LMSYS, AMD and Miles train Qwen3-30B-A3B SWE agent on MI355X

LMSYS Org, with AMD and the Miles team, trained Qwen3-30B-A3B as a software engineering agent on AMD MI355X GPUs using fully asynchronous agentic reinforcement learning, according to LMSYS Org's post on X. The team used SGLang, Megatron-LM and Harbor for the training.

LMSYS says SWE-bench Verified pass@1 rose from 7.3% to 27.5% on a 100-task subset, with four attempts per task, after SFT and RL.

LMSYS reports that RL steps dropped from 850 seconds to 284 seconds and rollout throughput rose about 4x, to roughly 3,480 trajectories per hour with 384 sandboxes running at once. The post says keeping long trajectories and hundreds of CPU sandboxes from stalling training required significant engineering work, and it links a full report dated October 22.