slime-user

Name: slime-user
Author: yzlnew

10installs10forks

How to install slime-user?

yzlnew/infra-skills/slime-user

$npx -y skills add yzlnew/infra-skills --skill slime-user

Installs into the current project.

Summary

Guide for using SLIME (LLM post-training framework for RL Scaling). Use when working with SLIME for reinforcement learning training of language models, including setup, configuration, training execution, multi-turn interactions, custom reward models, tool calling scenarios, or tr

Files · 1

View on GitHub

SKILL.md

1	# SLIME User Guide
2
3	SLIME is an LLM post-training framework for RL Scaling developed by THUDM. It supports various RL algorithms (GRPO, GSPO, PPO, Reinforce++), multiple training backends (Megatron, FSDP), and advanced features like multi-turn interactions, tool calling, and dynamic sampling.
4
5	## Quick Start Workflow
6
7	### For First-Time Users
8
9	1. Environment Setup
10	- Use Docker: `docker pull slimerl/slime:latest`
11	- Or build from source: See `docs/en/get_started/quick_start.md`
12	- Hardware: Supports H100/H200, B200 series
13
14	2. Download Model and Data
15	```bash
16	hf download Qwen/Qwen3-4B --local-dir /root/Qwen3-4B
17	hf download --repo-type dataset zhuzilin/dapo-math-17k --local-dir /root/dapo-math-17k
18	```
19
20	3. Convert Weights (Megatron backend only)
21	```bash
22	source scripts/models/qwen3-4B.sh
23	PYTHONPATH=/root/Megatron-LM python tools/convert_hf_to_torch_dist.py \
24	${MODEL_ARGS[@]} \
25	--hf-checkpoint /root/Qwen3-4B \
26	--save /root/Qwen3-4B_torch_dist
27	```
28
29	4. Run Training
30	```bash
31	bash scripts/run-qwen3-4B.sh
32	```
33
34	### For Experienced Users
35
36	When user needs specific functionality:
37	- Multi-turn/tool calling: Read [references/examples_reference.md](references/examples_reference.md) Search-R1 section
38	- Custom reward models: See custom RM pattern in examples reference
39	- FSDP instead of Megatron: Use `--train-backend fsdp`, skip weight conversion
40	- Large-scale training: See multi-node examples (GLM-4.5, DeepSeek-R1)
41	- Source code exploration: Check [references/source_code_reference.md](references/source_code_reference.md)
42
43	## Documentation Navigation
44
45	SLIME has extensive documentation. Use this guide to find what you need quickly.
46
47	### Essential Documentation (Read These First)
48	1. Quick Start Guide: `docs/en/get_started/quick_start.md` - Setup and first training run
49	2. Usage Guide: `docs/en/get_started/usage.md` - Comprehensive parameter reference
50	3. Example Docs: `docs/en/examples/qwen3-4B.md` or `docs/en/examples/glm4-9B.md`
51
