＋ Add to collection

Install (CLI)

npx skillmds add orchestra-research/miles-rl-training

Or Copy and Paste in Your Agent chat

Please install the Agent Skill "orchestra-research/miles-rl-training" for me from skillmd.com. A SkillMD skill is a SKILL.md file — plain markdown plus a few companion files — that you save into my skills directory. Use whichever fetch route fits your setup: (1) if you have a terminal, run npx -y skillmds add orchestra-research/miles-rl-training; (2) otherwise fetch the file manifest at https://api.skillmd.com/api/skills/orchestra-research/miles-rl-training/bundle?format=manifest and download each listed file from its "url" into miles-rl-training/ in your skills directory; (3) if the SkillMD MCP tools are connected, call skillmd_download with slug "orchestra-research/miles-rl-training" and save the returned files under miles-rl-training/. Show me the SKILL.md before saving if you'd like me to review it. If your skills directory is read-only (e.g. the Claude desktop app), save the files to a working folder instead and tell me — I'll add them through my app's skills settings.

README badge

Show this skill's registry status in your repo:

[![SkillMD](https://skillmd.com/badge/orchestra-research/miles-rl-training.svg)](https://skillmd.com/skills/orchestra-research/miles-rl-training)

miles-rl-training

Name: miles-rl-training
Rating: 4.5 (4 reviews)
Author: Orchestra Research

orchestra-research/miles-rl-training · Agent Skill (multi-file)

by Orchestra Research · bundle

Published Jul 5, 2026 · Last updated Jul 6, 2026

Train large-scale MoE models with FP8/INT4 low-precision RL, speculative decoding, and train-inference alignment using the miles framework.

SKILL.md

miles: Enterprise-Grade RL for Large-Scale Model Training

miles is a high-performance, enterprise-ready RL framework optimized for large-scale model post-training. Built as a production fork of slime, it addresses critical challenges in MoE training stability, low-precision training, and train-inference alignment.

When to Use miles

Choose miles when you need:

Training 1TB+ MoE models (DeepSeek V3, Qwen3-MoE)
FP8 or INT4 quantization-aware training
Bit-wise identical train-inference alignment
Speculative RL for maximum throughput
Production stability with enterprise support

Consider alternatives when:

You want the research-grade original → use slime
You need flexible backend swapping → use verl
You want PyTorch-native abstractions → use torchforge

Key Features

Low-Precision Training

Unified FP8: End-to-end FP8 for both inference and training
INT4 QAT: 1TB models on single-machine VRAM (H200)
Rollout Routing Replay (R3): Bit-wise expert alignment for MoE

Performance Optimizations

Speculative RL: 25%+ rollout speedup with online SFT draft models
Zero-Copy Weight Sync: CUDA IPC zero-copy mapping
Partial Rollout: Recycle half-finished trajectories

Train-Inference Alignment

TIS/MIS: Truncated/Masked Importance Sampling for off-policy correction
Kernel-level optimization: FlashAttention-3, DeepGEMM integration

Installation

# Recommended: Docker
docker pull radixark/miles:latest
docker run --rm --gpus all --ipc=host --shm-size=16g \
  -it radixark/miles:latest /bin/bash

# From source
git clone https://github.com/radixark/miles.git
cd miles
pip install -r requirements.txt
pip install -e .

Quick Start

miles inherits slime's configuration system. Basic training:

python train.py \
    --advantage-estimator grpo \
    --model-name qwen3-30b-a3b \
    --hf-checkpoint /path/to/qwen3-30b-a3b-hf \
    --rollout-batch-size 512 \
    --n-samples-per-prompt 8

Workflow 1: Large MoE Training

Use this workflow for training large MoE models like DeepSeek V3 or Qwen3-MoE.

