Kaleto commited on
Commit
b1e2a6e
·
verified ·
1 Parent(s): 3106665

Add README.md

Browse files
Files changed (1) hide show
  1. README.md +89 -0
README.md ADDED
@@ -0,0 +1,89 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ base_model: sensenova/SenseNova-Loop-9B
4
+ base_model_relation: quantized
5
+ library_name: vllm
6
+ pipeline_tag: text-generation
7
+ tags:
8
+ - nvfp4
9
+ - fp4
10
+ - modelopt
11
+ - vllm
12
+ - blackwell
13
+ - dgx-spark
14
+ - gb10
15
+ - sensenova
16
+ - sensenova-loop
17
+ - looped-transformer
18
+ - mhc
19
+ ---
20
+
21
+ # SenseNova-Loop-9B — NVFP4-MLP (W4A16, quality-first)
22
+
23
+ A 4-bit NVFP4 build of [sensenova/SenseNova-Loop-9B](https://huggingface.co/sensenova/SenseNova-Loop-9B) (the final release) that quantizes **only the MLPs** and leaves attention and the DeltaNet projections in BF16. It is made for **vLLM on NVIDIA Blackwell and the DGX Spark / GB10**, for when you want to stay as close to the original as an NVFP4 build allows.
24
+
25
+ On a single GB10 it runs **1.7× faster than BF16** (12.4 vs. 7.5 tok/s per stream, 55 tok/s across four), loads in 10.5 GiB, and in our cross-engine comparison it is closer to BF16 than the official Q4_K_M GGUF.
26
+
27
+ > 💚 **Who this is for:** Blackwell / DGX Spark owners who serve with vLLM and put accuracy first. Want more speed? The sibling [SenseNova-Loop-9B-NVFP4](https://huggingface.co/Kaleto/SenseNova-Loop-9B-NVFP4) decodes at 21 tok/s. On llama.cpp or a Mac, the community GGUF and MLX builds listed in the [sibling card](https://huggingface.co/Kaleto/SenseNova-Loop-9B-NVFP4#who-this-is-for-and-who-it-isnt) are the better fit.
28
+
29
+ ---
30
+
31
+ ## Quick facts
32
+
33
+ | | |
34
+ |---|---|
35
+ | **Base model** | [sensenova/SenseNova-Loop-9B](https://huggingface.co/sensenova/SenseNova-Loop-9B), revision `3e1ba6e8` |
36
+ | **Size** | 12.4 GB on disk (BF16: 19 GB) |
37
+ | **Format** | NVFP4, nvidia-modelopt 0.43 checkpoint, `quant_algo: W4A16_NVFP4` |
38
+ | **Quantized** | 96 linears: `mlp.{gate,up,down}_proj` in all 32 layers |
39
+ | **Kept as in the source** | all attention and DeltaNet projections, `lm_head`, `embed_tokens`, norms, mHC routers (FP32), `loop_gate`, `conv1d`, `A_log`/`dt_bias` (FP32), vision tower, MTP head (all byte-identical to the source) |
40
+ | **Calibration** | 64 × 1,024 tokens of mixed text, through the real looped forward (layers 4–27 visited twice) |
41
+
42
+ ## Speed on DGX Spark (GB10)
43
+
44
+ vLLM 0.31.0 with the `vllm-sensenova-loop` plugin, BF16 KV cache, greedy, ~4.8K-token prompt, 256 decoded tokens.
45
+
46
+ | Build | Weights in memory | Prefill tok/s | Decode, 1 stream | Decode, 4 streams (total) |
47
+ |---|---:|---:|---:|---:|
48
+ | BF16 (source) | 16.9 GiB | 1,785 | 7.5 tok/s | 31.1 tok/s |
49
+ | **this build** | **10.5 GiB** | 1,563 | **12.4 tok/s** | **54.6 tok/s** |
50
+ | [NVFP4](https://huggingface.co/Kaleto/SenseNova-Loop-9B-NVFP4) (full decoder) | 7.7 GiB | 1,492 | 21.2 tok/s | 81.1 tok/s |
51
+
52
+ ## Quality
53
+
54
+ All builds were scored on the same tokens in `llama-perplexity` style (2,048-token chunks, second half scored, full-vocabulary KLD against SenseNova's BF16 GGUF; 72,633 positions on one GB10). The [sibling card](https://huggingface.co/Kaleto/SenseNova-Loop-9B-NVFP4#quality) describes the method in full.
