Search results

SFT Packing: Missing cu_seqlens/max_seqlen broadcast in intermediate pipeline stages causes shape mismatch

bugSomething isn't workingSomething isn't working

community-request

Status: Open.

#4092 In NVIDIA/Megatron-LM;

· gaojingwei opened on Apr 1, 2026

Add MLA support for Muon optimizer

community-request

enhancementNew feature or requestNew feature or request

Status: Open.

#4091 In NVIDIA/Megatron-LM;

· oblivioncat opened on Apr 1, 2026

Hyperparameter Transfer beyond MuP

community-request

enhancementNew feature or requestNew feature or request

Status: Open.

#4088 In NVIDIA/Megatron-LM;

· plugyawn opened on Apr 1, 2026

[QUESTION]Memory increase issue introduced by Megatron-FSDP during training.

community-request

module: megatron-fsdp

questionFurther information is requestedFurther information is requested

Status: Open.

#4069 In NVIDIA/Megatron-LM;

· chinapanda opened on Mar 31, 2026

[QUESTION] KL Divergence Computation in DSA Uses a Mathematically Non-Equivalent Approximation

community-request

questionFurther information is requestedFurther information is requested

Status: Open.

#4055 In NVIDIA/Megatron-LM;

· umiswing opened on Mar 30, 2026

[QUESTION] MoE layer theoretical memory calculation needs to account for ETP/EDP.

community-request

module: moe

needs-follow-upIssue needs follow-upIssue needs follow-up

questionFurther information is requestedFurther information is requested

Status: Open.

#4050 In NVIDIA/Megatron-LM;

· lichenlu opened on Mar 30, 2026

Update megatron-core[dev] to support megatron-energon 7.x

bugSomething isn't workingSomething isn't working

community-request

Status: Open.

#4045 In NVIDIA/Megatron-LM;

· psycherun-kaiko opened on Mar 27, 2026

[Question] Enable GDN Packed Sequence Support for Context Parallelism in Qwen 3.5 series

community-request

questionFurther information is requestedFurther information is requested

Status: Open.

#4043 In NVIDIA/Megatron-LM;

· plmsmile opened on Mar 27, 2026

[Bug] HybridDeviceOptimizer KeyError with MoE models when using CPU offload

bugSomething isn't workingSomething isn't working

community-request

module: moe

Status: Open.

#4042 In NVIDIA/Megatron-LM;

· khazic opened on Mar 27, 2026

Improve docstrings and inline comments in example scripts

community-request

enhancementNew feature or requestNew feature or request

needs-follow-upIssue needs follow-upIssue needs follow-up

Status: Open.

#4040 In NVIDIA/Megatron-LM;

· DhineshPonnarasan opened on Mar 27, 2026

gpt3_mcore_te_tp2_zp_z3_resume_torch_dist_te_8experts2parallel_top2router broken after TE2.14

bugSomething isn't workingSomething isn't working

Status: Open.

#4033 In NVIDIA/Megatron-LM;

· ko3n1g opened on Mar 26, 2026

Feature Request: Attention Residuals

enhancementNew feature or requestNew feature or request

Status: Open.

#4016 In NVIDIA/Megatron-LM;

· sbhavani opened on Mar 24, 2026