Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Implementation for FP8/INT8 Rollout for RL training without performence drop.
| Date | Stars |
|---|---|
| 2026-07-24 | 307 |
| 2026-07-25 | 307 |
| 2026-07-28 | 307 |
| 2026-07-30 | 307 |
| 2026-08-06 | 307 |
Today
— stars today
This week
— stars this week
This month
— stars this month
Momentum
0.0
growth rate 0.00%/day
<h1 align="center">⚡ FlashRL ⚡</h1> <p align="center"><b>Fast RL training with Quantized Rollouts </b> (<a href="https://fengyao.notion.site/flash-rl">Blog</a>)</p> <p align="center"> <img src="https://img.shields.io/badge/license-MIT-blue.svg"> <img src="https://img.shields.io/badge/python-3.10+-blue"> <img src="https://img.shields.io/pypi/v/flash-llm-rl?color=green"> </p> <p align="center"> <a href="#-flashrl-">What is FlashRL?</a> • <a href="#-quick-start">Quick Start</a> • <a href="#-usage-guide">Usage Guide</a> • <a href="#-examples">Examples</a> • <a href="#-road-map">Road Map</a> • <a href="#-citation">Citation</a> </p> [FlashRL](https://fengyao.notion.site/flash-rl) patches the inference backend to generate RL rollouts in INT8 \& FP8, with accurate rollout logprob.  *Figure 1. **Left**: AIME accuracy of Qwen2.5-32B DAPO training with <span style="color: #ff7f0e;">INT8</span> and <span style="color: #1f77b4;">BF16</span> precisions for rollout generation using vLLM engine. **Right**: Training throughput (updates per hour) in the DAPO training (vLLM + BF16 FSDP).* ## ⚡ Quick Start ### Installation (Auto Patch) ```bash pip install flash-llm-rl # need to be installed in all nodes in multi-node training ``` (Optional) there are two options to verify the FlashRL install: 1) set `FLASHRL_LOGGING_LEVEL` to `DEBUG` and compare the log with the [provided ones](#examples); 2) for more details / debugging, please follow the [Tutorial](/tutorial/README.md). ### Manual Patch When the FlashRL patch conflicts with environments / other setups, an alternative approach is to: 1. cleanup FlashRL patch via `flashrl cleanup`, this will disable the auto patch; 2. manually add `import flash_rl` as in https://github.com/yaof20/verl/pull/1 ### Rollout Generation w. FP8 Quantization FlashRL is implemented as a plug-in-and-play manner, using [environment variables](#patcher) `FLASHRL_CONFIG` to control the quantization precision. Note that, due to quantization overhead, it is only recommended to use quantized rollout in the combination of large model (i.e., 14B+, preferrable 32B+) and long cot generation (DAPO training instead of GSM8K training). ```bash # for single-node job export FLASHRL_CONFIG=fp8 bash verl/examples/ppo_trainer/run_qwen2.5-32b.sh # alternatively, for multi-node jobs via `ray submit`, fp8 online quantization will be turned on via # > echo " FLASHRL_CONFIG: 'fp8'" | tee -a verl/trainer/runtime_env.yaml # add `FLASHRL_CONFIG: 'fp8'` to runtime env # > bash verl/recipe/dapo/run_dapo_qwen2.5_32b.sh # this can be any scripts ``` ### RL Logprob Patch Only Setting the config to `bf16` to extract precise logprob used in sampling without rollout quantization. This is useful for applying the [Truncated Importance Sampling](https://fengyao.notion.site/off-policy-rl?source=copy_link). ```bash # for single-node job export FLASHRL_CONFIG=bf16 bash verl/examples/ppo_trainer/run_qwen2.5-32b.sh # alternatively, for multi-node jobs via `ray submit`, RL Logprob Patch Only will be turned on via # > echo " FLASHRL_CONFIG: 'bf16'" | tee -a verl/trainer/runtime_env.yaml # add `FLASHRL_CONFIG: 'fp8'` to runtime env # > bash verl/recipe/dapo/run_dapo_qwen2.5_32b.sh # this can be any scripts ``` ## Usage Guide FlashRL has 3 major functionality, `profiling`, `configure helper`, and `patcher`. ### Profiling (optional for `fp8` and `bf16`) This step is not needed for the native `fp8` online quantization supported by `vLLM`, and the logprog-only path `bf16`, and is needed for `int8` or `fp8_channel` quantization. Specifically, profilling compares a `bf16` model and a quantized model to decide how the online quantization should be performed for an updated model. Please find below an example for `Qwen/Qwen2.5-32B` and `Qwen/Qwen2.5-0.5B-Instruct`. The quantized model can be any `w8a8`/`fp8` model produced by `llm-compressor`. Note that, Redhat AI provides various [quanti
Excerpt of 10,502 characters
Read on GitHubWould you bet a product on this? Bounded 0–100 and slow moving.
matched fp:a85c869666b19011, topic:vllm
matched fp:a85c869666b19011, readme:quantization, readme:quantized, desc:int8
matched fp:a85c869666b19011, topic:reinforcement-learning