miles
AI overview
Sign in and the AI will write an overview from our coverage.
Headlines · 9
- LMSYS trains Qwen3 coding agent with asynchronous RL on AMD MI355X
LMSYS Blog reports that its team trained Qwen3-30B-A3B-Instruct for coding-agent tasks using fully asynchronous agentic reinforcement learning on AMD Instinct MI355X, with Miles, SGLang, Megatron-LM, and Harbor. On a vendor-run, 100-task stratified subset of SWE-bench Verified, pass@1 rose from 7.3% to 27.5% after supervised fine-tuning and agentic RL. The system used separate inference and training nodes, capped trajectories at 64K tokens, and moved sandboxes to E2B cloud microVMs, supporting 384 concurrent trajectories and about 3,480 trajectories per hour.
LMSYS Blog · 🔥 0 - Moore Threads Open-Sources Megatron-MUSA-Titan for Large-Scale MoE Training
Moore Threads announced the open-sourcing of Megatron-MUSA-Titan, a large-model training infrastructure project adapted for the company’s MUSA platform and full-function GPUs. IT之家 reported that it targets trillion-parameter MoE pretraining and reinforcement learning on clusters with 10,000 GPUs. The project supports blockwise FP8 training and can work with the DeepEP and ACE asynchronous communication engines to overlap expert parallelism and computation. Moore Threads said the project had been used in continued trillion-parameter MoE pretraining on its 10,000-GPU-scale cluster and supports coordination with slime and miles, Megatron training, and SGLang sample generation. Its reported model FLOPS utilization and long-running stability were not independently verified in the supplied reports.
IT之家 AI · 🔥 0 - SGLang launches beta plugin for AWS Trainium inference and RL
The SGLang Team, in collaboration with AWS, announced a beta plugin for running LLM inference serving and reinforcement learning rollouts on AWS Trainium2 and Trainium3. Built on TorchNeuron, the beta initially covers Qwen3, Qwen3-MoE-30B-A3B, DeepSeek-V2-Lite, GLM-4, and GLM-5.2/5.3, with parallelism, Radix Cache, BF16, FP8, and an asynchronous scheduler. In an early end-to-end test, SGLang generated Qwen3-MoE-30B-A3B rollouts on Trainium3 while Miles trained a GRPO policy with Megatron on one H200 node; performance work remains ongoing.
LMSYS Blog · 🔥 0 - MiniMax M3.1 Flash adds SGLang and Miles integration
MiniMax said MiniMax M3.1 Flash can run with SGLang for inference and reinforcement learning through Miles. The company claimed it exceeded 1,000 tokens per second for batch-size-1 decoding with SGLang and was production-ready from day one. The integration was described as a joint effort involving MiniMax, NVIDIA, AMD, and RadixArk, but the report included no independent throughput measurement.
MiniMax_AI (X) · 🔥 0 - Modal Releases Open-Source Spindle Training Server
Modal released Spindle, an open-source training server compatible with Thinking Machines' Tinker API that runs in a user's Modal workspace. It supports full-parameter training and multi-LoRA for open models, while users control the server and pay for GPU time rather than tokens. Modal says its default configurations are tuned to maximize GPU utilization and that Spindle can support hundreds of concurrent training runs at similar performance and lower cost than closed-source APIs; the release remains in beta.
Modal Blog · 🔥 0 - Modal launches Modal Dojo, a library for production-grade LLM training
Modal announced Modal Dojo, a library it says provides the building blocks, tuned defaults and observability needed for production-grade LLM training, and said it is available now. The library builds on open-source frameworks Miles and Slime, to which Modal says it has contributed upstream patches, and ships an out-of-box observability dashboard covering loss and reward curves, rollout trajectories and substep profiling. Modal also says training can be started from a single prompt describing the model, dataset and target agent, and cites customer testimonials from Axiom Bio and Afterquery about long-context reinforcement-learning runs.
Modal Blog · 🔥 0 - SGLang and Miles add day-0 support for Muse Spark Open
The SGLang team says SGLang and Miles added day-0 support for Muse Spark Open. According to the post, Muse Spark Open is an open-weight multimodal mixture-of-experts model with an 816B-parameter text backbone, roughly 42 billion active parameters per token, a 1M-token context window, text, image, video and audio inputs, and checkpoints in NVFP4 and BF16. The post's vendor-run figures for the NVFP4 checkpoint on eight NVIDIA B300 GPUs using an OpenHands multi-turn agentic coding replay show 14.8K tokens/sec per GPU at 543 tokens/sec per user (p50) at concurrency 1, and 80.3K tokens/sec per GPU at 61 tokens/sec per user (p50) at concurrency 128.
LMSYS Blog · 🔥 0 - Modal explores mixed-precision, mixed-accelerator RL samplers
Modal published a technical post exploring whether large-model reinforcement learning can use samplers at different precisions and on different accelerators while keeping training stable. It says sampling already consumes 80–85% of GPUs in large RL runs, and reports fully asynchronous training of Qwen3.6-35B-A3B on agentic coding tasks using off-policy correction and autoscaling serverless GPUs. Samplers spanned six accelerator types — B200, B300, H100, H200, A100 and RTX PRO 6000 — feeding a BF16 trainer on B200, with evaluation on SWE-bench Pro V2 and its HARD-51 subset.
Modal Blog · 🔥 0 - Blog SGLang and Miles on NVIDIA Vera Rubin We had early access to two NVIDIA Vera Rubin nodes (8 GPUs) and used them to bring SGLang and Miles up on the platform: SGLang for inference and Miles for RL training. SGLang & Miles Team October 9, 2026
On this page Summary Inference: Serving Kimi K3 NVFP4 Attention MoE and communication kernels KDA RL training: Miles on NVIDIA Vera Rubin Agentic RL on the Vera CPU What's Next Acknowledgements Summary We had early access to two NVIDIA Vera Rubin nodes (8 GPUs) and used them to bring SGLang and Miles up on the platform: SGLang for inference and Miles for RL training. On the inference side, we tuned SGLang for Kimi K3 NVFP4: this post covers what we changed in the attention, MoE/communication and KDA kernels, and how much each change bought. On the training side, Miles, which uses SG
LMSYS Blog · 🔥 0
Experience and discussion from the community
Share my miles experienceAsk about miles
Nobody has shared their experience with miles yet.