tocmo0nlord/axolotl
1
0
Fork 0
Code Issues Pull Requests Actions 3 Packages Projects Releases Wiki Activity
2,446 Commits 319 Branches 32 Tags
383f220cfd658804f4c508a0686c988861ecffbe
Commit Graph

2446 Commits

This Branch
This Branch
All Branches
Author SHA1 Message Date
Wing Lian
383f220cfd build torch 2.9.0 base images (#3221) 2025-10-20 08:53:49 -04:00
NanoCode012
8bb871b5cf fix: deepspeed with context parallel (#3220) 2025-10-20 14:06:58 +07:00
Leonard
87565ecc05 Add chat_template.argilla_chat support for DPO datasets (#3202) 
* Add chat_template.argilla_chat support for DPO datasets

  Creates a new chat_template.argilla_chat prompt strategy for handling
  DPO datasets where chosen/rejected fields contain full conversations
  (messages + final response), following the pattern of chatml.argilla_chat
  and llama3.argilla_chat.

  - Add argilla_chat() function to chat_template.py
  - Add chat_template.argilla_chat to RLHF documentation
  - Add test coverage for argilla_chat with multiple tokenizers

  Dataset format:
  {
    "chosen": [
      {"role": "user", "content": "..."},
      {"role": "assistant", "content": "..."}
    ],
    "rejected": [
      {"role": "user", "content": "..."},
      {"role": "assistant", "content": "..."}
    ]
  }

* Fix chat_template.argilla_chat return value contract and add docstring

- Return (transform_fn, dataset_kwargs) tuple instead of bare transform_fn
- Add remove_columns specification for field_chosen and field_rejected
- Add comprehensive docstring with Args/Returns sections
- Update tests to unpack tuple return value

Addresses PR feedback to maintain consistency with chat_template.default()
and properly specify columns to remove after dataset transformation.

* Update tests/prompt_strategies/test_dpo_chat_templates.py

Co-authored-by: Wing Lian <wing.lian@gmail.com>

---------

Co-authored-by: Wing Lian <wing.lian@gmail.com>
 2025-10-17 17:00:26 +07:00
NanoCode012
93ba57396f fix: qwen3_vl attention config (#3216) 2025-10-17 10:35:03 +07:00
NanoCode012
aa1240acd8 fix: transformers deprecate load_in_Xbit in model_kwargs (#3205) 
* fix: transformers deprecate load_in_Xbit in model_kwargs

* fix: test to read from quantization_config kwarg

* fix: test

* fix: access

* fix: test weirdly entering incorrect config
 2025-10-16 16:07:27 +07:00
Wing Lian
4cdfdfebb5 upgrade transformers==4.57.1 and peft==0.23.1 (#3214) 2025-10-14 15:54:05 -04:00
github-actions[bot]
6e2f5ccf9f chore: update pre-commit hooks (#3211) [skip ci] 
Co-authored-by: djsaunde <1245942+djsaunde@users.noreply.github.com>
 2025-10-14 10:21:49 -04:00
NanoCode012
8c7f63cf97 fix: unpack cce imported incorrectly (#3212) [skip ci] 2025-10-13 17:19:15 +07:00
VED
cd856b45b1 feat:add support dataset_num_processes (#3129) [skip ci] 
* feat:add support dataset_num_processes

* chore

* required changes

* requested chnages

* required chnages

* required changes

* required changes

* elif get_default_process_count()

* add:del data

* Update cicd/Dockerfile.jinja

Co-authored-by: NanoCode012 <kevinvong@rocketmail.com>

* Update cicd/single_gpu.py

Co-authored-by: NanoCode012 <kevinvong@rocketmail.com>

---------

Co-authored-by: salman <salman.mohammadi@outlook.com>
Co-authored-by: NanoCode012 <kevinvong@rocketmail.com>
 2025-10-13 17:18:12 +07:00
salman
143dea4753 FSDPConfig (#3170) 2025-10-10 14:44:25 +01:00
Hitesh Sagtani
bc2ffb8204 fix: Enable KD plugin support for PEFT/LoRA adapters (#3207) 
- Fix _loss_function attribute not found on base model with PEFT
- Fix mismatched attribute name (loss_function vs _loss_function)
- Set _loss_function on unwrapped base model for PEFT
- Enable previously skipped test_llama_lora_kd test
- Add test config fixes for LoRA kernel compatibility

