Merge branch 'alekseymalakhov11-update-docs-and-poetry' of github.com…

…:turbo-llm/turbo-alignment into alekseymalakhov11-update-docs-and-poetry

CITATION.cff

@@ -5,8 +5,24 @@ message: >-
   metadata from this file.
 type: software
 authors:
-  - given-names: Surnachev
-    family-names: Nikita
+  - given-names: Alexander
+    family-names: Zamiralov
+  - given-names: Alexey
+    family-names: Gorbatovski
+  - given-names: Alexey
+    family-names: Malakhov
+  - given-names: Boris
+    family-names: Shaposhnikov
+  - given-names: Daniil
+    family-names: Taranets
+  - given-names: Elisey
+    family-names: Rykov
+  - given-names: Nikita
+    family-names: Surnachev
+  - given-names: Sergey
+    family-names: Kochetkov
+  - given-names: Yana
+    family-names: Shishkina
 repository-code: 'https://github.com/turbo-llm/turbo-alignment'
 abstract: "Turbo-Alignment is a library designed to streamline the fine-tuning and alignment of large language models, leveraging advanced techniques to enhance efficiency and scalability."
 keywords:
@@ -16,4 +32,4 @@ keywords:
   - ppo
   - rlhf
 license: Apache-2.0
-version: 0.0.3
+version: 0.0.3

README.md

@@ -130,32 +130,27 @@ python -m turbo_alignment train_kto --experiment_settings_path configs/exp/train
 <a name="-multimodal-train"></a>
 ## Multimodal train
 ⌛️  in progress..
-<!-- ## Multimodal Tasks 
-To start multimodal training, you should:
-- **Prepare the multimodal dataset**. See examples [here](docs/dataset_example.md#-multimodel-dataset).
-- **Preprocess the data (OPTIONAL)**. If you plan to run many experiments on the same dataset, you should preprocess it. The preprocessing stage includes reading `pixel_values` from images, encoding them with the specified encoder, and saving them in safetensors format. Later, during training, the pipeline will skip the stage of reading and encoding images and only extract prepared encodings from the safetensors files.
-- **Suitable config**: [llava.json](tests/fixtures/configs/train/multimodal/llama_llava_base_clip.json),[c_abs.json](tests/fixtures/configs/train/multimodal/llama_c_abs_clip_pickle.json) -->
+
 
 <a name="-rag-train"></a>
 ## RAG (Retrieval-Augmented Generation) 
-
 <a name="-sft-rag-train"></a>
 ### SFT-RAG
 - **📚 Dataset type**: prepare your dataset in `ChatDataset`, examples available [here](docs/dataset_example.md#-chat-dataset) format.
 - **📝 Configs Example**: [sft_with_retrieval_utility](configs/exp/train/sft/llama/sft_with_retrieval_utility.json)
-- **🖥️ CLI launch command**: ```bash
+- **🖥️ CLI launch command**: 
+```bash
 python -m turbo_alignment train_sft --experiment_settings_path configs/exp/train/sft/llama/sft_with_retrieval_utility.json
 ```
-
 <a name="-e2e-rag-train"></a>
 ### End2End-RAG
 - **📚 Dataset type**: prepare your dataset in `ChatDataset`, examples available [here](docs/dataset_example.md#-chat-dataset) format.
 - **📝 Configs Example**: [end2end_rag](configs/exp/train/rag/end2end_rag.json)
-- **🖥️ CLI launch command**: ```bash
+- **🖥️ CLI launch command**:
+```bash
 python -m turbo_alignment train_rag --experiment_settings_path configs/exp/train/rag/end2end_rag.json
 ```
 
-
 <a name="-inference"></a>
 # Inference
 <a name="-chat-inference"></a>
@@ -307,14 +302,15 @@ This project is licensed, see the [LICENSE](https://github.com/turbo-llm/turbo-a
 
 
 ## Citation
+Author names are listed in alphabetical order.
 
 ```bibtex
 @misc{surnachev2024turbo-alignment,
-  author = {Nikita Surnachev},
+  author = {Alexander Zamiralov and Alexey Gorbatovski and Alexey Malakhov and Boris Shaposhnikov and Daniil Taranets and Elisey Rykov and Nikita Surnachev and Sergey Kochetkov and Yana Shishkina},
   title = {turbo-alignment},
   year = {2024},
   publisher = {GitHub},
   journal = {GitHub repository},
   howpublished = {\url{https://github.com/turbo-llm/turbo-alignment}}
 }
-```
+```

pyproject.toml

@@ -18,7 +18,7 @@ peft = "0.8.2"
 typer ="^0.9.0"
 wandb ="^0.15.3"
 scipy = "^1.1.0"
-torch = "2.3.0"
+torch = "2.3.1"
 numpy = "^1.21"
 tomlkit ="^0.11.8"
 platformdirs ="^3.5.3"
@@ -50,9 +50,9 @@ albumentations = "^1.3.1"
 opencv-python = "^4.10.0.84"
 
 faiss-gpu = { version = "^1.7.2", optional = true }
-vllm = {version = "0.5.1", optional = true}
 deepspeed = { version = "0.12", optional = true }
 accelerate = { version = "0.29.0", optional = true }
+vllm = {version = "0.5.3", optional = true}
 
 
 [tool.poetry.dev-dependencies]

turbo_alignment/common/tf/loaders/model/model.py

@@ -2,13 +2,11 @@
 from peft import PeftModel, get_peft_model, prepare_model_for_int8_training
 from transformers import PreTrainedModel, PreTrainedTokenizerBase
 
-from turbo_alignment.common.tf.liger_kernels.monkey_patch_liger import (
-    apply_liger_kernel_to_gemma2,
-)
 from turbo_alignment.common.tf.loaders.model.registry import (
     PeftConfigRegistry,
     TransformersAutoModelRegistry,
 )
+from turbo_alignment.modeling.liger_kernels import apply_liger_kernel_to_gemma2
 from turbo_alignment.settings.model import (
     ModelForPeftSettings,
     PreTrainedAdaptersModelSettings,

turbo_alignment/modeling/liger_kernels/__init__.py

@@ -0,0 +1 @@
+from monkey_patch_liger import *

turbo_alignment/common/tf/liger_kernels/geglu.py → turbo_alignment/modeling/liger_kernels/geglu.py

@@ -5,7 +5,7 @@
 import triton
 import triton.language as tl
 
-from turbo_alignment.common.tf.liger_kernels.utils import (
+from turbo_alignment.modeling.liger_kernels.utils import (
     calculate_settings,
     compare_version,
     ensure_contiguous,

turbo_alignment/common/tf/liger_kernels/monkey_patch_liger.py → turbo_alignment/modeling/liger_kernels/monkey_patch_liger.py

@@ -1,9 +1,9 @@
 from transformers import PretrainedConfig, PreTrainedModel
 
 from turbo_alignment.common.logging import get_project_logger
-from turbo_alignment.common.tf.liger_kernels.cross_entropy import LigerCrossEntropyLoss
-from turbo_alignment.common.tf.liger_kernels.geglu import LigerGEGLUMLP
-from turbo_alignment.common.tf.liger_kernels.rope import liger_rotary_pos_emb
+from turbo_alignment.modeling.liger_kernels.cross_entropy import LigerCrossEntropyLoss
+from turbo_alignment.modeling.liger_kernels.geglu import LigerGEGLUMLP
+from turbo_alignment.modeling.liger_kernels.rope import liger_rotary_pos_emb
 
 logger = get_project_logger()