set download to use revision

.github/workflows/base.yml

@@ -36,12 +36,6 @@ jobs:
             python_version: "3.11"
             pytorch: 2.4.1
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
-          - cuda: "124"
-            cuda_version: 12.4.1
-            cudnn_version: ""
-            python_version: "3.11"
-            pytorch: 2.5.0
-            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
     steps:
       - name: Checkout
         uses: actions/checkout@v3

.github/workflows/main.yml

@@ -29,11 +29,6 @@ jobs:
             python_version: "3.11"
             pytorch: 2.4.1
             axolotl_extras:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.5.0
-            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -91,11 +86,6 @@ jobs:
             python_version: "3.11"
             pytorch: 2.4.1
             axolotl_extras:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.5.0
-            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout

.github/workflows/multi-gpu-e2e.yml

@@ -21,17 +21,10 @@ jobs:
             pytorch: 2.3.1
             axolotl_extras:
             num_gpus: 2
-          - cuda: 124
-            cuda_version: 12.4.1
+          - cuda: 121
+            cuda_version: 12.1.1
             python_version: "3.11"
-            pytorch: 2.4.1
-            axolotl_extras:
-            num_gpus: 2
-            nightly_build: "true"
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.5.0
+            pytorch: 2.3.1
             axolotl_extras:
             num_gpus: 2
             nightly_build: "true"

.github/workflows/nightlies.yml

@@ -28,11 +28,6 @@ jobs:
             python_version: "3.11"
             pytorch: 2.4.1
             axolotl_extras:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.5.0
-            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -90,11 +85,6 @@ jobs:
             python_version: "3.11"
             pytorch: 2.4.1
             axolotl_extras:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.5.0
-            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout

.github/workflows/pypi.yml

@@ -27,7 +27,7 @@ jobs:
         run: |
           pip3 install wheel packaging
           pip3 install -e .
-          pip3 install -r requirements-dev.txt -r requirements-tests.txt
+          pip3 install -r requirements-tests.txt
 
       - name: Extract tag name
         id: tag

.github/workflows/tests-nightly.yml

@@ -25,7 +25,7 @@ jobs:
       fail-fast: false
       matrix:
         python_version: ["3.10", "3.11"]
-        pytorch_version: ["2.3.1", "2.4.1", "2.5.0"]
+        pytorch_version: ["2.3.1", "2.4.1"]
     timeout-minutes: 20
 
     steps:
@@ -47,14 +47,13 @@ jobs:
           sed -i 's#^transformers.*#transformers @ git+https://github.com/huggingface/transformers.git@main#' requirements.txt
           sed -i 's#^peft.*#peft @ git+https://github.com/huggingface/peft.git@main#' requirements.txt
           sed -i 's#^accelerate.*#accelerate @ git+https://github.com/huggingface/accelerate.git@main#' requirements.txt
-          sed -i 's#^trl.*#trl @ git+https://github.com/huggingface/trl.git@main#' requirements.txt
 
       - name: Install dependencies
         run: |
           pip3 install --upgrade pip
           pip3 install --upgrade packaging
           pip3 install -U -e .
-          pip3 install -r requirements-dev.txt -r requirements-tests.txt
+          pip3 install -r requirements-tests.txt
 
       - name: Run tests
         run: |
@@ -96,13 +95,6 @@ jobs:
             num_gpus: 1
             axolotl_extras:
             nightly_build: "true"
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.5.0
-            num_gpus: 1
-            axolotl_extras:
-            nightly_build: "true"
     steps:
       - name: Checkout
         uses: actions/checkout@v4

.github/workflows/tests.yml

@@ -36,7 +36,7 @@ jobs:
       fail-fast: false
       matrix:
         python_version: ["3.10", "3.11"]
-        pytorch_version: ["2.3.1", "2.4.1", "2.5.0"]
+        pytorch_version: ["2.3.1", "2.4.1"]
     timeout-minutes: 20
 
     steps:
@@ -49,20 +49,16 @@ jobs:
           python-version: ${{ matrix.python_version }}
           cache: 'pip' # caching pip dependencies
 
-      - name: upgrade pip
-        run: |
-          pip3 install --upgrade pip
-          pip3 install --upgrade packaging setuptools wheel
-
       - name: Install PyTorch
         run: |
-          pip3 install torch==${{ matrix.pytorch_version }}
+          pip3 install torch==${{ matrix.pytorch_version }} --index-url https://download.pytorch.org/whl/cpu
 
       - name: Install dependencies
         run: |
-          pip3 show torch
+          pip3 install --upgrade pip
+          pip3 install --upgrade packaging
           pip3 install -U -e .
-          pip3 install -r requirements-dev.txt -r requirements-tests.txt
+          pip3 install -r requirements-tests.txt
 
       - name: Run tests
         run: |
@@ -76,7 +72,7 @@ jobs:
     if: github.repository_owner == 'axolotl-ai-cloud'
     # this job needs to be run on self-hosted GPU runners...
     runs-on: [self-hosted, modal]
-    timeout-minutes: 90
+    timeout-minutes: 60
     needs: [pre-commit, pytest]
 
     strategy:
@@ -101,12 +97,6 @@ jobs:
             pytorch: 2.4.1
             num_gpus: 1
             axolotl_extras:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.5.0
-            num_gpus: 1
-            axolotl_extras:
     steps:
       - name: Checkout
         uses: actions/checkout@v4

1991.yml

@@ -1,295 +0,0 @@
-base_model: Qwen/Qwen2.5-14B-Instruct
-model_type: AutoModelForCausalLM #nohup accelerate launch -m axolotl.cli.train /home/ubuntu/qwen2.5_14B.yml > training_output.log 2>&1 &
-tokenizer_type: AutoTokenizer
-trust_remote_code: true
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-datasets:
-  - path: tatsu-lab/alpaca
-    type: alpaca
-
-chat_template: chatml
-dataset_prepared_path:
-val_set_size: 0
-output_dir: ./outputs/out
-
-sequence_len: 2048
-sample_packing: true
-eval_sample_packing: true
-pad_to_sequence_len: true
-
-unfrozen_parameters:
-- ^lm_head.weight$
-- ^model.embed_tokens.weight$
-# input_layernorm layers
-- model.layers.0.input_layernorm
-- model.layers.1.input_layernorm
-- model.layers.2.input_layernorm
-- model.layers.3.input_layernorm
-- model.layers.4.input_layernorm
-- model.layers.5.input_layernorm
-- model.layers.6.input_layernorm
-- model.layers.7.input_layernorm
-- model.layers.8.input_layernorm
-- model.layers.9.input_layernorm
-- model.layers.10.input_layernorm
-- model.layers.11.input_layernorm
-- model.layers.12.input_layernorm
-- model.layers.13.input_layernorm
-- model.layers.14.input_layernorm
-- model.layers.15.input_layernorm
-- model.layers.16.input_layernorm
-- model.layers.17.input_layernorm
-- model.layers.18.input_layernorm
-- model.layers.19.input_layernorm
-- model.layers.20.input_layernorm
-- model.layers.21.input_layernorm
-- model.layers.22.input_layernorm
-- model.layers.23.input_layernorm
-# lm_head layers
-# mlp.down_proj layers
-- model.layers.1.mlp.down_proj
-- model.layers.35.mlp.down_proj
-- model.layers.38.mlp.down_proj
-- model.layers.37.mlp.down_proj
-- model.layers.36.mlp.down_proj
-- model.layers.15.mlp.down_proj
-- model.layers.11.mlp.down_proj
-- model.layers.12.mlp.down_proj
-- model.layers.34.mlp.down_proj
-- model.layers.44.mlp.down_proj
-- model.layers.45.mlp.down_proj
-- model.layers.9.mlp.down_proj
-- model.layers.41.mlp.down_proj
-- model.layers.33.mlp.down_proj
-- model.layers.43.mlp.down_proj
-- model.layers.40.mlp.down_proj
-- model.layers.13.mlp.down_proj
-- model.layers.8.mlp.down_proj
-- model.layers.39.mlp.down_proj
-- model.layers.10.mlp.down_proj
-- model.layers.14.mlp.down_proj
-- model.layers.16.mlp.down_proj
-- model.layers.31.mlp.down_proj
-- model.layers.32.mlp.down_proj
-# mlp.gate_proj layers
-- model.layers.1.mlp.gate_proj
-- model.layers.44.mlp.gate_proj
-- model.layers.46.mlp.gate_proj
-- model.layers.45.mlp.gate_proj
-- model.layers.43.mlp.gate_proj
-- model.layers.47.mlp.gate_proj
-- model.layers.42.mlp.gate_proj
-- model.layers.32.mlp.gate_proj
-- model.layers.27.mlp.gate_proj
-- model.layers.33.mlp.gate_proj
-- model.layers.28.mlp.gate_proj
-- model.layers.39.mlp.gate_proj
-- model.layers.41.mlp.gate_proj
-- model.layers.40.mlp.gate_proj
-- model.layers.30.mlp.gate_proj
-- model.layers.29.mlp.gate_proj
-- model.layers.31.mlp.gate_proj
-- model.layers.26.mlp.gate_proj
-- model.layers.37.mlp.gate_proj
-- model.layers.10.mlp.gate_proj
-- model.layers.38.mlp.gate_proj
-- model.layers.12.mlp.gate_proj
-- model.layers.36.mlp.gate_proj
-- model.layers.13.mlp.gate_proj
-# mlp.up_proj layers
-- model.layers.1.mlp.up_proj
-- model.layers.13.mlp.up_proj
-- model.layers.11.mlp.up_proj
-- model.layers.14.mlp.up_proj
-- model.layers.15.mlp.up_proj
-- model.layers.12.mlp.up_proj
-- model.layers.8.mlp.up_proj
-- model.layers.16.mlp.up_proj
-- model.layers.9.mlp.up_proj
-- model.layers.19.mlp.up_proj
-- model.layers.10.mlp.up_proj
-- model.layers.7.mlp.up_proj
-- model.layers.17.mlp.up_proj
-- model.layers.20.mlp.up_proj
-- model.layers.21.mlp.up_proj
-- model.layers.18.mlp.up_proj
-- model.layers.38.mlp.up_proj
-- model.layers.37.mlp.up_proj
-- model.layers.39.mlp.up_proj
-- model.layers.42.mlp.up_proj
-- model.layers.41.mlp.up_proj
-- model.layers.27.mlp.up_proj
-- model.layers.28.mlp.up_proj
-- model.layers.34.mlp.up_proj
-# model.norm layers
-# post_attention_layernorm layers
-- model.layers.0.post_attention_layernorm
-- model.layers.1.post_attention_layernorm
-- model.layers.2.post_attention_layernorm
-- model.layers.3.post_attention_layernorm
-- model.layers.4.post_attention_layernorm
-- model.layers.5.post_attention_layernorm
-- model.layers.6.post_attention_layernorm
-- model.layers.7.post_attention_layernorm
-- model.layers.8.post_attention_layernorm
-- model.layers.9.post_attention_layernorm
-- model.layers.10.post_attention_layernorm
-- model.layers.11.post_attention_layernorm
-- model.layers.12.post_attention_layernorm
-- model.layers.13.post_attention_layernorm
-- model.layers.14.post_attention_layernorm
-- model.layers.15.post_attention_layernorm
-- model.layers.16.post_attention_layernorm
-- model.layers.17.post_attention_layernorm
-- model.layers.18.post_attention_layernorm
-- model.layers.19.post_attention_layernorm
-- model.layers.20.post_attention_layernorm
-- model.layers.21.post_attention_layernorm
-- model.layers.22.post_attention_layernorm
-- model.layers.23.post_attention_layernorm
-# self_attn.k_proj layers
-- model.layers.47.self_attn.k_proj
-- model.layers.39.self_attn.k_proj
-- model.layers.41.self_attn.k_proj
-- model.layers.37.self_attn.k_proj
-- model.layers.35.self_attn.k_proj
-- model.layers.44.self_attn.k_proj
-- model.layers.38.self_attn.k_proj
-- model.layers.14.self_attn.k_proj
-- model.layers.7.self_attn.k_proj
-- model.layers.12.self_attn.k_proj
-- model.layers.11.self_attn.k_proj
-- model.layers.32.self_attn.k_proj
-- model.layers.10.self_attn.k_proj
-- model.layers.8.self_attn.k_proj
-- model.layers.9.self_attn.k_proj
-- model.layers.6.self_attn.k_proj
-- model.layers.45.self_attn.k_proj
-- model.layers.42.self_attn.k_proj
-- model.layers.5.self_attn.k_proj
-- model.layers.40.self_attn.k_proj
-- model.layers.33.self_attn.k_proj
-- model.layers.0.self_attn.k_proj
-- model.layers.34.self_attn.k_proj
-- model.layers.13.self_attn.k_proj
-# self_attn.o_proj layers
-- model.layers.12.self_attn.o_proj
-- model.layers.5.self_attn.o_proj
-- model.layers.14.self_attn.o_proj
-- model.layers.16.self_attn.o_proj
-- model.layers.20.self_attn.o_proj
-- model.layers.13.self_attn.o_proj
-- model.layers.11.self_attn.o_proj
-- model.layers.4.self_attn.o_proj
-- model.layers.6.self_attn.o_proj
-- model.layers.19.self_attn.o_proj
-- model.layers.7.self_attn.o_proj
-- model.layers.18.self_attn.o_proj
-- model.layers.8.self_attn.o_proj
-- model.layers.38.self_attn.o_proj
-- model.layers.15.self_attn.o_proj
-- model.layers.17.self_attn.o_proj
-- model.layers.9.self_attn.o_proj
-- model.layers.10.self_attn.o_proj
-- model.layers.21.self_attn.o_proj
-- model.layers.28.self_attn.o_proj
-- model.layers.32.self_attn.o_proj
-- model.layers.35.self_attn.o_proj
-- model.layers.39.self_attn.o_proj
-- model.layers.3.self_attn.o_proj
-# self_attn.q_proj layers
-- model.layers.1.self_attn.q_proj
-- model.layers.2.self_attn.q_proj
-- model.layers.3.self_attn.q_proj
-- model.layers.44.self_attn.q_proj
-- model.layers.29.self_attn.q_proj
-- model.layers.45.self_attn.q_proj
-- model.layers.43.self_attn.q_proj
-- model.layers.32.self_attn.q_proj
-- model.layers.38.self_attn.q_proj
-- model.layers.19.self_attn.q_proj
-- model.layers.42.self_attn.q_proj
-- model.layers.34.self_attn.q_proj
-- model.layers.36.self_attn.q_proj
-- model.layers.40.self_attn.q_proj
-- model.layers.26.self_attn.q_proj
-- model.layers.20.self_attn.q_proj
-- model.layers.39.self_attn.q_proj
-- model.layers.28.self_attn.q_proj
-- model.layers.35.self_attn.q_proj
-- model.layers.41.self_attn.q_proj
-- model.layers.33.self_attn.q_proj
-- model.layers.25.self_attn.q_proj
-- model.layers.30.self_attn.q_proj
-- model.layers.27.self_attn.q_proj
-# self_attn.v_proj layers
-- model.layers.0.self_attn.v_proj
-- model.layers.7.self_attn.v_proj
-- model.layers.39.self_attn.v_proj
-- model.layers.31.self_attn.v_proj
-- model.layers.15.self_attn.v_proj
-- model.layers.10.self_attn.v_proj
-- model.layers.32.self_attn.v_proj
-- model.layers.41.self_attn.v_proj
-- model.layers.6.self_attn.v_proj
-- model.layers.33.self_attn.v_proj
-- model.layers.42.self_attn.v_proj
-- model.layers.29.self_attn.v_proj
-- model.layers.14.self_attn.v_proj
-- model.layers.9.self_attn.v_proj
-- model.layers.35.self_attn.v_proj
-- model.layers.38.self_attn.v_proj
-- model.layers.13.self_attn.v_proj
-- model.layers.30.self_attn.v_proj
-- model.layers.5.self_attn.v_proj
-- model.layers.34.self_attn.v_proj
-- model.layers.28.self_attn.v_proj
-- model.layers.37.self_attn.v_proj
-- model.layers.27.self_attn.v_proj
-- model.layers.11.self_attn.v_proj
-# model.embed_tokens layers
-
-
-gradient_accumulation_steps: 2
-micro_batch_size: 2
-num_epochs: 3
-optimizer: adamw_torch_fused
-lr_scheduler: linear
-learning_rate: 5e-6
-
-train_on_inputs: false
-group_by_length: false
-bf16: auto
-fp16:
-tf32: false
-
-plugins:
-  - axolotl.integrations.liger.LigerPlugin
-liger_rope: true
-liger_rms_norm: true
-liger_swiglu: true
-liger_fused_linear_cross_entropy: true
-
-gradient_checkpointing: unsloth
-gradient_checkpointing_kwargs:
-  use_reentrant: false
-early_stopping_patience:
-resume_from_checkpoint:
-local_rank:
-logging_steps: 1
-xformers_attention:
-flash_attention: true
-
-warmup_steps: 10
-evals_per_epoch: 2
-saves_per_epoch: 1
-save_total_limit: 4
-debug:
-deepspeed: deepspeed_configs/zero3_bf16.json
-weight_decay: 0.05
-special_tokens:
-  eos_token: <|im_end|>

README.md

@@ -121,7 +121,7 @@ Features:
 
 Get started with Axolotl in just a few steps! This quickstart guide will walk you through setting up and running a basic fine-tuning task.
 
