Supported Models
SkyRL provides support for end-to-end training with the FSDP and Megatron training backends (both using vLLM for inference), as well as support for a JAX backend (for both training and inference) via the Tinker API server.
The table below lists the model families we have tested on SkyRL. Each family has its own page with example scripts, hardware requirements and known issues.
| Model | Sizes | Full FT | LoRA | Low precision | Min. hardware | Status |
|---|---|---|---|---|---|---|
| Beam | 501B-A23B | ✅ | ✅ | MXFP8 | 2x8 B200 | ✅ |
| GLM-5.3 | 753B | ✅ | ✅ | MXFP8, FP8 rollouts | 4x8 B300 | ✅ |
| GLM-5.3-Flash | 313B-A17B | ✅ | ✅ | 1x8 B300 | ✅ | |
| Kimi K2.7 | 1T | ❌ | ✅ | INT4 QAT only | 2x8 B300 | ✅ |
| Kimi K2.5/K2.6 | 1T | ❌ | ✅ | INT4 QAT only | 2x8 B300 | Same arch as K2.7, should work |
| Nemotron-3.5 | Lightning-30B-A3B | ✅ | 1x4 H100 | ✅ | ||
| Nemotron-3 | Nano-4B, Nano-30B-A3B | ✅ | 1x8 H100 | ✅ | ||
| Qwen3.8-Flash-Next | ~180B MoE | ✅ | ✅ | 4x8 B200 | ✅ | |
| Qwen3.6 | 27B, 35B-A3B | ✅ | ✅ | INT4 QAT | 1x8 H200 | ✅ |
| Qwen-3.5 | 0.8B-27B, 35B-A3B/122B-A10B/397B-A17B | ✅ | ✅ | FP8, MXFP8 | 1x4 GPUs | ✅ |
| Qwen3 | 0.6B-32B, 30B-A3B/235B-A22B | ✅ | ✅ | 1x4 GPUs | ✅ |
Blank means we haven't tested it yet. ❌ means it is known not to work.
Other models with example scripts
| Model | Backend | Full FT | LoRA | Min. hardware | Example scripts |
|---|---|---|---|---|---|
| GLM-4.7-Flash | Megatron | ✅ | 1x8 GPUs | run_megatron_grpo_glm4_7_30b.sh | |
| GLM-4.5-Air | Megatron | ✅ | 2x8 H100 | run_megatron_rdt_dpep_glm45_air_lora.sh | |
| Moonlight-16B-A3B | Megatron | ✅ | 1x8 H100 | run_megatron_moonlight.sh, router_replay/ | |
| MiMo-7B-RL | Megatron | ✅ | 1x8 H100 | run_megatron_dapo_mimo_7b_rl.sh, MTP spec decoding in spec_decode/ | |
| Nemotron-Mini-4B-Instruct | Megatron | ✅ | 1x8 GPUs | run_megatron_nemotron_mini_4b.sh | |
| gpt-oss-20b | FSDP | ✅ | 1x4 GPUs | run_gsm8k_gptoss.sh | |
| Qwen3-VL-2B/8B | FSDP, Megatron (SFT) | ✅ | ✅ | 1x8 GPUs | geometry3k/, visgym/, run_sft_megatron_vlm.sh |
| Qwen3-Coder-30B-A3B | FSDP | ✅ | 2x8 GPUs | run_mini_swe_30B.sh | |
| Qwen2.5 (0.5B-32B, Coder, Math) | FSDP | ✅ | ✅ | 1x4 GPUs | Most of the examples/ (GSM8K, DAPO, search, text-to-SQL, ...) |
| Qwen1.5-MoE-A2.7B | FSDP | ✅ | 1x8 GPUs | run_qwen1_5_MoE_A2_7B.sh |
FSDP
Any model that is supported via HuggingFace transformers + vLLM is supported via the FSDP backend. The FSDP backend also comes with support for Ulysses-style sequence parallelism, making it a viable option for training on dense models with up to 32B parameters and 32K/64K context lengths.
Megatron
The Megatron backend comes with support for 5D parallelism (DP/TP/CP/PP/EP) and allows for efficient scaling of large MoE models (100B+). Support for mapping from HuggingFace model definitions to Megatron GPTModels is provided by NVIDIA's Megatron-Bridge library. All of the models in the table above are supported via Megatron-Bridge + vLLM.
Other models supported via Megatron-Bridge and vLLM should run on SkyRL as well. Contributions and bug fixes for additional model families are welcome!
JAX
The JAX backend supports the following models:
- Qwen-3.5 Dense
- Deepseek-V3
- Llama-3
- Qwen3 Dense/MoE
Known limitations
- Vision-language checkpoints (Kimi K2.x, GLM-5.3) are trained text-only (
language_model_only=true), the vision tower stays frozen in vLLM. Qwen-3.5 VL is the exception and can train the vision tower. - MTP heads are dropped by default. Supported models can train them with
trainer.mtp.enabled=true, see MTP training and speculative decoding. - For FP8/INT4 release checkpoints the Megatron trainer needs a BF16 copy. See the model page for which checkpoint to use.