ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
3,327 commits 391 branches 58 tags 304 MiB
Commit graph

3,327 commits

This branch
This branch
All branches
Author SHA1 Message Date
Datta Nimmaturi
407d47b1ac Fix missing RotaryEmbeddingConfigMixin 2026-01-29 11:15:45 +05:30
Erland366
5c016f57ca Fix missing RotaryEmbeddingConfigMixin 2026-01-28 20:55:43 +00:00
pre-commit-ci[bot]
9e9a542b1b [pre-commit.ci] auto fixes from pre-commit.com hooks 
for more information, see https://pre-commit.ci
 2026-01-27 10:46:18 +00:00
Datta Nimmaturi
1f7facfcc6 Recognise GLM as MoE 2026-01-27 10:46:07 +00:00
Datta Nimmaturi
c1ca837297 Fix top_k on trl GRPO 2026-01-26 07:22:21 +00:00
Datta Nimmaturi
0e0bf10c90 No double compile for qwen3moe 2026-01-22 06:49:08 +00:00
pre-commit-ci[bot]
842f586749 [pre-commit.ci] auto fixes from pre-commit.com hooks 
for more information, see https://pre-commit.ci
 2026-01-21 03:58:35 +00:00
Datta Nimmaturi
4b4f47cde7 Merge remote-tracking branch 'origin/nightly' into moe_kernels 2026-01-21 03:58:24 +00:00
Datta Nimmaturi
94b744ef5e [transformers] [v5] remove unused hybridcache (#3910) 
* remote unused hybridcache

* cleanup
 2026-01-20 11:41:02 +05:30
Datta Nimmaturi
efd9d3ba5e Fixes for VL MoE and v5 transformers 2026-01-18 12:19:26 +00:00
Daniel Han
ecd10f2e55 Update pyproject.toml 2026-01-15 07:00:25 -08:00
Datta Nimmaturi
1107626da2 Revert "non TMA [T4]" 
This reverts commit 56a72c677a.
 2026-01-15 14:08:44 +00:00
Daniel Han
6edbfbc435 Update _utils.py 2026-01-15 05:09:26 -08:00
pluesclues
2164423ea6 Merge pull request #3628 from pluesclues/alternative_compute_chunked_loss 
Chunk Across Batch and Context length for logprob calculations for grpo
 2026-01-15 05:01:19 -08:00
Datta Nimmaturi
56a72c677a non TMA [T4] 2026-01-15 12:39:17 +00:00
Datta Nimmaturi
6be2291619 fixup derp 2026-01-15 12:14:05 +00:00
Daniel Han
832fffa40a Merge pull request #3895 from Datta0/rl_ref_trl 
[trl] use non lora model as base for RL
 2026-01-15 03:33:09 -08:00
Datta Nimmaturi
658284f0a8 Revert "use non lora model as base for RL" 
This reverts commit 5c73c69b87.
 2026-01-15 11:26:04 +00:00
pre-commit-ci[bot]
f5dde984a1 [pre-commit.ci] auto fixes from pre-commit.com hooks 
for more information, see https://pre-commit.ci
 2026-01-15 11:25:11 +00:00
Datta Nimmaturi
7f6dc63dc8 use non lora model as base for RL 2026-01-15 11:23:21 +00:00
Datta Nimmaturi
5c73c69b87 use non lora model as base for RL 2026-01-15 11:19:56 +00:00
Daniel Han
b55a2e30eb Merge pull request #3879 from ducviet00/fix-gc 
Disable gradient checkpointing when explicitly off for vision
 2026-01-14 04:32:02 -08:00
Michael Han
45eeae95c5 Update template.md 2026-01-14 03:45:35 -08:00
Datta Nimmaturi
894828cfb8 make dy contiguous 2026-01-14 11:10:59 +00:00
Datta Nimmaturi
1d86812b7c Make autotune default 2026-01-13 05:40:55 +00:00
Daniel Han
9adebb0dcf Merge pull request #3880 from f14-bertolotti/f14-wrong-ndim 
wrong number of dimensions
 2026-01-12 21:32:48 -08:00
Daniel Han
0f0b870781 Apply suggestion from @danielhanchen 2026-01-12 21:32:20 -08:00
Daniel Han
22ff11315e Merge pull request #3881 from unslothai/pre-commit-ci-update-config 
[pre-commit.ci] pre-commit autoupdate
 2026-01-12 21:29:58 -08:00
pre-commit-ci[bot]
2f8c4d962b [pre-commit.ci] pre-commit autoupdate 
updates:
- [github.com/astral-sh/ruff-pre-commit: v0.14.10 → v0.14.11](https://github.com/astral-sh/ruff-pre-commit/compare/v0.14.10...v0.14.11)
 2026-01-12 19:08:13 +00:00
Francesco Bertolotti
eaf3f932e0 wrong number of dimensions 2026-01-12 16:19:43 +01:00
Datta Nimmaturi
31016b4d3d LoRA for MoE 2026-01-12 13:43:20 +00:00
pre-commit-ci[bot]
1ac5c7288d [pre-commit.ci] auto fixes from pre-commit.com hooks 
for more information, see https://pre-commit.ci
 2026-01-12 11:59:18 +00:00
Datta Nimmaturi
a2f990db1e disable autotune 2026-01-12 11:57:20 +00:00
Duc-Viet Hoang
5b422f7a06 Complete disable gradient_checkpointing for vision when use_gradient_checkpointing=False 2026-01-12 10:03:54 +07:00
Daniel Han
bf7b5b06a1 Merge pull request #3865 from ykaitao/ktyang_configure_embedding_for_training 
reduce code duplication by _offload_frozen_module_for_training
 2026-01-09 21:02:55 -08:00
danielhanchen
6465496ab2 fix: use peft.utils.other for ModulesToSaveWrapper import 
ModulesToSaveWrapper was removed from peft.tuners.tuners_utils in PEFT
0.16.0. The class has been available in peft.utils.other since at least
PEFT 0.7.1, which is the minimum version Unsloth requires.

