Chengyue Wu
commited on
Commit
·
19930b4
1
Parent(s):
5040112
update with LFS support
Browse files- .gitattributes +2 -0
- README.md +148 -3
- assets/benchmark_results.png +3 -0
- assets/throughput.png +3 -0
- assets/training_recipe.png +3 -0
- assets/visualization_animation.gif +3 -0
- modeling.py +1 -4
.gitattributes
CHANGED
@@ -34,3 +34,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
36 |
tokenizer.json filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
36 |
tokenizer.json filter=lfs diff=lfs merge=lfs -text
|
37 |
+
*.gif filter=lfs diff=lfs merge=lfs -text
|
38 |
+
*.png filter=lfs diff=lfs merge=lfs -text
|
README.md
CHANGED
@@ -1,3 +1,148 @@
|
|
1 |
-
---
|
2 |
-
license: apache-2.0
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: apache-2.0
|
3 |
+
language:
|
4 |
+
- en
|
5 |
+
base_model:
|
6 |
+
- Qwen/Qwen2.5-7B-Instruct
|
7 |
+
---
|
8 |
+
|
9 |
+
# Fast-dLLM v2 (7B) — Efficient Block-Diffusion LLM
|
10 |
+
|
11 |
+
## 📖 Introduction
|
12 |
+
|
13 |
+
Autoregressive (AR) large language models (LLMs) have achieved remarkable performance across a wide range of natural language tasks, yet their **inherent sequential decoding limits inference efficiency**.
|
14 |
+
|
15 |
+
We present **Fast-dLLM v2** — a carefully designed **block diffusion language model (dLLM)** that efficiently adapts a pretrained AR model (**Qwen2.5-7B-Instruct**) into a diffusion-style decoder for **parallel text generation**.
|
16 |
+
|
17 |
+
### ✨ Key Innovations
|
18 |
+
- **Block Diffusion Mechanism + Complementary Attention Mask**
|
19 |
+
Enables **blockwise bidirectional context modeling** without sacrificing AR objectives.
|
20 |
+
- **Hierarchical Caching**
|
21 |
+
- **Block-level cache**: Stores historical context representations across blocks.
|
22 |
+
- **Sub-block cache**: Parallel decoding within partially generated blocks.
|
23 |
+
- **Token Shift Mechanism**
|
24 |
+
Retains autoregressive characteristics while supporting bidirectional context within blocks.
|
25 |
+
- **Parallel Decoding Pipeline**
|
26 |
+
Achieves up to **2.5× speedup** over standard AR decoding **without compromising quality**.
|
27 |
+
|
28 |
+
> 🚀 Fast-dLLM v2 uses **only ~1B tokens** for fine-tuning — a **500× reduction** vs. full-attention diffusion LLMs (Dream: 580B tokens) — while **matching or surpassing AR baselines** in accuracy.
|
29 |
+
|
30 |
+

|
31 |
+
|
32 |
+
---
|
33 |
+
|
34 |
+
## 🛠 Model Overview
|
35 |
+
- **Type**: Block Diffusion Language Model (dLLM)
|
36 |
+
- **Base Model**: `Qwen/Qwen2.5-7B-Instruct`
|
37 |
+
- **Architecture**: Transformer w/ RoPE, SwiGLU activation, RMSNorm, Attention QKV bias
|
38 |
+
- **Params**: ~7B
|
39 |
+
- **Layers**: 28
|
40 |
+
- **Attention Heads**: 28 (Q), 4 (KV, GQA)
|
41 |
+
- **Block Diffusion Size**: 32 tokens
|
42 |
+
- **Key Feature**: Parallel **block-wise decoding** + **hierarchical caching (block-level & sub-block)**
|
43 |
+
|
44 |
+
---
|
45 |
+
|
46 |
+
## 📦 Installation
|
47 |
+
You will need `transformers`, `torch`, and our **custom generation function**:
|
48 |
+
|
49 |
+
```bash
|
50 |
+
pip install transformers torch numpy
|
51 |
+
```
|
52 |
+
|
53 |
+
---
|
54 |
+
|
55 |
+
## 🚀 Quickstart
|
56 |
+
|
57 |
+
```python
|
58 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
59 |
+
|
60 |
+
model_name = "Efficient-Large-Model/Fast_dLLM_7B"
|
61 |
+
|
62 |
+
model = AutoModelForCausalLM.from_pretrained(
|
63 |
+
model_name,
|
64 |
+
torch_dtype="auto",
|
65 |
+
device_map="auto",
|
66 |
+
trust_remote_code=True
|
67 |
+
)
|
68 |
+
|
69 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
|
70 |
+
|
71 |
+
prompt = "Give me a short introduction to large language model."
