Qwen3.8

This page describes the initial NeMo RL support for the dense Qwen/Qwen3.8-27B model.

Support Status

Qwen/Qwen3.8-27B is Functionally Ready for text-only GRPO on the Megatron (MBridge) training backend, with vLLM inference. The shipped recipe is a short functional smoke test. It validates model loading, rollout generation, weight refit, log-probability computation, and an optimizer step; it is not a long-run convergence recipe.

What’s Supported

Model	Modality	Training backend	Parallelism	Inference
Qwen/Qwen3.8-27B	LLM (dense)	Megatron	TP + PP + CP	vLLM

The model follows the same dense hybrid-attention integration path as the supported Qwen3.5 dense models. Context parallelism requires sequence packing, which is enabled in the example recipe below.

Note

AutoModel dependencies now include the Qwen3.5-family dense adapter via NeMo RL PR #3498. Qwen3.8 AutoModel validation and an example recipe remain tracked in issue #3675.

Example Recipes

The recipe below is an example starting point. Recipe YAML files under examples/configs/recipes/ are the source of truth; check the YAML file for the authoritative settings.

Model	Modality	Algorithm	Backend	Scale	Recipe
Qwen3.8-27B	LLM	GRPO	Megatron	2n8g	grpo-qwen3.8-27b-2n8g-megatron-tp4pp2cp2.yaml

Choose a Recipe

27B GRPO (Megatron)

Use the Megatron recipe to validate the setup, launch mechanics, logging, and checkpointing.

uv run examples/run_grpo.py \
  --config examples/configs/recipes/llm/grpo-qwen3.8-27b-2n8g-megatron-tp4pp2cp2.yaml

This is a functional validation recipe rather than a long-run convergence recipe; validate longer training separately for the target workload.

100-Step Functional Validation Results

The reference curves below were produced by running the example recipe for more than 100 steps.

The recipe’s OpenMathInstruct-2 dataset is relatively easy for Qwen3.8-27B, and the maximum sequence length is limited to 4,096 tokens. Consequently, the reward curve does not show a pronounced upward trend. These curves are provided only as a functional validation reference for training beyond 100 steps, not as evidence of long-run convergence.