Ahmad Kiswani

Calling GridEncoder with empty positions `positions.shape=[0, 3]` causes CUDA to enter an invalid state silently. Subsequent calls to CUDA kernels would fail with the following error: ``` RuntimeError: CUDA error:...

feat: Support qwen3-next, mcore path

7

# What does this PR do ? Add qwen3-next support, recipes and tests to mcore path. # Issues https://github.com/NVIDIA-NeMo/RL/issues/1431 # Usage SFT: ```bash uv run --extra mcore --extra vllm --extra...

fix: remove sft-qwen2.5-fsdp2tp8sp from nighlies

1

# What does this PR do ? sequence parallel + tp_size >1 is currently broken in torch==2.8.0, no point in running the test as it will raise an exception ##...

feat: Enable simulated user for multi-turn GRPO

1

Replacing PR #732 Waiting for runs to confirm the convergence graph attached in the original PR # What does this PR do ? Add an simple example on multi-turn GRPO...

Support Qwen-3 next in MCore path

2

An issue to track supporting Qwen-3 next in MCore path

CLI Architecture Improvement Proposal

1

# CLI Architecture Redesign **Use cases, pain points, and background** - The current CLI has 17+ separate entry points (`ng_run`, `ng_test`, `ng_collect_rollouts`, etc.) - The commands don't have a consistent...