How to use Ada Flash: what you can do with it

3 use cases

• Measure Ada Flash decoding speed against baselines on GSM8K

  Run the one-command Ada Flash benchmark on a GSM8K split to get throughput numbers for your own GPUs.

  • REPO ZinYY/AdaFlash
  • Software and DevOps
  • Intermediate
  View prompt

  In the ZinYY/AdaFlash repo, set up the environment from the README, prepare the gsm8k split, and run scripts/tools/benchmark.sh with MODEL_PATH=Qwen/Qwen3-8B and DRAFT_MODEL_PATH=AdaFlash/Qwen3-8B-PerfectBlend for <your GPU machine>. Report the throughput output from the benchmark. Verify that the server started with the patched SGLang before trusting any number.

  Share Email me this

• Serve Qwen3-8B with an Ada Flash drafter and test it over HTTP

  Launch an Ada Flash SGLang server for Qwen3-8B and send requests to its /generate endpoint.

  • REPO ZinYY/AdaFlash
  • Software and DevOps
  • Intermediate
  View prompt

  In the ZinYY/AdaFlash repo, start the AdaFlash server with scripts/serve/serve_thresh_head.sh for Qwen/Qwen3-8B and then run python bin/benchmark.py against its /generate endpoint. Expected output is a benchmark summary from the client. Verify that the server is ready before sending requests and that the drafter is AdaFlash/Qwen3-8B-PerfectBlend.

  Share Email me this

• Train an Ada Flash drafter on your own response data

  Regenerate training data, initialize the adaptive length head, then run asynchronous on-policy training.

  • REPO ZinYY/AdaFlash
  • Software and DevOps
  • Advanced
  View prompt

  In the ZinYY/AdaFlash repo, prepare a dataset, regenerate responses with Qwen/Qwen3-8B, initialize the adaptive length head, and start scripts/pipeline/run_async_pipeline_adaflash.sh with inference GPU 0 and training GPUs 1,2. Expected output is versioned drafter weights written under the swap directory and logs under the log directory. Verify that the initialized draft path exists before starting the async run.

  Share Email me this

Quick start with Ada Flash

1. Check you have what it needs: A Linux machine with NVIDIA GPUs and CUDA, since the README sets CUDA devices and separate GPU sets for serving and training; Conda and Python 3.12, as used in the README install steps; Access to the Qwen3-8B target model and the AdaFlash/DFlash checkpoints on Hugging Face.
2. Open ZinYY/AdaFlash on GitHub and read the README and license.
3. Create the environment with conda create -n adaflash python=3.12 -y and then conda activate adaflash.
4. Install the patched SGLang with pip install "sglang[all] @ git+https://github.com/sgl-project/sglang.git@refs/pull/34171/head#subdirectory=python" and then pip install -r requirements.txt.
5. Prepare the benchmark split with python bin/prepare_data.py --dataset gsm8k.
6. Set export MODEL_PATH=Qwen/Qwen3-8B, export DRAFT_MODEL_PATH=AdaFlash/Qwen3-8B-PerfectBlend, export DATASET=gsm8k, export NUM_SAMPLES=1024, and export CONCURRENCY=64.

These steps come from our top use case, Measure Ada Flash decoding speed against baselines on GSM8K, which also has a copy-paste agent prompt.

Using the Deduped MCP, get the setup steps for ZinYY/AdaFlash

claude mcp add --transport http deduped https://dedupedai.com/mcp

Best AI repos for Software and DevOps

About this repo

• Who it is forBuilt for ML engineers and researchers who serve or train Qwen3-family models and want to measure or improve speculative decoding speed.
• Why it is hot133 stars, launched 81 days ago
• LanguagePython
• Sourcegithub.com/ZinYY/AdaFlash · Back to Hot repos

Requirements

• A Linux machine with NVIDIA GPUs and CUDA, since the README sets CUDA devices and separate GPU sets for serving and training
• Conda and Python 3.12, as used in the README install steps
• Access to the Qwen3-8B target model and the AdaFlash/DFlash checkpoints on Hugging Face
• The patched SGLang version from the linked SGLang PR, because a stock SGLang release may not expose the required arguments

Related repos

• Trended earlier REPO SYSTRAN/faster-whisper Faster Whisper faster-whisper is a Python reimplementation of OpenAI's Whisper built on the CTranslate2 inference engine. It is up to 4 times faster than openai/whisper at the same accuracy while using less memory, supports 8-bit quantization on CPU and GPU, batched transcription, word-level timestamps, Distil-Whisper checkpoints and a built-in Silero VAD filter.
• Trended earlier REPO datawhalechina/zero-to-sglang Zero to SGLang A free, open course from Datawhale and RadixArk that teaches LLM inference by building a mini-sglang from scratch and then reading the real SGLang source.
• Trended earlier REPO whitecircle/halo Halo Halo is an open-source framework from White Circle for training large language and multimodal models with pre-training, SFT, preference optimization, and multi-turn RL on Hugging Face models.