-**Requirements**: Nvidia GPU (Ampere architecture or newer for `bf16` and Flash Attention), Python >=3.10 and PyTorch >=2.3.1.
+**Requirements**: Python >=3.10 and Pytorch >=2.1.1.
 
 ```bash
 git clone https://github.com/axolotl-ai-cloud/axolotl
@@ -383,7 +383,7 @@ See [examples](examples) for quick start. It is recommended to duplicate and mod
         - typescript
       type: ... # unimplemented custom format
 
-      # fastchat conversation (deprecation soon, use chat_template https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/conversation.html#chat_template)
+      # fastchat conversation
       # See 'conversation' options: https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py
     - path: ...
       type: sharegpt

cicd/Dockerfile.jinja

@@ -23,11 +23,11 @@ RUN git fetch origin +$GITHUB_REF && \
     git checkout FETCH_HEAD
 
 # If AXOLOTL_EXTRAS is set, append it in brackets
+RUN pip install causal_conv1d
 RUN if [ "$NIGHTLY_BUILD" = "true" ] ; then \
         sed -i 's#^transformers.*#transformers @ git+https://github.com/huggingface/transformers.git@main#' requirements.txt; \
         sed -i 's#^peft.*#peft @ git+https://github.com/huggingface/peft.git@main#' requirements.txt; \
         sed -i 's#^accelerate.*#accelerate @ git+https://github.com/huggingface/accelerate.git@main#' requirements.txt; \
-        sed -i 's#^trl.*#trl @ git+https://github.com/huggingface/trl.git@main#' requirements.txt; \
     fi
 
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
@@ -37,7 +37,7 @@ RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
     fi
 
 # So we can test the Docker image
-RUN pip install -r requirements-dev.txt -r requirements-tests.txt
+RUN pip install -r requirements-tests.txt
 
 # fix so that git fetch/pull from remote works
 RUN git config remote.origin.fetch "+refs/heads/*:refs/remotes/origin/*" && \

cicd/cicd.sh

@@ -1,6 +1,6 @@
 #!/bin/bash
 set -e
 
-pytest -n4 --ignore=tests/e2e/ /workspace/axolotl/tests/
+pytest --ignore=tests/e2e/ /workspace/axolotl/tests/
 pytest -n1 --dist loadfile -v /workspace/axolotl/tests/e2e/patched/ /workspace/axolotl/tests/e2e/integrations/
 pytest --ignore=tests/e2e/patched/ --ignore=tests/e2e/multigpu/ --ignore=tests/e2e/integrations/ /workspace/axolotl/tests/e2e/

cicd/multigpu.py

@@ -64,7 +64,7 @@ def run_cmd(cmd: str, run_folder: str):
 @stub.function(
     image=cicd_image,
     gpu=GPU_CONFIG,
-    timeout=60 * 60,
+    timeout=45 * 60,
     cpu=8.0,
     memory=131072 * N_GPUS,
 )

cicd/tests.py

@@ -65,7 +65,7 @@ def run_cmd(cmd: str, run_folder: str):
 @stub.function(
     image=cicd_image,
     gpu=GPU_CONFIG,
-    timeout=60 * 60,
+    timeout=45 * 60,
     cpu=8.0,
     memory=131072,
 )

deepspeed_configs/zero3_bf16.json

@@ -14,6 +14,15 @@
   "bf16": {
     "enabled": true
   },
+  "fp16": {
+    "enabled": "auto",
+    "auto_cast": false,
+    "loss_scale": 0,
+    "initial_scale_power": 32,
+    "loss_scale_window": 1000,
+    "hysteresis": 2,
+    "min_loss_scale": 1
+  },
   "gradient_accumulation_steps": "auto",
   "gradient_clipping": "auto",
   "train_batch_size": "auto",

deepspeed_configs/zero3_bf16_cpuoffload_all.json

@@ -24,6 +24,15 @@
   "bf16": {
     "enabled": true
   },
+  "fp16": {
+    "enabled": "auto",
+    "auto_cast": false,
+    "loss_scale": 0,
+    "initial_scale_power": 32,
+    "loss_scale_window": 1000,
+    "hysteresis": 2,
+    "min_loss_scale": 1
+  },
   "gradient_accumulation_steps": "auto",
   "gradient_clipping": "auto",
   "train_batch_size": "auto",

deepspeed_configs/zero3_bf16_cpuoffload_params.json

@@ -20,6 +20,15 @@
   "bf16": {
     "enabled": true
   },
+  "fp16": {
+    "enabled": "auto",
+    "auto_cast": false,
+    "loss_scale": 0,
+    "initial_scale_power": 32,
+    "loss_scale_window": 1000,
+    "hysteresis": 2,
+    "min_loss_scale": 1
+  },
   "gradient_accumulation_steps": "auto",
   "gradient_clipping": "auto",
   "train_batch_size": "auto",

docker/Dockerfile

@@ -20,6 +20,7 @@ RUN git clone --depth=1 https://github.com/axolotl-ai-cloud/axolotl.git
 WORKDIR /workspace/axolotl
 
 # If AXOLOTL_EXTRAS is set, append it in brackets
+RUN pip install causal_conv1d
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
         pip install -e .[deepspeed,flash-attn,optimizers,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \

docs/config.qmd

@@ -83,7 +83,7 @@ lora_on_cpu: true
 datasets:
   # HuggingFace dataset repo | s3://,gs:// path | "json" for local dataset, make sure to fill data_files
   - path: vicgalle/alpaca-gpt4
-    # The type of prompt to use for training. [alpaca, sharegpt, gpteacher, oasst, reflection]
+  # The type of prompt to use for training. [alpaca, sharegpt, gpteacher, oasst, reflection]
     type: alpaca # format | format:<prompt_style> (chat/instruct) | <prompt_strategies>.load_<load_fn>
     ds_type: # Optional[str] (json|arrow|parquet|text|csv) defines the datatype when path is a file
     data_files: # Optional[str] path to source data files
@@ -124,48 +124,6 @@ datasets:
       # For `completion` datsets only, uses the provided field instead of `text` column
       field:
 
-  # Using chat template
-  - path: ...
-    # Set type to `chat_template` to use this strategy
-    type: chat_template
-    # Specify the name of the chat template to use
-    # The name of the chat template to use for training, following values are supported:
-    # - tokenizer_default: Uses the chat template that is available in the tokenizer_config.json. If the chat template is not available in the tokenizer, it will raise an error. This is the default.
-    # - alpaca/inst/chatml/gemma/cohere/llama3/phi_3/deepseek_v2/jamba: These chat templates are available in the axolotl codebase at src/axolotl/utils/chat_templates.py
-    # - tokenizer_default_fallback_*: where * is the name of the chat template to fallback to if the tokenizer does not have a chat template else default to tokenizer. E.g. tokenizer_default_fallback_chatml.
-    # - jinja: Uses a custom jinja template for the chat template. The custom jinja template should be provided in the chat_template_jinja field.
-    chat_template: tokenizer_default
-    # Custom jinja template for chat template. This will be only used if `chat_template` is set to `jinja` or empty (in which case chat_template is automatically set to `jinja`).
-    chat_template_jinja:
-    # The key in the data example that contains the messages. Default is "messages".
-    field_messages: messages
-    # The key in the message turn that contains the role. Default is "role".
-    message_field_role: role
-    # The key in the message turn that contains the content. Default is "content".
-    message_field_content: content
-    # Optional[Dict[str, List]]. Roles mapping for the messages.
-    roles:
-      user: ["human", "user"]
-      assistant: ["gpt", "assistant", "ai"]
-      system: ["system"]
-
-    ## NOTE: Leaving the below empty will default to using the simple legacy tokenization strategy where only last message is trained on.
-
-    # Optional[List[str]]. Roles to train on. The tokens from these roles will be considered for the loss.
-    roles_to_train: ["gpt", "assistant"]
-    # Optional[str]. Which EOS tokens to train on in the conversation. Possible values are:
-    # - all: train on all EOS tokens
-    # - turn: train on the EOS token at the end of each trainable turn
-    # - last: train on the last EOS token in the conversation
-    train_on_eos: last
-    # The key in the message turn that indicates via boolean whether tokens of a turn should be considered for training. Useful to selectively train on certain turns besides the `roles_to_train`.
-    message_field_training: training
-    # The key in the message turn that contains the training details. Useful to selectively train on certain tokens in a turn.
-    # The value of the key is a List[Dict] containing `begin_offset` (start character index in content), `end_offset` (end character index in content), and `train` (boolean whether to train).
-    # See example at `docs/dataset-formats/conversation.qmd`
-    message_field_training_detail: train_detail
-
-
 # If false, the datasets will not be shuffled and will keep their original order in `datasets`.
 # The same applies to the `test_datasets` option and the `pretraining_dataset` option. Default is true.
 shuffle_merged_datasets: true
@@ -184,16 +142,9 @@ test_datasets:
 # use RL training: 'dpo', 'ipo', 'kto'
 rl:
 
-# The name of the chat template to use for training, following values are supported:
-# - tokenizer_default: Uses the chat template that is available in the tokenizer_config.json. If the chat template is not available in the tokenizer, it will raise an error. This is the default value.
-# - alpaca/inst/chatml/gemma/cohere/llama3/phi_3/deepseek_v2/jamba: These chat templates are available in the axolotl codebase at src/axolotl/utils/chat_templates.py
-# - tokenizer_default_fallback_*: where * is the name of the chat template to fallback to. E.g. tokenizer_default_fallback_chatml. This is useful when the chat template is not available in the tokenizer.
-# - jinja: Uses a custom jinja template for the chat template. The custom jinja template should be provided in the chat_template_jinja field.
-# The selected chat template will be saved to the tokenizer_config.json for easier inferencing
-# Note: It is recommended to set train_on_inputs to true when using a chat template that is different from the model's default chat template.
-chat_template: tokenizer_default
-# custom jinja template for chat template. This will be only used if chat_template is set to `jinja` or `null` (in which case chat_template is automatically set to `jinja`). Default is null.
-chat_template_jinja: null
+# Saves the desired chat template to the tokenizer_config.json for easier inferencing
+# Currently supports chatml and inst (mistral/mixtral)
+chat_template: chatml
 # Changes the default system message
 default_system_message: You are a helpful assistant. Please give a long and detailed answer. # Currently only supports chatml.
 # Axolotl attempts to save the dataset as an arrow after packing the data together so
@@ -315,7 +266,6 @@ wandb_log_model: # "checkpoint" to log model to wandb Artifacts every `save_step
 # mlflow configuration if you're using it
 mlflow_tracking_uri: # URI to mlflow
 mlflow_experiment_name: # Your experiment name
-mlflow_run_name: # Your run name
 hf_mlflow_log_artifacts:  # set to true to copy each saved checkpoint on each save to mlflow artifact registry
 
 # Comet configuration if you're using it
@@ -364,7 +314,7 @@ max_steps:
 
 eval_table_size: # Approximate number of predictions sent to wandb depending on batch size. Enabled above 0. Default is 0
 eval_max_new_tokens: # Total number of tokens generated for predictions sent to wandb. Default is 128
-eval_causal_lm_metrics: # HF evaluate metrics used during evaluation. Default is ["sacrebleu", "comet", "ter", "chrf", "perplexity"]
+eval_causal_lm_metrics: # HF evaluate metrics used during evaluation. Default is ["sacrebleu", "comet", "ter", chrf]
 
 loss_watchdog_threshold: # High loss value, indicating the learning has broken down (a good estimate is ~2 times the loss at the start of training)
 loss_watchdog_patience: # Number of high-loss steps in a row before the trainer aborts (default: 3)

docs/dataset-formats/conversation.qmd

@@ -6,8 +6,6 @@ order: 3
 
 ## sharegpt
 
-UPDATE: ShareGPT is being deprecated in the next release. Please see `chat_template` section below.
-
 conversations where `from` is `human`/`gpt`. (optional: first row with role `system` to override default system prompt)
 
 ```{.json filename="data.jsonl"}
@@ -71,138 +69,3 @@ creates a chat where bot is asked to tell a joke, then explain why the joke is f
 ```{.json filename="data.jsonl"}
 {"conversations": [{"title": "...", "text": "...", "explanation": "..."}]}
 ```
-
-
-## chat_template
-
-Chat Template strategy uses a jinja2 template that converts a list of messages into a prompt. Support using tokenizer's template, a supported template, or custom jinja2.
-
-```{.json filename="data.jsonl"}
-{"conversations": [{"role": "...", "content": "..."}]}
-```
-
-See `config.qmd` for full configs and supported templates.
-
-### Migrating from sharegpt
-
-Most configs can be adapted as follows:
-
-```yaml
-# old
-chat_template: chatml
-datasets:
-  - path: ...
-    type: sharegpt
-    conversation: chatml
-
-# new (if using tokenizer's chat_template)
-datasets:
-  - path: ...
-    type: chat_template
-
-    field_messages: conversations
-    message_field_role: from
-    message_field_content: value
-
-# new (if setting a new chat_template like chatml, gemma, etc)
-chat_template: chatml
-datasets:
-  - path: ...
-    type: chat_template
-
-    field_messages: conversations
-    message_field_role: from
-    message_field_content: value
-```
-
-We recommend checking the below examples for other usecases.
-
-### Examples
-
-1. Using the default chat template in the tokenizer_config.json on OpenAI messages format, training on only last message.
-
-```yaml
-datasets:
-  - path: ...
-    type: chat_template
-```
-
-2. Using the `gemma` chat template to override the tokenizer_config.json's chat template on OpenAI messages format, training on all assistant messages.
-
-```yaml
-chat_template: gemma # this overwrites the tokenizer's chat_template
-datasets:
-  - path: ...
-    type: chat_template
-    roles_to_train: ["assistant"]
-```
-
-3. Using the tokenizer_config.json's chat template or `chatml` as fallback if the former's chat template does not exist, on OpenAI messages format, training on all assistant messages.
-
-```yaml
-chat_template: tokenizer_default_fallback_chatml # this overwrites the tokenizer's chat_template
-datasets:
-  - path: ...
-    type: chat_template
-    roles_to_train: ["assistant"]
-```
-
-4. Using a custom jinja template on OpenAI messages format, training on all assistant messages.
-
-```yaml
-# chat_template: jinja # `jinja` will be implied if the `chat_template_jinja` is set and this field is empty
-chat_template_jinja: "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'system') %}{{'<|system|>' + '\n' + message['content'] + '<|end|>' + '\n'}}{% elif (message['role'] == 'user') %}{{'<|user|>' + '\n' + message['content'] + '<|end|>' + '\n' + '<|assistant|>' + '\n'}}{% elif message['role'] == 'assistant' %}{{message['content'] + '<|end|>' + '\n'}}{% endif %}{% endfor %}"
-
-datasets:
-  - path: ...
-    type: chat_template
-    roles_to_train: ["assistant"]
-```
-
-5. (Advanced) Using fine-grained control over tokens and turns to train in a conversation
-
-For a data sample that looks like:
-
-```{.json filename="data.jsonl"}
-{
-  "conversations": [
-    {"from": "system", "value": "You are an AI assistant.", "train": false},
-    {"from": "human", "value": "Hello", "train": false},
-    {"from": "assistant", "value": "Hello", "train": true},
-    {"from": "human", "value": "How are you?", "train": true},
-    {
-      "from": "assistant",
-      "value": "I'm doing very well, thank you!",
-      "train_detail": [
-        {"begin_offset": 0, "end_offset": 8, "train": false},
-        {"begin_offset": 9, "end_offset": 18, "train": true},
-        {"begin_offset": 19, "end_offset": 30, "train": false},
-      ],
-    },
-    {
-        "from": "human",
-        "value": "I'm doing very well, thank you!",
-        "train": true,
-    },
-    {"from": "assistant", "value": "Hi there!", "train": true}
-  ]
-}
-```
-
-The configuration would look like:
-
-```yaml
-datasets:
-  - path: ...
-    type: chat_template
-    chat_template: tokenizer_default
-    field_messages: conversations
-    message_field_role: from
-    message_field_content: value
-    roles_to_train: []
-    train_on_eos: turn
-    message_field_training: train
-    message_field_training_detail: train_detail
-```
-
-Tip: It is not necessary to use both `message_field_training` and `message_field_training_detail` at a time.

examples/gemma2/reward-model.yaml

@@ -1,63 +0,0 @@
-base_model: google/gemma-2-2b
-model_type: AutoModelForSequenceClassification
-tokenizer_type: AutoTokenizer
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-reward_model: true
-chat_template: gemma
-datasets:
-  - path: argilla/distilabel-intel-orca-dpo-pairs
-    type: bradley_terry.chat_template
-val_set_size: 0.0
-output_dir: ./outputs/out
-remove_unused_columns: false
-
-sequence_len: 2048
-sample_packing: false
-eval_sample_packing: false
-pad_to_sequence_len: true
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-
-gradient_accumulation_steps: 4
-micro_batch_size: 2
-num_epochs: 4
-optimizer: adamw_bnb_8bit
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-train_on_inputs: false
-group_by_length: false
-bf16: true
-fp16:
-tf32: true
-
-gradient_checkpointing: true
-gradient_checkpointing_kwargs:
-  use_reentrant: false
-early_stopping_patience:
-resume_from_checkpoint:
-local_rank:
-logging_steps: 1
-xformers_attention:
-flash_attention: true
-
-warmup_ratio: 0.1
-evals_per_epoch:
-eval_table_size:
-eval_max_new_tokens: 128
-saves_per_epoch: 1
-debug:
-deepspeed:
-weight_decay: 0.0
-fsdp:
-fsdp_config:
-special_tokens:

examples/llama-3/instruct-dpo-lora-8b.yml

@@ -11,6 +11,7 @@ rl: dpo
 datasets:
   - path: fozziethebeat/alpaca_messages_2k_dpo_test
     type: chat_template.default
+    chat_template: llama3
     field_messages: conversation
     field_chosen: chosen
     field_rejected: rejected

examples/llama-3/instruct-lora-8b.yml

@@ -10,6 +10,7 @@ chat_template: llama3
 datasets:
   - path: fozziethebeat/alpaca_messages_2k_test
     type: chat_template
+    chat_template: llama3
     field_messages: messages
     message_field_role: role
     message_field_content: content

examples/llama-3/qlora-1b.yml

@@ -1,77 +0,0 @@
-base_model: meta-llama/Llama-3.2-1B
-
-load_in_8bit: false
-load_in_4bit: true
-strict: false
-
-datasets:
-  - path: teknium/GPT4-LLM-Cleaned
-    type: alpaca
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.1
-output_dir: ./outputs/qlora-out
-
-adapter: qlora
-lora_model_dir:
-
-sequence_len: 2048
-sample_packing: true
-eval_sample_packing: true
-pad_to_sequence_len: true
-
-lora_r: 32
-lora_alpha: 16
-lora_dropout: 0.05
-lora_target_linear: true
-lora_fan_in_fan_out:
-lora_target_modules:
-  - gate_proj
-  - down_proj
-  - up_proj
-  - q_proj
-  - v_proj
-  - k_proj
-  - o_proj
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 4
-micro_batch_size: 2
-num_epochs: 1
-optimizer: adamw_bnb_8bit
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-train_on_inputs: false
-group_by_length: false
-bf16: auto
-fp16:
-tf32: false
-
-gradient_checkpointing: true
-early_stopping_patience:
-resume_from_checkpoint:
-local_rank:
-logging_steps: 1
-xformers_attention:
-flash_attention: true
-
-loss_watchdog_threshold: 5.0
-loss_watchdog_patience: 3
-
-warmup_steps: 10
-evals_per_epoch: 4
-eval_table_size:
-eval_max_new_tokens: 128
-saves_per_epoch: 1
-debug:
-deepspeed:
-weight_decay: 0.0
-fsdp:
-fsdp_config:
-special_tokens:
-  pad_token: "<|end_of_text|>"

requirements-dev.txt

@@ -2,4 +2,3 @@ pre-commit
 black
 mypy
 types-requests
-tbparse

requirements.txt

@@ -1,12 +1,12 @@
 --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
 packaging==23.2
-peft==0.13.2
-transformers==4.46.0
-tokenizers>=0.20.1
-bitsandbytes==0.44.1
-accelerate==1.0.1
-datasets==3.0.1
-deepspeed==0.15.3
+peft==0.13.0
+transformers==4.45.1
+tokenizers>=0.19.1
+bitsandbytes==0.44.0
+accelerate==0.34.2
+datasets==2.21.0
+deepspeed==0.14.4
 pydantic==2.6.3
 addict
 fire
@@ -16,7 +16,7 @@ flash-attn==2.6.3
 sentencepiece
 wandb
 einops
-xformers>=0.0.23.post1
+xformers==0.0.28.post1
 optimum==1.16.2
 hf_transfer
 colorama
@@ -43,7 +43,7 @@ s3fs>=2024.5.0
 gcsfs>=2024.5.0
 # adlfs
 
