Search results

Reduce downstream Megatron patching for RL use cases

enhancementNew feature or requestNew feature or request

module: rl

Status: Open.

#4590 In NVIDIA/Megatron-LM;

· sbhavani opened on May 2, 2026

Nemotron 3 Nano Omni Training support

enhancementNew feature or requestNew feature or request

Status: Open.

#4588 In NVIDIA/Megatron-LM;

· cuichenx opened on May 2, 2026

Megatron-Fsdp checkpoint fails on Qwen3.5 with tensor and expert paralelism

bugSomething isn't workingSomething isn't working

community-request

waiting-on-maintainersWaiting on maintainers to respondWaiting on maintainers to respond

Status: Open.

#4553 In NVIDIA/Megatron-LM;

· pavelgein opened on Apr 30, 2026

support mtp sft during rl train

community-request

enhancementNew feature or requestNew feature or request

waiting-on-maintainersWaiting on maintainers to respondWaiting on maintainers to respond

Status: Open.

#4549 In NVIDIA/Megatron-LM;

· liuzhenhai93 opened on Apr 30, 2026

Activation recompute can be skipped on the final instance of the recomputed modules to improve performance.

enhancementNew feature or requestNew feature or request

Status: Open.

#4546 In NVIDIA/Megatron-LM;

· cspades opened on Apr 30, 2026

Reduce boilerplate around MultiStorageClient feature checks

Status: Open.

#4507 In NVIDIA/Megatron-LM;

· maanug-nv opened on Apr 28, 2026

[Feature Request] PackedSeqParams: Add construction helpers for THD packed-sequence metadata

community-request

enhancementNew feature or requestNew feature or request

waiting-on-maintainersWaiting on maintainers to respondWaiting on maintainers to respond

Status: Open.

#4493 In NVIDIA/Megatron-LM;

· bbuschkaemper opened on Apr 28, 2026

there is not keyword argument 'timeout' in torch.distributed.barrier()

bugSomething isn't workingSomething isn't working

community-request

waiting-on-maintainersWaiting on maintainers to respondWaiting on maintainers to respond

Status: Open.

#4480 In NVIDIA/Megatron-LM;

· zhaoyinglia opened on Apr 27, 2026

AttributeError: 'dict' object has no attribute 'device' in forward_step_calc_loss for Multimodal (MiMo) models +EP

bugSomething isn't workingSomething isn't working

community-request

waiting-on-customerWaiting on the original author to respondWaiting on the original author to respond

Status: Open.

#4474 In NVIDIA/Megatron-LM;

· chenhongyu2048 opened on Apr 26, 2026

DeepSeek-V4 training support

enhancementNew feature or requestNew feature or request

Status: Open.

#4468 In NVIDIA/Megatron-LM;

· sbhavani opened on Apr 24, 2026

SSM sharded tensor merge can OOM during checkpoint loading

bugSomething isn't workingSomething isn't working

community-request

waiting-on-maintainersWaiting on maintainers to respondWaiting on maintainers to respond

Status: Open.

#4441 In NVIDIA/Megatron-LM;

· returnL opened on Apr 23, 2026

[Feature Request] Add initial overlap_p2p_comm support for non-interleaved pipeline parallelism

community-request

enhancementNew feature or requestNew feature or request

waiting-on-maintainersWaiting on maintainers to respondWaiting on maintainers to respond

Status: Open.

#4385 In NVIDIA/Megatron-LM;

· cky-dev opened on Apr 20, 2026