Fixes https://github.com/axolotl-ai-cloud/axolotl/issues/3206
 2025-10-10 08:57:00 -04:00
NanoCode012
153edcfe79 fix(doc): add act checkpointing migration to fsdp2 docs (#3193) [skip ci] 2025-10-10 10:57:50 +07:00
Wing Lian
08b8fa62cc only calculate packed ds length once if using a large world size (#3210) 2025-10-09 14:18:46 -04:00
Wing Lian
3a5c97e6e5 use can_device_access_peer for P2P checks (#3209) [skip ci] 
* use can_device_access_peer for P2P checks

* also log warn when automatically setting NCCL_P2P_DISABLE=1
 2025-10-09 14:17:31 -04:00
VED
37f78c8592 add chat_template_jinja to wandb (#3192) [skip ci] 
* add chat_template_jinja to wandb

* temp_ct_file.flush()

* Update src/axolotl/utils/callbacks/__init__.py

Co-authored-by: Wing Lian <wing.lian@gmail.com>

* Update src/axolotl/utils/callbacks/__init__.py

Co-authored-by: Wing Lian <wing.lian@gmail.com>

* Apply suggestion from @winglian

---------

Co-authored-by: Wing Lian <wing.lian@gmail.com>
 2025-10-09 12:05:54 -04:00
NanoCode012
ab63b92c38 feat: add lfm2 family and latest moe model (#3208) 
* feat: add lfm2 family and latest moe model

* fix: use ml-cross-entropy for lfm2 examples
 2025-10-09 10:47:41 -04:00
Manh Nguyen
6f8ce024d1 Remove check_torch_compile_deepspeed (#3195) [skip ci] 
Signed-off-by: nguyen599 <pnvmanh2123@gmail.com>
 2025-10-08 11:27:01 -04:00
Wing Lian
d0e9c3c1c5 When using Ray use prepare for dataloader fixes (#3198) 
* make sure to use ray prepare for dataloader fixes

* ray tests use 2.7.0+

* don't call init_distributed w ray and deepspeed

* handle dict deepspeed config

* better handling of dict deepspeed config

* use json.dumps

* guard to_dict

* wrap import for optional ray
 2025-10-08 10:43:41 -04:00
github-actions[bot]
4c3488cc9f chore: update pre-commit hooks (#3160) [skip ci] 
Co-authored-by: djsaunde <1245942+djsaunde@users.noreply.github.com>
 2025-10-08 08:58:02 -04:00
Wing Lian
130637a3fa upgrade transformers to 4.57.0 (#3201) 
* upgrade transformers to 4.57.0

* remove deprecated autoawq and use latest peft

* remove autoawq from setuptools script

* fix imports

* make sure torchvision is installed

* remove support for BetterTransformer

* skip fsdp_qlora_prequant test

* more robust error reporting
 2025-10-08 08:43:46 -04:00
VED
377c510e95 sleep model support (#3135) 
Co-authored-by: salman <salman.mohammadi@outlook.com>
 2025-10-08 12:39:21 +01:00
Wing Lian
409cfb8a87 deprecate torch 2.6.0 support (#3197) [skip ci] 2025-10-07 11:23:41 -04:00
Wing Lian
ce74c20109 don't cache pip install (#3194) 
* don't cache pip install

* no cache dir for disk space for sdist too
 2025-10-01 11:11:39 -04:00
VED
a6bfbe3400 torch_dtype -> dtype (#3177) 
* torch_dtype -> dtype

* torch_dtype -> dtype
 2025-10-01 15:02:51 +07:00
Dan Saunders
f4376748f3 debug log: multiprocess race condition fix (#3188) 2025-09-26 15:07:39 -04:00
Dan Saunders
740d5a1d31 doc fix (#3187) 2025-09-26 09:55:15 -04:00
Grant Holmes (Ren)
850c1a5f8d Add FSDP v2 swap memory support + QLoRA compatibility fixes (#3167) 
Co-authored-by: salman <salman.mohammadi@outlook.com>
 2025-09-26 10:23:59 +01:00
NanoCode012
7fa8ac40cd Feat(cce): add qwen3_vl, qwen3_vl_moe, granitemoeshared, granitemoehybrid, and upgraded all cce patches (#3178) 
* feat: upgrade cce with patches for transformers 4.56