This fixes the ImportError when using PEFT >= 0.16.0.
 2026-01-09 23:24:39 +00:00
Kaitao Yang
0bff0ffbe5 reduce code duplication by _offload_frozen_module_for_training 2026-01-09 06:07:38 -08:00
Daniel Han
0e54b817af Merge pull request #3869 from hnxnq7/fix-kaggle-telemetry-detection 
Fix Kaggle telemetry misclassification when COLAB_ keys exist
 2026-01-08 17:29:23 -08:00
pre-commit-ci[bot]
3ce1060dd1 [pre-commit.ci] auto fixes from pre-commit.com hooks 
for more information, see https://pre-commit.ci
 2026-01-09 00:33:01 +00:00
Rachel Li
1193c9f526 Fix Kaggle telemetry detection & address review feedback 
- Fix Kaggle misclassification by prioritizing filesystem markers over env vars
- Preserve telemetry pings when statistics is explicitly provided
- Replace bare except with except Exception
- Minor cleanup based on automated review feedback
 2026-01-08 19:32:33 -05:00
Rachel Li
84701c55ff Fix telemetry ping regression for explicit statistics 
Fixed Codex regression: keep snapshot_download pings for explicit statistics values; detection only runs when statistics is None. Also replaced bare except.
 2026-01-08 19:20:24 -05:00
pre-commit-ci[bot]
e7d68f3e57 [pre-commit.ci] auto fixes from pre-commit.com hooks 
for more information, see https://pre-commit.ci
 2026-01-09 00:04:59 +00:00
Rachel Li
e13160ddc8 Update _utils.py 
fixed indentation
 2026-01-08 19:04:30 -05:00
pre-commit-ci[bot]
67bef80b1f [pre-commit.ci] auto fixes from pre-commit.com hooks 
for more information, see https://pre-commit.ci
 2026-01-08 23:49:55 +00:00
Rachel Li
1cdf751f8e Fix Kaggle telemetry misclassification when COLAB_ keys exist 
Problem: Kaggle notebook environments can expose both KAGGLE_* and COLAB_* environment keys. _get_statistics currently checks COLAB_ before KAGGLE_, causing Kaggle sessions to be labeled colab/colabpro.

Prefer filesystem markers (e.g. /kaggle/working, /content + /opt/colab) before env-key heuristics, then fall back to the existing env-key checks. This avoids misclassification when providers leak overlapping env vars.

Kaggle test notebook: https://www.kaggle.com/code/hnxnq07/kaggle-stats-gathering-test
 2026-01-08 18:44:22 -05:00
Daniel Han
ff8f5cd328 Merge pull request #3612 from Vangmay/feature/raw-text-dataprep 
Feature/raw text dataprep
 2026-01-08 03:38:15 -08:00
pre-commit-ci[bot]
3ac4f3c213 [pre-commit.ci] auto fixes from pre-commit.com hooks 
for more information, see https://pre-commit.ci
 2026-01-08 11:35:21 +00:00
Daniel Han
24bbe8a97a Fix bugs and add improvements to RawTextDataLoader 
- Fix test file: use return_tokenized instead of return_tensors
- Fix test file: use text_dataset instead of undefined dataset variable
- Move parameter validation to constructor (fail fast on invalid params)
- Add labels field in tokenized output for causal LM training
- Add empty file handling with clear error message
- Add tests for constructor validation and labels field
 2026-01-08 11:35:00 +00:00
Daniel Han
84ae73789c Merge pull request #3863 from unslothai/fix/fbgemm-cutlass-errors-sm100 
Fix FBGEMM/CUTLASS errors on SM100 (Blackwell) GPUs
 2026-01-08 03:19:53 -08:00
pre-commit-ci[bot]
41b7fe0c67 [pre-commit.ci] auto fixes from pre-commit.com hooks 
for more information, see https://pre-commit.ci
 2026-01-08 04:15:17 +00:00