-trl @ git+https://github.com/huggingface/trl.git@31d02cfb795284591a084416b9dcb7bef5d08924
+trl==0.9.6
 zstandard==0.22.0
 fastcore
 
@@ -52,5 +52,3 @@ lm_eval==0.4.4
 langdetect==1.0.9
 immutabledict==4.2.0
 antlr4-python3-runtime==4.13.2
-
-torchao==0.5.0

requirements_env.txt

@@ -1,315 +0,0 @@
-accelerate==0.34.1
-addict==2.4.0
-aiofiles==23.2.1
-aiohttp==3.9.0
-aiosignal==1.3.1
-aiostream==0.5.2
-alembic==1.13.1
-annotated-types==0.6.0
-annoy==1.17.3
-ansible==6.7.0
-ansible-core==2.13.13
-ansible-vault==2.1.0
-anyio==3.7.1
-appdirs==1.4.4
-art==6.0
-asgiref==3.7.2
-async-timeout==4.0.2
-attrdict==2.0.1
-attrs==22.2.0
-awscli==1.32.75
--e git+ssh://git@github.com/OpenAccess-AI-Collective/axolotl.git@6e354682e3c1735d3f7fb9e362280c38e922260f#egg=axolotl
-backoff==2.2.1
-base58==2.1.1
-beartype==0.17.2
-bitnet==0.2.1
-bitsandbytes==0.42.0
-bittensor==6.7.0
-black==23.7.0
-blinker==1.7.0
-boto3==1.34.75
-botocore==1.34.75
-cachetools==5.3.3
-cachy==0.1.1
-certifi==2023.7.22
-cffi==1.16.0
-cfgv==3.3.1
-chai-guanaco==1.2.4
-charset-normalizer==3.2.0
-cleo==0.6.8
-click==8.1.7
-cloudpickle==2.0.0
-cohere==4.11.2
-colorama==0.4.4
-coloredlogs==15.0.1
-CoLT5-attention==0.10.20
-contextlib2==21.6.0
-contourpy==1.2.0
-cryptography==41.0.3
-cycler==0.12.1
-cytoolz==0.12.3
-databricks-cli==0.18.0
-dataclasses-json==0.5.7
-datasets==2.11.0
-ddt==1.6.0
-decorator==5.1.1
-deepspeed==0.15.0
-# Editable Git install with no remote (dialogpt==0.1)
--e /Users/wing/Projects/ml/dialogpt/src
-dill==0.3.6
-distlib==0.3.6
-docker==7.0.0
-docker-pycreds==0.4.0
-docstring-parser==0.15
-docutils==0.16
-ecdsa==0.18.0
-einops==0.7.0
-einops-exts==0.0.4
-einx==0.1.3
-entrypoints==0.4
-eth-hash==0.6.0
-eth-keys==0.5.0
-eth-typing==4.0.0
-eth-utils==2.3.1
-evaluate==0.4.0
-exceptiongroup==1.1.1
-fastapi==0.109.2
-fastcore==1.5.29
-ffmpy==0.4.0
-filelock==3.12.2
--e git+https://github.com/NousResearch/finetuning-subnet.git@24e9407d6b4430a7ca39d344692f89ce5a97d27e#egg=finetuning_subnet
-fire==0.5.0
-first==2.0.2
-flake8==7.0.0
-Flask==3.0.1
-fonttools==4.47.2
-frozendict==2.4.1
-frozenlist==1.3.3
-fschat @ git+https://github.com/lm-sys/FastChat.git@27a05b04a35510afb1d767ae7e5990cbd278f8fe
-fsspec==2023.6.0
-fuzzywuzzy==0.18.0
-gitdb==4.0.10
-GitPython==3.1.31
-google-pasta==0.2.0
-gradio==4.42.0
-gradio_client==1.3.0
-greenlet==2.0.2
-grpclib==0.4.7
-gunicorn==21.2.0
-h11==0.14.0
-h2==4.1.0
-hpack==4.0.0
-httpcore==0.17.3
-httpx==0.24.1
-huggingface-hub==0.23.4
-humanfriendly==10.0
-hyperframe==6.0.1
-identify==2.5.24
-idna==3.4
-immutables==0.20
-importlib-metadata==6.7.0
-importlib-resources==6.1.1
-inflection==0.5.1
-iniconfig==2.0.0
-itsdangerous==2.1.2
-Jinja2==3.1.2
-jmespath==1.0.1
-joblib==1.3.2
-jsonlines==3.1.0
-jsonschema==2.6.0
-kiwisolver==1.4.5
-langchain==0.0.144
-Levenshtein==0.24.0
-libcst==1.1.0
-liger-kernel==0.0.0
-lion-pytorch==0.1.2
-llama-cpp-python==0.1.36
-llvmlite==0.40.1
-local-attention==1.9.0
-loguru==0.7.0
-Mako==1.3.2
-Markdown==3.5.2
-markdown-it-py==3.0.0
-markdown2==2.4.10
-MarkupSafe==2.1.2
-marshmallow==3.19.0
-marshmallow-enum==1.5.1
-matplotlib==3.8.2
-mccabe==0.7.0
-mdurl==0.1.2
-MEGABYTE-pytorch==0.0.7
--e git+https://github.com/cg123/mergekit.git@53c5f414774a0558b8d84858fb6374bc93a8f1c1#egg=mergekit
-mlflow==2.10.0
-modal==0.62.77
-more-itertools==10.2.0
-mpmath==1.2.1
-msgpack==1.0.7
-msgpack-numpy-opentensor==0.5.0
-multidict==6.0.4
-multiprocess==0.70.14
-munch==2.5.0
-mypy==1.3.0
-mypy-extensions==1.0.0
-nest-asyncio==1.6.0
-netaddr==0.10.1
-networkx==3.0rc1
-nh3==0.2.14
-nodeenv==1.8.0
-nomic==2.0.2
-numba==0.57.1
-numexpr==2.8.4
-numpy==1.24.4
-oauthlib==3.2.2
-openai==0.27.4
-openapi==1.1.0
-openapi-schema-pydantic==1.2.4
-optimum==1.8.6
-orjson==3.10.7
-packaging==23.1
-pandas==2.0.0
-parameterized==0.9.0
-password-strength==0.0.3.post2
-pastel==0.1.1
-pathos==0.3.0
-pathspec==0.11.1
-pathtools==0.1.2
-peft==0.11.1
-pendulum==3.0.0
-Pillow==9.5.0
-pip-tools==1.11.0
-platformdirs==3.2.0
-pluggy==1.4.0
-poetry==0.7.1
-pox==0.3.2
-ppft==1.7.6.6
-pre-commit==3.3.2
-prettytable==3.10.0
-prompt-toolkit==3.0.39
-protobuf==3.20.2
-protobuf3-to-dict==0.1.5
-psutil==5.9.5
-psycopg==3.1.18
-PuLP==2.8.0
-py==1.11.0
-py-bip39-bindings==0.1.11
-py-cpuinfo==9.0.0
-py-ed25519-zebra-bindings==1.0.1
-py-sr25519-bindings==0.2.0
-pyarrow==11.0.0
-pyasn1==0.6.0
-pycodestyle==2.11.1
-pycparser==2.21
-pycryptodome==3.20.0
-pydantic==2.5.3
-pydantic_core==2.14.6
-pydub==0.25.1
-pyfiglet==0.8.post1
-pyflakes==3.2.0
-Pygments==2.15.1
-PyJWT==2.8.0
-pylev==1.4.0
-PyNaCl==1.5.0
-pynvml==11.5.0
-pyparsing==2.4.7
-pyrsistent==0.14.11
-pytest==8.0.2
-pytest-asyncio==0.23.4
-python-dateutil==2.8.2
-python-dotenv==1.0.1
-python-Levenshtein==0.24.0
-python-multipart==0.0.9
-pytz==2023.3
-PyYAML==6.0.1
-querystring-parser==1.2.4
-rapidfuzz==3.6.1
-regex==2023.6.3
-requests==2.31.0
-requests-toolbelt==0.8.0
-resolvelib==0.8.1
-responses==0.18.0
-retry==0.9.2
-rich==13.7.0
-rsa==4.7.2
-ruff==0.6.3
-s3transfer==0.10.1
-safetensors==0.4.5
-sagemaker==2.148.0
-scalecodec==1.2.7
-schedulefree==1.2.1
-schema==0.7.5
-scikit-learn==1.4.0
-scipy==1.9.3
-seaborn==0.13.2
-semantic-version==2.10.0
-sentencepiece==0.2.0
-sentry-sdk==1.19.1
-setproctitle==1.3.2
-shellingham==1.5.4
-shortuuid==1.0.11
-shtab==1.6.5
-sigtools==4.0.1
-six==1.16.0
-skypilot==0.4.1
-smdebug-rulesconfig==1.0.1
-smmap==5.0.0
-sniffio==1.3.0
-SQLAlchemy==1.4.47
-sqlparse==0.4.4
-starlette==0.36.3
-substrate-interface==1.5.2
-svgwrite==1.4.3
-sympy==1.11.1
-synchronicity==0.6.7
-tabulate==0.9.0
-tblib==1.7.0
-tenacity==8.2.2
-tensor-parallel==2.0.0
-termcolor==2.2.0
-text2art==0.2.0
-threadpoolctl==3.2.0
-tiktoken==0.6.0
-time-machine==2.14.1
-timm==0.9.16
-tokenizers==0.19.1
-tokenmonster==1.1.12
-toml==0.9.6
-tomli==2.0.1
-tomlkit==0.12.0
-toolz==0.12.1
-torch==2.2.0
-torchdata==0.6.1
-torchdiffeq==0.2.3
-TorchFix==0.4.0
-torchtext==0.15.2
-torchvision==0.17.0
-tqdm==4.66.2
-transformers==4.44.2
-trl==0.9.6
-typer==0.12.5
-types-certifi==2021.10.8.3
-types-requests==2.31.0.20240125
-types-setuptools==69.0.0.20240125
-types-toml==0.10.8.7
-typing==3.7.4.3
-typing-inspect==0.8.0
-typing_extensions==4.9.0
-tyro==0.5.18
-tzdata==2023.3
-unique-names-generator==1.0.2
-urllib3==2.2.2
-uvicorn==0.22.0
-vector_quantize_pytorch==1.14.1
-virtualenv==20.23.0
-voyager==2.0.2
-wandb==0.16.2
-watchfiles==0.21.0
-wavedrom==2.0.3.post3
-wcwidth==0.2.6
-websocket-client==1.7.0
-websockets==12.0
-Werkzeug==3.0.1
-wonderwords==2.2.0
-xxhash==3.2.0
-yarl==1.8.2
-zetascale==2.2.7
-zipp==3.15.0

scripts/chat_datasets.py

@@ -1,60 +0,0 @@
-"""
-helper script to parse chat datasets into a usable yaml
-"""
-import click
-import yaml
-from datasets import load_dataset
-
-
-@click.command()
-@click.argument("dataset", type=str)
-@click.option("--split", type=str, default="train")
-def parse_dataset(dataset=None, split="train"):
-    ds_cfg = {}
-    ds_cfg["path"] = dataset
-    ds_cfg["split"] = split
-    ds_cfg["type"] = "chat_template"
-    ds_cfg["chat_template"] = "<<<Replace based on your model>>>"
-
-    dataset = load_dataset(dataset, split=split)
-    features = dataset.features
-    feature_keys = features.keys()
-    field_messages = None
-    for key in ["conversation", "conversations", "messages"]:
-        if key in feature_keys:
-            field_messages = key
-            break
-    if not field_messages:
-        raise ValueError(
-            f'No conversation field found in dataset: {", ".join(feature_keys)}'
-        )
-    ds_cfg["field_messages"] = field_messages
-
-    message_fields = features["conversations"][0].keys()
-    message_field_role = None
-    for key in ["from", "role"]:
-        if key in message_fields:
-            message_field_role = key
-            break
-    if not message_field_role:
-        raise ValueError(
-            f'No role field found in messages: {", ".join(message_fields)}'
-        )
-    ds_cfg["message_field_role"] = message_field_role
-
-    message_field_content = None
-    for key in ["content", "text", "value"]:
-        if key in message_fields:
-            message_field_content = key
-            break
-    if not message_field_content:
-        raise ValueError(
-            f'No content field found in messages: {", ".join(message_fields)}'
-        )
-    ds_cfg["message_field_content"] = message_field_content
-
-    print(yaml.dump({"datasets": [ds_cfg]}))
-
-
-if __name__ == "__main__":
-    parse_dataset()

setup.py

@@ -30,9 +30,6 @@ def parse_requirements():
 
     try:
         xformers_version = [req for req in _install_requires if "xformers" in req][0]
-        torchao_version = [req for req in _install_requires if "torchao" in req][0]
-        autoawq_version = [req for req in _install_requires if "autoawq" in req][0]
-
         if "Darwin" in platform.system():
             # don't install xformers on MacOS
             _install_requires.pop(_install_requires.index(xformers_version))
@@ -52,18 +49,11 @@ def parse_requirements():
             else:
                 raise ValueError("Invalid version format")
 
-            if (major, minor) >= (2, 5):
-                _install_requires.pop(_install_requires.index(xformers_version))
-                _install_requires.pop(_install_requires.index(autoawq_version))
-            elif (major, minor) >= (2, 4):
+            if (major, minor) >= (2, 4):
                 if patch == 0:
                     _install_requires.pop(_install_requires.index(xformers_version))
                     _install_requires.append("xformers>=0.0.27")
-                else:
-                    _install_requires.pop(_install_requires.index(xformers_version))
-                    _install_requires.append("xformers==0.0.28.post1")
-            elif (major, minor) >= (2, 3):
-                _install_requires.pop(_install_requires.index(torchao_version))
+            if (major, minor) >= (2, 3):
                 if patch == 0:
                     _install_requires.pop(_install_requires.index(xformers_version))
                     _install_requires.append("xformers>=0.0.26.post1")
@@ -71,16 +61,15 @@ def parse_requirements():
                     _install_requires.pop(_install_requires.index(xformers_version))
                     _install_requires.append("xformers>=0.0.27")
             elif (major, minor) >= (2, 2):
-                _install_requires.pop(_install_requires.index(torchao_version))
                 _install_requires.pop(_install_requires.index(xformers_version))
                 _install_requires.append("xformers>=0.0.25.post1")
             else:
-                _install_requires.pop(_install_requires.index(torchao_version))
                 _install_requires.pop(_install_requires.index(xformers_version))
                 _install_requires.append("xformers>=0.0.23.post1")
 
     except PackageNotFoundError:
         pass
+
     return _install_requires, _dependency_links
 
 
@@ -109,7 +98,6 @@ setup(
         ],
         "mamba-ssm": [
             "mamba-ssm==1.2.0.post1",
-            "causal_conv1d",
         ],
         "auto-gptq": [
             "auto-gptq==0.5.1",

src/axolotl/cli/__init__.py

@@ -30,7 +30,7 @@ from axolotl.common.cli import TrainerCliArgs, load_model_and_tokenizer
 from axolotl.integrations.base import PluginManager
 from axolotl.logging_config import configure_logging
 from axolotl.train import TrainDatasetMeta
-from axolotl.utils.chat_templates import get_chat_template
+from axolotl.utils.chat_templates import chat_templates
 from axolotl.utils.comet_ import setup_comet_env_vars
 from axolotl.utils.config import (
     normalize_cfg_datasets,
@@ -272,7 +272,7 @@ def do_inference_gradio(
             importlib.import_module("axolotl.prompters"), prompter
         )
     elif cfg.chat_template:
-        chat_template_str = get_chat_template(cfg.chat_template)
+        chat_template_str = chat_templates(cfg.chat_template)
 
     model = model.to(cfg.device, dtype=cfg.torch_dtype)
 
@@ -462,12 +462,7 @@ def load_datasets(
         processor=processor,
     )
 
-    if (
-        cli_args.debug
-        or cfg.debug
-        or cli_args.debug_text_only
-        or int(cli_args.debug_num_examples) > 0
-    ):
+    if cli_args.debug or cfg.debug:
         LOG.info("check_dataset_labels...")
         check_dataset_labels(
             train_dataset.select(

src/axolotl/cli/preprocess.py

@@ -27,7 +27,6 @@ from axolotl.prompt_strategies.sharegpt import (
     register_chatml_template,
     register_llama3_template,
 )
-from axolotl.utils.trainer import disable_datasets_caching
 
 LOG = logging.getLogger("axolotl.cli.preprocess")
 
@@ -71,11 +70,10 @@ def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs):
         LOG.warning(msg)
         parsed_cfg.dataset_prepared_path = DEFAULT_DATASET_PREPARED_PATH
 
