Search results

[BUG] ChainedOptimizer applies global grad-norm clipping to Muon (orthogonalizing) param groups, silently stalling training

community-request

Status: Open.

#5394 In NVIDIA/Megatron-LM;

· yuchenwang3 opened on Jun 17, 2026

Feature Request: DSA/DSv4 Indexer Replay for RL

enhancementNew feature or requestNew feature or request

module: rl

Status: Open.

#5384 In NVIDIA/Megatron-LM;

· sbhavani opened on Jun 16, 2026

Hybrid CP local_cp_size=1 can shard RoPE with the static CP group

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.

#5352 In NVIDIA/Megatron-LM;

· dbfancier opened on Jun 16, 2026

Feature Request: Preconditioned Matrix Optimizers

community-request

enhancementNew feature or requestNew feature or request

Status: Open.

#5345 In NVIDIA/Megatron-LM;

· plugyawn opened on Jun 15, 2026

_reduce silently returns unreduced tensor when input is non-contiguous

bugSomething isn't workingSomething isn't working

community-request

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

Status: Open.

#5337 In NVIDIA/Megatron-LM;

· Pearblossom-M opened on Jun 14, 2026

Feature Request: Configurable Communication Scheduling Policies for MoE Training

community-request

enhancementNew feature or requestNew feature or request

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

Status: Open.

#5330 In NVIDIA/Megatron-LM;

· nathon-lee opened on Jun 13, 2026

[bug] Gemma4 with cfg.model.recompute_granularity = "full" is not working

bugSomething isn't workingSomething isn't working

community-request

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

Status: Open.

#5323 In NVIDIA/Megatron-LM;

· warpuv opened on Jun 12, 2026

Feature Request: MiniMax Sparse Attention (MSA)

Status: Open.

#5322 In NVIDIA/Megatron-LM;

· sbhavani opened on Jun 12, 2026

[Bug][DSv4-Hybrid] apply_rope_fusion=True causes NaN at iter 2 in BF16 mock pretrain (DSv4HybridSelfAttention)

community-request

Status: Open.

#5317 In NVIDIA/Megatron-LM;

· Micuks opened on Jun 12, 2026

FT local checkpointing: corrupt replica-retrieved local ckpt (deterministic NaN at load+3 iters) has no integrity validation / global-ckpt fallback; rerun_state_machine loops forever under ft_launcher in-job restart

community-request

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

Status: Open.

#5281 In NVIDIA/Megatron-LM;

· cxiong-ship-it opened on Jun 11, 2026

DeepSeek V4 FP8 training some parameters need keep bfloat16

bugSomething isn't workingSomething isn't working

community-request

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

Status: Open.

#5267 In NVIDIA/Megatron-LM;

· ZHUI opened on Jun 10, 2026

Cannot save optimizer when swiglu and FSDP

bugSomething isn't workingSomething isn't working

community-request

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

Status: Open.

#5265 In NVIDIA/Megatron-LM;

· pavelgein opened on Jun 10, 2026