All issues

Search results

[BUG] MFSDP v2 with HybridEP: intermittent bad gradients in eager training

bugSomething isn't workingSomething isn't working

MFSDPv2

module: megatron-fsdp

Status: Open.

Bug

#7400 In NVIDIA/Megatron-LM;

· wujingyue opened on Sep 16, 2026

[Bug] Fine-grained activation offloading: the inline reload in tensor_pop races its own D2H, corrupting the warmup iteration

bugSomething isn't workingSomething isn't working

community-request

Status: Open.

#7398 In NVIDIA/Megatron-LM;

· GarlGuo opened on Sep 16, 2026

[BUG] Shared expert overlap delays independent expert wgrad until input-gradient merge

community-request

Status: Open.

#7389 In NVIDIA/Megatron-LM;

· fkuner opened on Sep 16, 2026

[ENHANCEMENT] Shared-prefix training with context parallelism for exact agent trajectories

enhancementNew feature or requestNew feature or request

module: rl

Status: Open.

#7386 In NVIDIA/Megatron-LM;

· arnavkomaragiri opened on Sep 16, 2026

[REGRESSION][dev] DeepSeek-V4 CSA repeated recompilation on forward for THD packed sequences.

community-request

Status: Open.

#7340 In NVIDIA/Megatron-LM;

· lkhphuc opened on Sep 15, 2026

[BUG] get_grad_norm_fp32 / clip_grad_by_total_norm_fp32 pass a mixed bf16/fp32 gradient list to one TE multi_tensor launch -> illegal memory access (precision-aware optimizer + natively fp32 params)

community-request

Status: Open.

#7338 In NVIDIA/Megatron-LM;

· linmuchuiyang opened on Sep 15, 2026

Make training initialization and runtime services consume configs

enhancementNew feature or requestNew feature or request

module: training

Status: Open.

Task

#7325 In NVIDIA/Megatron-LM;

· cuichenx opened on Sep 14, 2026

Separate configuration preparation from runtime initialization

enhancementNew feature or requestNew feature or request

module: training

Status: Open.

Task

#7324 In NVIDIA/Megatron-LM;

· cuichenx opened on Sep 14, 2026

[Bug][MCore dev 0e6ac576] Intermittent CUDA runtime failures in DeepSeek-V4-0731 fused CSA/DSA with packed THD + CP32 on SM90

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.

#7292 In NVIDIA/Megatron-LM;

· li126com opened on Sep 14, 2026

[ENHANCEMENT] Reduce torch_dist checkpoint preparation overhead for prepended-axis tensors

community-request

enhancementNew feature or requestNew feature or request

Status: Open.

#7275 In NVIDIA/Megatron-LM;

· suimenno3002 opened on Sep 12, 2026

Move PEFT model components into Megatron Core

enhancementNew feature or requestNew feature or request

Status: Open.

Task

#7271 In NVIDIA/Megatron-LM;

· maanug-nv opened on Sep 11, 2026

Add adapter-only checkpointing to Megatron-LM

enhancementNew feature or requestNew feature or request

Status: Open.

Task

#7270 In NVIDIA/Megatron-LM;

· maanug-nv opened on Sep 11, 2026