-    with disable_datasets_caching():
-        if parsed_cfg.rl:  # and parsed_cfg.rl != "orpo":
-            load_rl_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)
-        else:
-            load_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)
+    if parsed_cfg.rl:  # and parsed_cfg.rl != "orpo":
+        load_rl_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)
+    else:
+        load_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)
 
     if parsed_cli_args.download:
         model_name = parsed_cfg.base_model

src/axolotl/common/cli.py

@@ -23,7 +23,7 @@ class TrainerCliArgs:
 
     debug: bool = field(default=False)
     debug_text_only: bool = field(default=False)
-    debug_num_examples: int = field(default=0)
+    debug_num_examples: int = field(default=5)
     inference: bool = field(default=False)
     merge_lora: bool = field(default=False)
     prompter: Optional[str] = field(default=None)

src/axolotl/core/chat/format/chatml.py

@@ -1,34 +0,0 @@
-"""
-ChatML transformation functions for MessageContents
-"""
-from typing import Optional
-
-from ..messages import MessageContents, Messages
-from .shared import wrap_tools
-
-
-def format_message(
-    message: Messages,
-    message_index: Optional[int] = None,  # pylint: disable=unused-argument
-) -> Messages:
-    if message.is_chat_formatted:
-        return message
-
-    # prepend the role prefix within a MessageContents to message.content
-    message.content.insert(
-        0,
-        MessageContents(
-            type="text",
-            value=f"<|im_start|>{message.role}\n",
-            weight=0,
-        ),
-    )
-    message.content.append(
-        MessageContents(type="text", value="<|im_end|>", weight=message.weight)
-    )
-    message.content.append(MessageContents(type="text", value="\n", weight=0))
-
-    message = wrap_tools(message)
-
-    message.is_chat_formatted = True
-    return message

src/axolotl/core/chat/format/llama3x.py

@@ -1,45 +0,0 @@
-"""
-Llama 3.x chat formatting functions for MessageContents
-"""
-from typing import Optional
-
-from ..messages import MessageContents, Messages
-from .shared import wrap_tools
-
-
-def format_message(message: Messages, message_index: Optional[int] = None) -> Messages:
-    if message.is_chat_formatted:
-        return message
-
-    message_role = message.role
-    if message.role == "tool":
-        message_role = "ipython"
-
-    # prepend the role prefix within a MessageContents to message.content
-    message.content.insert(
-        0,
-        MessageContents(
-            type="text",
-            value=f"<|start_header_id|>{message_role}<|end_header_id|>\n\n",
-            weight=0,
-        ),
-    )
-
-    message.content.append(
-        MessageContents(type="text", value="<|eot_id|>", weight=message.weight)
-    )
-
-    message = wrap_tools(message)
-
-    if message_index == 0:
-        message.content.insert(
-            0,
-            MessageContents(
-                type="text",
-                value="<|begin_of_text|>",
-                weight=0,
-            ),
-        )
-
-    message.is_chat_formatted = True
-    return message

src/axolotl/core/chat/format/shared.py

@@ -1,47 +0,0 @@
-"""
-shared functions for format transforms
-"""
-from axolotl.core.chat.messages import MessageContents, Messages
-
-
-def wrap_tools(message: Messages):
-    # loop over message.content by index to find tool calls, we need to wrap each with tags,
-    # so be wary of indexing issues when changing the list while iterating.
-    # iterate over the range in reverse order to avoid index shifting
-    for i in range(len(message.content) - 1, -1, -1):
-        if message.content[i].type == "tool_call":
-            # append a </tool_call> MessageContents text tag after
-            message.content.insert(
-                i + 1,
-                MessageContents(
-                    type="text", value="</tool_call>\n", weight=message.weight
-                ),
-            )
-            # make sure the actual tool call content ends with a newline
-            message.content[i].has_newline = True
-            # prepend a <tool_call> MessageContents text tag before
-            message.content.insert(
-                i,
-                MessageContents(
-                    type="text", value="<tool_call>\n", weight=message.weight
-                ),
-            )
-        elif message.content[i].type == "tool_response":
-            # append a </tool_call> MessageContents text tag after
-            message.content.insert(
-                i + 1,
-                MessageContents(
-                    type="text", value="</tool_response>\n", weight=message.weight
-                ),
-            )
-            # make sure the actual tool response content ends with a newline
-            message.content[i].has_newline = True
-            # prepend a <tool_call> MessageContents text tag before
-            message.content.insert(
-                i,
-                MessageContents(
-                    type="text", value="<tool_response>\n", weight=message.weight
-                ),
-            )
-
-    return message

src/axolotl/core/chat/messages.py

@@ -1,230 +0,0 @@
-"""
-internal message representations of chat messages
-"""
-import json
-from enum import Enum
-from typing import Any, Callable, List, Optional, Union
-
-from pydantic import BaseModel
-from transformers import PreTrainedTokenizer
-
-
-class MessageRoles(str, Enum):
-    """
-    Message roles for the system, user, assistant, and tools
-    """
-
-    system = "system"  # pylint: disable=invalid-name
-    user = "user"  # pylint: disable=invalid-name
-    assistant = "assistant"  # pylint: disable=invalid-name
-    tool = "tool"  # pylint: disable=invalid-name
-    ipython = (  # pylint: disable=invalid-name
-        # for responses from builtin tools
-        "ipython"
-    )
-
-
-class MessageContentTypes(str, Enum):
-    """
-    Message content types for text, image, audio, tool calls, and tool responses
-    """
-
-    special_token = "special_token"  # pylint: disable=invalid-name  # nosec B105
-    text = "text"  # pylint: disable=invalid-name
-    image = "image"  # pylint: disable=invalid-name
-    audio = "audio"  # pylint: disable=invalid-name
-    tool_call = "tool_call"  # pylint: disable=invalid-name  # to differentiate regular responses from tool calls from the assistant
-    tool_response = "tool_response"  # pylint: disable=invalid-name
-
-
-class SpecialToken(str, Enum):
-    """
-    Special tokens for beginning of string and end of string
-    """
-
-    bos_token = "bos_token"  # pylint: disable=invalid-name  # nosec B105
-    eos_token = "eos_token"  # pylint: disable=invalid-name  # nosec B105
-
-
-class ToolCallFunction(BaseModel):
-    """
-    Tool call function with name and arguments
-    """
-
-    name: str
-    arguments: dict[str, str]
-
-
-class Tool(BaseModel):
-    """
-    Tool with description, function, and parameters
-    """
-
-    description: str
-    function: ToolCallFunction
-    parameters: dict[str, str]  # .properties
-
-
-class ToolCallContents(BaseModel):
-    """
-    Tool call contents with name, arguments, and optional id
-    """
-
-    name: str
-    arguments: dict[str, Union[str, int]]
-    id: Optional[str] = None  # pylint: disable=invalid-name
-
-    def __str__(self) -> str:
-        data = {"name": self.name, "arguments": self.arguments}
-        if self.id is not None:
-            data["id"] = self.id
-        return json.dumps(data)
-
-
-class ToolResponseContents(BaseModel):
-    """
-    Tool response contents with name, content, and optional id
-    """
-
-    name: str
-    content: Union[str, dict[str, Union[str, int, float]]]
-    id: Optional[str] = None  # pylint: disable=invalid-name
-
-    def __str__(self) -> str:
-        data = {"name": self.name, "content": self.content}
-        if self.id is not None:
-            data["id"] = self.id
-        return json.dumps(data)
-
-
-class MessageContents(BaseModel):
-    """
-    Message contents with type, value, metadata, weight, newline, and end of contents
-    """
-
-    type: Union[str, MessageContentTypes]
-    value: Union[str, ToolCallContents, ToolResponseContents, SpecialToken]
-    meta: Optional[dict[str, Any]] = None  # support additional arbitrary metadata
-    weight: Optional[Union[int, float]] = None
-    has_newline: bool = False
-    eoc: bool = False  # end of contents
-
-    def __str__(self) -> str:
-        str_val = str(self.value)
-        if self.has_newline and not str_val.endswith("\n"):
-            str_val += "\n"
-        return str_val
-
-
-class Messages(BaseModel):
-    """
-    Messages with role, content, metadata, weight, and chat formatting
-    """
-
-    role: Union[MessageRoles, str]  # allows for arbitrary roles
-    content: List["MessageContents"]
-    meta: Optional[dict[str, Any]] = None  # support additional arbitrary metadata
-    weight: Optional[Union[int, float]] = None
-    is_chat_formatted: bool = False
-
-    def __str__(self) -> str:
-        return "".join(str(c) for c in self.content)
-
-    def tokenized(
-        self, tokenizer: PreTrainedTokenizer, ignore_index=-100
-    ) -> dict[str, List[int]]:
-        # iterate over the contents, tokenizing the concatenated string values up to the current MessageContents
-        # returns a dictionary mapping w input_ids, attention_mask, and labels
-        input_ids: List[int] = []
-        labels: List[int] = []
-        pending_input_ids: List[int] = []
-        pending_weight = self.weight
-        running_content = ""
-        for _, msg_content in enumerate(self.content):
-            # TODO also handle non-text content types
-            if msg_content.type in [
-                MessageContentTypes.text.value,
-                MessageContentTypes.tool_call.value,
-                MessageContentTypes.tool_response.value,
-            ]:
-                running_content += str(msg_content)
-                tok_results = tokenizer(running_content, add_special_tokens=False)
-                tok_input_ids = tok_results["input_ids"]
-                if pending_input_ids:
-                    new_pending_inputs = tok_input_ids[
-                        len(input_ids) : len(input_ids) + len(pending_input_ids)
-                    ]
-                    if new_pending_inputs != pending_input_ids:
-                        # logging.warning("tokenization mismatch from concatenation.")
-                        pending_input_ids = new_pending_inputs
-                    input_ids.extend(pending_input_ids)
-                    if pending_weight:
-                        labels.extend(pending_input_ids)
-                    else:
-                        labels.extend([ignore_index] * len(pending_input_ids))
-                pending_input_ids = tok_results["input_ids"][len(input_ids) :]
-                pending_weight = self.weight and msg_content.weight not in [0, 0.0]
-        input_ids.extend(pending_input_ids)
-        if pending_weight:
-            labels.extend(pending_input_ids)
-        else:
-            labels.extend([ignore_index] * len(pending_input_ids))
-        attention_mask = [1] * len(input_ids)
-        return {
-            "input_ids": input_ids,
-            "attention_mask": attention_mask,
-            "labels": labels,
-        }
-
-
-class Chats(BaseModel):
-    """
-    top level data structure for chat conversations
-    """
-
-    conversation: List[Messages]
-
-    def __str__(self) -> str:
-        return "".join(str(c) for c in self.conversation)
-
-    def tokenized(
-        self, tokenizer: Callable[[str], dict[str, List[int]]], ignore_index=-100
-    ) -> dict[str, List[int]]:
-        input_ids = []
-        attention_mask = []
-        labels = []
-        for msg in self.conversation:
-            msg_results = msg.tokenized(tokenizer, ignore_index)
-            input_ids.extend(msg_results["input_ids"])
-            attention_mask.extend(msg_results["attention_mask"])
-            labels.extend(msg_results["labels"])
-        return {
-            "input_ids": input_ids,
-            "attention_mask": attention_mask,
-            "labels": labels,
-        }
-
-
-class ChatFormattedChats(Chats):
-    """
-    Chat formatted chats with formatter and optional train on inputs
-    """
-
-    formatter: Callable  # [[Union[dict, Chats]], Chats]
-    train_on_inputs: bool = False
-
-    def model_post_init(self, __context):
-        for i, msg in enumerate(self.conversation):
-            self.conversation[i] = self.formatter(msg, message_index=i)
-            if self.train_on_inputs:
-                self.conversation[i].weight = 1
-
-
-class PreferenceChats(BaseModel):
-    """
-    representation for preference data for chat
-    """
-
-    prompt: List[Messages]
-    chosen: Messages
-    rejected: Messages

src/axolotl/core/datasets/chat.py

@@ -1,55 +0,0 @@
-"""
-chat dataset module
-"""
-import os
-from typing import Callable, Optional, Union
-
-from datasets import Dataset
-from transformers import PreTrainedTokenizer
-
-from axolotl.core.chat.messages import ChatFormattedChats
-
-
-class TokenizedChatDataset(Dataset):
-    """
-    Tokenized chat dataset
-    """
-
-    def __init__(
-        self,
-        data: Dataset,
-        model_transform: Union[PreTrainedTokenizer, Callable],
-        *args,
-        message_transform: Optional[Callable] = None,
-        formatter=None,
-        process_count: Optional[int] = None,
-        keep_in_memory: Optional[bool] = False,
-        **kwargs,
-    ):
-        def map_fn(ex):
-            if message_transform is not None:
-                ex = message_transform(ex)
-            if formatter is not None:
-                ex = ChatFormattedChats(
-                    formatter=formatter,
-                    **ex,
-                )
-            else:
-                ex = ChatFormattedChats(
-                    **ex,
-                )
-            return ex.tokenized(model_transform)
-
-        process_or_cpu_count: int = (
-            process_count or os.cpu_count()  # type: ignore[assignment]
-        )
-        num_proc = min(64, process_or_cpu_count)
-        features = data.features.keys()
-        tokenized_data = data.map(
-            map_fn,
-            num_proc=num_proc,
-            keep_in_memory=keep_in_memory,
-            remove_columns=features,
-            desc="Tokenizing Chats",
-        )
-        super().__init__(tokenized_data.data, *args, **kwargs)

src/axolotl/core/datasets/transforms/chat_builder.py

@@ -1,150 +0,0 @@
-"""
-This module contains a function that builds a transform that takes a row from the dataset and converts it to a Chat.
-"""
-from typing import Any, Mapping, Union
-
-
-def chat_message_transform_builder(  # pylint: disable=dangerous-default-value
-    train_on_inputs=False,
-    conversations_field: str = "conversations",
-    message_field_role: Union[str, list[str]] = ["role", "from"],  # commonly "role"
-    message_field_content: Union[str, list[str]] = [
-        "value",
-        "text",
-        "content",
-    ],  # commonly "content"
-    message_field_training: Union[str, list[str]] = [
-        "train",
-        "weight",
-    ],  # commonly "weight"
-):
-    """Builds a transform that takes a row from the dataset and converts it to a Chat
-
-    Args:
-        train_on_inputs (bool, optional):
-            If True, the transform will train on the inputs. If False, the transform will train on the targets.
-            Defaults to False.
-        conversations_field (str, optional):
-            The field name of the conversations. Defaults to "conversations".
-        message_field_role (str | list[str], optional):
-            The field name of the role. Defaults to "role".
-        message_field_content (str | list[str], optional):
-            The field name of the message content. Defaults to "content".
-        message_field_training (str | list[str], optional):
-            The field name of the train/weight. Defaults to "weight".
-
-    Returns:
-        Callable:
-            A function that takes a list of conversations and returns a list of messages.
-    """
-
-    message_field_role = (
-        [message_field_role]
-        if isinstance(message_field_role, str)
-        else message_field_role
-    )
-    message_field_content = (
-        [message_field_content]
-        if isinstance(message_field_content, str)
-        else message_field_content
-    )
-    message_weight_fields = (
-        [message_field_training]
-        if isinstance(message_field_training, str)
-        else message_field_training
-    )
-
-    role_value_mappings = {
-        "system": "system",
-        "user": "user",
-        "human": "user",
-        "assistant": "assistant",
-        "gpt": "assistant",
-        "tool": "tool",
-        "ipython": "ipython",
-    }
-    if train_on_inputs:
-        role_default_weights_mappings = {
-            "system": 1,
-            "user": 1,
-            "assistant": 1,
-            "tool": 1,
-            "ipython": 1,
-        }
-    else:
-        role_default_weights_mappings = {
-            "system": 0,
-            "user": 0,
-            "assistant": 1,
-            "tool": 0,
-            "ipython": 0,
-        }
-
-    def transform_builder(sample: Mapping[str, Any]):
-        if conversations_field not in sample:
-            raise ValueError(f"Field '{conversations_field}' not found in sample.")
-        # if none of the role fields are in the message, raise an error
-        if not any(
-            role in sample[conversations_field][0] for role in message_field_role
-        ):
-            raise ValueError("No role field found in message.")
-        role_field = next(
-            role
-            for role in message_field_role
-            if role in sample[conversations_field][0]
-        )
-        if not any(
-            field in sample[conversations_field][0] for field in message_field_content
-        ):
-            raise ValueError("No message_content field found in message.")
-        message_content_field = next(
-            field
-            for field in message_field_content
-            if field in sample[conversations_field][0]
-        )
-        if not any(
-            field in sample[conversations_field][0] for field in message_field_training
-        ):
-            message_weight_field = None
-        else:
-            message_weight_field = next(
-                field
-                for field in message_weight_fields
-                if field in sample[conversations_field][0]
-            )
-
-        messages = []
-        for message in sample[conversations_field]:
-            role = role_value_mappings[message[role_field]]
-            weight = (
-                int(message[message_weight_field])
-                if message_weight_field
-                else role_default_weights_mappings[role]
-            )
-
-            # TODO if "tool_calls" in message[message_content_field]: then convert tool call to ToolCallContents
-            if isinstance(message[message_content_field], str):
-                messages.append(
-                    {
-                        "role": role,
-                        "content": [
-                            {
-                                "type": "text",
-                                "value": message[message_content_field],
-                            }
-                        ],
-                        "weight": weight,
-                    }
-                )
-            else:
-                messages.append(
-                    {
-                        "role": role,
-                        "content": message[message_content_field],
-                        "weight": weight,
-                    }
-                )
-
-        return {"conversation": messages}
-
-    return transform_builder

src/axolotl/core/trainer_builder.py

@@ -7,7 +7,6 @@ import abc
 import gc
 import importlib
 import importlib.util
-import inspect
 import logging
 import math
 import os
@@ -28,6 +27,7 @@ from torch.optim.lr_scheduler import OneCycleLR
 from torch.utils.data import BatchSampler, DataLoader, RandomSampler, SequentialSampler
 from transformers import (
     EarlyStoppingCallback,
+    PreTrainedModel,
     Trainer,
     TrainerCallback,
     TrainingArguments,
@@ -43,10 +43,8 @@ from trl import (
     KTOTrainer,
     ORPOConfig,
     ORPOTrainer,
-    RewardConfig,
-    RewardTrainer,
 )
-from trl.trainer.utils import RewardDataCollatorWithPadding, pad_to_length
+from trl.trainer.utils import pad_to_length
 
 from axolotl.monkeypatch.multipack import SUPPORTED_MULTIPACK_MODEL_TYPES
 from axolotl.monkeypatch.relora import ReLoRACallback, ReLoRAScheduler
@@ -63,7 +61,7 @@ from axolotl.utils.callbacks import (
     log_prediction_callback_factory,
 )
 from axolotl.utils.callbacks.lisa import lisa_callback_factory
-from axolotl.utils.chat_templates import get_chat_template
+from axolotl.utils.chat_templates import chat_templates
 from axolotl.utils.collators import (
     BatchSamplerDataCollatorForSeq2Seq,
     DataCollatorForSeq2Seq,
@@ -303,13 +301,6 @@ class AxolotlCPOConfig(AxolotlTrainingMixins, CPOConfig):
     )
 
