8 Pull requests merged by 7 people

• Refactor Tokenizer -> BaseTokenizer

  #1333 merged Jul 10, 2025

• Migrating flux checkpointing to hf api

  #1377 merged Jul 10, 2025

• [SimpleFSDP] Add support for hsdp+tp

  #1343 merged Jul 10, 2025

• basic validator implementation

  #1362 merged Jul 10, 2025

• [float8 moe] Support TP

  #1375 merged Jul 9, 2025

• call init_weights before generation

  #1371 merged Jul 8, 2025

• fix: correct CONFIG_FILE path in multinode_trainer script

  #1374 merged Jul 8, 2025

• dp2ep Expert Parallel

  #1324 merged Jul 8, 2025

4 Pull requests opened by 4 people

• Add option to exclude low flop mms from every-other-mm sac policy

  #1372 opened Jul 8, 2025

• [DSV3] Adding deepseek-v3 model into torchtitan

  #1373 opened Jul 8, 2025

• [llama3] add configurations for Llama 3 1B and 3B models

  #1376 opened Jul 9, 2025

• add float8 support

  #1378 opened Jul 10, 2025

2 Issues closed by 1 person

• [Evaluation] Minimal support for downstream tasks

  #883 closed Jul 10, 2025

• [Bug] Potential bugs in "_grouped_mm" in Llama4 MoE codes

  #1237 closed Jul 8, 2025

3 Issues opened by 3 people

• Llama4 conversion scripts error on aux-loss-free params

  #1379 opened Jul 10, 2025

• Puzzling collectives in TP ( SP to be exact)

  #1369 opened Jul 7, 2025

• Support more models such as Qwen

  #1368 opened Jul 4, 2025

14 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• [DSV3] Add PP support for DSV3

  #1345 commented on Jul 10, 2025 • 8 new comments

• Add AMD GPU node for integration test

  #1241 commented on Jul 10, 2025 • 7 new comments

• OOM recovery under multi-node FSDP/HSDP

  #1329 commented on Jul 7, 2025 • 0 new comments

• DeepSeek V3 Support

  #760 commented on Jul 8, 2025 • 0 new comments

• Llama 4 issue tracking

  #1118 commented on Jul 8, 2025 • 0 new comments

• Data loader's state_dict is being lost between being saved and loaded when the dataset loops

  #1357 commented on Jul 9, 2025 • 0 new comments

• How to adapt HuggingFace or other models for TorchTitan

  #1322 commented on Jul 10, 2025 • 0 new comments

• Issue reproducing Float8 performance benchmark

  #1344 commented on Jul 10, 2025 • 0 new comments

• [RFC] validation and evaluation in torchtitan

  #1210 commented on Jul 10, 2025 • 0 new comments

• compile: turn off fullgraph=True to support llama4

  #1182 commented on Jul 8, 2025 • 0 new comments

• Enable ROCm CI support.

  #1260 commented on Jul 9, 2025 • 0 new comments

• Add support for saving HF format tensors with DCP

  #1351 commented on Jul 11, 2025 • 0 new comments

• [benchmark] add h200 bench

  #1361 commented on Jul 7, 2025 • 0 new comments

• [WIP] Compile for dp2ep

  #1365 commented on Jul 10, 2025 • 0 new comments