Prerequisites Checklist

H100/H200 GPUs with FP8 support
MoE model (DeepSeek V3, Qwen3-MoE)
Docker environment with miles

Step 1: Environment Setup

# FP8 block scaling (recommended for stability)
export NVTE_FP8_BLOCK_SCALING_FP32_SCALES=1
export CUDA_DEVICE_MAX_CONNECTIONS=1

Step 2: Configure Training

python train.py \
    --actor-num-gpus-per-node 8 \
    --rollout-num-gpus 8 \
    --hf-checkpoint /path/to/deepseek-v3 \
    --advantage-estimator grpo \
    --tensor-model-parallel-size 8 \
    --expert-model-parallel-size 4 \
    --prompt-data /path/to/data.jsonl \
    --num-rollout 3000

Verification Checklist

Model loads without errors
Routing decisions are consistent
No NaN/Inf in loss values

Workflow 2: Speculative RL Training

Use this workflow for maximum rollout throughput with EAGLE speculative decoding.

How Speculative RL Works

Small draft model generates candidate tokens
Target model verifies in parallel
Draft model updated via online SFT to track policy

Step 1: Enable Speculative Decoding

miles supports EAGLE speculative decoding via SGLang:

python train.py \
    --actor-num-gpus-per-node 8 \
    --hf-checkpoint /path/to/target-model \
    --sglang-speculative-algorithm EAGLE \
    --sglang-speculative-num-steps 3 \
    --sglang-speculative-eagle-topk 1 \
    --sglang-speculative-num-draft-tokens 4 \
    --sglang-speculative-draft-model-path /path/to/draft-model \
    --advantage-estimator grpo \
    --prompt-data /path/to/data.jsonl

Step 2: Enable Online MTP Training (Optional)

For online SFT of draft model during training:

--mtp-num-layers 1 \
--enable-mtp-training \
--mtp-loss-scaling-factor 0.2

Note: Online MTP training requires a torch dist checkpoint with MTP weights. Add --mtp-num-layers 1 during checkpoint conversion from HuggingFace.

Expected Speedup

Standard rollout: Baseline
Speculative RL: 25-40% faster rollout
With partial rollout: Additional 10-15% throughput

Configuration Reference

miles inherits all slime arguments. See slime API Reference for the complete list.

Cluster Resources (from slime)

--actor-num-nodes 1
--actor-num-gpus-per-node 8
--rollout-num-gpus 8
--rollout-num-gpus-per-engine 2
--colocate

Megatron Parallelism (from slime)

--tensor-model-parallel-size 8
--pipeline-model-parallel-size 2
--expert-model-parallel-size 4    # MoE expert parallelism

Speculative Decoding (miles-specific)

--sglang-speculative-algorithm EAGLE
--sglang-speculative-num-steps 3
--sglang-speculative-eagle-topk 1
--sglang-speculative-num-draft-tokens 4
--sglang-enable-draft-weights-cpu-backup
--sglang-speculative-draft-model-path /your/draft/model/path

Online MTP Training (miles-specific)

--mtp-num-layers 1
--enable-mtp-training
--mtp-loss-scaling-factor 0.2

Key Features (Conceptual)

The following features are documented in miles but specific CLI flags may vary. Consult the miles repository for latest configuration.

Unified FP8 Pipeline

End-to-end FP8 sampling and training that eliminates quantization-induced discrepancy causing RL collapse in MoE models.

Rollout Routing Replay (R3)

Records expert routing decisions during SGLang inference and replays them during Megatron training for bit-wise expert alignment.

How R3 Works:

During SGLang inference, expert routing decisions are recorded
Routing decisions stored in sample.rollout_routed_experts
During Megatron training, routing is replayed instead of recomputed
Ensures identical expert selection between train and inference

INT4 Quantization-Aware Training

Enables single-machine deployment of 1TB+ models (e.g., on H200).