 
-@dataclass
-class AxolotlRewardConfig(AxolotlTrainingMixins, RewardConfig):
-    """
-    Reward config for Reward training
-    """
-
-
 class SchedulerMixin(Trainer):
     """
     Mixin class for scheduler setup in CausalTrainer.
@@ -407,10 +398,12 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
     def __init__(
         self,
         *_args,
+        num_epochs=1,
         bench_data_collator=None,
         eval_data_collator=None,
         **kwargs,
     ):
+        self.num_epochs = num_epochs
         self.bench_data_collator = bench_data_collator
         self.eval_data_collator = eval_data_collator
         super().__init__(*_args, **kwargs)
@@ -666,9 +659,7 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
         return DataLoader(bench_dataset, **dataloader_params)
         # return self.accelerator.prepare(DataLoader(bench_dataset, **dataloader_params))
 
-    def compute_loss(
-        self, model, inputs, return_outputs=False, num_items_in_batch=None
-    ):
+    def compute_loss(self, model, inputs, return_outputs=False):
         # use one's weighted cross entropy loss calc
         # if self.args.sample_packing:
         #     labels = inputs.pop("labels")
@@ -676,18 +667,8 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
         #     loss = trainer_weighted_loss(outputs, labels, shift_labels=True)
         #     return (loss, outputs) if return_outputs else loss
         if self.args.orpo_alpha:
-            return self.orpo_compute_loss(
-                model,
-                inputs,
-                return_outputs=return_outputs,
-                num_items_in_batch=num_items_in_batch,
-            )
-        return super().compute_loss(
-            model,
-            inputs,
-            return_outputs=return_outputs,
-            num_items_in_batch=num_items_in_batch,
-        )
+            return self.orpo_compute_loss(model, inputs, return_outputs=return_outputs)
+        return super().compute_loss(model, inputs, return_outputs=return_outputs)
 
     @staticmethod
     def orpo_concatenate_inputs(inputs, label_pad_token=-100, pad_token=0, device=None):
@@ -783,13 +764,7 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
         ).squeeze(2)
         return torch.mul(per_token_logps, mask).sum(dim=1) / mask.sum(dim=1)
 
-    def orpo_compute_loss(
-        self,
-        model,
-        inputs,
-        return_outputs=False,
-        num_items_in_batch=None,  # pylint: disable=unused-argument
-    ):
+    def orpo_compute_loss(self, model, inputs, return_outputs=False):
         concat_inputs = AxolotlTrainer.orpo_concatenate_inputs(
             inputs,
             label_pad_token=-100,
@@ -895,13 +870,13 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
         for key, value in metrics.items():
             self._stored_metrics[train_eval][key].append(value)
 
-    def _save_checkpoint(self, model, trial):
+    def _save_checkpoint(self, model, trial, metrics=None):
         # make sure the checkpoint dir exists, since trainer is flakey
         checkpoint_folder = f"{PREFIX_CHECKPOINT_DIR}-{self.state.global_step}"
         run_dir = self._get_output_dir(trial=trial)
         output_dir = os.path.join(run_dir, checkpoint_folder)
         os.makedirs(output_dir, exist_ok=True)
-        return super()._save_checkpoint(model, trial)
+        return super()._save_checkpoint(model, trial, metrics=metrics)
 
 
 class AxolotlMambaTrainer(AxolotlTrainer):
@@ -916,7 +891,6 @@ class AxolotlMambaTrainer(AxolotlTrainer):
         model,
         inputs,
         return_outputs=False,  # pylint: disable=unused-argument
-        num_items_in_batch=None,  # pylint: disable=unused-argument
     ):
         input_ids = inputs.pop("input_ids")
         lm_logits = model(input_ids).logits
@@ -1024,32 +998,18 @@ class AxolotlDPOTrainer(SchedulerMixin, DPOTrainer):
         return super().push_to_hub(*args, **kwargs)
 
     def tokenize_row(
-        self,
-        features,
-        processing_class,
-        max_prompt_length,
-        max_completion_length,
-        add_special_tokens,
+        self, feature, model: Optional[Union[PreTrainedModel, torch.nn.Module]] = None
     ) -> Dict:
-        res = super().tokenize_row(
-            features,
-            processing_class,
-            max_prompt_length,
-            max_completion_length,
-            add_special_tokens,
-        )
-        if processing_class.bos_token_id is None and res["prompt_input_ids"][0] is None:
+        res = super().tokenize_row(feature, model=model)
+        if self.tokenizer.bos_token_id is None and res["prompt_input_ids"][0] is None:
             for key in res.keys():
                 res[key] = res[key][1:]
         return res
 
     def training_step(
-        self,
-        model: nn.Module,
-        inputs: Dict[str, Union[torch.Tensor, Any]],
-        num_items_in_batch=None,
+        self, model: nn.Module, inputs: Dict[str, Union[torch.Tensor, Any]]
     ) -> torch.Tensor:
-        loss: torch.Tensor = super().training_step(model, inputs, num_items_in_batch)
+        loss: torch.Tensor = super().training_step(model, inputs)
         gc.collect()
         torch.cuda.empty_cache()
         return loss
@@ -1079,14 +1039,6 @@ class AxolotlCPOTrainer(SchedulerMixin, CPOTrainer):
     tag_names = ["axolotl", "cpo"]
 
 
-class AxolotlRewardTrainer(SchedulerMixin, RewardTrainer):
-    """
-    Extend the base RewardTrainer for axolotl helpers
-    """
-
-    tag_names = ["axolotl", "reward"]
-
-
 class TrainerBuilderBase(abc.ABC):
     """
     Base class for trainer builder
@@ -1152,17 +1104,12 @@ class TrainerBuilderBase(abc.ABC):
                 SaveAxolotlConfigtoWandBCallback(self.cfg.axolotl_config_path)
             )
         if self.cfg.use_mlflow and is_mlflow_available():
-            from transformers.integrations.integration_utils import MLflowCallback
-
             from axolotl.utils.callbacks.mlflow_ import (
                 SaveAxolotlConfigtoMlflowCallback,
             )
 
-            callbacks.extend(
-                [
-                    SaveAxolotlConfigtoMlflowCallback(self.cfg.axolotl_config_path),
-                    MLflowCallback,
-                ]
+            callbacks.append(
+                SaveAxolotlConfigtoMlflowCallback(self.cfg.axolotl_config_path)
             )
         if self.cfg.use_comet and is_comet_available():
             from axolotl.utils.callbacks.comet_ import SaveAxolotlConfigtoCometCallback
@@ -1267,8 +1214,6 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             return ReLoRATrainer
         if self.cfg.model_config_type == "mamba":
             return AxolotlMambaTrainer
-        if self.cfg.reward_model:
-            return AxolotlRewardTrainer
         return AxolotlTrainer
 
     def build(self, total_num_steps):
@@ -1500,12 +1445,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             report_to.append("comet_ml")
 
         training_arguments_kwargs["report_to"] = report_to
-        if self.cfg.use_wandb:
-            training_arguments_kwargs["run_name"] = self.cfg.wandb_name
-        elif self.cfg.use_mlflow:
-            training_arguments_kwargs["run_name"] = self.cfg.mlflow_run_name
-        else:
-            training_arguments_kwargs["run_name"] = None
+        training_arguments_kwargs["run_name"] = (
+            self.cfg.wandb_name if self.cfg.use_wandb else None
+        )
         training_arguments_kwargs["optim"] = (
             self.cfg.optimizer if self.cfg.optimizer else "adamw_hf"
         )
@@ -1594,7 +1536,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         training_arguments_kwargs["model_type"] = self.cfg.model_config_type
         training_arguments_kwargs["pretraining"] = bool(self.cfg.pretraining_dataset)
         if self.cfg.chat_template:
-            training_arguments_kwargs["chat_template"] = get_chat_template(
+            training_arguments_kwargs["chat_template"] = chat_templates(
                 self.cfg.chat_template
             )
 
@@ -1608,9 +1550,6 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
 
         trainer_kwargs = {}
 
-        if self.cfg.reward_model:
-            trainer_kwargs["max_length"] = self.cfg.sequence_len
-
         if self.cfg.optimizer in [
             "optimi_adamw",
             "ao_adamw_4bit",
@@ -1654,13 +1593,10 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 "accelerator_config"
             ] = self.cfg.accelerator_config
 
-        training_args_cls = (
-            AxolotlTrainingArguments
-            if not self.cfg.reward_model
-            else AxolotlRewardConfig
-        )
-        training_args = training_args_cls(  # pylint: disable=unexpected-keyword-arg
-            **training_arguments_kwargs,
+        training_args = (
+            AxolotlTrainingArguments(  # pylint: disable=unexpected-keyword-arg
+                **training_arguments_kwargs,
+            )
         )
         training_args = self.hook_post_create_training_args(training_args)
 
@@ -1682,37 +1618,27 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             # https://docs.nvidia.com/deeplearning/performance/dl-performance-matrix-multiplication/index.html
             data_collator_kwargs["pad_to_multiple_of"] = 64
 
-        if self.cfg.reward_model:
-            data_collator_kwargs["max_length"] = self.cfg.sequence_len
-
         trainer_cls = self._get_trainer_cls()
         trainer_kwargs, trainer_cls = self.hook_pre_create_trainer(
             trainer_kwargs, trainer_cls
         )
-        if eval_data_collator := self.build_collator(
-            training_args, is_eval=True, **data_collator_kwargs
-        ):
-            if not self.cfg.reward_model:
-                trainer_kwargs["eval_data_collator"] = eval_data_collator
-        if not self.cfg.reward_model:
-            trainer_kwargs["bench_data_collator"] = transformers.DataCollatorForSeq2Seq(
-                self.tokenizer,
-                return_tensors="pt",
-                **data_collator_kwargs,
-            )
-        sig = inspect.signature(trainer_cls)
-        if "processing_class" in sig.parameters.keys():
-            trainer_kwargs["processing_class"] = self.tokenizer
-        else:
-            trainer_kwargs["tokenizer"] = self.tokenizer
-
         trainer = trainer_cls(
             model=self.model,
             train_dataset=self.train_dataset,
             eval_dataset=self.eval_dataset,
             args=training_args,
+            tokenizer=self.tokenizer,
             data_collator=self.build_collator(training_args, **data_collator_kwargs),
+            eval_data_collator=self.build_collator(
+                training_args, is_eval=True, **data_collator_kwargs
+            ),
+            bench_data_collator=transformers.DataCollatorForSeq2Seq(
+                self.tokenizer,
+                return_tensors="pt",
+                **data_collator_kwargs,
+            ),
             callbacks=self.get_callbacks(),
+            num_epochs=self.cfg.num_epochs,
             **trainer_kwargs,
         )
         trainer = self.hook_post_create_trainer(trainer)
@@ -1746,14 +1672,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 V2BatchSamplerDataCollatorForSeq2Seq,
                 BatchSamplerDataCollatorForSeq2Seq,
                 DataCollatorForSeq2Seq,
-                RewardDataCollatorWithPadding,
             ]
         ]
-        if self.cfg.reward_model:
-            collator = RewardDataCollatorWithPadding
-            if "max_length" in kwargs:
-                kwargs.pop("max_length")
-        elif use_batch_sampler_collator:
+        if use_batch_sampler_collator:
             if self.cfg.model_config_type in SUPPORTED_MULTIPACK_MODEL_TYPES:
                 collator = V2BatchSamplerDataCollatorForSeq2Seq
             elif (
@@ -1955,7 +1876,7 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             dpo_trainer_kwargs["max_length"] = self.cfg.sequence_len
             dpo_trainer_kwargs["max_target_length"] = None
             dpo_trainer_kwargs["max_prompt_length"] = self.cfg.sequence_len
-            dpo_trainer_kwargs["generate_during_eval"] = self.cfg.use_wandb
+            dpo_trainer_kwargs["generate_during_eval"] = True
         elif self.cfg.rl == "orpo":
             trainer_cls = AxolotlORPOTrainer
             trainer_cls_args = [self.model]
@@ -1967,17 +1888,11 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             trainer_cls_args = [self.model]
         else:
             raise ValueError(f"Unsupported RL: {self.cfg.rl}")
-
-        sig = inspect.signature(trainer_cls)
-        if "processing_class" in sig.parameters.keys():
-            dpo_trainer_kwargs["processing_class"] = self.tokenizer
-        else:
-            dpo_trainer_kwargs["tokenizer"] = self.tokenizer
-
         dpo_trainer = trainer_cls(
             *trainer_cls_args,
             args=training_args,
             train_dataset=self.train_dataset,
+            tokenizer=self.tokenizer,
             callbacks=self.get_callbacks(),
             **dpo_trainer_kwargs,
         )

src/axolotl/monkeypatch/llama_attn_hijack_flash.py

@@ -22,6 +22,7 @@ from transformers.models.llama.modeling_llama import (
     apply_rotary_pos_emb,
     repeat_kv,
 )
+from xformers.ops import SwiGLU
 
 from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids, set_module_name
 
@@ -43,19 +44,7 @@ except ImportError:
 LOG = logging.getLogger("axolotl")
 
 
-def is_xformers_available() -> bool:
-    try:
-        import xformers  # pylint: disable=unused-import  # noqa: F401
-
-        return True
-    except ImportError:
-        return False
-
-
 def is_xformers_swiglu_available() -> bool:
-    if not is_xformers_available():
-        return False
-
     from xformers.ops.common import get_xformers_operator
 
     try:
@@ -68,11 +57,6 @@ def is_xformers_swiglu_available() -> bool:
 
 
 def replace_llama_mlp_with_swiglu(model):
-    if is_xformers_swiglu_available():
-        from axolotl.monkeypatch.xformers_ import FusedMLP
-    else:
-        raise RuntimeError("xformers SwiGLU not available for this environment")
-
     for name, module in model.named_modules():
         if isinstance(module, LlamaMLP):
             mlp = FusedMLP(
@@ -197,6 +181,49 @@ class FusedAttention(LlamaAttention):
         set_module_name(model, name, new_attn)
 
 
+class FusedMLP(torch.nn.Module):
+    """
+    Fused MLP layer for incrementally improved training efficiency
+    """
+
+    def __init__(
+        self,
+        config,
+        gate_proj: torch.nn.Linear,
+        up_proj: torch.nn.Linear,
+        down_proj: torch.nn.Linear,
+    ):
+        super().__init__()
+        self.config = config
+        self.swiglu = SwiGLU(
+            in_features=config.hidden_size,
+            hidden_features=config.intermediate_size,
+            bias=False,
+            _pack_weights=True,
+        )
+        # overwrite initialized weights with pretrained weights
+        self.swiglu.w12.weight.data = torch.cat(
+            (gate_proj.weight.data, up_proj.weight.data), dim=0
+        )
+        self.swiglu.w3.weight.data = down_proj.weight.data
+
+    def _post_training(self, model, name):
+        w1, w2 = torch.split(  # pylint: disable=invalid-name
+            self.swiglu.w12.weight.data, self.config.intermediate_size, dim=0
+        )
+
+        # Assign the split weights back to the original layers
+        new_mlp = LlamaMLP(self.config)
+        new_mlp.gate_proj.weight.data = w1
+        new_mlp.up_proj.weight.data = w2
+        new_mlp.down_proj.weight.data = self.swiglu.w3.weight.data
+
+        set_module_name(model, name, new_mlp)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:  # pylint: disable=invalid-name
+        return self.swiglu(x)
+
+
 # Disable the transformation of the attention mask in LlamaModel as the flash attention
 # requires the attention mask to be the same as the key_padding_mask
 def _prepare_decoder_attention_mask(

src/axolotl/monkeypatch/multipack.py

@@ -27,18 +27,15 @@ SUPPORTED_MULTIPACK_MODEL_TYPES = [
 ]
 
 
-# def patch_for_multipack(model_type, model_name=None, is_remote_code=False):
-def patch_for_multipack(model_type, model_name=None, has_remote_code=False):
+def patch_for_multipack(model_type, model_name=None, is_remote_code=False):
     if model_type == "gemmoe":
         patch_remote(model_name, ".configuration_gemmoe", ".modeling_gemmoe")
     elif model_type == "deepseek_v2":
         patch_remote(model_name, ".configuration_deepseek", ".modeling_deepseek")
-    # elif hasattr(transformers, "modeling_flash_attention_utils") and not is_remote_code:
-    elif hasattr(transformers, "modeling_flash_attention_utils"):
-        if not has_remote_code:
-            transformers.modeling_flash_attention_utils._get_unpad_data = (  # pylint: disable=protected-access
-                get_unpad_data
-            )
+    elif hasattr(transformers, "modeling_flash_attention_utils") and not is_remote_code:
+        transformers.modeling_flash_attention_utils._get_unpad_data = (  # pylint: disable=protected-access
+            get_unpad_data
+        )
         if model_type == "mixtral" and is_deepspeed_zero3_enabled():
             patch_mixtral_moe_forward_zero3()
         return

src/axolotl/monkeypatch/unsloth_.py

@@ -16,6 +16,26 @@ from transformers.models.llama.modeling_llama import (
 
 LOG = get_logger("axolotl.monkeypatch.unsloth")
 
+ORIGINAL_CEL_CODE = """# Shift so that tokens < n predict n
+        shift_logits = logits[..., :-1, :].contiguous()
+        shift_labels = labels[..., 1:].contiguous()
+        # Flatten the tokens
+        loss_fct = CrossEntropyLoss()
+        shift_logits = shift_logits.view(-1, self.config.vocab_size)
+        shift_labels = shift_labels.view(-1)
+        # Enable model parallelism
+        shift_labels = shift_labels.to(shift_logits.device)
+        loss = loss_fct(shift_logits, shift_labels)
+"""
+
+PATCHED_CEL_CODE = """shift_logits = logits[..., :-1, :].contiguous()
+        shift_labels = labels[..., 1:].contiguous()
+        loss = fast_cross_entropy_loss(
+            logits = shift_logits,
+            labels = shift_labels,
+        )
+"""
+
 ORIGINAL_QKV_CODE = """
     query_states = self.q_proj(hidden_states)
     key_states = self.k_proj(hidden_states)
@@ -60,6 +80,12 @@ def get_forward_code() -> str:
     return forward
 