|
72 |
+
messages = [
|
73 |
+
{"role": "system", "content": "You are a helpful assistant."},
|
74 |
+
{"role": "user", "content": prompt}
|
75 |
+
]
|
76 |
+
|
77 |
+
text = tokenizer.apply_chat_template(
|
78 |
+
messages,
|
79 |
+
tokenize=False,
|
80 |
+
add_generation_prompt=True
|
81 |
+
)
|
82 |
+
inputs = tokenizer([text], return_tensors="pt").to(model.device)
|
83 |
+
|
84 |
+
# Fast-dLLM v2 parallel decoding
|
85 |
+
gen_ids = model.generate(
|
86 |
+
inputs["input_ids"],
|
87 |
+
tokenizer=tokenizer,
|
88 |
+
max_new_tokens=512,
|
89 |
+
small_block_size=8,
|
90 |
+
threshold=0.9,
|
91 |
+
)
|
92 |
+
|
93 |
+
response = tokenizer.decode(
|
94 |
+
gen_ids[0][inputs["input_ids"].shape[1]:],
|
95 |
+
skip_special_tokens=True
|
96 |
+
)
|
97 |
+
print(response)
|
98 |
+
```
|
99 |
+
|
100 |
+
---
|
101 |
+
|
102 |
+
## 📊 Performance & Benchmarks
|
103 |
+
|
104 |
+
### ▶ Real-time Throughput
|
105 |
+
Fast-dLLM v2 offers **up to 2.54× higher throughput** than Qwen2.5-7B-Instruct, **without loss in quality**.
|
106 |
+
|
107 |
+

|
108 |
+
|
109 |
+
---
|
110 |
+
|
111 |
+
### 🏆 Benchmark Results
|
112 |
+
We compare Fast-dLLM v2 against AR baselines and previous diffusion LLMs on diverse tasks:
|
113 |
+
HumanEval, MBPP (code), GSM8K, Math (reasoning), IFEval (instruction), MMLU, GPQA (knowledge QA).
|
114 |
+
|
115 |
+
- **1B group**: Fast-dLLM v2 (7B) achieves **best average score: 45.0**.
|
116 |
+
- **7B group**: Fast-dLLM v2 (7B) achieves **best average score: 60.3**, surpassing LLaDA and Dream models.
|
117 |
+
|
118 |
+

|
119 |
+
|
120 |
+
---
|
121 |
+
|
122 |
+
## 📜 Citation
|
123 |
+
|
124 |
+
If you use Fast-dLLM v2 in your research or products, please cite:
|
125 |
+
|
126 |
+
```bibtex
|
127 |
+
@misc{wu2025fastdllmv2efficientblockdiffusion,
|
128 |
+
title={Fast-dLLM v2: Efficient Block-Diffusion LLM},
|
129 |
+
author={Chengyue Wu and Hao Zhang and Shuchen Xue and Shizhe Diao and Yonggan Fu and Zhijian Liu and Pavlo Molchanov and Ping Luo and Song Han and Enze Xie},
|
130 |
+
year={2025},
|
131 |
+
eprint={2509.26328},
|
132 |
+
archivePrefix={arXiv},
|
133 |
+
primaryClass={cs.CL},
|
134 |
+
url={https://arxiv.org/abs/2509.26328},
|
135 |
+
}
|
136 |
+
```
|
137 |
+
|
138 |
+
---
|
139 |
+
|
140 |
+
## 📄 License
|
141 |
+
Released under **Apache 2.0**, following the base Qwen2.5 license.
