Search results

MLA pads V on the thd path unconditionally, which costs up to 87% of a step at long sequence

community-request

Status: Open.

#6240 In NVIDIA/Megatron-LM;

· bzantium opened on Aug 4, 2026

[BUG] Loss logging double-counts the first iteration after checkpoint resume

community-request

Status: Open.

#6237 In NVIDIA/Megatron-LM;

· Burceoneo12345 opened on Aug 4, 2026

[GTP] Documentation update

nemotron

Status: Open.

Documentation

#6236 In NVIDIA/Megatron-LM;

· fanshiqing opened on Aug 4, 2026

ValueError: can't optimize a non-leaf Tensor when using --optimizer-cpu-offload (HybridDeviceOptimizer) together with mark_keep_in_fp32 parameters (e.g. DeepSeek V4 CSA ape/attn_sink, mHC alpha_*/bias)

bugSomething isn't workingSomething isn't working

community-request

Status: Open.

#6212 In NVIDIA/Megatron-LM;

· HowardZorn opened on Aug 3, 2026

[ENHANCEMENT] Docs: add parallel_state to the curated Core APIs guide (api-guide/core)

community-request

enhancementNew feature or requestNew feature or request

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

Status: Open.

#6175 In NVIDIA/Megatron-LM;

· Faisal-Alsrheed opened on Jul 31, 2026

[Prototype] MatrixFSDP-style TensorAtomic placement for MFSDP v2 Muon

Status: Open.

#6161 In NVIDIA/Megatron-LM;

· wujingyue opened on Jul 30, 2026

Support CP > 1 with inter-document masking in pretraining

Status: Open.

#6156 In NVIDIA/Megatron-LM;

· sbhavani opened on Jul 30, 2026

Support CUDA graph capture with symmetric memory in MFSDP v2

Status: Open.

Enhancement

#6154 In NVIDIA/Megatron-LM;

· wujingyue opened on Jul 30, 2026

Verify MFSDP v2 works with MCore's GraphableMegatronModule

Status: Open.

Task

#6152 In NVIDIA/Megatron-LM;

· wujingyue opened on Jul 30, 2026

Integrate NeMo Automodel Qwen3.5-35B-A3B recipe with MFSDP v2

Status: Open.

Enhancement

#6139 In NVIDIA/Megatron-LM;

· wujingyue opened on Jul 30, 2026

[GTP] Scale efficiency

Status: Open.

Task

#6136 In NVIDIA/Megatron-LM;

· fanshiqing opened on Jul 30, 2026

MoE router padding mask fails with expert bias during training

Status: Open.

#6111 In NVIDIA/Megatron-LM;

· cuichenx opened on Jul 29, 2026