 
+def check_cel_is_patchable() -> bool:
+    forward = get_forward_code()
+    forward, _ = detab_code(forward)
+    return ORIGINAL_CEL_CODE in forward
+
+
 def get_self_attn_code() -> str:
     forward = inspect.getsource(LlamaFlashAttention2.forward)
     return forward
@@ -72,31 +98,48 @@ def check_self_attn_is_patchable() -> bool:
 
 
 def integrate_cross_entropy_loss_patch(model_type: str = "llama") -> None:
-    from unsloth.kernels.cross_entropy_loss import fast_cross_entropy_loss
-
-    def UnslothForCausalLMLoss(  # pylint: disable=invalid-name
-        logits,
-        labels,
-        vocab_size: int,  # pylint: disable=unused-argument
-        num_items_in_batch: int = None,
-        ignore_index: int = -100,  # pylint: disable=unused-argument
-        **kwargs,  # pylint: disable=unused-argument
-    ):
-        # Upcast to float if we need to compute the loss to avoid potential precision issues
-        logits = logits.float()
-        # Shift so that tokens < n predict n
-        shift_logits = logits[..., :-1, :].contiguous()
-        shift_labels = labels[..., 1:].contiguous()
-
-        loss = fast_cross_entropy_loss(
-            logits=shift_logits, labels=shift_labels, n_items=num_items_in_batch
-        )
-        return loss
-
     if model_type == "llama":
-        from transformers.loss import loss_utils
+        forward = get_forward_code()
+        LlamaForCausalLM._original_forward = forward  # pylint: disable=protected-access
+        forward, _ = detab_code(forward)
+        assert ORIGINAL_CEL_CODE in forward, "Original forward code not found"
 
-        loss_utils.ForCausalLMLoss = UnslothForCausalLMLoss  # type: ignore[assignment]
+        forward = forward.replace(
+            "@add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)", ""
+        )
+        forward = forward.replace(
+            "@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)",
+            "",
+        )
+        forward = forward.replace(ORIGINAL_CEL_CODE, PATCHED_CEL_CODE)
+        forward = forward.replace(
+            "def forward(",
+            "def fast_cross_entropy_loss_forward(",
+            1,
+        )
+
+        # load imports necessary
+        import transformers.models.llama.modeling_llama
+
+        items_to_import = []
+        for item in dir(transformers.models.llama.modeling_llama):
+            if item in forward:
+                items_to_import.append(item)
+
+        exec(  # pylint: disable=exec-used  # nosec B102
+            "from unsloth.kernels.cross_entropy_loss import fast_cross_entropy_loss",
+            globals(),
+        )
+
+        exec(  # pylint: disable=exec-used  # nosec B102
+            "from transformers.models.llama.modeling_llama import ("
+            + ", ".join(x for x in items_to_import)
+            + ")",
+            globals(),
+        )
+        exec(forward, globals())  # pylint: disable=exec-used  # nosec B102
+        LOG.info("patching unsloth fast_cross_entropy_loss", main_process_only=True)
+        LlamaForCausalLM.forward = fast_cross_entropy_loss_forward  # pylint: disable=undefined-variable  # noqa: F821
     else:
         raise ValueError("Unsupported model type")
 

src/axolotl/monkeypatch/xformers_/__init__.py

@@ -1,51 +0,0 @@
-"""
-Fused MLP layer for incrementally improved training efficiency
-"""
-import torch
-from transformers.models.llama.modeling_llama import LlamaMLP
-from xformers.ops import SwiGLU
-
-from axolotl.monkeypatch.utils import set_module_name
-
-
-class FusedMLP(torch.nn.Module):
-    """
-    Fused MLP layer for incrementally improved training efficiency
-    """
-
-    def __init__(
-        self,
-        config,
-        gate_proj: torch.nn.Linear,
-        up_proj: torch.nn.Linear,
-        down_proj: torch.nn.Linear,
-    ):
-        super().__init__()
-        self.config = config
-        self.swiglu = SwiGLU(
-            in_features=config.hidden_size,
-            hidden_features=config.intermediate_size,
-            bias=False,
-            _pack_weights=True,
-        )
-        # overwrite initialized weights with pretrained weights
-        self.swiglu.w12.weight.data = torch.cat(
-            (gate_proj.weight.data, up_proj.weight.data), dim=0
-        )
-        self.swiglu.w3.weight.data = down_proj.weight.data
-
-    def _post_training(self, model, name):
-        w1, w2 = torch.split(  # pylint: disable=invalid-name
-            self.swiglu.w12.weight.data, self.config.intermediate_size, dim=0
-        )
-
-        # Assign the split weights back to the original layers
-        new_mlp = LlamaMLP(self.config)
-        new_mlp.gate_proj.weight.data = w1
-        new_mlp.up_proj.weight.data = w2
-        new_mlp.down_proj.weight.data = self.swiglu.w3.weight.data
-
-        set_module_name(model, name, new_mlp)
-
-    def forward(self, x: torch.Tensor) -> torch.Tensor:  # pylint: disable=invalid-name
-        return self.swiglu(x)

src/axolotl/prompt_strategies/__init__.py

@@ -11,10 +11,6 @@ LOG = logging.getLogger("axolotl.prompt_strategies")
 
 def load(strategy, tokenizer, cfg, ds_cfg, processor=None):
     try:
-        if strategy == "messages":
-            from .messages import load as messages_load
-
-            return messages_load(tokenizer, cfg, ds_cfg, processor=processor)
         load_fn = "load"
         if strategy.split(".")[-1].startswith("load_"):
             load_fn = strategy.split(".")[-1]
@@ -35,5 +31,4 @@ def load(strategy, tokenizer, cfg, ds_cfg, processor=None):
         return None
     except Exception as exc:  # pylint: disable=broad-exception-caught
         LOG.error(f"Failed to load prompt strategy `{strategy}`: {str(exc)}")
-        raise exc
-    return None
+        return None

src/axolotl/prompt_strategies/bradley_terry/README.md

@@ -1,10 +0,0 @@
-### example yaml
-
-```yaml
-chat_template: gemma
-datasets:
-  - path: argilla/distilabel-intel-orca-dpo-pairs
-    type: bradley_terry.chat_template
-val_set_size: 0.0
-output_dir: ./outputs/out
-```

src/axolotl/prompt_strategies/bradley_terry/__init__.py

@@ -1,35 +0,0 @@
-"""Module to load prompt strategies."""
-
-import importlib
-import inspect
-import logging
-
-from axolotl.prompt_strategies.user_defined import UserDefinedDatasetConfig
-
-LOG = logging.getLogger("axolotl.prompt_strategies.bradley_terry")
-
-
-def load(strategy, tokenizer, cfg, ds_cfg):
-    # pylint: disable=duplicate-code
-    try:
-        load_fn = "load"
-        if strategy.split(".")[-1].startswith("load_"):
-            load_fn = strategy.split(".")[-1]
-            strategy = ".".join(strategy.split(".")[:-1])
-        mod = importlib.import_module(
-            f".{strategy}", "axolotl.prompt_strategies.bradley_terry"
-        )
-        func = getattr(mod, load_fn)
-        load_kwargs = {}
-        if strategy == "user_defined":
-            load_kwargs["ds_cfg"] = UserDefinedDatasetConfig(**ds_cfg)
-        else:
-            sig = inspect.signature(func)
-            if "ds_cfg" in sig.parameters:
-                load_kwargs["ds_cfg"] = ds_cfg
-        return func(tokenizer, cfg, **load_kwargs)
-    except ModuleNotFoundError:
-        return None
-    except Exception as exc:  # pylint: disable=broad-exception-caught
-        LOG.error(f"Failed to load prompt strategy `{strategy}`: {str(exc)}")
-        return None

src/axolotl/prompt_strategies/bradley_terry/chat_template.py

@@ -1,102 +0,0 @@
-"""
-Bradley-Terry model with chat template prompt strategy.
-"""
-
-import logging
-from typing import Any, Dict, Optional
-
-from axolotl.prompt_strategies.chat_template import (
-    ChatTemplatePrompter,
-    ChatTemplateStrategy,
-)
-from axolotl.utils.chat_templates import get_chat_template_from_config
-
-# Configure the logger
-LOG = logging.getLogger("axolotl.prompt_strategies.bradley_terry.chat_template")
-LOG.setLevel(logging.INFO)
-
-
-class BTChatTemplateStrategy(ChatTemplateStrategy):
-    """
-    Bradley-Terry reward model pairwise chat template prompt strategy.
-    """
-
-    def tokenize_prompt(self, prompt):
-        """
-
-        :param prompt: the actual row of data from the underlying dataset
-        :return:
-        """
-
-        self.messages = "chosen_messages"
-        # pylint: disable=duplicate-code
-        prompt[self.messages] = []
-        if prompt["system"]:
-            prompt[self.messages].append(
-                {"role": "system", "content": prompt["system"]}
-            )
-        prompt[self.messages].append({"role": "user", "content": prompt["input"]})
-        prompt[self.messages].append({"role": "assistant", "content": prompt["chosen"]})
-        chosen_tokenized = super().tokenize_prompt(prompt)
-
-        self.messages = "rejected_messages"
-        # pylint: disable=duplicate-code
-        prompt[self.messages] = []
-        if prompt["system"]:
-            prompt[self.messages].append(
-                {"role": "system", "content": prompt["system"]}
-            )
-        prompt[self.messages].append({"role": "user", "content": prompt["input"]})
-        prompt[self.messages].append(
-            {"role": "assistant", "content": prompt["rejected"]}
-        )
-        rejected_tokenized = super().tokenize_prompt(prompt)
-
-        return {
-            "input_ids_chosen": chosen_tokenized["input_ids"],
-            "attention_mask_chosen": chosen_tokenized["attention_mask"],
-            "labels_chosen": 1.0,
-            "input_ids_rejected": rejected_tokenized["input_ids"],
-            "attention_mask_rejected": rejected_tokenized["attention_mask"],
-            "labels_rejected": 0.0,
-        }
-
-
-def load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
-    ds_cfg = ds_cfg or {}
-    chat_template_string = get_chat_template_from_config(
-        cfg=cfg, ds_cfg=ds_cfg, tokenizer=tokenizer
-    )
-
-    prompter_params = {
-        "tokenizer": tokenizer,
-        "chat_template": chat_template_string,
-        "message_field_role": ds_cfg.get("message_field_role", "role"),
-        "message_field_content": ds_cfg.get("message_field_content", "content"),
-        "message_field_training": ds_cfg.get("message_field_training", None),
-        "message_field_training_detail": ds_cfg.get(
-            "message_field_training_detail", None
-        ),
-        "roles": ds_cfg.get("roles"),
-        "drop_system_message": ds_cfg.get("drop_system_message", False),
-        # we need to add one for detecting sequences with exceeding the `sequence_len` limit.
-        "max_length": cfg.sequence_len + 1
-        if not cfg.reward_model
-        else cfg.sequence_len,
-    }
-
-    strategy_params = {
-        "train_on_inputs": cfg.train_on_inputs,
-        "sequence_len": cfg.sequence_len,
-        "roles_to_train": ds_cfg.get("roles_to_train", []),
-        "train_on_eos": ds_cfg.get("train_on_eos", None),
-    }
-
-    strategy = BTChatTemplateStrategy(
-        ChatTemplatePrompter(**prompter_params), tokenizer=tokenizer, **strategy_params
-    )
-
-    if "field_messages" in ds_cfg and hasattr(strategy, "messages"):
-        strategy.messages = ds_cfg["field_messages"]
-
-    return strategy

src/axolotl/prompt_strategies/bradley_terry/llama3.py

@@ -1,27 +0,0 @@
-"""
-chatml transforms for datasets with system, input, chosen, rejected to match llama3 chat template
-"""
-
-
-def icr(
-    cfg,
-    **kwargs,
-):  # pylint: disable=possibly-unused-variable,unused-argument
-    """
-    chatml transforms for datasets with system, input, chosen, rejected
-    ex. https://huggingface.co/datasets/argilla/distilabel-intel-orca-dpo-pairs
-    """
-
-    def transform_fn(sample):
-        if "system" in sample and sample["system"]:
-            prompt = (
-                f"<|start_header_id|>system<|end_header_id|>\n\n{sample['system']}<|eot_id|>"
-                f"<|start_header_id|>user<|end_header_id|>\n\n{sample['input']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
-            )
-        else:
-            prompt = f"<|start_header_id|>user<|end_header_id|>\n\n{sample['input']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
-        sample["chosen"] = prompt + f"{sample['chosen']}<|eot_id|>"
-        sample["rejected"] = prompt + f"{sample['rejected']}<|eot_id|>"
-        return sample
-
-    return transform_fn

src/axolotl/prompt_strategies/chat_template.py

@@ -9,7 +9,7 @@ from transformers import ProcessorMixin
 
 from axolotl.prompt_tokenizers import PromptTokenizingStrategy
 from axolotl.prompters import IGNORE_TOKEN_ID, Prompter
-from axolotl.utils.chat_templates import get_chat_template_from_config
+from axolotl.utils.chat_templates import chat_templates
 
 # Configure the logger
 LOG = logging.getLogger("axolotl")
@@ -403,16 +403,11 @@ class ChatTemplateStrategy(PromptTokenizingStrategy):
 
 
 def load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None, processor=None):
-    # pylint: disable=duplicate-code
     ds_cfg = ds_cfg or {}
-    chat_template_string = get_chat_template_from_config(
-        cfg=cfg, ds_cfg=ds_cfg, tokenizer=tokenizer
-    )
-    LOG.info(f"Using chat template:\n---\n{chat_template_string!s}\n---")
 
     prompter_params = {
         "tokenizer": tokenizer,
-        "chat_template": chat_template_string,
+        "chat_template": chat_templates(ds_cfg.get("chat_template", "chatml")),
         "message_field_role": ds_cfg.get("message_field_role", "role"),
         "message_field_content": ds_cfg.get("message_field_content", "content"),
         "message_field_training": ds_cfg.get("message_field_training", None),

src/axolotl/prompt_strategies/dpo/chat_template.py

@@ -2,16 +2,15 @@
 DPO prompt strategies for using tokenizer chat templates.
 """
 
-from axolotl.utils.chat_templates import extract_chat_template_args, get_chat_template
+from axolotl.utils.chat_templates import chat_templates
 
 
 def default(
     cfg, dataset_idx=0, **kwargs
 ):  # pylint: disable=possibly-unused-variable,unused-argument
     ds_cfg = cfg["datasets"][dataset_idx]
-    chat_template_choice, chat_template_jinja = extract_chat_template_args(
-        cfg=cfg, ds_cfg=ds_cfg
-    )
+    chat_template_str = chat_templates(cfg.chat_template)
+
     field_messages = ds_cfg.get("field_messages", "messages")
     field_chosen = ds_cfg.get("field_chosen", "chosen")
     field_rejected = ds_cfg.get("field_rejected", "rejected")
@@ -31,12 +30,6 @@ def default(
             role_map[source] = target
 
     def transform_fn(sample, tokenizer=None):
-        chat_template_string = get_chat_template(
-            user_choice=chat_template_choice,
-            jinja_template=chat_template_jinja,
-            tokenizer=tokenizer,
-        )
-
         messages = sample[field_messages]
         messages = [
             {
@@ -53,29 +46,28 @@ def default(
             "role": role_map[sample[field_rejected][field_message_role]],
             "content": sample[field_rejected][field_message_content],
         }
-        dummy_user_message = {"role": "user", "content": "[[dummy_message]]"}
 
         result = {}
         result["prompt"] = tokenizer.apply_chat_template(
             messages,
             add_generation_prompt=True,
-            chat_template=chat_template_string,
+            chat_template=chat_template_str,
             tokenize=False,
         )
 
         result["chosen"] = tokenizer.apply_chat_template(
-            [dummy_user_message, chosen],
+            [chosen],
             add_generation_prompt=False,
-            chat_template=chat_template_string,
+            chat_template=chat_template_str,
             tokenize=False,
         )
         chosen_strip_index = result["chosen"].find(chosen["content"])
         result["chosen"] = result["chosen"][chosen_strip_index:].rstrip()
 
         result["rejected"] = tokenizer.apply_chat_template(
-            [dummy_user_message, rejected],
+            [rejected],
             add_generation_prompt=False,
-            chat_template=chat_template_string,
+            chat_template=chat_template_str,
             tokenize=False,
         )
         rejected_strip_index = result["rejected"].find(rejected["content"])

src/axolotl/prompt_strategies/messages/__init__.py

@@ -1,34 +0,0 @@
-"""Module to load message prompt strategies."""
-
-import importlib
-import inspect
-import logging
-
-LOG = logging.getLogger("axolotl.prompt_strategies.messages")
-
-
-def load(tokenizer, cfg, ds_cfg, processor=None):
-    try:
-        strategy = ds_cfg.get("input_transform", "chat")
-        # pylint: disable=duplicate-code
-        load_fn = "load"
-        if strategy.split(".")[-1].startswith("load_"):
-            load_fn = strategy.split(".")[-1]
-            strategy = ".".join(strategy.split(".")[:-1])
-        mod = importlib.import_module(
-            f".{strategy}", "axolotl.prompt_strategies.messages"
-        )
-        func = getattr(mod, load_fn)
-        load_kwargs = {}
-        sig = inspect.signature(func)
-        if "ds_cfg" in sig.parameters:
-            load_kwargs["ds_cfg"] = ds_cfg
-        if "processor" in sig.parameters:
-            load_kwargs["processor"] = processor
-        return func(tokenizer, cfg, **load_kwargs)
-    except ModuleNotFoundError:
-        return None
-    except Exception as exc:  # pylint: disable=broad-exception-caught
-        LOG.error(f"Failed to load prompt strategy `{strategy}`: {str(exc)}")
-        raise exc
-    return None