|
142 |
+
|
143 |
+
---
|
144 |
+
|
145 |
+
## 🔗 Resources
|
146 |
+
- 📄 [Paper](https://arxiv.org/abs/2509.26328)
|
147 |
+
- 💻 [Code](https://github.com/NVlabs/Fast-dLLM)
|
148 |
+
- 🤗 [HuggingFace Model](https://huggingface.co/Efficient-Large-Model/Fast_dLLM_7B)
|
assets/benchmark_results.png
ADDED
![]() |
Git LFS Details
|
assets/throughput.png
ADDED
![]() |
Git LFS Details
|
assets/training_recipe.png
ADDED
![]() |
Git LFS Details
|
assets/visualization_animation.gif
ADDED
![]() |
Git LFS Details
|
modeling.py
CHANGED
@@ -555,7 +555,6 @@ class Fast_dLLM_QwenForCausalLM(Fast_dLLM_QwenPreTrainedModel, GenerationMixin):
|
|
555 |
top_p=0.95,
|
556 |
temperature=0,
|
557 |
use_block_cache=False,
|
558 |
-
block_cache_refresh_interval=16,
|
559 |
**kwargs
|
560 |
):
|
561 |
num_blocks = max_new_tokens // block_size
|
@@ -581,7 +580,6 @@ class Fast_dLLM_QwenForCausalLM(Fast_dLLM_QwenPreTrainedModel, GenerationMixin):
|
|
581 |
x_init = torch.cat([input_ids, x_init], dim=1)
|
582 |
|
583 |
x_t = x_init.clone()
|
584 |
-
step = 0
|
585 |
block_past_key_values = None
|
586 |
while True:
|
587 |
if stop_token in x_t[:, prompt_length:]:
|
@@ -612,7 +610,7 @@ class Fast_dLLM_QwenForCausalLM(Fast_dLLM_QwenPreTrainedModel, GenerationMixin):
|
|
612 |
break
|
613 |
|
614 |
if use_block_cache:
|
615 |
-
if
|
616 |
output = self.forward(input_ids=x_t[:, -block_size:], use_cache=True, past_key_values=past_key_values, update_past_key_values=False, use_block_cache=True)
|
617 |
logits, block_past_key_values = output.logits, output.block_past_key_values
|
618 |
logits = torch.cat([logits[:, :1, :], logits[:, :-1, :]], dim=1)
|
@@ -638,7 +636,6 @@ class Fast_dLLM_QwenForCausalLM(Fast_dLLM_QwenPreTrainedModel, GenerationMixin):
|
|
638 |
|
639 |
x_t[:, start:end][unmask_idx] = x_1[unmask_idx]
|
640 |
|
641 |
-
step += 1
|
642 |
input_ids = x_t
|
643 |
# Truncate stop_token
|
644 |
if stop_token in input_ids[:, original_input_length:]:
|
|
|
555 |
top_p=0.95,
|
556 |
temperature=0,
|
557 |
use_block_cache=False,
|
|
|
558 |
**kwargs
|
559 |
):
|
560 |
num_blocks = max_new_tokens // block_size
|
|
|
580 |
x_init = torch.cat([input_ids, x_init], dim=1)
|
581 |
|
582 |
x_t = x_init.clone()
|
|
|
583 |
block_past_key_values = None
|
584 |
while True:
|
585 |
if stop_token in x_t[:, prompt_length:]:
|
|
|
610 |
break
|
611 |
|
612 |
if use_block_cache:
|
613 |
+
if block_past_key_values is None or (x_t[:, -block_size+small_block_start_idx] == mask_id).any():
|
614 |
output = self.forward(input_ids=x_t[:, -block_size:], use_cache=True, past_key_values=past_key_values, update_past_key_values=False, use_block_cache=True)
|
615 |
logits, block_past_key_values = output.logits, output.block_past_key_values
|
616 |
logits = torch.cat([logits[:, :1, :], logits[:, :-1, :]], dim=1)
|
|
|
636 |
|
637 |
x_t[:, start:end][unmask_idx] = x_1[unmask_idx]
|
638 |
|
|
|
639 |
input_ids = x_t
|
640 |
# Truncate stop_token
|
641 |
if stop_token in input_ids[:, original_input_length:]:
|