Posts by Rishi Madduri
Scaling RL with verl on AMD Instinct MI355X: Async Walkthrough and Sync Benchmark
- 18 August 2026
Reinforcement learning (RL) for large language models (LLMs) alternates between two phases: generation (rollout), where the current policy produces responses, and training, where those responses are used to update the policy. In verl, the key design choices are when these phases run relative to each other (synchronously or with overlap) and where they run (colocated on the same GPUs or on separate GPU pools). This blog first explains the differences between the two modes and when to use each.
FlashInfer on ROCm: High‑Throughput Prefill Attention via AITER
- 06 April 2026
The explosive growth of large language models (LLMs) like DeepSeek-R1, Llama 3, and Qwen 3 has created an urgent need for efficient inference solutions. As these models scale to billions of parameters and context lengths extend to hundreds of thousands of tokens, the attention mechanism becomes a critical bottleneck, consuming substantial memory for key-value (KV) caches and requiring significant compute for each token generated.
Enabling FlashInfer on ROCm for Accelerated LLM Serving
- 01 October 2025
FlashInfer is an innovative framework designed to accelerate inference of large language models (LLMs). Given the explosive growth and adoption of models like DeepSeek R1, Llama 3, and Qwen 3, efficient inference is critical to meet the demands of real-world deployment. However, challenges such as GPU memory bottlenecks, throughput limitations, and latency remain significant hurdles for deploying these models at scale.
Efficient MoE training on AMD ROCm: How-to use MegaBlocks on AMD GPUs
- 23 March 2025
Training massive deep-learning models requires a balance of efficiency and scalability. In the context of the Transformers architecture, Mixture of Experts (MoE) models are massive machine learning architectures characterized for dividing tasks among multiple specialized sub-networks or “experts”. A gating network determines the expert to which a given input should be routed, enabling the model to handle complex tasks more efficiently by using the specialized capabilities of each expert. This dynamic routing mechanism allows MoE models to scale efficiently, activating only a subset of the network for each input, therefore reducing computational load while maintaining high model capacity.