src/axolotl/prompt_strategies/messages/chat.py

@@ -1,84 +0,0 @@
-"""
-Chat dataset wrapping strategy for new internal messages representations
-"""
-from typing import Any, Callable, Dict, Optional
-
-from axolotl.core.datasets.chat import TokenizedChatDataset
-from axolotl.core.datasets.transforms.chat_builder import chat_message_transform_builder
-from axolotl.prompt_tokenizers import DatasetWrappingStrategy
-
-
-class ChatMessageDatasetWrappingStrategy(DatasetWrappingStrategy):
-    """
-    Chat dataset wrapping strategy for new internal messages representations
-    """
-
-    def __init__(
-        self,
-        processor,
-        message_transform=None,
-        formatter=None,
-        **kwargs,  # pylint: disable=unused-argument
-    ):
-        """
-        :param processor: tokenizer or image processor
-        :param kwargs:
-        """
-        self.processor = processor
-        self.dataset = None
-        self.message_transform = message_transform
-        self.formatter = formatter
-
-    def wrap_dataset(
-        self,
-        dataset,
-        process_count: Optional[int] = None,
-        keep_in_memory: Optional[bool] = False,
-        **kwargs,  # pylint: disable=unused-argument
-    ):
-        self.dataset = TokenizedChatDataset(
-            dataset,
-            message_transform=self.message_transform,
-            model_transform=self.processor,
-            formatter=self.formatter,
-            process_count=process_count,
-            keep_in_memory=keep_in_memory,
-        )
-        return self.dataset
-
-
-def load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
-    ds_cfg = ds_cfg or {}
-
-    field_messages = ds_cfg.get("field_messages")
-    message_field_role = ds_cfg.get("message_field_role")
-    message_field_content = ds_cfg.get("message_field_content")
-    message_field_training = ds_cfg.get("message_field_training")
-
-    builder_kwargs = {}
-    if field_messages:
-        builder_kwargs["conversations_field"] = field_messages
-    if message_field_role:
-        builder_kwargs["message_field_role"] = message_field_role
-    if message_field_content:
-        builder_kwargs["message_field_content"] = message_field_content
-    if message_field_training:
-        builder_kwargs["message_field_training"] = message_field_training
-
-    chat_template = ds_cfg.get("chat_template", cfg.get("chat_template", "chatml"))
-    format_message = (
-        lambda x: x  # noqa E731  # pylint: disable=unnecessary-lambda-assignment
-    )
-    if chat_template == "chatml":
-        from axolotl.core.chat.format.chatml import format_message  # noqa F811
-    if chat_template.startswith("llama3"):
-        from axolotl.core.chat.format.llama3x import format_message  # noqa F811
-    message_transform: Callable = chat_message_transform_builder(
-        train_on_inputs=ds_cfg.get("train_on_inputs", False),
-        **builder_kwargs,
-    )
-    strategy = ChatMessageDatasetWrappingStrategy(
-        tokenizer, message_transform=message_transform, formatter=format_message
-    )
-
-    return strategy

src/axolotl/prompt_strategies/orpo/chat_template.py

@@ -5,7 +5,7 @@ from pydantic import BaseModel
 
 from axolotl.prompt_tokenizers import IGNORE_INDEX, PromptTokenizingStrategy
 from axolotl.prompters import Prompter
-from axolotl.utils.chat_templates import get_chat_template_from_config
+from axolotl.utils.chat_templates import chat_templates
 
 
 class Message(BaseModel):
@@ -28,13 +28,18 @@ def load(
     """
     chatml transforms for datasets with system, input, chosen, rejected
     """
-    chat_template_string = get_chat_template_from_config(
-        cfg=cfg, ds_cfg=ds_cfg, tokenizer=tokenizer
-    )
-    tokenizer.chat_template = chat_template_string
+
+    chat_template = chat_templates("chatml")
+    if ds_cfg and "chat_template" in ds_cfg:
+        chat_template = ds_cfg["chat_template"]
+        try:
+            chat_template = chat_templates(chat_template)
+        except ValueError:
+            pass
+    tokenizer.chat_template = chat_template
 
     return ORPOTokenizingStrategy(
-        ORPOPrompter(chat_template_string, tokenizer),
+        ORPOPrompter(chat_template, tokenizer),
         tokenizer,
         cfg.train_on_inputs,
         cfg.sequence_len,
@@ -243,30 +248,28 @@ class ORPOPrompter(Prompter):
 def argilla(cfg, **kwargs):  # pylint: disable=possibly-unused-variable,unused-argument
     dataset_parser = ORPODatasetParsingStrategy()
 
+    chat_template_str = chat_templates(cfg.chat_template)
+
     def transform_fn(sample, tokenizer=None):
         res = {}
 
-        chat_template_string = get_chat_template_from_config(
-            cfg=cfg, tokenizer=tokenizer
-        )
-
         res["prompt"] = tokenizer.apply_chat_template(
             [msg.model_dump() for msg in dataset_parser.get_prompt(sample).messages],
             add_generation_prompt=True,
-            chat_template=chat_template_string,
+            chat_template=chat_template_str,
             tokenize=False,
         )
         prompt_str_len = len(res["prompt"])
         res["chosen"] = tokenizer.apply_chat_template(
             [msg.model_dump() for msg in dataset_parser.get_chosen(sample).messages],
             add_generation_prompt=False,
-            chat_template=chat_template_string,
+            chat_template=chat_template_str,
             tokenize=False,
         )[prompt_str_len:]
         res["rejected"] = tokenizer.apply_chat_template(
             [msg.model_dump() for msg in dataset_parser.get_rejected(sample).messages],
             add_generation_prompt=False,
-            chat_template=chat_template_string,
+            chat_template=chat_template_str,
             tokenize=False,
         )[prompt_str_len:]
 

src/axolotl/prompt_strategies/sharegpt.py

@@ -61,9 +61,6 @@ def build_loader(
     default_conversation: Optional[str] = None,
 ):
     def _load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
-        LOG.warning(
-            "sharegpt type support will be deprecated in the next release of Axolotl. Please use chat_template instead. https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/conversation.html#chat_template",
-        )
         conversation = (
             ds_cfg["conversation"]
             if ds_cfg and "conversation" in ds_cfg

src/axolotl/prompt_tokenizers.py

@@ -30,12 +30,6 @@ class InvalidDataException(Exception):
     """
 
 
-class DatasetWrappingStrategy(abc.ABC):
-    """
-    Abstract class for wrapping datasets for Chat Messages
-    """
-
-
 class PromptTokenizingStrategy(abc.ABC):
     """
     Abstract class for tokenizing strategies

src/axolotl/train.py

@@ -10,6 +10,7 @@ from typing import Optional, Tuple, Union
 
 import torch
 import transformers.modelcard
+from accelerate import Accelerator
 from accelerate.logging import get_logger
 from accelerate.utils import save_fsdp_model
 from datasets import Dataset
@@ -96,11 +97,12 @@ def train(
     if cfg.adapter:
         msg += " and peft_config..."
     LOG.debug(msg)
+    # we wait unitl the last possible moment to setup Accelerator
+    Accelerator()
     model, peft_config = load_model(
         cfg, tokenizer, processor=processor, inference=cli_args.inference
     )
-    if model.generation_config is not None:
-        model.generation_config.do_sample = True
+    model.generation_config.do_sample = True
 
     model_ref = None
     if cfg.rl and cfg.rl != "orpo":
@@ -260,10 +262,8 @@ def train(
 
     if not cfg.hub_model_id:
         try:
-            trainer.create_model_card(
-                model_name=cfg.output_dir.lstrip("./").encode("utf-8").decode("utf-8")
-            )
-        except (AttributeError, UnicodeDecodeError):
+            trainer.create_model_card(model_name=cfg.output_dir.lstrip("./"))
+        except AttributeError:
             pass
     elif cfg.hub_model_id:
         # defensively push to the hub to ensure the model card is updated

src/axolotl/utils/collators/mm_chat.py

@@ -4,7 +4,6 @@ Collators for multi-modal chat messages and packing
 from dataclasses import dataclass
 from typing import Any, Dict, List, Optional, Union
 
-from PIL import Image
 from transformers import PreTrainedTokenizerBase, ProcessorMixin
 from transformers.data.data_collator import DataCollatorMixin
 from transformers.utils import PaddingStrategy
@@ -53,12 +52,7 @@ class MultiModalChatDataCollator(DataCollatorMixin):
             )
             for example in examples
         ]
-        images = [
-            Image.open(example["images"])
-            if isinstance(example["images"], str)
-            else example["images"]
-            for example in examples
-        ]
+        images = [example["images"] for example in examples]
 
         if max_images > 0:
             images = [img_batch[:max_images] for img_batch in images]

src/axolotl/utils/config/__init__.py

@@ -228,7 +228,6 @@ def normalize_cfg_datasets(cfg):
                         f"updating dataset {ds_cfg.path} with `chat_template: {cfg.chat_template}` to match your chat_template"
                     )
                     cfg.datasets[idx].chat_template = cfg.chat_template
-                    cfg.datasets[idx].chat_template_jinja = cfg.chat_template_jinja
 
 
 def validate_config(cfg: DictDefault, capabilities: Optional[dict] = None):

src/axolotl/utils/config/models/input/v0_4_1/__init__.py

@@ -8,16 +8,9 @@ import logging
 import os
 from enum import Enum
 from importlib.metadata import version
-from typing import Annotated, Any, Dict, List, Literal, Optional, Tuple, Union
+from typing import Any, Dict, List, Literal, Optional, Tuple, Union
 
-from pydantic import (
-    BaseModel,
-    Field,
-    StringConstraints,
-    conlist,
-    field_validator,
-    model_validator,
-)
+from pydantic import BaseModel, Field, conlist, field_validator, model_validator
 from transformers import SchedulerType
 from transformers.training_args import OptimizerNames
 
@@ -28,37 +21,6 @@ LOG = logging.getLogger("axolotl.utils.config.models.input")
 SUPPORTED_METRICS = {"sacrebleu", "comet", "ter", "chrf", "perplexity"}
 
 
-class RLType(str, Enum):
-    """RL trainer type configuration subset"""
-
-    dpo = "dpo"  # pylint: disable=invalid-name
-    ipo = "ipo"  # pylint: disable=invalid-name
-    orpo = "orpo"  # pylint: disable=invalid-name
-    kto = "kto"  # pylint: disable=invalid-name
-    simpo = "simpo"  # pylint: disable=invalid-name
-
-
-class ChatTemplate(str, Enum):
-    """Chat templates configuration subset"""
-
-    alpaca = "alpaca"  # pylint: disable=invalid-name
-    chatml = "chatml"  # pylint: disable=invalid-name
-    mistral_v1 = "mistral_v1"  # pylint: disable=invalid-name
-    mistral_v2v3 = "mistral_v2v3"  # pylint: disable=invalid-name
-    mistral_v3_tekken = "mistral_v3_tekken"  # pylint: disable=invalid-name
-    gemma = "gemma"  # pylint: disable=invalid-name
-    cohere = "cohere"  # pylint: disable=invalid-name
-    llama3 = "llama3"  # pylint: disable=invalid-name
-    llama3_2_vision = "llama3_2_vision"  # pylint: disable=invalid-name
-    phi_3 = "phi_3"  # pylint: disable=invalid-name
-    phi_35 = "phi_35"  # pylint: disable=invalid-name
-    deepseek_v2 = "deepseek_v2"  # pylint: disable=invalid-name
-    jamba = "jamba"  # pylint: disable=invalid-name
-    jinja = "jinja"  # pylint: disable=invalid-name
-    qwen_25 = "qwen_25"  # pylint: disable=invalid-name
-    tokenizer_default = "tokenizer_default"  # pylint: disable=invalid-name
-
-
 class DeprecatedParameters(BaseModel):
     """configurations that are deprecated"""
 
@@ -140,22 +102,14 @@ class SFTDataset(BaseModel):
     path: Optional[str] = None
     split: Optional[str] = None
     type: Optional[Union[str, UserDefinedPrompterType]] = None
-    input_transform: Optional[str] = None
     shards: Optional[int] = None
     conversation: Optional[str] = None
-    # Do not make this too strict or it will break the validator to choose different dataset class
-    chat_template: Optional[
-        Union[
-            ChatTemplate,
-            str,
-        ]
-    ] = None
-    chat_template_jinja: Optional[str] = None
+    chat_template: Optional[str] = None
     data_files: Optional[Union[str, List[str]]] = None
-    input_format: Optional[str] = None
     name: Optional[str] = None
     ds_type: Optional[str] = None
     train_on_split: Optional[str] = None
+
     field: Optional[str] = None
     field_human: Optional[str] = None
     field_model: Optional[str] = None
@@ -166,31 +120,11 @@ class SFTDataset(BaseModel):
     message_field_training_detail: Optional[str] = None
     roles_to_train: Optional[List[str]] = None
     train_on_eos: Optional[str] = None
+
     roles: Optional[Dict[str, List[str]]] = None
     drop_system_message: Optional[bool] = None
+
     trust_remote_code: Optional[bool] = False
-    revision: Optional[str] = None
-
-    @model_validator(mode="before")
-    @classmethod
-    def check_chat_template_config(cls, data):
-        # Set chat_template to tokenizer_default if not set
-        if data.get("type") == "chat_template" and not data.get("chat_template"):
-            data["chat_template"] = ChatTemplate.tokenizer_default
-
-        # if chat_template is set to jinja, chat_template_jinja is required
-        if data.get("chat_template") == ChatTemplate.jinja and not data.get(
-            "chat_template_jinja"
-        ):
-            raise ValueError(
-                "chat_template_jinja is required when chat_template is set to jinja"
-            )
-
-        # If chat_template_jinja is set, set chat_template to jinja
-        if data.get("chat_template_jinja") and not data.get("chat_template"):
-            data["chat_template"] = ChatTemplate.jinja
-
-        return data
 
 
 class UserDefinedDPOType(BaseModel):
@@ -212,7 +146,6 @@ class DPODataset(BaseModel):
     split: Optional[str] = None
     type: Optional[Union[UserDefinedDPOType, str]] = None
     data_files: Optional[List[str]] = None
-    revision: Optional[str] = None
 
 
 class UserDefinedKTOType(BaseModel):
@@ -234,7 +167,32 @@ class KTODataset(BaseModel):
     type: Optional[Union[UserDefinedKTOType, str]] = None
     data_files: Optional[List[str]] = None
     trust_remote_code: Optional[bool] = False
-    revision: Optional[str] = None
+
+
+class RLType(str, Enum):
+    """RL trainer type configuration subset"""
+
+    dpo = "dpo"  # pylint: disable=invalid-name
+    ipo = "ipo"  # pylint: disable=invalid-name
+    orpo = "orpo"  # pylint: disable=invalid-name
+    kto = "kto"  # pylint: disable=invalid-name
+    simpo = "simpo"  # pylint: disable=invalid-name
+
+
+class ChatTemplate(str, Enum):
+    """Chat templates configuration subset"""
+
+    alpaca = "alpaca"  # pylint: disable=invalid-name
+    chatml = "chatml"  # pylint: disable=invalid-name
+    inst = "inst"  # pylint: disable=invalid-name
+    gemma = "gemma"  # pylint: disable=invalid-name
+    cohere = "cohere"  # pylint: disable=invalid-name
+    llama3 = "llama3"  # pylint: disable=invalid-name
+    llama3_2_vision = "llama3_2_vision"  # pylint: disable=invalid-name
+    phi_3 = "phi_3"  # pylint: disable=invalid-name
+    phi_35 = "phi_35"  # pylint: disable=invalid-name
+    deepseek_v2 = "deepseek_v2"  # pylint: disable=invalid-name
+    jamba = "jamba"  # pylint: disable=invalid-name
 
 
 class LoftQConfig(BaseModel):
@@ -486,7 +444,6 @@ class MLFlowConfig(BaseModel):
     use_mlflow: Optional[bool] = None
     mlflow_tracking_uri: Optional[str] = None
     mlflow_experiment_name: Optional[str] = None
-    mlflow_run_name: Optional[str] = None
     hf_mlflow_log_artifacts: Optional[bool] = None
 
 
@@ -583,10 +540,8 @@ class AxolotlInputConfig(
     resume_from_checkpoint: Optional[str] = None
     auto_resume_from_checkpoints: Optional[bool] = None
     resize_token_embeddings_to_32x: Optional[bool] = None
-    mean_resizing_embeddings: Optional[bool] = False
 
     rl: Optional[RLType] = None
-    reward_model: Optional[bool] = None
 
     datasets: Optional[conlist(Union[SFTDataset, DPODataset, KTODataset], min_length=1)] = None  # type: ignore
     test_datasets: Optional[conlist(Union[SFTDataset, DPODataset, KTODataset], min_length=1)] = None  # type: ignore
@@ -753,13 +708,7 @@ class AxolotlInputConfig(
     gpu_memory_limit: Optional[Union[int, str]] = None
     low_cpu_mem_usage: Optional[bool] = None
 
-    chat_template: Optional[
-        Union[
-            ChatTemplate,
-            Annotated[str, StringConstraints(pattern="^tokenizer_default_fallback_")],
-        ]
-    ] = None
-    chat_template_jinja: Optional[str] = None
+    chat_template: Optional[ChatTemplate] = None
     default_system_message: Optional[str] = None
 
     fix_untrained_tokens: Optional[bool] = None
@@ -868,23 +817,6 @@ class AxolotlInputConfig(
 
         return data
 
-    @model_validator(mode="before")
-    @classmethod
-    def check_chat_template_config(cls, data):
-        # if chat_template is set to jinja, chat_template_jinja is required
-        if data.get("chat_template") == ChatTemplate.jinja and not data.get(
-            "chat_template_jinja"
-        ):
-            raise ValueError(
-                "chat_template_jinja is required when chat_template is set to jinja"
-            )
-
-        # If chat_template_jinja is set, set chat_template to jinja
-        if data.get("chat_template_jinja") and not data.get("chat_template"):
-            data["chat_template"] = ChatTemplate.jinja
-
-        return data
-
     @model_validator(mode="before")
     @classmethod
     def check_sample_packing_wo_flash(cls, data):
@@ -915,17 +847,6 @@ class AxolotlInputConfig(
             )
         return data
 
-    @model_validator(mode="before")
-    @classmethod
-    def hint_reward_model_pad(cls, data):
-        if data.get("reward_model") and not data.get("pad_to_sequence_len"):
-            LOG.warning(
-                "`pad_to_sequence_len: true` is recommended when using reward_model"
-            )
-            if data.get("pad_to_sequence_len") is None:
-                data["pad_to_sequence_len"] = True
-        return data
-
     @model_validator(mode="before")
     @classmethod
     def check_gas_bsz(cls, data):

src/axolotl/utils/data/rl.py

@@ -90,7 +90,6 @@ def load_prepare_dpo_datasets(cfg):
                 ds = load_dataset(  # pylint: disable=invalid-name
                     ds_cfg["path"],
                     split=ds_cfg["split"],
-                    revision=ds_cfg.get("revision", None),
                 )
                 split_datasets.insert(i, ds)
 

src/axolotl/utils/data/sft.py

@@ -19,12 +19,10 @@ from transformers import PreTrainedTokenizerBase
 from axolotl.common.const import DEFAULT_DATASET_PREPARED_PATH
 from axolotl.datasets import TokenizedPromptDataset
 from axolotl.prompt_strategies import load
-from axolotl.prompt_strategies.bradley_terry import load as bradley_terry_load
 from axolotl.prompt_tokenizers import (
     AlpacaMultipleChoicePromptTokenizingStrategy,
     AlpacaPromptTokenizingStrategy,
     AlpacaReflectionPTStrategy,
-    DatasetWrappingStrategy,
     GPTeacherPromptTokenizingStrategy,
     JeopardyPromptTokenizingStrategy,
     OpenAssistantPromptTokenizingStrategy,
@@ -460,7 +458,7 @@ def load_tokenized_prepared_datasets(
             else:
                 LOG.debug("NOT shuffling merged datasets")
 
