isilicon commited on
Commit
f5bbb46
·
verified ·
1 Parent(s): 68b8031

Preserve attributed pinned files and exact-configuration evaluation

Browse files
Files changed (5) hide show
  1. README.md +22 -0
  2. SOURCE_MANIFEST.json +38 -0
  3. UPSTREAM_README.md +152 -0
  4. config.json +57 -0
  5. model.safetensors +3 -0
README.md ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ base_model: "z-lab/Qwen3.8-27B-DFlash2"
4
+ tags:
5
+ - preservation-mirror
6
+ - mango
7
+ ---
8
+ # Preservation mirror: z-lab/Qwen3.8-27B-DFlash2
9
+
10
+ Mirrored by **isilicon** from [z-lab/Qwen3.8-27B-DFlash2](https://huggingface.co/z-lab/Qwen3.8-27B-DFlash2/tree/50307d4c4cde6860d4eee73e2547cd786fe8e8a4) at commit `50307d4c4cde6860d4eee73e2547cd786fe8e8a4`. Original training, edits and quantization belong to the upstream creators. This mirror does not claim new training or authorship.
11
+
12
+ Files retain their original selected paths and bytes. This repository contains the selected quantization/dependency, not every quantization from the source. See `SOURCE_MANIFEST.json` for every path, size and checksum, and `UPSTREAM_README.md` for original model information, attribution and caveats when supplied upstream. Original available license/notice files are retained. The source model card declares the license shown above; base-model conditions still apply.
13
+
14
+ Keep original subfolders when downloading (for example `4-bit/` for MLX). Tokenizers, projectors, draft models and offloaded tensor dependencies may live in separate mirrors. A weight-only mirror is not a claim of standalone runtime readiness.
15
+
16
+ ## Evaluation status
17
+
18
+ No standalone full benchmark score is claimed for these exact files. This is an auxiliary dependency or a downloaded/queued candidate. Scores from another quantization or the main model do not transfer to this repository.
19
+
20
+ ## Preservation status
21
+
22
+ `MIRROR_STATUS.json`, when present, records destination-file verification against the pinned source metadata. This is not a fresh independent-device full restore/load test. An HF mirror remains dependent on Hugging Face availability and policies.
SOURCE_MANIFEST.json ADDED
@@ -0,0 +1,38 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 1,
3
+ "mirror_owner": "isilicon",
4
+ "source_repo": "z-lab/Qwen3.8-27B-DFlash2",
5
+ "source_revision": "50307d4c4cde6860d4eee73e2547cd786fe8e8a4",
6
+ "license_declared_by_source": "apache-2.0",
7
+ "source_base_model": [
8
+ "Qwen/Qwen3.8-27B"
9
+ ],
10
+ "selected_files": [
11
+ {
12
+ "source_path": "config.json",
13
+ "path": "config.json",
14
+ "bytes": 1239,
15
+ "sha256": null,
16
+ "git_blob": "79279cc5665bced6f3cdaa2095a2ffe819497b2e"
17
+ },
18
+ {
19
+ "source_path": "model.safetensors",
20
+ "path": "model.safetensors",
21
+ "bytes": 3848817896,
22
+ "sha256": "67fc76d68dc5a9415511a4f394ef744d67510cd20e93b37cc2cc7d28e4bab65c",
23
+ "git_blob": "a988e2ea15d5cb6553f37a7f140304fb42fa8c10"
24
+ },
25
+ {
26
+ "source_path": "README.md",
27
+ "path": "UPSTREAM_README.md",
28
+ "bytes": 5397,
29
+ "sha256": null,
30
+ "git_blob": "1ee0eb2b2f81767f40521b61245211e5400b81bc"
31
+ }
32
+ ],
33
+ "scope": "Selected pinned files and original available notices; not all upstream quants.",
34
+ "dependency_for": [
35
+ "qwen-mlx"
36
+ ],
37
+ "local_integrity_receipt_sha256": "708a7154f00fe34d2f5bf6e2dc51a948a4aae3af0b4425a942c60155bb2804f2"
38
+ }
UPSTREAM_README.md ADDED
@@ -0,0 +1,152 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ library_name: transformers
4
+ pipeline_tag: text-generation
5
+ base_model:
6
+ - Qwen/Qwen3.8-27B
7
+ inference: false
8
+ tags:
9
+ - dflash2
10
+ - speculative-decoding
11
+ - block-diffusion
12
+ - draft-model
13
+ - sglang
14
+ - vllm
15
+ ---
16
+
17
+ # Qwen3.8-27B-DFlash2
18
+
19
+ [Blog](https://inco.ai/blog/dflash2/) | [GitHub](https://github.com/z-lab/dflash)
20
+
21
+ This repository contains the DFlash 2 draft model for
22
+ [`Qwen/Qwen3.8-27B`](https://huggingface.co/Qwen/Qwen3.8-27B).
