Qwen3.8-27B-ColdFusion-GAIN-Blackwell-DFlash2-Ultra-V1.0

2.4× faster per task · up to 200+ tok/s · HumanEval 84% → 94% · GSM8K 94% → 96%

A 27B hybrid text + image model set up to run fast on one RTX 5090 (32 GB, sm_120) with llama.cpp CUDA and DFlash2 speculative decoding. 262K context.

Update 2026-09-11 — new measured data: CUDA graphs are worth +33 % and can stay enabled (this page previously advised disabling them, which costs a third of the decode throughput), a reasoning-effort sweep on GSM8K, the real NVFP4-vs-Q4_K_M quality/speed trade, a full speculative-decoding knob sweep, and the multi-agent cache fix. Everything below is measured on the same box, with the conditions stated next to each table.

The gain, concretely — time and speed

Same machine (RTX 5090 32 GB + 128 GB), same prompts, greedy. Base = unsloth/Qwen3.8-27B-GGUF UD-Q4_K_M as-is (no speculation).

A full HumanEval-50 run: 9m25s → 4m00s. A GSM8K-50 run: 5m16s → 2m06s.

Accuracy — held, not sacrificed

NEW — Reasoning effort: measured on a stratified GSM8K-100

reasoning_effort is a chat-template kwarg. It changes how much thinking the model does before it answers, which dominates wall-clock on real tasks far more than any kernel tuning.

Benchmark: 100 GSM8K problems, 5 slices across the whole 1319-problem test set (offsets 0/300/600/900/1200, 20 each), temperature 0, answers checked numerically, max_tokens 3072. Config: this model + DFlash2, -np 1 -c 65536 -fa on, KV q8_0.

Read this before picking a mode:

• medium is strictly better than low here: +2 points and 6% faster (it wrote fewer tokens). The default of this model is xhigh; if you never set the kwarg you are running the slowest mode for no measured accuracy benefit on this benchmark.
• Use --chat-template-kwargs "{\"reasoning_effort\":\"medium\"}" on the server so clients get it by default. A client that sends reasoning_effort explicitly overrides it.
• Keep max_tokens ≥ 4096 on hard tasks in low/medium; and note that a low effort can occasionally loop without concluding (observed once on a debugging task: 4096 tokens of reasoning, 0 tokens of answer — the same task answered fine in medium).
• A 4-task hard bench (temp 0, bug hunt / scheduling / system diagnosis / Bayes) done during the same session: GAIN answered 7 of 8 faster than Fable (e.g. 9.5 s vs 11.8 s on the scheduling problem, both finding the optimal 13 h makespan; Bayes posterior 1.94% correct in both; the diagnosis task was answered by GAIN and timed out in Fable low). Both models got the ambiguous diagnosis task wrong — this is not a "perfect" model, it is a fast one.

NEW — NVFP4 vs Q4_K_M: the actual trade

Both builds of the same finetune, same drafter, same everything else. Perplexity: 24 chunks × 2048 ctx, KV in f16 to isolate the weight effect, llama-perplexity on a mixed code/markdown corpus.

+15.5% decode for +2.7% perplexity. Measured twice, interleaved, same window.

Why NVFP4 costs quality at all — it is structural, not a bad quant: on Blackwell the NVFP4 path is W4A4 (the activations are also FP4) while Q4_K_M runs A16. See llama.cpp PR #24364 ("force W4A8 path for NVFP4_W4A16 layers on Blackwell, where NVFP4 normally uses the native W4A4 path"). So high-quality NVFP4 recipes need either W4A8/A16 dispatch or more high-precision tensors — which costs back the bytes you saved.

Related measurements from the same session (Fable, requantized to NVFP4 by hand):

Quantizer notes worth knowing if you build your own: the imatrix-aware NVFP4 scale search is upstream PR #25153 (discrete search over ±12 UE4M3 codes around amax/6, optionally importance-weighted). A weighted-LSQ / coordinate-descent scale solve was tried and is worse (+3.0% perplexity here), matching the PR author's own finding. The single best quality-per-byte move found was putting output.weight in Q4_K rather than Q8_0: it removes 3.6 ms from every draft phase (the MTP/DFlash head is read once per draft step) for +0.14% perplexity.

NEW — CUDA graphs: +33 %, and they can stay ON

This page used to advise GGML_CUDA_DISABLE_GRAPHS=1 to dodge a CUDA illegal-memory-access crash seen on some builds. Measured on this box, that advice costs a third of the decode throughput.

Agent profile (temperature 0, reasoning_effort=medium, 400-token code generation, best of 2 runs, two rounds, repeatability ±0.2 %):

+33 % decode, and the entire gain sits in the verify pass: 33.4 → 24.1 ms per step, i.e. ~9 ms of every forward pass was pure kernel-launch latency (an eager 27B graph is on the order of a thousand launches; a captured graph replays them in a single submission). The draft pass barely moves (4.66 → 3.91 ms) because the DFlash2 drafter is not graph-captured — so there is still ~2–3 ms of launch overhead left on the table if the drafter's graph were captured too.