Memory Savings with INT4:

Model Size	BF16 VRAM	INT4 VRAM	Reduction
70B	140GB	45GB	3.1x
235B	470GB	150GB	3.1x
671B	1.3TB	420GB	3.1x

Train-Inference Alignment

miles achieves "exactly 0 KL divergence" between training and inference through:

Flash Attention 3
DeepGEMM
Batch-invariant kernels from Thinking Machines Lab
torch.compile integration

Sample Data Structure

miles uses the same Sample dataclass as slime with the rollout_routed_experts field for MoE routing replay:

@dataclass
class Sample:
    prompt: str | list[dict]
    tokens: list[int]
    response: str
    reward: float | dict
    loss_mask: list[int]
    status: Status
    metadata: dict
    rollout_log_probs: list[float]
    rollout_routed_experts: list[list[int]]  # MoE routing for R3

See slime API Reference for the complete Sample definition.

Common Issues and Solutions

Issue: FP8 Training Collapse

Symptoms: Loss explodes, NaN values

Solutions:

Use block scaling: export NVTE_FP8_BLOCK_SCALING_FP32_SCALES=1
Reduce learning rate: --lr 5e-7
Ensure MoE routing is consistent between train/inference

Issue: Speculative Draft Drift

Symptoms: Low acceptance rate over time

Solutions:

Enable online MTP training to keep draft model aligned
Reduce speculative steps: --sglang-speculative-num-steps 2
Use CPU backup: --sglang-enable-draft-weights-cpu-backup

Issue: Train-Inference Mismatch

Symptoms: Policy divergence, reward collapse

Solutions:

Use TIS for off-policy correction: --use-tis --tis-threshold 0.9
Verify log probs match between SGLang and Megatron
Enable R3 for MoE models

Supported Models

Family	Models	MoE Support
DeepSeek	R1, V3, V3.2	Full
Qwen	2, 2.5, 3 (including MoE)	Full
Llama	3, 3.1, 3.3, 4	Dense only
Gemma	2, 3, 3N	Dense only
GLM	4.5, 4.6, 4.7	Dense only
MiniMax	M2, M2.1	Full

Resources

GitHub: https://github.com/radixark/miles
Introduction Blog: https://lmsys.org/blog/2025-11-19-miles/
Slime (upstream): https://github.com/THUDM/slime
SGLang: https://github.com/sgl-project/sglang

---
name: miles-rl-training
description: Train large-scale MoE models with FP8/INT4 low-precision RL, speculative decoding, and train-inference alignment using the miles framework.
license: MIT
---

# miles: Enterprise-Grade RL for Large-Scale Model Training

miles is a high-performance, enterprise-ready RL framework optimized for large-scale model post-training. Built as a production fork of slime, it addresses critical challenges in MoE training stability, low-precision training, and train-inference alignment.

## When to Use miles

**Choose miles when you need:**
- Training 1TB+ MoE models (DeepSeek V3, Qwen3-MoE)
- FP8 or INT4 quantization-aware training
- Bit-wise identical train-inference alignment
- Speculative RL for maximum throughput
- Production stability with enterprise support

**Consider alternatives when:**
- You want the research-grade original → use **slime**
- You need flexible backend swapping → use **verl**
- You want PyTorch-native abstractions → use **torchforge**

## Key Features

### Low-Precision Training
- **Unified FP8**: End-to-end FP8 for both inference and training
- **INT4 QAT**: 1TB models on single-machine VRAM (H200)
- **Rollout Routing Replay (R3)**: Bit-wise expert alignment for MoE

### Performance Optimizations
- **Speculative RL**: 25%+ rollout speedup with online SFT draft models
- **Zero-Copy Weight Sync**: CUDA IPC zero-copy mapping
- **Partial Rollout**: Recycle half-finished trajectories

### Train-Inference Alignment
- **TIS/MIS**: Truncated/Masked Importance Sampling for off-policy correction
- **Kernel-level optimization**: FlashAttention-3, DeepGEMM integration

