDA1 on the cached second turn (leftover FA workspace stays resident). 256K Q8 loads but idle free is ~598 MiB.

Capacity mode is 256K with Q4_0 KV (./run.sh qwen-long in the lab scripts). Q4 KV PPL was +0.67% vs Q8 on an earlier measurement. Do not raise Q8 past 131072 on 2× P100.

Why / how (short)

GP100 is not GP104: two schedulers/SM, no DP4A, HFMA2 is 2× FP32 only as half2, I2F/F2F are quarter-rate, HBM2 ~606 GB/s measured, PCIe only. Q4_K is the same 4.5 bpw as G64 and a ~2.3× decode-round tax here because its scales are extracted every row iteration. Q4_0’s 18-byte block fails 4-byte alignment on half its blocks. G64 is Q4_1’s aligned (d,m) + nibbles stretched to 64 weights (36 B / 64 w). Same A16 HFMA2 inner product; fewer DRAM sectors.

A decode round on this pair is accept (~10 µs) + draft on CUDA1 (GPU0 idle) + issue-bound verify. ~600 launches/round are real; their cost is 0.64 ms. CUDA graphs replay launches and measured flat. The legal speed clock is predicted_per_token_ms, not the old wall-clock slot harness.

Full write-up, with the Pascal Tuning Guide, Arafa, DECODE_TIMELINE.md trust tags, and the lab architecture reviews: DESIGN.md.

To G64 a different model (not this 27B pack): CONVERT.md — scripts/g64/convert_to_g64.py inspect|convert|serve on the fork. Not Qwen-specific. Default mix is G64 body / Q8_0 head / Q4_1 token_embd.

Architecture

qwen35, 27B, 65 blocks (48 gated-attn+SSM, 17 full-attn), d_model 5120, FFN 17408, vocab 248320, context 262144, native MTP on blk.64. Dense.

HFMA2 takes Q4_1 / Q4_1_G64 / Q4_K / Q8_0 with ne[0]%32==0, ne[1]%4==0, weight ne[2]==1. Every quantized 2-D weight here passes.

Related work (not this stack)

Minima (2026-09) reports NVFP4 W4A4 on Qwen3.8-27B with vLLM + FP8 KV on Blackwell SM120. The overlapping claim is that GDN linears are the easy half to quantize — this G64 pack already 4-bit those weights, gates included. It is not NVFP4, not FP8 KV, and P100 cannot run that recipe. Do not copy those numbers onto this card.

Provenance

License

Apache 2.0, inherited from Qwen/Qwen3.8-27B. Local inference and research. Abliterated / uncensored — not a safety-tuned deployment model.