shawnzzzzz/Qwen3-30B-A3B-GRPO-W4A4-NoOverlong-Step800

Qwen3-30B-A3B GRPO W4A4 No-Overlong — Step 800

This is the standalone Hugging Face export of the final policy checkpoint from the following NeMo-RL training run:

grpo-qwen3-30ba3b-base-8n4g-dapo512-20k-nvfp4-pertoken-r2-20260806-0of3-tokenmean-allmlp-nooverlong-dequant-te3049-v1

Training configuration

• Base model: Qwen/Qwen3-30B-A3B-Base
• Algorithm: GRPO
• Training data: DAPO Math 17K
• Final checkpoint: step 800 of 800
• Low-precision policy training: per-token NVFP4 W4A4, all MLP layers
• Overlong policy: disabled, so this control is aligned with the no-overlong stability-loss experiments
• Stability methods: 0 of 3 enabled (Length, Segment, and Alignment are off)
• Transformer Engine: dequantization fix based on TE 3049
• Final recorded validation accuracy: 0.44140625

Export format

The NeMo-RL Megatron distributed checkpoint was converted with Megatron Bridge to a standard Transformers checkpoint. The exported weights are stored as 16 BF16 safetensors shards. Optimizer and dataloader state are intentionally not included; this repository is intended for inference, evaluation, or subsequent fine-tuning rather than exact optimizer-state training resumption.

Minimal loading example

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "shawnzzzzz/Qwen3-30B-A3B-GRPO-W4A4-NoOverlong-Step800"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)