> ## Documentation Index
> Fetch the complete documentation index at: https://miles.radixark.com/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Platforms

> Hardware-specific tutorials. Most users want NVIDIA H/B; AMD MI300X is supported via ROCm.

Miles runs on NVIDIA H/B-series and AMD MI300X with the same launch scripts. Each platform page covers driver versions, build flags, and the FP8 / ROCm quirks you need to know before kicking off a job.

<CardGroup cols={2}>
  <Card title="NVIDIA GPUs" icon="microchip" href="/docs/platforms/nvidia">
    The default GB300 / GB200 / B200 / B100 / H200 / H100 with FP8, NVLink, and InfiniBand.
  </Card>

  <Card title="AMD GPUs" icon="microchip" href="/docs/platforms/amd">
    ROCm 6.3+ with patches for virtual memory management. Same launch scripts.
  </Card>
</CardGroup>

## Supported features by GPU

| Feature                     | H / B-series  | A100 | MI300X            |
| --------------------------- | ------------- | ---- | ----------------- |
| BF16 training               | ✅             | ✅    | ✅                 |
| FP8 GEMM                    | ✅             | -    | ✅                 |
| INT4 W4A16 QAT              | ✅             | ✅    | ⚠️                |
| Speculative decoding        | ✅             | ✅    | ✅                 |
| Rollout Routing Replay (R3) | ✅             | ✅    | ✅                 |
| P2P weight transfer (RDMA)  | ✅ IB / RoCEv2 | ✅    | ✅ Infinity Fabric |
| Megatron CP                 | ✅             | ✅    | ✅                 |
| Deterministic inference     | ✅             | ✅    | ⚠️                |

## Storage and network

Independent of GPU vendor, you'll want:

* **Shared filesystem** for multi-node — NFS, GPFS, Lustre. Reads dominate writes during
  training; provision read bandwidth.
* **High-bandwidth interconnect** — IB (NVIDIA), RoCEv2 (NVIDIA), Slingshot, or
  Infinity Fabric (AMD). 200+ GB/s per node is typical for trillion-param training.
* **NVMe local scratch** for SGLang radix cache and Ray spill — at least 1 TB.