## Installation

```bash
# Recommended: Docker
docker pull radixark/miles:latest
docker run --rm --gpus all --ipc=host --shm-size=16g \
  -it radixark/miles:latest /bin/bash

# From source
git clone https://github.com/radixark/miles.git
cd miles
pip install -r requirements.txt
pip install -e .
```

## Quick Start

miles inherits slime's configuration system. Basic training:

```bash
python train.py \
    --advantage-estimator grpo \
    --model-name qwen3-30b-a3b \
    --hf-checkpoint /path/to/qwen3-30b-a3b-hf \
    --rollout-batch-size 512 \
    --n-samples-per-prompt 8
```

---

## Workflow 1: Large MoE Training

Use this workflow for training large MoE models like DeepSeek V3 or Qwen3-MoE.

### Prerequisites Checklist
- [ ] H100/H200 GPUs with FP8 support
- [ ] MoE model (DeepSeek V3, Qwen3-MoE)
- [ ] Docker environment with miles

### Step 1: Environment Setup

```bash
# FP8 block scaling (recommended for stability)
export NVTE_FP8_BLOCK_SCALING_FP32_SCALES=1
export CUDA_DEVICE_MAX_CONNECTIONS=1
```

### Step 2: Configure Training

```bash
python train.py \
    --actor-num-gpus-per-node 8 \
    --rollout-num-gpus 8 \
    --hf-checkpoint /path/to/deepseek-v3 \
    --advantage-estimator grpo \
    --tensor-model-parallel-size 8 \
    --expert-model-parallel-size 4 \
    --prompt-data /path/to/data.jsonl \
    --num-rollout 3000
```

### Verification Checklist
- [ ] Model loads without errors
- [ ] Routing decisions are consistent
- [ ] No NaN/Inf in loss values

---

## Workflow 2: Speculative RL Training

Use this workflow for maximum rollout throughput with EAGLE speculative decoding.

### How Speculative RL Works

1. Small draft model generates candidate tokens
2. Target model verifies in parallel
3. Draft model updated via online SFT to track policy

### Step 1: Enable Speculative Decoding

miles supports EAGLE speculative decoding via SGLang:

```bash
python train.py \
    --actor-num-gpus-per-node 8 \
    --hf-checkpoint /path/to/target-model \
    --sglang-speculative-algorithm EAGLE \
    --sglang-speculative-num-steps 3 \
    --sglang-speculative-eagle-topk 1 \
    --sglang-speculative-num-draft-tokens 4 \
    --sglang-speculative-draft-model-path /path/to/draft-model \
    --advantage-estimator grpo \
    --prompt-data /path/to/data.jsonl
```

### Step 2: Enable Online MTP Training (Optional)

For online SFT of draft model during training:

```bash
--mtp-num-layers 1 \
--enable-mtp-training \
--mtp-loss-scaling-factor 0.2
```

**Note**: Online MTP training requires a torch dist checkpoint with MTP weights. Add `--mtp-num-layers 1` during checkpoint conversion from HuggingFace.

### Expected Speedup

- **Standard rollout**: Baseline
- **Speculative RL**: 25-40% faster rollout
- **With partial rollout**: Additional 10-15% throughput

---

## Configuration Reference

miles inherits all slime arguments. See [slime API Reference](../slime/references/api-reference.md) for the complete list.

### Cluster Resources (from slime)

```bash
--actor-num-nodes 1
--actor-num-gpus-per-node 8
--rollout-num-gpus 8
--rollout-num-gpus-per-engine 2
--colocate
```

### Megatron Parallelism (from slime)

```bash
--tensor-model-parallel-size 8
--pipeline-model-parallel-size 2
--expert-model-parallel-size 4    # MoE expert parallelism
```

### Speculative Decoding (miles-specific)

```bash
--sglang-speculative-algorithm EAGLE
--sglang-speculative-num-steps 3
--sglang-speculative-eagle-topk 1
--sglang-speculative-num-draft-tokens 4
--sglang-enable-draft-weights-cpu-backup
--sglang-speculative-draft-model-path /your/draft/model/path
```

### Online MTP Training (miles-specific)

```bash
--mtp-num-layers 1
--enable-mtp-training
--mtp-loss-scaling-factor 0.2
```

---

## Key Features (Conceptual)

The following features are documented in miles but specific CLI flags may vary. Consult the miles repository for latest configuration.