23
+ It is not a standalone language model: it runs inside a speculative
24
+ decoding server and drafts tokens for the target model to verify. This repository is a mirror
25
+ of [`incoai/Qwen3.8-27B-DFlash2`](https://huggingface.co/incoai/Qwen3.8-27B-DFlash2).
26
+
27
+ DFlash 2 is a block-diffusion drafter for speculative decoding. It predicts
28
+ a whole block of tokens in a single pass and keeps the top candidates at
29
+ every position. A lightweight selector then traces one coherent path through them.
30
+ Two-tap dynamic convolutions in the backbone keep the draft from decaying
31
+ toward the end of the block. Decoding is lossless: greedy output
32
+ matches the target model exactly, and sampling preserves its distribution.
33
+
34
+ <div align="center">
35
+ <img src="assets/dflash2-figure.png" alt="DFlash 2: parallel block drafting with a candidate path selector" width="100%">
36
+ </div>
37
+
38
+ ## Quick Start
39
+
40
+ Serve with [SGLang](https://github.com/sgl-project/sglang):
41
+
42
+ ```bash
43
+ pip install "sglang[all] @ git+https://github.com/sgl-project/sglang.git#subdirectory=python"
44
+
45
+ python -m sglang.launch_server \
46
+ --model-path Qwen/Qwen3.8-27B \
47
+ --speculative-algorithm DFLASH \
48
+ --speculative-draft-model-path incoai/Qwen3.8-27B-DFlash2 \
49
+ --speculative-num-draft-tokens 8
50
+ ```
51
+
52
+ Or with [vLLM](https://github.com/vllm-project/vllm):
53
+
54
+ ```bash
55
+ pip install -U "vllm @ git+https://github.com/vllm-project/vllm.git@refs/pull/52816/head"
56
+
57
+ vllm serve Qwen/Qwen3.8-27B \
58
+ --speculative-config '{
59
+ "method": "dflash",
60
+ "model": "incoai/Qwen3.8-27B-DFlash2",
61
+ "num_speculative_tokens": 7
62
+ }'
63
+ ```
64
+
65
+ See the [blog post](https://inco.ai/blog/dflash2/) for other engines and more details.
66
+
67
+ ## Evaluation
68
+
69
+ - Runtime: SGLang on one NVIDIA H200, with FlashAttention 3 for target and draft attention
70
+ - Speculation block size: 8 (7 draft tokens per verification step)
71
+ - Sampling: Qwen3.8's officially recommended parameters (temperature 1.0, top-p 0.95, top-k 20), with `xhigh` reasoning effort
72
+ - Maximum new tokens: 4096
73
+ - Prompts: benchmark formatting from [`z-lab/dflash`](https://github.com/z-lab/dflash)
74
+
75
+ We compare autoregressive decoding, Qwen3.8's built-in seven-token MTP,
76
+ a community DSpark drafter
77
+ ([`RadixArk/Qwen3.8-27B-DSpark`](https://huggingface.co/RadixArk/Qwen3.8-27B-DSpark)),
78
+ and DFlash 2. All speculative methods propose seven draft tokens per
79
+ verification step.
80
+
81
+ ### Acceptance Length
82
+
83
+ Acceptance length is the per-request mean of completion tokens divided by verification steps.
84
+ Higher is better.
85
+
86
+ | Task | MTP | DSpark | DFlash 2 |
87
+ | :--- | ---: | ---: | ---: |
88
+ | GSM8K | 5.02 | 4.36 | **5.46** |
89
+ | MATH-500 | 4.72 | 3.92 | **5.28** |
90
+ | HumanEval | 3.91 | 3.30 | **4.39** |
91
+ | MBPP | 3.99 | 3.51 | **4.79** |
92
+ | MT-Bench | 3.74 | 3.01 | **4.10** |
93
+
94
+ ### Throughput
95
+
96
+ Throughput is total output tokens divided by end-to-end wall time.
97
+ Each cell shows `output tok/s (speedup vs. autoregressive)`.
