OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
AI Infrastructure & Serving

Flash Rl

Flash Rl is a Python library that patches an inference backend to generate INT8 or FP8 reinforcement-learning rollouts while retaining rollout log-probabilities. Its documented workflows connect the patcher to vLLM-based training, including profiling and YAML configuration for some quantization modes.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Flash Rl — OpenViST project profile in AI Infrastructure & Serving.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Consider Flash Rl when you already have an RL training pipeline and want to configure quantized rollouts through a patched inference backend. The README particularly recommends this approach for large models and long chain-of-thought generation, where it says quantization overhead is more suitable.

Who it is for

  • Researchers and engineers maintaining reinforcement-learning training pipelines.
  • Teams using vLLM for rollout generation and looking to configure quantized rollout paths.
  • Practitioners who need sampling log-probabilities without quantizing rollouts, using the documented BF16 mode.

Documented uses

  • Configure FP8 or INT8 rollout generation in a reinforcement-learning training workflow.
  • Profile a BF16 model against a quantized model and generate the YAML configuration used by the patcher for INT8 or FP8-channel paths.
  • Extract precise sampling log-probabilities through the BF16 mode without quantizing rollouts.
Before you choose

Requirements & limits

The public GitHub repository lists an MIT license. The README specifies Python 3.10 or later and requires installation on all nodes for multi-node training.

  • Native vLLM FP8 and BF16 log-probability-only paths do not require profiling or configuration generation; INT8 and FP8-channel paths do.
  • The project warns that quantization overhead makes small-model or short-generation workloads less suitable for quantized rollouts.
  • If automatic patching conflicts with the environment, the README documents cleanup followed by a manual import-based patch.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Prepare an INT8 rollout configuration

Profile a model and create the YAML configuration used by Flash Rl's patcher for an INT8 rollout setup.

Before you start

  • Python 3.10 or later.
  • A BF16 model and a quantized w8a8 or FP8 model; the README examples use Qwen models.
  • Flash Rl installed in the training environment.

Steps

  1. Run flashrl profile with the BF16 model, quantized model, output profile path, and --fn int8.

  2. Run flashrl setup with the quantized model, generated profile path, --fn int8, and desired YAML output path.

  3. Set FLASHRL_CONFIG to the generated YAML path for the patcher to use.

Expected outcome

A profile and YAML configuration are prepared for Flash Rl's INT8 patcher.

Guide limitations

  • Profiling and configuration generation are required for INT8; the README says they are not needed for native vLLM FP8 or the BF16 log-probability-only path.
  • The README recommends quantized rollouts for large models and long chain-of-thought generation because of quantization overhead.
Documentation 1 ↗
Source-based practical guide

Select native FP8 or BF16 log-probability mode

Set the documented Flash Rl mode for a single-node verl training script.

Before you start

  • Python 3.10 or later and Flash Rl installed.
  • An existing single-node verl training script, as in the README examples.

Steps

  1. Set FLASHRL_CONFIG=fp8 to select the documented native FP8 rollout path, or set FLASHRL_CONFIG=bf16 for the log-probability-only path.

  2. Run the existing training script, following the README's example command for the selected mode.

Expected outcome

The patcher is configured for FP8 rollouts or BF16 log-probability extraction, depending on the selected value.

Guide limitations

  • The BF16 mode does not quantize rollouts.
  • The README presents these commands as examples using verl scripts; it does not establish compatibility with arbitrary training setups.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
15 October 2026
Original website title
yaof20/Flash-RL · Official homepage title · retained in its source language
How we review sourcesSuggest a correction