### Unified FP8 Pipeline

End-to-end FP8 sampling and training that eliminates quantization-induced discrepancy causing RL collapse in MoE models.

### Rollout Routing Replay (R3)

Records expert routing decisions during SGLang inference and replays them during Megatron training for bit-wise expert alignment.

**How R3 Works**:
1. During SGLang inference, expert routing decisions are recorded
2. Routing decisions stored in `sample.rollout_routed_experts`
3. During Megatron training, routing is replayed instead of recomputed
4. Ensures identical expert selection between train and inference

### INT4 Quantization-Aware Training

Enables single-machine deployment of 1TB+ models (e.g., on H200).

**Memory Savings with INT4**:

| Model Size | BF16 VRAM | INT4 VRAM | Reduction |
|------------|-----------|-----------|-----------|
| 70B | 140GB | 45GB | 3.1x |
| 235B | 470GB | 150GB | 3.1x |
| 671B | 1.3TB | 420GB | 3.1x |

### Train-Inference Alignment

miles achieves "exactly 0 KL divergence" between training and inference through:
- Flash Attention 3
- DeepGEMM
- Batch-invariant kernels from Thinking Machines Lab
- `torch.compile` integration

---

## Sample Data Structure

miles uses the same `Sample` dataclass as slime with the `rollout_routed_experts` field for MoE routing replay:

```python
@dataclass
class Sample:
    prompt: str | list[dict]
    tokens: list[int]
    response: str
    reward: float | dict
    loss_mask: list[int]
    status: Status
    metadata: dict
    rollout_log_probs: list[float]
    rollout_routed_experts: list[list[int]]  # MoE routing for R3
```

See [slime API Reference](../slime/references/api-reference.md) for the complete Sample definition.

---

## Common Issues and Solutions

### Issue: FP8 Training Collapse

**Symptoms**: Loss explodes, NaN values

**Solutions**:
- Use block scaling: `export NVTE_FP8_BLOCK_SCALING_FP32_SCALES=1`
- Reduce learning rate: `--lr 5e-7`
- Ensure MoE routing is consistent between train/inference

### Issue: Speculative Draft Drift

**Symptoms**: Low acceptance rate over time

**Solutions**:
- Enable online MTP training to keep draft model aligned
- Reduce speculative steps: `--sglang-speculative-num-steps 2`
- Use CPU backup: `--sglang-enable-draft-weights-cpu-backup`

### Issue: Train-Inference Mismatch

**Symptoms**: Policy divergence, reward collapse

**Solutions**:
- Use TIS for off-policy correction: `--use-tis --tis-threshold 0.9`
- Verify log probs match between SGLang and Megatron
- Enable R3 for MoE models

---

## Supported Models

| Family | Models | MoE Support |
|--------|--------|-------------|
| DeepSeek | R1, V3, V3.2 | Full |
| Qwen | 2, 2.5, 3 (including MoE) | Full |
| Llama | 3, 3.1, 3.3, 4 | Dense only |
| Gemma | 2, 3, 3N | Dense only |
| GLM | 4.5, 4.6, 4.7 | Dense only |
| MiniMax | M2, M2.1 | Full |

---

## Resources

- **GitHub**: https://github.com/radixark/miles
- **Introduction Blog**: https://lmsys.org/blog/2025-11-19-miles/
- **Slime (upstream)**: https://github.com/THUDM/slime
- **SGLang**: https://github.com/sgl-project/sglang

Files

This skill is a package of 3 files. Install with the command above, or download the folder.