* feat: add missing models to cce readme
 2025-09-26 12:11:29 +07:00
Dan Saunders
f9748c4dc5 Cp fix (#3182) 
* patch transformers to allow CP + FA2

* nits

* only patch in CP > 1 case
 2025-09-25 12:03:50 -04:00
miketung
33975ce4bc feat(qwen3-next): Adds targeting of shared expert and attention modules (#3183) 
* Adds targetting of shared expert and attention modules in each layer

* Update VRAM usage

---------

Co-authored-by: Mike Tung <mike@diffbot.com>
 2025-09-25 17:06:16 +07:00
陈华杰
e8b962d47f feat: support training with JSON string tool arguments (#3136) 
* feat: support training with JSON string tool arguments; fix PyArrow data type inconsistent error

* feat: raise error for tool call arguments decode

* Add test_chat_templates_tool_call_string_arguments.py

Add test for string arguments

* fix: change to correct qwen3 tokenizer

* fix: update docs to clarify arguments json

* chore: lint

* fix: duplicate

* chore: revert

* feat: add error to faq

* fix: remove duplicate fixture

---------

Co-authored-by: caoqinping <caoqinping@lixiang.com>
Co-authored-by: gamersover-blog <1611885128@qq.com>
Co-authored-by: NanoCode012 <nano@axolotl.ai>
 2025-09-25 12:06:21 +07:00
NanoCode012
856ff12171 feat(doc): add optimizations table of content to our improvements (#3175) [skip ci] 
* chore: format

* feat: add usage for alst

* chore: wording

* feat: add optimizations doc

* Apply suggestion from @SalmanMohammadi

Co-authored-by: salman <salman.mohammadi@outlook.com>

* Update docs/dataset-formats/index.qmd

Co-authored-by: salman <salman.mohammadi@outlook.com>

* feat: add alst, act offloading, nd parallelism, use relative links, and fix format

* chore: comments

---------

Co-authored-by: salman <salman.mohammadi@outlook.com>
 2025-09-24 16:13:49 -04:00
Dan Saunders
6bc959342b remove unused dep (#3180) 2025-09-24 13:18:44 -04:00
NanoCode012
b3b92687c4 chore: rename gemma3 270m config (#3174) 2025-09-24 13:48:38 +07:00
NanoCode012
55d1be2ae6 fix: unify default for conversations_field [skip-e2e] (#3070) 
* fix: unify default for conversations_field

* fix: suggestion to remove defaults
 2025-09-23 21:22:15 +07:00
NanoCode012
08d831c3d5 Feat: add qwen3-next (w packing+cce) (#3150) 
* feat: upgrade cce for qwen3-next

* feat: add sample qwen3 config

* feat: add packing patch for chunk_gated_delta_rule

* feat: add qwen3 link

* fix: tuple name

* feat: add tested qwen3 config

* fix: improve log

* feat: add patch for fla without packing

* fix: remove fla patch for standard mode

* feat: enable packing

* feat: add qwen3-next tests

* chore: move tests
 2025-09-23 11:31:15 +07:00
AlexHT Hung
7be8740c5c fix(rl): pass max_prompt_len to training args as max_prompt_length (#3113) 
* pass max_prompt_len to training args as max_prompt_length

* Update rl.py

* refactor

* format

* fix: default for max_prompt_length

* fix: defaults for trainer

---------

Co-authored-by: NanoCode012 <nano@axolotl.ai>
 2025-09-19 17:34:28 +07:00
NanoCode012
c51d6b06c3 feat: add apertus model and cce (#3144) [skip ci] 
* feat: add apertus, glm4v, glm4v_moe cce

* fix: arcee docs

* feat: add apertus

* feat: added vram usage

* fix: add apertus note

* feat: update doc on apertus xielu

* fix: add monkeypatch for xielu activation issue

* fix: simplify env

* feat: pin commit

* feat: add packing

* chore: move patch calling

* Update examples/apertus/README.md

Co-authored-by: salman <salman.mohammadi@outlook.com>

* Update examples/apertus/README.md

Co-authored-by: salman <salman.mohammadi@outlook.com>

* Update examples/apertus/README.md

Co-authored-by: salman <salman.mohammadi@outlook.com>

---------

Co-authored-by: salman <salman.mohammadi@outlook.com>
 2025-09-19 17:34:04 +07:00
NanoCode012
09959fac70 Feat: add Magistral Small 2509 and native mistral3 tokenizer support (#3165) 
* feat: update mistral common

