z-lab/Qwen3.5-35B-A3B-DFlash
Qwen3.5-35B-A3B-DFlash Paper(https://arxiv.org/abs/2602.06036) | Github(https://github.com/z-lab/dflash) | Blog(https://z-lab.ai/projects/dflash) This DFlash draft model is a joint retrain from Z-Lab(https://z-lab.ai) and Modal(https://modal.com), t...
Thành tựu nổi bật
Những danh hiệu mới nhất của dự án
#1 Repository Của Tuần
#1 Repository Của Tuần
#1 Repository Của Tuần
🚀 Qwen3.5-35B-A3B-DFlash — Nháp mô hình khối diffusion tăng tốc speculative decoding
DFlash dùng block diffusion draft model đề xuất đa token, target model xác minh — tăng throughput tới 3.71x.
Đây là mô hình nháp DFlash cho Qwen/Qwen3.5-35B-A3B, do Z-Lab và Modal cùng huấn luyện lại. Không phải mô hình ngôn ngữ độc lập — chỉ dùng để ghép với target model trong server speculative decoding.
✨ Điểm nổi bật
- Joint retrain bởi Z-Lab & Modal: huấn luyện lại từ Qwen3.5-35B-A3B với sequence length 40k và sliding-window attention.
- Block diffusion draft model: đề xuất nhiều token đồng thời thay vì tuần tự như MTP.
- Speculative decoding: target model xác minh đề xuất, tăng throughput nhưng giữ nguyên phân phối output.
- Hỗ trợ SGLang: cấu hình
--speculative-algorithm DFLASHvới block size 8 hoặc 16. - Benchmark mạnh: 3.71x tốc độ ở concurrency 1, 2.89x ở concurrency 32 trên B200.
- Hỗ trợ vLLM: đang chờ PR merge.
💡 Vì sao đáng chú ý
DFlash khác biệt với MTP bằng cách dùng block diffusion thay vì autoregressive chain — giảm số bước xác minh và tăng accept length. Kết quả: trên GSM8K, DFlash block=16 đạt 939.2 tok/s (3.03x) so với baseline 310.0 tok/s. Phù hợp triển khai trên B200 với SGLang + FA4/TRT-LLM attention.
#Qwen3.5 #DFlash #SpeculativeDecoding #BlockDiffusion #SGLang #TiniX
Similar Projects
Discover other similar projects and AI models.
Qwen3-0.6B Qwen3 Highlights Qwen3 is the latest generation of large language models in Qwen series, offering a comprehensive suite of dense and mixture-of-experts (MoE) models. Built upon extensive training, Qwen3 delivers groundbreaking advancement...
DeepSeek-R1 Paper Link👁️ 1. Introduction We introduce our first-generation reasoning models, DeepSeek-R1-Zero and DeepSeek-R1. DeepSeek-R1-Zero, a model trained via large-scale reinforcement learning (RL) without supervised fine-tuning (SFT) as a p...
Qwen3.6-35B-A3B-DFlash Paper(https://arxiv.org/abs/2602.06036) | Github(https://github.com/z-lab/dflash) | Blog(https://z-lab.ai/projects/dflash) This DFlash draft model is a joint retrain from Z-Lab(https://z-lab.ai) and Modal(https://modal.com), t...
Model Information The Meta Llama 3.1 collection of multilingual large language models (LLMs) is a collection of pretrained and instruction tuned generative models in 8B, 70B and 405B sizes (text in/text out). The Llama 3.1 instruction tuned text onl...
GPT-2 Test the whole generation capabilities here: https://transformer.huggingface.co/doc/gpt2-large Pretrained model on English language using a causal language modeling (CLM) objective. It was introduced in this paper(https://d4mucfpksywv.cloudfro...
Model Details Meta developed and released the Meta Llama 3 family of large language models (LLMs), a collection of pretrained and instruction tuned generative text models in 8 and 70B sizes. The Llama 3 instruction tuned models are optimized for dia...
Try gpt-oss · Guides · Model card · OpenAI blog Welcome to the gpt-oss series, OpenAI’s open-weight models(https://openai.com/open-models) designed for powerful reasoning, agentic tasks, and versatile developer use cases. We’re releasing two flavors...
Try gpt-oss · Guides · Model card · OpenAI blog Welcome to the gpt-oss series, OpenAI’s open-weight models(https://openai.com/open-models) designed for powerful reasoning, agentic tasks, and versatile developer use cases. We’re releasing two flavors...
Model Details Meta developed and released the Meta Llama 3 family of large language models (LLMs), a collection of pretrained and instruction tuned generative text models in 8 and 70B sizes. The Llama 3 instruction tuned models are optimized for dia...