📄SKILL.md entry
📁references
📄api-reference.md 4.0 KB
📄troubleshooting.md 5.7 KB

Orchestra-Research/AI-Research-SKILLs/tree/main/06-post-training/miles commit · 773a52944b

▌
slime-rl-training · orchestra-research bundle
Post-train LLMs with reinforcement learning using the slime framework, which integrates Megatron-LM for training and SGLang for rollout generation.

10.4k
repo stars
▌
nemo-mbridge-recipe-recommender · nvidia bundle
Indexes Megatron Bridge recipes and recommends the best starting config based on model, GPU count, and training goal.

2.2k
repo stars
▌
slime-rl-training · lord1egypt
Guides LLM post-training with RL using slime, a Megatron+SGLang framework for training GLM, Qwen, DeepSeek, and Llama models with GRPO, async, and multi-turn workflows.

2
repo stars
▌
nemo-automodel-model-onboarding · nvidia bundle
Guides implementation of new model architectures in NeMo AutoModel through five phases: discovery, implementation, registration, validation, and testing.

2.2k
repo stars
▌
stable-baselines3 · k-dense-ai bundle
Train reinforcement learning agents using PPO, SAC, DQN, TD3, DDPG, and A2C algorithms with a scikit-learn-like API. Supports custom Gymnasium environments, vectorized environments, callbacks, and model persistence.

30.2k
repo stars
▌
trl-training · huggingface
Train and fine-tune transformer language models using TRL (Transformers Reinforcement Learning) with support for SFT, DPO, GRPO, KTO, RLOO, and reward model training via CLI commands.

10.8k
repo stars

Frequently asked questions

How do I install the miles-rl-training skill?

Run npx skillmds add orchestra-research/miles-rl-training in your terminal (requires Node.js), paste this page's agent-chat prompt into Claude, Cursor, or any MCP-connected agent, or download the SKILL.md file and copy it into your agent's skills directory.

What does the miles-rl-training skill do?

Train large-scale MoE models with FP8/INT4 low-precision RL, speculative decoding, and train-inference alignment using the miles framework. It is listed under AI & ML, Model Training & Fine-tuning on SkillMD.

Is miles-rl-training safe to use?

SkillMD's automated safety review verdict for this skill is CAUTION. Independent scanners report: SkillSpector: PASS, Skill Scanner: PASS. Capability flags: makes network calls. SkillMD never runs a skill's scripts for you; review the SKILL.md before installing.

Which AI agents work with miles-rl-training?

This skill is tagged as working with Claude Code, Claude.ai, OpenAI Codex. SKILL.md is an open format, so most agents that read a skills directory can load it too.

Is miles-rl-training free to use?

Yes. Installing skills from SkillMD is free. This skill is licensed under MIT.

Who published miles-rl-training?

Orchestra Research (@orchestra-research) published this skill. Their other Agent Skills are listed on their SkillMD profile.

miles-rl-training

miles: Enterprise-Grade RL for Large-Scale Model Training

When to Use miles

Key Features

Low-Precision Training

Performance Optimizations

Train-Inference Alignment

Installation

Quick Start

Workflow 1: Large MoE Training

Prerequisites Checklist

Step 1: Environment Setup

Step 2: Configure Training

Verification Checklist

Workflow 2: Speculative RL Training

How Speculative RL Works

Step 1: Enable Speculative Decoding

Step 2: Enable Online MTP Training (Optional)

Expected Speedup

Configuration Reference

Cluster Resources (from slime)

Megatron Parallelism (from slime)

Speculative Decoding (miles-specific)

Online MTP Training (miles-specific)

Key Features (Conceptual)

Unified FP8 Pipeline

Rollout Routing Replay (R3)

INT4 Quantization-Aware Training

Train-Inference Alignment

Sample Data Structure

Common Issues and Solutions

Issue: FP8 Training Collapse

Issue: Speculative Draft Drift

Issue: Train-Inference Mismatch

Supported Models

Resources

Files

Related

Frequently asked questions