55
+
56
+ | Build | Size | Mean KLD ↓ | Same top token ↑ |
57
+ |---|---:|---:|---:|
58
+ | [AtomicChat AD-Q4_K_M](https://huggingface.co/AtomicChat/SenseNova-Loop-9B-GGUF) (GGUF, imatrix) | 5.96 GB | **0.054** | **92.6 %** |
59
+ | **this build** | 12.4 GB | 0.087 | 90.0 % |
60
+ | [SenseNova Q4_K_M](https://huggingface.co/sensenova/SenseNova-Loop-9B-GGUF) (GGUF, official) | 5.82 GB | 0.111 | 89.2 % |
61
+ | [balllvin NVFP4](https://huggingface.co/balllvin/SenseNova-Loop-9B-NVFP4) (QAD) | 8.9 GB | 0.139 | 87.3 % |
62
+ | [NVFP4](https://huggingface.co/Kaleto/SenseNova-Loop-9B-NVFP4) (full decoder) | 9.4 GB | 0.151 | 86.8 % |
63
+
64
+ Per text type (mean KLD, this build): news 0.202 · wiki EN 0.077 · wiki DE 0.100 · code 0.086 · chat 0.033 · role-play 0.058.
65
+
66
+ To be fair to the GGUFs: AtomicChat's importance-matrix file is more accurate at half the size. Use it on llama.cpp. NVFP4's advantage is that it runs natively on Blackwell tensor cores in vLLM, which on a GB10 currently means about 3× the decode speed of the llama.cpp fork.
67
+
68
+ ## Running it with vLLM
69
+
70
+ Needs the out-of-tree plugin **[`vllm-sensenova-loop`](https://github.com/KaletoAI/vllm-sensenova-loop)**, which packages SenseTime's vLLM implementation from PR [#64832](https://github.com/vllm-project/vllm/pull/64832) and adds ModelOpt NVFP4 support. It was tested with `vllm/vllm-openai:v0.31.0` on GB10.
71
+
72
+ ```bash
73
+ pip install --no-deps https://github.com/KaletoAI/vllm-sensenova-loop/archive/refs/heads/master.tar.gz # inside your vLLM 0.31 environment or image
74
+
75
+ vllm serve Kaleto/SenseNova-Loop-9B-NVFP4-MLP \
76
+ --served-model-name sensenova-loop-9b-nvfp4-mlp \
77
+ --reasoning-parser qwen3 \
78
+ --enable-auto-tool-choice --tool-call-parser qwen3_coder \
79
+ --max-model-len 32768 --max-num-seqs 3 \
80
+ --enable-chunked-prefill --max-num-batched-tokens 8192
81
+ ```
82
+
83
+ The [sibling card](https://huggingface.co/Kaleto/SenseNova-Loop-9B-NVFP4#running-it-with-vllm) covers thinking on/off, memory and the current plugin limits: text only and no MTP yet. On sampling, add `min_p: 0.05` or `top_k: 20` to SenseNova's temperature 1.0 / top_p 0.95 to keep the long tail clean.
84
+
85
+ ## Credits & license
86
+
87
+ Thanks to [SenseNova / SenseTime](https://huggingface.co/sensenova) for the model and the vLLM implementation, to NVIDIA for modelopt, and to the community builders whose work is compared above.
88
+
89
+ Apache-2.0, inherited from [sensenova/SenseNova-Loop-9B](https://huggingface.co/sensenova/SenseNova-Loop-9B). This is an independent community build, not affiliated with SenseNova.