Stability (the original concern): a sustained run with graphs ON and 2 slots alternating (the worst case for graph re-capture) pushed 7200 tokens at 121.1 tok/s with zero CUDA errors and a healthy server, on the stock llama-cuda build. Together with the rest of this session (200+ timed runs, a 100-problem GSM8K sweep, perplexity runs, multi-slot tests) on a second build, graphs never caused a failure here.

Leave graphs enabled (do not set the variable). If you do hit an illegal memory access on your build, GGML_CUDA_DISABLE_GRAPHS=1 is a fallback — you are trading 33 % of throughput for a build-specific bug, and it is worth reporting upstream instead.

NEW — Speculative decoding: what actually helps

Agent profile: temperature 0, reasoning_effort=medium, 400-token code generation, best of 2 runs, two full rounds (repeatability was ±0.2%).

Two findings that matter more than the flags:

1. n_max = 7 is already the optimum (it is the drafter's block size; higher values are clamped, lower values break its block path — n=4 is slower).
2. The verify pass is sub-linear: ~1.1 ms per additional verified token at ~24 ms for the first column. Extra drafted columns are nearly free — the ceiling is the acceptance rate, not the draft cost. Mean accepted length measured: 2.08 tokens/step on code, 2.9–3.6 on agent/reasoning prompts. This is why drafter quality matters and drafter speed barely does.

NEW — Multi-agent / long sessions (the biggest single win)

If several agents share the server (e.g. an IDE agent plus a CLI agent), the default prompt cache is too small and the KV of the idle slot is dropped — the agent pays a full re-prefill when it comes back:

slot.prompt_clear: id 1 | erasing 0 tokens     # every turn

Fix (host RAM, zero VRAM cost):

--cache-ram 32768 --no-cache-idle-slots

Measured on a 140k-token context pair of agents: agent return 8.62 s → 0.32 s, and in a full run 78 s → ~1 s. -np 1 is the default recommendation; use -np 2/-np 4 only with the flag below, since the drafter's context is not unified across slots.

Two validated configurations (VRAM measured with the NVFP4 model + DFlash2)

--kv-unified-per-slot 131072 is the flag that actually guarantees 131k per slot: the server log then reads sizing KV pool to n_parallel * kv_unified_per_slot = 2 * 131072 = 262144 and n_slots = 2, n_ctx_slot = 131072. Without it, -np 2 alone divides the pool.

If you want the +3.9% from an f16 target KV, reduce the context instead: f16 at 131 072 costs 8.4 GB of KV versus 9.1 GB for q8_0 at 262 144 — i.e. the f16 build at half context uses less memory than the q8_0 build at full context, and decodes faster.

Files

Built from

• Base: Qwen/Qwen3.8-27B
• Finetune: DavidAU/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1 (reasoning-oriented)
• NVFP4 quant: esatapedico — MID-HIGH tier (NVFP4 backbone, Q8_0 head/embed)
• Drafter: z-lab/Qwen3.8-27B-DFlash2 — block-diffusion, lossless

DFlash2 accepts ~4–6 tokens per step. It's trained on the base Qwen3.8-27B, so it works with any quant of it. Speculative decoding is lossless (rejection sampling).

Run (llama.cpp CUDA)

# CUDA graphs stay ON (default) -> +33 % decode. Only if your build hits the old
# "CUDA illegal memory access" crash: set GGML_CUDA_DISABLE_GRAPHS=1 (costs ~33 %).
# set GGML_CUDA_DISABLE_GRAPHS=1    # Windows;  export on Linux/macOS

llama-server.exe \
  -m  ...-NVFP4.gguf \
  -md ...-draft-DFlash2-Q4_K_M.gguf \
  --spec-type draft-dflash --spec-draft-n-max 7 \
  -ngl all -np 1 -c 262144 -fa on \
  -ctk q8_0 -ctv q8_0 \
  --spec-draft-type-k f16 --spec-draft-type-v f16 \
  --chat-template-kwargs "{\"reasoning_effort\":\"medium\"}" \
  --cache-ram 32768 --no-cache-idle-slots \
  --alias qwen38-ultra:latest --reasoning-format auto
# add --mmproj ...-mmproj-BF16.gguf for image input (disables prompt-cache reuse)

--spec-draft-n-max 7 is the drafter's maximum (block_size 8); higher values are clamped. The draft KV is f16 (+1.1% measured); the target KV stays q8_0 to keep 262K context. Two-slot variant: replace the slot/context flags with -np 2 --kv-unified --kv-unified-per-slot 131072 and pick a different --alias (comma-separated aliases are supported: --alias "qwen38-ultra:latest,my-second-slot").

Notes

• Thinking model — set max_tokens ≥ 4096 on hard tasks.
• 262K context works; needles checked at 187K.
• CUDA graphs are ON by default here and measured +33 % (94.6 → 126.1 tok/s), stable across sustained 2-slot runs. GGML_CUDA_DISABLE_GRAPHS=1 is a last-resort fallback, not a default.
• Ollama: see the provided Modelfile.
• Reproducing the numbers above: decode figures come from the lmx protocol (temperature 0.7, 256-token generations) and from agent-profile runs at temperature 0; accuracy figures are GSM8K with numeric answer checking at temperature 0. Only interleaved A/B comparisons are trusted — separate measurement windows drift by up to 10% on this machine depending on background load (browser, torrent client), which is what makes single-run speed claims unreliable.