52	For detailed navigation of all documentation, see [references/doc_navigation.md](references/doc_navigation.md).
53
54	### Common Tasks → Documentation Mapping
55
56	\| Task \| Documentation \|
57	\|------\|---------------\|
58	\| First-time setup \| `docs/en/get_started/quick_start.md` \|
59	\| Understanding parameters \| `docs/en/get_started/usage.md` \|
60	\| Basic training (8 GPUs) \| `docs/en/examples/qwen3-4B.md` \|
61	\| Multi-turn tool use \| `examples/search-r1/` \|
62	\| Custom generation logic \| `docs/en/get_started/customization.md` \|
63	\| Multi-node training \| `docs/en/examples/glm4.5-355B-A32B.md` \|
64	\| FSDP backend \| `docs/en/get_started/usage.md` (FSDP section) \|
65	\| VLM training \| `examples/geo3k_vlm/` \|
66	\| Troubleshooting \| `docs/en/get_started/qa.md` \|
67
68	## Core Concepts
69
70	### Training Loop
71	SLIME uses a "Rollout → Train" loop:
72	1. Rollout: Generate responses using SGLang inference
73	2. Reward: Compute rewards using reward model
74	3. Train: Update model weights using Megatron/FSDP
75	4. Repeat for `--num-rollout` iterations
76
77	### Key Constraint
78	```
79	rollout-batch-size × n-samples-per-prompt = global-batch-size × num-steps-per-rollout
80	```
81
82	### Resource Allocation Modes
83
84	Colocated (training and inference share GPUs):
85	```bash
86	--actor-num-nodes 1 \
87	--actor-num-gpus-per-node 8 \
88	--colocate \
89	--sglang-mem-fraction-static 0.7
90	```
91
92	Disaggregated (separate GPUs for training/inference):
93	```bash
94	--actor-num-nodes 1 \
95	--actor-num-gpus-per-node 4 \
96	--rollout-num-gpus 4
97	```
98
99	## Parameter Quick Reference
100
101	### Essential Parameters
102
103	Model Loading:
104	- `--hf-checkpoint`: HuggingFace model path (for SGLang and FSDP)
105	- `--ref-load`: Megatron reference model checkpoint
106	- `--load`: Megatron actor checkpoint (resume training)
107	- `--save`: Save path for checkpoints
108
109	Data:
110	- `--prompt-data`: JSONL dataset path
111	- `--input-key`: Field name for prompts (default: "prompt")
112	- `--label-key`: Field name for labels (default: "label")
113	- `--metadata-key`: Field name for metadata (default: "metadata")
114	- `--apply-chat-template`: Apply tokenizer chat template
115
116	Rollout:
117	- `--rollout-batch-size`: Prompts per rollout
118	- `--n-samples-per-prompt`: Responses per prompt
119	- `--rollout-max-response-len`: Max response length
120	- `--rollout-temperature`: Sampling temperature
121
122	Training:
123	- `--num-rollout`: Total training iterations
124	- `--num-steps-per-rollout`: Optimizer steps per rollout (default: 1)
125	- `--global-batch-size`: Samples per optimizer step
126	- `--advantage-e

