Views
No views yet
[!TIP] Quantized on 2026-04-13 withlm_head/linear_attn(GDN) /mtp/visual/embed_tokenspreserved in BF16.
[!WARNING] NVIDIA DGX Spark (GB10 SM121) — Driver 590.48+ / CUDA 13.1+As of April 2026, NVFP4 software support on SM121 is still incomplete. The native W4A4 compute path is not yet functional on this hardware — the runtime silently falls back to W4A16 (BF16 activations), negating the theoretical throughput advantage of FP4.If accuracy and inference speed are your priority, we recommend the INT4 AutoRound version: 👉 YuYu1015/Huihui-Qwen3.5-35B-A3B-Claude-4.6-Opus-abliterated-int4-AutoRoundINT4 AutoRound leverages the mature W4A16 Marlin kernel path on DGX Spark, offering more thorough calibration (~99.5% quality retention) and significantly more stable performance. The full potential of NVFP4 will only be unlocked once NVIDIA delivers complete W4A4 kernel support for SM121.
| Item | Value |
|---|---|
| Architecture | Dense + GDN (Mamba) + Attention hybrid |
| Base model | Qwen/Qwen3.5-9B |
| Fine-tuned by | huihui-ai (Claude 4.6 Opus distillation + abliteration) |
| Quantized by | YuYu1015 |
| Model size | ~9.6 GB (NVFP4, vs ~19 GB BF16 original) |
| Context length | Up to 262,144 tokens |
| Thinking mode | Supported (enable_thinking: true/false) |
| Tool calling | Supported (qwen3_coder parser) |
| MTP | Built-in MTP weights included (preserved in BF16) |
| Item | Value |
|---|---|
| Method | llm-compressor (main branch, PR #2608) |
| Scheme | NVFP4 (E2M1 + FP8 per-group scaling, group size 16) |
| Format | compressed-tensors (main branch) |
| Calibration dataset | HuggingFaceH4/ultrachat_200k (train_sft split) |
| Calibration samples | 512 |
| Calibration sequence length | 2048 |
| Hardware | NVIDIA DGX Spark (GB10, 128GB unified memory) |
| Environment | transformers>=5.0 + llm-compressor main (Qwen3.5 qwen3_5 model_type requires tf5) |
| Layer | Reason |
|---|---|
lm_head | Output head, sensitive to quantization noise |
re:.*linear_attn.* | GDN/DeltaNet (Mamba) layers — may output zeros if quantized |
re:.*mtp\..* | Multi-Token Prediction weights |
re:.*visual\..* | Vision encoder |
re:.*embed_tokens$ | Input embeddings |
1vllm serve /path/to/model \
2 --quantization compressed-tensors \
3 --served-model-name qwen3.5-9b \
4 --reasoning-parser qwen3 \
5 --enable-auto-tool-choice \
6 --tool-call-parser qwen3_coder \
7 --kv-cache-dtype auto \
8 --gpu-memory-utilization 0.90 \
9 --max-model-len 65536 \
10 --enable-prefix-caching \
11 --enable-chunked-prefill \
12 --trust-remote-code \
13 --language-model-onlycvt.e2m1x2 instruction)--kv-cache-dtype auto--language-model-only skips vision encoder profiling for text-only inferencesudo sh -c 'echo 3 > /proc/sys/vm/drop_caches'[!TIP] 2026-04-13 量化上傳,lm_head/linear_attn(GDN) /mtp/visual/embed_tokens保留 BF16。
[!WARNING] NVIDIA DGX Spark (GB10 SM121) 使用者 — Driver 590.48+ / CUDA 13.1+截至 2026 年 4 月,NVFP4 在 SM121 上的軟體支援仍不完整。原生 W4A4 運算路徑尚未在此硬體上就緒——執行時會靜默退回 W4A16(BF16 activation),FP4 的理論吞吐量優勢無法發揮。若精度與推理速度為首要考量,建議改用 INT4 AutoRound 版本: 👉 YuYu1015/Huihui-Qwen3.5-35B-A3B-Claude-4.6-Opus-abliterated-int4-AutoRoundINT4 AutoRound 在 DGX Spark 上使用成熟的 W4A16 Marlin kernel 路徑,校準更完整(品質保留約 99.5%),效能顯著更穩定。待 NVIDIA 為 SM121 提供完整的 W4A4 kernel 支援後,NVFP4 的真正優勢才能發揮。
| 項目 | 數值 |
|---|---|
| 架構 | Dense + GDN (Mamba) + Attention 混合 |
| 基礎模型 | Qwen/Qwen3.5-9B |
| 微調者 | huihui-ai(Claude 4.6 Opus 蒸餾 + abliteration) |
| 量化者 | YuYu1015 |
| 模型大小 | ~9.6 GB(NVFP4,原版 BF16 約 19 GB) |
| Context 長度 | 最高 262,144 tokens |
| 思考模式 | 支援(enable_thinking: true/false) |
| 工具呼叫 | 支援(qwen3_coder parser) |
| MTP | 內建 MTP 權重(保留 BF16) |
| 項目 | 數值 |
|---|---|
| 方法 | llm-compressor(main 分支,PR #2608) |
| 方案 | NVFP4(E2M1 + FP8 逐群縮放,群組大小 16) |
| 格式 | compressed-tensors(main 分支) |
| 校準資料集 | HuggingFaceH4/ultrachat_200k (train_sft 分割) |
| 校準樣本數 | 512 |
| 校準序列長度 | 2048 |
| 量化硬體 | NVIDIA DGX Spark(GB10, 128GB 統一記憶體) |
| 環境 | transformers>=5.0 + llm-compressor main(Qwen3.5 qwen3_5 model_type 需要 tf5) |
| 層 | 原因 |
|---|---|
lm_head | 輸出頭,對量化雜訊敏感 |
re:.*linear_attn.* | GDN/DeltaNet (Mamba) 層,量化後可能輸出零 |
re:.*mtp\..* | Multi-Token Prediction 權重 |
re:.*visual\..* | 視覺編碼器 |
re:.*embed_tokens$ | 輸入嵌入 |
1vllm serve /path/to/model \
2 --quantization compressed-tensors \
3 --served-model-name qwen3.5-9b \
4 --reasoning-parser qwen3 \
5 --enable-auto-tool-choice \
6 --tool-call-parser qwen3_coder \
7 --kv-cache-dtype auto \
8 --gpu-memory-utilization 0.90 \
9 --max-model-len 65536 \
10 --enable-prefix-caching \
11 --enable-chunked-prefill \
12 --trust-remote-code \
13 --language-model-onlycvt.e2m1x2 指令)--kv-cache-dtype auto--language-model-only 跳過視覺編碼器 profiling,加速純文字推理啟動sudo sh -c 'echo 3 > /proc/sys/vm/drop_caches'