NanoCode012

feat: add chat_template kwargs

2

# Description Closes https://github.com/axolotl-ai-cloud/axolotl/issues/2658 ## Motivation and Context ## How has this been tested? Untested! ## Screenshots (if appropriate) ## Types of changes ## Social Handles (Optional)

feat(doc): add info on how to use dapo / dr grpo and misc doc fixes

5

# Description - Updated grpo trl docs in config qmd ~- Added `log_sweep` as possible scheduler in schema (it was supported long ago, but not allowed in schema)~ - Clarified...

Fix: RL base feature parity

9

# Description RL trainer was not loading some basic configs like `logging_steps` etc. This PR consolidates the setting of these params and cleans them up. This PR also fixes a...

(WIP) Feat: Add wizard CLI to create yaml config

1

# Description Still a WIP, but pushed to share the user journey and take feedbacks. This requires Attention Enum PR to be merged first. ## Motivation and Context Assist users...

FutureWarning: FSDP.state_dict_type() and FSDP.set_state_dict_type() are being deprecated. Please use APIs, get_state_dict() and set_state_dict(), which can support different parallelisms, FSDP1, FSDP2, DDP.

1

### ⚠️ Please check that this feature request hasn't been suggested before. - [x] I searched previous [Ideas in Discussions](https://github.com/axolotl-ai-cloud/axolotl/discussions/categories/ideas) didn't find any similar feature requests. - [x] I searched...

[Feat] Log Flops to wandb using callback

3

### ⚠️ Please check that this feature request hasn't been suggested before. - [X] I searched previous [Ideas in Discussions](https://github.com/OpenAccess-AI-Collective/axolotl/discussions/categories/ideas) didn't find any similar feature requests. - [X] I searched...

Fix: pin same vllm version for all torch

4

# Description VLLM should support torch2.5-2.7 equally. Prev, torch2.5 would install 0.7.4 which would break GRPO as seen in #2601 ## Motivation and Context ## How has this been tested?...

fix: lm_head is a view or related view modified for CCE

1

# Description Found out from a lot of tests that `lm_head` has some issues with linearized model. Can't even `.clone()` it. Chose to detach then reattach the gradient. The conversion...

# Description Separated from #2504 Deepseekv3 not working, but added to be used as future ref code for any potential contributors ## Motivation and Context ## How has this been...

feat: update xformers for new attention interface

2

# Description Updated xformers llama attention patch for transformers v4.50.0 ## Motivation and Context ## How has this been tested? ## Screenshots (if appropriate) ## Types of changes ## Social...