Files changed (1) hide show
  1. README.md +67 -0
README.md ADDED
@@ -0,0 +1,67 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ library_name: speculators
3
+ base_model:
4
+ - qwen3/qwen3-8b
5
+ license: apache-2.0
6
+ tags:
7
+ - speculative-decoding
8
+ - dflash
9
+ - speculators
10
+ ---
11
+
12
+ # RedHatAI/Qwen3-8B-speculator.dflash
13
+
14
+ This is a DFlash speculator model for [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B).
15
+
16
+ ## Training Details
17
+
18
+ This model was trained using the [Speculators](https://github.com/vllm-project/speculators) library on a subset of [Magpie-Align/Magpie-Llama-3.1-Pro-300K-Filtered](https://huggingface.co/datasets/Magpie-Align/Magpie-Llama-3.1-Pro-300K-Filtered) and the `train_sft` split of [HuggingFaceH4/ultrachat_200k](https://huggingface.co/datasets/HuggingFaceH4/ultrachat_200k). Responses were regenerated by Qwen3-8B (with reasoning). Training compute for this model was sponsored by [Modal](https://modal.com).
19
+
20
+ ## Model Specifications
21
+
22
+ | | |
23
+ |---|---|
24
+ | **Base Model** | Qwen/Qwen3-8B |
25
+ | **Chat Template** | Qwen/Qwen3-8B (use `/chat/completions` endpoint) |
26
+ | **Format** | Safetensors |
27
+ | **License** | Apache 2.0 |
28
+ | **Validation Hardware** | Nvidia A100 |
29
+
30
+ ## Deployment
31
+
32
+ ```bash
33
+ # Install vLLM from the required PR
34
+ pip install git+https://github.com/vllm-project/vllm.git@refs/pull/41880/head
35
+
36
+ # Deploy with speculative decoding
37
+ vllm serve Qwen/Qwen3-8B \
38
+ --tensor-parallel-size 1 \
39
+ --max-model-len 16384 \
40
+ --speculative-config '{
41
+ "model": "RedHatAI/Qwen3-8B-speculator.dflash",
42
+ "num_speculative_tokens": 7,
43
+ "method": "dflash"
44
+ }'
45
+ ```
46
+
47
+ ## Preliminary Evaluations
48
+
49
+ Per-position token acceptance rates across datasets:
50
+ (with reasoning enabled)
51
+
52
+ | Dataset | Pos 1 | Pos 2 | Pos 3 | Pos 4 | Pos 5 | Pos 6 | Pos 7 | Avg Length |
53
+ |---------|-------|-------|-------|-------|-------|-------|-------|------------|
54
+ | HumanEval | 79.9% | 58.0% | 40.3% | 27.0% | 17.8% | 11.3% | 6.8% | 3.410 |
55
+ | math_reasoning | 82.2% | 62.7% | 46.2% | 33.5% | 23.4% | 15.8% | 9.9% | 3.740 |
56
+ | qa | 68.9% | 42.6% | 25.0% | 14.4% | 8.1% | 4.4% | 2.3% | 2.660 |
57
+ | question | 73.0% | 47.6% | 30.1% | 18.9% | 11.7% | 7.1% | 4.1% | 2.930 |
58
+ | rag | 71.1% | 44.8% | 27.0% | 15.7% | 8.9% | 4.9% | 2.5% | 2.750 |
59
+ | summarization | 65.5% | 36.1% | 19.0% | 9.5% | 4.7% | 2.3% | 1.1% | 2.380 |
60
+ | tool_call | 71.3% | 44.6% | 25.8% | 14.4% | 7.8% | 4.1% | 2.1% | 2.700 |
61
+ | translation | 63.8% | 38.4% | 22.1% | 11.8% | 6.1% | 3.2% | 1.5% | 2.470 |
62
+ | writing | 73.2% | 47.7% | 30.1% | 18.9% | 11.8% | 7.2% | 4.2% | 2.930 |
63
+
64
+
65
+ ## References
66
+
67
+ **Paper**: [DFlash: Block Diffusion for Flash Speculative Decoding](https://arxiv.org/abs/2602.06036)