Views
No views yet

| Item | Value |
|---|---|
| Architecture | Dense (27B), 64 layers — 16 × (3 × Gated DeltaNet → FFN + 1 × Gated Attention → FFN) + MTP |
| Base model | Qwen/Qwen3.6-27B |
| Fine-tuned by | huihui-ai (abliteration) |
| Quantized by | YuYu1015 |
| Model size | 16.96 GB (vs ~55.6 GB BF16 original) — 5.21 BPW average |
| Format | GGUF (single file), UD-style mixed precision via imatrix |
| Context length | Up to 262,144 tokens (tested at 65,536) |
| Thinking mode | Supported (enable_thinking: true/false) |
| MTP | Built-in MTP head preserved at Q8_0 (drafter for --spec-type draft-mtp) |
| Multimodal | Vision tower dropped by convert_hf_to_gguf.py (GGUF is text-only, saves ~1 GB) |
| llama.cpp | b9200+ required (older builds miss hybrid GDN + MTP loading) |
| Tensor | Type | Reason |
|---|---|---|
FFN (gate/up/down) | Q4_K_M | Bulk weight, high quantization tolerance |
Attention Q/K/V | Q6_K | Sensitive to noise, upranked from base |
| Output projection | Q6_K | UD convention |
| Token embedding | Q6_K | UD convention |
| MTP / NextN head | Q8_0 | Drafter — Q8_0 keeps acceptance near-lossless |
convert_hf_to_gguf.py picks up all 4 blk.64.nextn.* tensors directly — no Qwen-official MTP graft needed.Qwen3_5ForConditionalGeneration (multimodal), but GGUF only packs the LLM portion. Text-only inference works out of the box.1llama-server \
2 -m Huihui-Qwen3.6-27B-abliterated-UD-Q4_K_XL-MTP.gguf \
3 -fit off \
4 -c 65536 \
5 -np 1 \
6 -fa on \
7 -ngl 99 \
8 --cache-type-k q8_0 \
9 --cache-type-v q8_0 \
10 --spec-type draft-mtp \
11 --spec-draft-n-max 2 \
12 --host 0.0.0.0 --port 8080| Flag | Why |
|---|---|
-fit off | Newer llama.cpp's auto-fit aborts on Qwen3.6 hybrid GDN+MTP layout |
-np 1 | MTP path is single-slot only |
--spec-type draft-mtp | New flag (older versions used --spec-type mtp) |
--spec-draft-n-max 2 | Empirically the sweet spot — n=3 regresses ~22% due to hybrid bug |
--cache-type-k/v q8_0 | KV cache compression, no measurable quality loss on this model |
| Configuration | Throughput |
|---|---|
| No speculative decoding | 39.15 tok/s |
MTP --spec-draft-n-max 1 | 55 tok/s |
MTP --spec-draft-n-max 2 | 60 tok/s (best) |
MTP --spec-draft-n-max 3 | 55 tok/s (-22% vs n=2 ideal due to hybrid rollback bug) |
Note on the CPU: the test rig uses an old Intel Xeon E5-2696 v3 (2.30 GHz, 2015). Modern CPUs (Zen 4 / Raptor Lake or newer) consistently push 30–50% higher tok/s on the same RTX 3090 because llama.cpp's hybrid GDN path still touches CPU per step. These numbers are a floor, not a ceiling.
failed to fit params to free device memory, n_gpu_layers already set by user → add -fit off.missing tensor 'blk.64.ssm_conv1d.weight' → upgrade to b9200+.| 項目 | 數值 |
|---|---|
| 架構 | Dense(27B),64 層 — 16 × (3 × Gated DeltaNet → FFN + 1 × Gated Attention → FFN) + MTP |
| 基礎模型 | Qwen/Qwen3.6-27B |
| 微調者 | huihui-ai(abliteration) |
| 量化者 | YuYu1015 |
| 模型大小 | 16.96 GB(原版 BF16 約 55.6 GB)— 平均 5.21 BPW |
| 格式 | GGUF(單檔),UD 風格混合精度,imatrix 校準 |
| Context 長度 | 最高 262,144 tokens(實測 65,536) |
| 思考模式 | 支援(enable_thinking: true/false) |
| MTP | 內建 MTP head 以 Q8_0 保留(作為 --spec-type draft-mtp 的草稿器) |
| 多模態 | convert_hf_to_gguf.py 自動跳過視覺塔(GGUF 純文字,省 ~1 GB) |
| llama.cpp | 必須 b9200+(舊版不支援 hybrid GDN + MTP 載入) |
| Tensor | 類型 | 理由 |
|---|---|---|
FFN(gate/up/down) | Q4_K_M | 大宗權重,量化容忍度高 |
Attention Q/K/V | Q6_K | 對雜訊敏感,由基礎類型升級 |
| Output projection | Q6_K | UD 慣例 |
| Token embedding | Q6_K | UD 慣例 |
| MTP / NextN head | Q8_0 | 草稿器 — Q8_0 幾乎無損維持 acceptance |
convert_hf_to_gguf.py 可直接抓到 blk.64.nextn.* 共 4 個 tensor — 不需嫁接 Qwen 官方 MTP。Qwen3_5ForConditionalGeneration(多模態),但 GGUF 只包 LLM 部分。純文字推理直接可用。1llama-server \
2 -m Huihui-Qwen3.6-27B-abliterated-UD-Q4_K_XL-MTP.gguf \
3 -fit off \
4 -c 65536 \
5 -np 1 \
6 -fa on \
7 -ngl 99 \
8 --cache-type-k q8_0 \
9 --cache-type-v q8_0 \
10 --spec-type draft-mtp \
11 --spec-draft-n-max 2 \
12 --host 0.0.0.0 --port 8080| Flag | 用途 |
|---|---|
-fit off | 新版 llama.cpp 的 auto-fit 對 Qwen3.6 hybrid GDN+MTP 會 abort |
-np 1 | MTP 路徑只能單 slot |
--spec-type draft-mtp | 新版 flag(舊版叫 --spec-type mtp) |
--spec-draft-n-max 2 | 實測甜蜜點 — n=3 因 hybrid bug 回退約 22% |
--cache-type-k/v q8_0 | KV cache 壓縮,此模型上無可量測品質損失 |
| 設定 | 速度 |
|---|---|
| 無投機解碼 | 39.15 tok/s |
MTP --spec-draft-n-max 1 | 55 tok/s |
MTP --spec-draft-n-max 2 | 60 tok/s(最佳) |
MTP --spec-draft-n-max 3 | 55 tok/s(hybrid rollback bug 拖累,較 n=2 理論值 -22%) |
CPU 說明: 測試機使用 2015 年的舊 Intel Xeon E5-2696 v3(2.30 GHz)。換成現代 CPU(Zen 4 / Raptor Lake 以上)在同樣 RTX 3090 上通常可再提升 30–50% tok/s,因為 llama.cpp 的 hybrid GDN 路徑每步仍需 CPU 介入。這個數字是下限,不是上限。
failed to fit params to free device memory, n_gpu_layers already set by user → 加 -fit off。missing tensor 'blk.64.ssm_conv1d.weight' → 升級到 b9200+。