98
+
99
+ #### Concurrency 1
100
+
101
+ | Task | Autoregressive | MTP | DSpark | DFlash 2 |
102
+ | :--- | ---: | ---: | ---: | ---: |
103
+ | GSM8K | 68.9 | 178.5 (2.59×) | 185.3 (2.69×) | **236.1 (3.43×)** |
104
+ | MATH-500 | 69.0 | 172.8 (2.51×) | 174.5 (2.53×) | **230.7 (3.34×)** |
105
+ | HumanEval | 69.0 | 151.9 (2.20×) | 159.9 (2.32×) | **214.6 (3.11×)** |
106
+ | MBPP | 69.0 | 153.1 (2.22×) | 163.3 (2.37×) | **226.9 (3.29×)** |
107
+ | MT-Bench | 68.9 | 134.9 (1.96×) | 137.6 (2.00×) | **184.0 (2.67×)** |
108
+
109
+ #### Concurrency 8
110
+
111
+ | Task | Autoregressive | MTP | DSpark | DFlash 2 |
112
+ | :--- | ---: | ---: | ---: | ---: |
113
+ | GSM8K | 467.2 | 1,022.1 (2.19×) | 1,040.8 (2.23×) | **1,328.7 (2.84×)** |
114
+ | MATH-500 | 480.0 | 1,023.5 (2.13×) | 1,025.8 (2.14×) | **1,368.3 (2.85×)** |
115
+ | HumanEval | 483.4 | 934.2 (1.93×) | 956.5 (1.98×) | **1,291.5 (2.67×)** |
116
+ | MBPP | 478.0 | 938.1 (1.96×) | 974.1 (2.04×) | **1,328.0 (2.78×)** |
117
+ | MT-Bench | 480.5 | 835.2 (1.74×) | 802.3 (1.67×) | **1,090.2 (2.27×)** |
118
+
119
+ #### Concurrency 32
120
+
121
+ | Task | Autoregressive | MTP | DSpark | DFlash 2 |
122
+ | :--- | ---: | ---: | ---: | ---: |
123
+ | GSM8K | 1,329.8 | 1,381.1 (1.04×) | 1,506.5 (1.13×) | **1,922.5 (1.45×)** |
124
+ | MATH-500 | 1,505.8 | 1,415.6 (0.94×) | 1,429.0 (0.95×) | **1,951.8 (1.30×)** |
125
+ | HumanEval | 1,546.5 | 1,296.8 (0.84×) | 1,330.1 (0.86×) | **1,799.0 (1.16×)** |
126
+ | MBPP | 1,507.7 | 1,314.9 (0.87×) | 1,361.3 (0.90×) | **1,886.8 (1.25×)** |
127
+ | MT-Bench | 1,507.4 | 1,159.7 (0.77×) | 1,115.5 (0.74×) | **1,525.3 (1.01×)** |
128
+
129
+ ## Citation
130
+
131
+ If you find DFlash 2 useful, please cite:
132
+
133
+ ```bibtex
134
+ @misc{inco2026dflash2,
135
+ title = {{DFlash 2: Keep Drafting Parallel}},
136
+ author = {{Inco AI}},
137
+ year = {2026},
138
+ month = {August},
139
+ url = {https://inco.ai/blog/dflash2/}
140
+ }
141
+ ```
142
+
143
+ Please also cite the original DFlash paper:
144
+
145
+ ```bibtex
146
+ @inproceedings{chen2026dflash,
147
+ title = {{DFlash: Block Diffusion for Flash Speculative Decoding}},
148
+ author = {Chen, Jian and Liang, Yesheng and Liu, Zhijian},
149
+ booktitle = {International Conference on Machine Learning (ICML)},
150
+ year = {2026}
151
+ }
152
+ ```
config.json ADDED
@@ -0,0 +1,57 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "DFlash2DraftModel"
4
+ ],
5
+ "attention_bias": false,
6
+ "attention_dropout": 0.0,
7
+ "bos_token_id": null,
8
+ "is_causal": false,
9
+ "dflash_config": {
10
+ "block_size": 8,
11
+ "conv_group_size": 16,
12
+ "conv_kernel_size": 2,
13
+ "mask_token_id": 248070,
14
+ "selector_rank": 256,
15
+ "selector_top_k": 16,
16
+ "target_layer_ids": [
17
+ 5,
18
+ 19,
19
+ 33,
20
+ 47,
21
+ 61
22
+ ]
23
+ },
24
+ "dtype": "bfloat16",
25
+ "eos_token_id": 248044,
26
+ "head_dim": 128,
27
+ "hidden_act": "silu",
28
+ "hidden_size": 5120,
29
+ "initializer_range": 0.02,
30
+ "intermediate_size": 17408,
31
+ "layer_types": [
32
+ "sliding_attention",
33
+ "sliding_attention",
34
+ "sliding_attention",
35
+ "sliding_attention",
36
+ "sliding_attention"
37
+ ],
38
+ "max_position_embeddings": 262144,
39
+ "max_window_layers": 5,
40
+ "model_type": "qwen3",
41
+ "num_attention_heads": 32,
42
+ "num_hidden_layers": 5,
43
+ "num_key_value_heads": 8,
44
+ "num_target_layers": 64,
45
+ "pad_token_id": 248044,
46
+ "rms_norm_eps": 1e-06,
47
+ "rope_parameters": {
48
+ "rope_theta": 10000000,
49
+ "rope_type": "default"
50
+ },
51
+ "sliding_window": 2048,
52
+ "tie_word_embeddings": false,
53
+ "transformers_version": "5.15.0",
54
+ "use_cache": true,
55
+ "use_sliding_window": true,
56
+ "vocab_size": 248320
57
+ }
model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:67fc76d68dc5a9415511a4f394ef744d67510cd20e93b37cc2cc7d28e4bab65c
3
+ size 3848817896