DeepSeek-V4 Flash RL Runs on AMD MI355X
ying11231 · x · 2026-07-11
lmsysorg published a blog post detailing how to run the training and rollout of DeepSeek-V4 Flash RL end-to-end on AMD Instinct MI355X GPUs, utilizing Miles and ROCm for alignment.
Key results highlighted in the article include:
- Validated 100+ optimization steps across four MI355X nodes
- The train-rollout log-prob gap is constrained around 0.09
- On AIME-2024, pass@1 increased from 0.39 to 0.49, and pass@8 improved from 0.53 to 0.67
- Utilized FP8 rollout + BF16 actor, supporting datatype-aware online weight updates
- Achieved a stable TP1 / PP4 / EP4 layout on ROCm without collective stalls
- Demonstrated consistency across training and inference for hybrid attention, mHC mixing, and hash-routed MoE
Overall, this is an engineering-focused infrastructure piece emphasizing how multi-GPU training/inference stacks, numerical formats, and parallel layouts work in tandem.
More from Infra
- China’s AI arms race is increasingly defined by chips, data centers, and open models — BenBajarin · 2026-07-22
- Agent search bottlenecks are now about variance, not raw latency — rohanpaul_ai · 2026-07-22
- Gavin Baker argues Nvidia may be one of open source AI’s biggest supporters — GavinSBaker · 2026-07-22
- AI Power Demand Exposes US Energy Gap, Urging Shift from Scarcity to Abundance — bradneuberg · 2026-07-22
- Gavin Baker says Nvidia’s $630B figure would be system revenue, not all Nvidia’s — GavinSBaker · 2026-07-22
- A Firecracker-based platform says it can host 6,000 AI agents on one 256 GB server — maritime_sh · 2026-07-22