Search results

Torchtitan CI gap Between ROCm & CUDA

module: rocm

Status: Open.

Task

#2098 In pytorch/torchtitan;

· akashveramd opened on Dec 2, 2025

[Possible PR discuss] Automatic Pipeline Parallel Partitioning Considering Embedding Layers

Status: Open.

Feature

#2093 In pytorch/torchtitan;

· McmillanTAC opened on Dec 1, 2025

question of <code>_op_sac_save_list</code> for op-sac

Status: Open.

Task

#2091 In pytorch/torchtitan;

· rakkit opened on Nov 28, 2025

save optimizer state with HF

Status: Open.

#2090 In pytorch/torchtitan;

· silingtong123 opened on Nov 28, 2025

SimpleFSDP + DSV3 + FlexAttention issue

Status: Open.

Bug

#2089 In pytorch/torchtitan;

· ruisizhang123 opened on Nov 26, 2025

Add PyTorch version compatibility documentation and validation

Status: Open.

#2088 In pytorch/torchtitan;

· Changomango0903 opened on Nov 26, 2025

Relationship between DS, TP,CP and EP

Status: Open.

#2087 In pytorch/torchtitan;

· githubsgi opened on Nov 25, 2025

mxfp8 MoE train is slower for DeepSeekV3 16b and Qwen models

Status: Open.

#2086 In pytorch/torchtitan;

· Yerniyaz opened on Nov 25, 2025

Context Parallel for Qwen3

Status: Open.

Feature

#2077 In pytorch/torchtitan;

· unavailableun opened on Nov 24, 2025

Slow Dataloader should use num_worker &gt; 1

Status: Open.

Feature

#2073 In pytorch/torchtitan;

· hypnopump opened on Nov 21, 2025

NCCL kernels take longer when composing CUDAGraph with SimpleFSDP

Status: Open.

#2071 In pytorch/torchtitan;

· syed-ahmed opened on Nov 20, 2025

Llama4 Debug model - RuntimeError: No available kernel. Aborting execution.

Status: Open.

#2065 In pytorch/torchtitan;

· githubsgi opened on Nov 20, 2025