* feat: add mistral3processor

* fix: loading

* fix: cast pixel_values to fp32

* fix: image tensor conversion

* feat: add FA2 support for pixtral based models

* fix: update mistral small 3.1 to use native tokenizer

* fix: install tips

* fix: improve info on sample dataset files

* chore: move mistral configs into subfolders

* fix: remove unneeded patch

* fix: indent

* feat: add integration tests

* chore: move

* feat: add magistral 2509 docs and example

* fix: convert tensor to bool

* feat: expand tests

* chore: move tests
 2025-09-18 15:42:20 +07:00
Dan Saunders
4065bc14c6 Debug log, logging improvements (#3159) 
* simplify logging

* remove comment

* progress on debug.log

* add debug-level logger for file log

* simplify

* case insensitivity; 3rd party logging improvements

* simplify

* fix

* tests

* lint

* nits

* nit

* Update tests/test_utils_tee.py

Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com>

* cleanup / comments

* fix

* oops

---------

Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com>
 2025-09-17 13:27:03 -04:00
salman
e5c427f6de qat doc updates (#3162) [skip-ci] 2025-09-17 10:38:15 +01:00
Wing Lian
86d6ee7c05 upgrade trl and accelerate (#3161) 
* upgrade trl==0.23.0

* upgrade accelerate patch fix

* add hints when using gradient_checkpointing with DPO

* set gradient-checpointing properly
 2025-09-16 14:53:01 -04:00
Wing Lian
d4cff1b7bb improve setting of NCCL_P2P_DISABLE on runpod (#3132) [skip ci] 
* improve setting of NCCL_P2P_DISABLE on runpod

* use recs from review
 2025-09-16 14:52:45 -04:00
Wing Lian
1ef6c196f7 setup env vars for ray train for FSDP (#3130) [skip ci] 2025-09-16 14:52:29 -04:00
salman
58d67bf98d Migrate QAT API; fix axolotl quantize for QAT-ed models; add NVFP4 (#3107) 2025-09-12 10:55:50 +01:00
salman
0401a15888 SEO go brrr (#3153) [skip-ci] 2025-09-12 10:55:11 +01:00
NanoCode012
fcfc13d710 feat(doc): update thinking and chat_template notes (#3114) [skip ci] 
* feat: update thinking and chat_template notes

* fix: grammar
 2025-09-12 14:45:18 +07:00
salman
9406c0c488 log before eval step (#3148) [skip-ci] 2025-09-11 11:19:30 +01:00
Dan Saunders
1b53c49e1a text diffusion training plugin (#3067) 
* diffusion training plugin

* cleanup

* nits

* fixes + improvements

* add back in reinit_weights (clobbered?); masking / pretrain fixes

* nits

* cleanup; tests draft

* sample generation, tests fixes

* fixes

* nits

* add inference support; add auto-mask token support

* nits

* nits

* progress

* simplify logging

* lint

* prefix args with diffusion_

* coderabbito

* tests fix

* nit

* nits

* cleanup + nits

* nits

* fix SFT sample gen

* fixes

* fix

* comments

* comments

* lint

* reward model lora fix

* cleanup; fix pretraining_dataset case

* gradio inference

* update cfgs

* update cfgs

* train, generation parity, cleanup

* fix

* simplify

* test

* test fix
 2025-09-10 20:27:00 -04:00
NanoCode012
b71482cec5 Feat: add hunyuan v1 (#3016) 
* feat: add hunyuan cce support

* feat: update cce docs

* feat: add multipack support for granite and hunyuan

* feat: add hunyuan docs and example config

* feat: update readme instructions to include CCE installation

* fix: chat template log appearing despite tokenizer already having template

* feat: add vram usage

* fix: remove duplicate cce install

* fix: use latest commit of PR in case rebased/pushed

* Revert "fix: use latest commit of PR in case rebased/pushed"

This reverts commit 8b60aa00de.

* feat: update doc as upstream merged
 2025-09-10 09:03:30 +07:00