People who liked this also viewed

6 picks

Type

slime-user

10installs10forks

View on GitHub

How to install slime-user?

yzlnew/infra-skills/slime-user

$npx -y skills add yzlnew/infra-skills --skill slime-user

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# SLIME User Guide
2
3	SLIME is an LLM post-training framework for RL Scaling developed by THUDM. It supports various RL algorithms (GRPO, GSPO, PPO, Reinforce++), multiple training backends (Megatron, FSDP), and advanced features like multi-turn interactions, tool calling, and dynamic sampling.
4
5	## Quick Start Workflow
6
7	### For First-Time Users
8
9	1. Environment Setup
10	- Use Docker: `docker pull slimerl/slime:latest`
11	- Or build from source: See `docs/en/get_started/quick_start.md`
12	- Hardware: Supports H100/H200, B200 series
13
14	2. Download Model and Data
15	```bash
16	hf download Qwen/Qwen3-4B --local-dir /root/Qwen3-4B
17	hf download --repo-type dataset zhuzilin/dapo-math-17k --local-dir /root/dapo-math-17k
18	```
19
20	3. Convert Weights (Megatron backend only)
21	```bash
22	source scripts/models/qwen3-4B.sh
23	PYTHONPATH=/root/Megatron-LM python tools/convert_hf_to_torch_dist.py \
24	${MODEL_ARGS[@]} \
25	--hf-checkpoint /root/Qwen3-4B \
26	--save /root/Qwen3-4B_torch_dist
27	```
28
29	4. Run Training
30	```bash
31	bash scripts/run-qwen3-4B.sh
32	```
33
34	### For Experienced Users
35
36	When user needs specific functionality:
37	- Multi-turn/tool calling: Read [references/examples_reference.md](references/examples_reference.md) Search-R1 section
38	- Custom reward models: See custom RM pattern in examples reference
39	- FSDP instead of Megatron: Use `--train-backend fsdp`, skip weight conversion
40	- Large-scale training: See multi-node examples (GLM-4.5, DeepSeek-R1)
41	- Source code exploration: Check [references/source_code_reference.md](references/source_code_reference.md)
42
43	## Documentation Navigation
44
45	SLIME has extensive documentation. Use this guide to find what you need quickly.
46
47	### Essential Documentation (Read These First)
48	1. Quick Start Guide: `docs/en/get_started/quick_start.md` - Setup and first training run
49	2. Usage Guide: `docs/en/get_started/usage.md` - Comprehensive parameter reference
50	3. Example Docs: `docs/en/examples/qwen3-4B.md` or `docs/en/examples/glm4-9B.md`
51
52	For detailed navigation of all documentation, see [references/doc_navigation.md](references/doc_navigation.md).
53
54	### Common Tasks → Documentation Mapping
55
56	\| Task \| Documentation \|
57	\|------\|---------------\|
58	\| First-time setup \| `docs/en/get_started/quick_start.md` \|
59	\| Understanding parameters \| `docs/en/get_started/usage.md` \|
60	\| Basic training (8 GPUs) \| `docs/en/examples/qwen3-4B.md` \|
61	\| Multi-turn tool use \| `examples/search-r1/` \|
62	\| Custom generation logic \| `docs/en/get_started/customization.md` \|
63	\| Multi-node training \| `docs/en/examples/glm4.5-355B-A32B.md` \|
64	\| FSDP backend \| `docs/en/get_started/usage.md` (FSDP section) \|
65	\| VLM training \| `examples/geo3k_vlm/` \|
66	\| Troubleshooting \| `docs/en/get_started/qa.md` \|
67
68	## Core Concepts
69
70	### Training Loop
71	SLIME uses a "Rollout → Train" loop:
72	1. Rollout: Generate responses using SGLang inference
73	2. Reward: Compute rewards using reward model
74	3. Train: Update model weights using Megatron/FSDP
75	4. Repeat for `--num-rollout` iterations
76
77	### Key Constraint
78	```
79	rollout-batch-size × n-samples-per-prompt = global-batch-size × num-steps-per-rollout
80	```
81
82	### Resource Allocation Modes
83
84	Colocated (training and inference share GPUs):
85	```bash
86	--actor-num-nodes 1 \
87	--actor-num-gpus-per-node 8 \
88	--colocate \
89	--sglang-mem-fraction-static 0.7
90	```
91
92	Disaggregated (separate GPUs for training/inference):
93	```bash
94	--actor-num-nodes 1 \
95	--actor-num-gpus-per-node 4 \
96	--rollout-num-gpus 4
97	```
98
99	## Parameter Quick Reference
100
101	### Essential Parameters
102
103	Model Loading:
104	- `--hf-checkpoint`: HuggingFace model path (for SGLang and FSDP)
105	- `--ref-load`: Megatron reference model checkpoint
106	- `--load`: Megatron actor checkpoint (resume training)
107	- `--save`: Save path for checkpoints
108
109	Data:
110	- `--prompt-data`: JSONL dataset path
111	- `--input-key`: Field name for prompts (default: "prompt")
112	- `--label-key`: Field name for labels (default: "label")
113	- `--metadata-key`: Field name for metadata (default: "metadata")
114	- `--apply-chat-template`: Apply tokenizer chat template
115
116	Rollout:
117	- `--rollout-batch-size`: Prompts per rollout
118	- `--n-samples-per-prompt`: Responses per prompt
119	- `--rollout-max-response-len`: Max response length
120	- `--rollout-temperature`: Sampling temperature
121
122	Training:
123	- `--num-rollout`: Total training iterations
124	- `--num-steps-per-rollout`: Optimizer steps per rollout (default: 1)
125	- `--global-batch-size`: Samples per optimizer step
126	- `--advantage-e

People who liked this also viewed

6 picks

Type