-        if cfg.sample_packing and not cfg.skip_prepare_dataset:
+        if not cfg.skip_prepare_dataset:
             dataset, _ = process_datasets_for_packing(cfg, dataset, None)
 
         if cfg.local_rank == 0 and not cfg.skip_prepare_dataset:
@@ -575,7 +573,7 @@ def get_dataset_wrapper(
     d_base_type,
     dataset,
     d_prompt_style=None,
-    processor=None,  # pylint: disable=unused-argument
+    processor=None,
 ):
     dataset_wrapper = None
     dataset_prompter = None
@@ -610,10 +608,8 @@ def get_dataset_wrapper(
         )
     elif cfg.skip_prepare_dataset:
         dataset_wrapper = dataset
-    elif ds_strategy := config_dataset.type.startswith(
-        "bradley_terry"
-    ) and bradley_terry_load(
-        config_dataset.type.split(".", 1)[1], tokenizer, cfg, config_dataset
+    elif ds_strategy := load(
+        config_dataset.type, tokenizer, cfg, config_dataset, processor=processor
     ):
         dataset_prompter = UnsupportedPrompter()
         dataset_wrapper = TokenizedPromptDataset(
@@ -621,18 +617,6 @@ def get_dataset_wrapper(
             dataset,
             **ds_kwargs,
         )
-    elif ds_strategy := load(
-        config_dataset.type, tokenizer, cfg, config_dataset, processor=processor
-    ):
-        if isinstance(ds_strategy, DatasetWrappingStrategy):
-            dataset_wrapper = ds_strategy.wrap_dataset(dataset, **ds_kwargs)
-        else:
-            dataset_prompter = UnsupportedPrompter()
-            dataset_wrapper = TokenizedPromptDataset(
-                ds_strategy,
-                dataset,
-                **ds_kwargs,
-            )
     elif d_base_type == "alpaca":
         dataset_prompter = AlpacaPrompter(d_prompt_style)
         ds_strategy = AlpacaPromptTokenizingStrategy(

src/axolotl/utils/mlflow_.py

@@ -16,7 +16,3 @@ def setup_mlflow_env_vars(cfg: DictDefault):
     # Enable mlflow if experiment name is present
     if cfg.mlflow_experiment_name and len(cfg.mlflow_experiment_name) > 0:
         cfg.use_mlflow = True
-
-    # Enable logging hf artifacts in mlflow if value is truthy
-    if cfg.hf_mlflow_log_artifacts is True:
-        os.environ["HF_MLFLOW_LOG_ARTIFACTS"] = "true"

src/axolotl/utils/samplers/multipack.py

@@ -133,8 +133,6 @@ class MultipackBatchSampler(BatchSampler):
         self.eff_total_used = 0
         self.eff_total_slots = 0
 
-        self.len_across_ranks = None
-
     def set_epoch(self, epoch: int):
         self.epoch = epoch
 
@@ -197,14 +195,15 @@ class MultipackBatchSampler(BatchSampler):
             LOG.info(f"gather_len_batches: {repr(estimates)}")
             return math.floor(0.998 * min(estimates))
 
-        min_len_batches = reduce_and_broadcast(lambda: num, calc_min_len)
+        min_len_batches = reduce_and_broadcast(
+            lambda: num,
+            calc_min_len,
+        )
         return min_len_batches
 
     def __len__(self):
-        if not self.len_across_ranks:
-            len_batches = self.num_batches()
-            self.len_across_ranks = self.gather_len_batches(len_batches)
-        return self.len_across_ranks
+        len_batches = self.num_batches()
+        return self.gather_len_batches(len_batches)
 
     def _len_est(self):
         efficiency = (

src/axolotl/utils/trainer.py

@@ -11,7 +11,7 @@ import numpy as np
 import torch
 import torch.cuda
 from accelerate.logging import get_logger
-from datasets import disable_caching, enable_caching
+from datasets import set_caching_enabled
 from torch.utils.data import DataLoader, RandomSampler
 from transformers.utils import is_torch_bf16_gpu_available
 
@@ -87,10 +87,10 @@ def trainer_weighted_loss(model_output, labels, shift_labels=True):
 @contextmanager
 def disable_datasets_caching():
     try:
-        disable_caching()
+        set_caching_enabled(False)
         yield
     finally:
-        enable_caching()
+        set_caching_enabled(True)
 
 
 def add_position_ids(sample):
@@ -306,11 +306,7 @@ def process_pretraining_datasets_for_packing(
 
 
 def calculate_total_num_steps(cfg, train_dataset, update=True):
-    if (
-        not cfg.total_num_tokens
-        and not cfg.skip_prepare_dataset
-        and not cfg.reward_model
-    ):
+    if not cfg.total_num_tokens and not cfg.skip_prepare_dataset:
         total_num_tokens = np.sum(
             train_dataset.data.column("input_ids")
             .to_pandas()
@@ -327,7 +323,6 @@ def calculate_total_num_steps(cfg, train_dataset, update=True):
         not skip_estimates
         and not cfg.total_supervised_tokens
         and not cfg.skip_prepare_dataset
-        and not cfg.reward_model
     ):
         total_supervised_tokens = (
             train_dataset.data.column("labels")

tests/core/chat/test_messages.py

@@ -1,197 +0,0 @@
-"""
-Tests for the chat messages module
-"""
-import unittest
-
-import pytest
-from transformers import AddedToken, AutoTokenizer
-
-from axolotl.core.chat.format.chatml import format_message
-from axolotl.core.chat.messages import ChatFormattedChats, Chats
-
-
-@pytest.fixture(scope="session", name="llama_tokenizer")
-def llama_tokenizer_fixture():
-    return AutoTokenizer.from_pretrained("NousResearch/Meta-Llama-3.1-8B")
-
-
-@pytest.fixture(scope="session", name="chatml_tokenizer")
-def llama_tokenizer_w_chatml(llama_tokenizer):
-    llama_tokenizer.add_special_tokens(
-        {
-            "eos_token": AddedToken(
-                "<|im_end|>", rstrip=False, lstrip=False, normalized=False
-            )
-        }
-    )
-    llama_tokenizer.add_tokens(
-        [
-            AddedToken("<|im_start|>", rstrip=False, lstrip=False, normalized=False),
-        ]
-    )
-
-    return llama_tokenizer
-
-
-@pytest.fixture(scope="session", name="chat_msgs")
-def chat_msgs_fixture():
-    return {
-        "conversation": [
-            {
-                "role": "system",
-                "content": [
-                    {"type": "text", "value": "You are a helpful assistant."},
-                ],
-            },
-            {
-                "role": "user",
-                "content": [
-                    {"type": "text", "value": "What is today's stock price of Apple?"},
-                ],
-            },
-            {
-                "role": "assistant",
-                "content": [
-                    {
-                        "type": "tool_call",
-                        "value": {
-                            "name": "get_date",
-                            "arguments": {},
-                        },
-                    },
-                    {
-                        "type": "tool_call",
-                        "value": {
-                            "name": "get_stock_price",
-                            "arguments": {"symbol": "AAPL"},
-                        },
-                    },
-                ],
-                "weight": 1,
-            },
-            {
-                "role": "tool",
-                "content": [
-                    {
-                        "type": "tool_response",
-                        "value": {
-                            "name": "get_date",
-                            "content": {"date": "2024-09-09"},
-                        },
-                    },
-                    {
-                        "type": "tool_response",
-                        "value": {
-                            "name": "get_stock_price",
-                            "content": {"symbol": "AAPL", "price": 123.45},
-                        },
-                    },
-                ],
-            },
-            {
-                "role": "assistant",
-                "content": [
-                    {
-                        "type": "text",
-                        "value": "The stock price of Apple is $123.45.\n",
-                        "weight": 0,
-                    },
-                    {
-                        "type": "text",
-                        "value": "<reflection>The original query asked for today's stock price of Apple. This implies they also wanted the date included in the response.</reflection>",
-                    },
-                    {
-                        "type": "text",
-                        "value": "The stock price of Apple on September 9, 2024 is $123.45.",
-                    },
-                ],
-                "weight": 1,
-            },
-        ]
-    }
-
-
-class TestMessagesCase:
-    """
-    Test cases for the chat messages module
-    """
-
-    def test_tool_call_stringify(self, chat_msgs):
-        chat_msgs_as_obj = Chats(**chat_msgs)
-        assert '{"name": "get_stock_price", "arguments": {"symbol": "AAPL"}}' == str(
-            chat_msgs_as_obj.conversation[2].content[1].value
-        )
-
-    def test_chatml_formatted_wrapper(self, chat_msgs):
-        chat_msg_formatted = ChatFormattedChats(**chat_msgs, formatter=format_message)
-        target_chatml = """<|im_start|>system
-You are a helpful assistant.<|im_end|>
-<|im_start|>user
-What is today's stock price of Apple?<|im_end|>
-<|im_start|>assistant
-<tool_call>
-{"name": "get_date", "arguments": {}}
-</tool_call>
-<tool_call>
-{"name": "get_stock_price", "arguments": {"symbol": "AAPL"}}
-</tool_call>
-<|im_end|>
-<|im_start|>tool
-<tool_response>
-{"name": "get_date", "content": {"date": "2024-09-09"}}
-</tool_response>
-<tool_response>
-{"name": "get_stock_price", "content": {"symbol": "AAPL", "price": 123.45}}
-</tool_response>
-<|im_end|>
-<|im_start|>assistant
-The stock price of Apple is $123.45.
-<reflection>The original query asked for today's stock price of Apple. This implies they also wanted the date included in the response.</reflection>The stock price of Apple on September 9, 2024 is $123.45.<|im_end|>\n"""
-        assert target_chatml == str(chat_msg_formatted)
-
-    def test_chatml_formatting_tool_call(self, chat_msgs):
-        chat_msgs_as_obj = Chats(**chat_msgs)
-        target_chatml_turn2 = """<|im_start|>assistant\n<tool_call>\n{"name": "get_date", "arguments": {}}\n</tool_call>\n<tool_call>\n{"name": "get_stock_price", "arguments": {"symbol": "AAPL"}}\n</tool_call>\n<|im_end|>\n"""
-        assert target_chatml_turn2 == str(
-            format_message(chat_msgs_as_obj.conversation[2])
-        )
-
-    def test_train_labels(self, chatml_tokenizer, chat_msgs):
-        chat_msg_formatted = ChatFormattedChats(**chat_msgs, formatter=format_message)
-        tokenized = chat_msg_formatted.conversation[2].tokenized(chatml_tokenizer)
-        # fmt: off
-        target_labels = [
-            -100, -100, -100,  # role
-            27, 14506, 13735, 397, 5018, 609, 794,
-            330, 456, 4257, 498, 330, 16774, 794, 4792, 534, 524,
-            14506, 13735, 397, 27, 14506, 13735, 397, 5018, 609, 794,
-            330, 456, 31641, 9217, 498, 330, 16774, 794, 5324, 19314,
-            794, 330, 84016, 43, 96742, 524, 14506, 13735, 397,
-            128256,  # <|im_end|>
-            -100  # trailing newline
-        ]
-        # fmt: on
-        assert tokenized["labels"] == target_labels
-
-    def test_train_labels_2(self, chatml_tokenizer, chat_msgs):
-        # also test if indivudal contents are set not to train
-        chat_msg_formatted = ChatFormattedChats(**chat_msgs, formatter=format_message)
-        tokenized = chat_msg_formatted.conversation[4].tokenized(chatml_tokenizer)
-        # fmt: off
-        target_labels = [
-            -100, -100, -100,  # role
-            -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,  # initial response
-            27, 78098, 16761, 4113, 3319, 4691, 369, 3432, 596, 5708, 3430,
-            315, 8325, 13, 1115, 24897, 814, 1101, 4934, 279, 2457,
-            5343, 304, 279, 2077, 4005, 78098, 16761, 5708, 3430, 315,
-            8325, 389, 6250, 220, 24, 11, 220, 2366, 19, 374, 400,
-            4513, 13, 1774, 13,
-            128256,  # <|im_end|>
-            -100,  # trailing newline
-        ]
-        # fmt: on
-        assert tokenized["labels"] == target_labels
-
-
-if __name__ == "__main__":
-    unittest.main()

tests/e2e/multigpu/test_eval.py

@@ -1,155 +0,0 @@
-"""
-E2E tests for multigpu eval
-"""
-import logging
-import os
-import unittest
-from pathlib import Path
-
-import yaml
-from accelerate.test_utils import execute_subprocess_async
-
-from axolotl.utils.dict import DictDefault
-
-from ..utils import with_temp_dir
-
-LOG = logging.getLogger("axolotl.tests.e2e.multigpu")
-os.environ["WANDB_DISABLED"] = "true"
-
-AXOLOTL_ROOT = Path(__file__).parent.parent.parent.parent
-
-
-class TestMultiGPUEval(unittest.TestCase):
-    """
-    Test case for MultiGPU Eval Sample Packing
-    """
-
-    @with_temp_dir
-    def test_eval_sample_packing(self, temp_dir):
-        # pylint: disable=duplicate-code
-        cfg = DictDefault(
-            {
-                "base_model": "JackFram/llama-68m",
-                "load_in_8bit": False,
-                "load_in_4bit": True,
-                "strict": False,
-                "sequence_len": 2048,
-                "adapter": "qlora",
-                "sample_packing": True,
-                "eval_sample_packing": True,
-                "pad_to_sequence_len": True,
-                "lora_r": 8,
-                "lora_alpha": 16,
-                "lora_dropout": 0.05,
-                "lora_target_linear": True,
-                "lora_modules_to_save": ["embed_tokens", "lm_head"],
-                "val_set_size": 0.1,
-                "special_tokens": {"pad_token": "<|end_of_text|>"},
-                "datasets": [
-                    {
-                        "path": "teknium/GPT4-LLM-Cleaned",
-                        "type": "alpaca",
-                    },
-                ],
-                "num_epochs": 1,
-                "max_steps": 5,
-                "micro_batch_size": 2,
-                "gradient_accumulation_steps": 4,
-                "output_dir": temp_dir,
-                "learning_rate": 0.00001,
-                "optimizer": "adamw_8bit",
-                "lr_scheduler": "cosine",
-                "flash_attention": True,
-                "loss_watchdog_threshold": 5.0,
-                "loss_watchdog_patience": 3,
-                "bf16": "auto",
-                "warmup_steps": 1,
-                "evals_per_epoch": 2,
-                "eval_max_new_tokens": 128,
-                "saves_per_epoch": 1,
-                "logging_steps": 1,
-                "weight_decay": 0.0,
-            }
-        )
-
-        # write cfg to yaml file
-        Path(temp_dir).mkdir(parents=True, exist_ok=True)
-        with open(Path(temp_dir) / "config.yaml", "w", encoding="utf-8") as fout:
-            fout.write(yaml.dump(cfg.to_dict(), Dumper=yaml.Dumper))
-
-        execute_subprocess_async(
-            [
-                "accelerate",
-                "launch",
-                "--num-processes",
-                "2",
-                "-m",
-                "axolotl.cli.train",
-                str(Path(temp_dir) / "config.yaml"),
-            ]
-        )
-
-    @with_temp_dir
-    def test_eval(self, temp_dir):
-        # pylint: disable=duplicate-code
-        cfg = DictDefault(
-            {
-                "base_model": "JackFram/llama-68m",
-                "load_in_8bit": False,
-                "load_in_4bit": True,
-                "strict": False,
-                "sequence_len": 2048,
-                "adapter": "qlora",
-                "sample_packing": True,
-                "eval_sample_packing": False,
-                "pad_to_sequence_len": True,
-                "lora_r": 8,
-                "lora_alpha": 16,
-                "lora_dropout": 0.05,
-                "lora_target_linear": True,
-                "lora_modules_to_save": ["embed_tokens", "lm_head"],
-                "val_set_size": 0.1,
-                "special_tokens": {"pad_token": "<|end_of_text|>"},
-                "datasets": [
-                    {
-                        "path": "teknium/GPT4-LLM-Cleaned",
-                        "type": "alpaca",
-                    },
-                ],
-                "num_epochs": 1,
-                "max_steps": 5,
-                "micro_batch_size": 2,
-                "gradient_accumulation_steps": 4,
-                "output_dir": temp_dir,
-                "learning_rate": 0.00001,
-                "optimizer": "adamw_8bit",
-                "lr_scheduler": "cosine",
-                "flash_attention": True,
-                "loss_watchdog_threshold": 5.0,
-                "loss_watchdog_patience": 3,
-                "bf16": "auto",
-                "warmup_steps": 1,
-                "evals_per_epoch": 2,
-                "eval_max_new_tokens": 128,
-                "saves_per_epoch": 1,
-                "logging_steps": 1,
-                "weight_decay": 0.0,
-            }
-        )
-
-        # write cfg to yaml file
-        Path(temp_dir).mkdir(parents=True, exist_ok=True)
-        with open(Path(temp_dir) / "config.yaml", "w", encoding="utf-8") as fout:
-            fout.write(yaml.dump(cfg.to_dict(), Dumper=yaml.Dumper))
-
-        execute_subprocess_async(
-            [
-                "accelerate",
-                "launch",
-                "--num-processes",
-                "2",
-                "-m",
-                "axolotl.cli.train",
-                str(Path(temp_dir) / "config.yaml"),
-            ]
-        )

tests/e2e/multigpu/test_llama.py

@@ -19,8 +19,6 @@ from ..utils import with_temp_dir
 LOG = logging.getLogger("axolotl.tests.e2e.multigpu")
 os.environ["WANDB_DISABLED"] = "true"
 
-AXOLOTL_ROOT = Path(__file__).parent.parent.parent.parent
-
 
 @pytest.fixture(scope="session", autouse=True)
 def download_model():
@@ -348,115 +346,3 @@ class TestMultiGPULlama(unittest.TestCase):
                 str(Path(temp_dir) / "config.yaml"),
             ]
         )
-
-    @with_temp_dir
-    def test_ds_zero3_packed(self, temp_dir):
-        # pylint: disable=duplicate-code
-        cfg = DictDefault(
-            {
-                "base_model": "TinyLlama/TinyLlama_v1.1",
-                "tokenizer_type": "LlamaTokenizer",
-                "sample_packing": True,
-                "eval_sample_packing": False,
-                "pad_to_sequence_len": True,
-                "sequence_len": 2048,
-                "val_set_size": 0.05,
-                "special_tokens": {
-                    "unk_token": "<unk>",
-                    "bos_token": "<s>",
-                    "eos_token": "</s>",
-                },
-                "datasets": [
-                    {
-                        "path": "tatsu-lab/alpaca",
-                        "type": "alpaca",
-                    },
-                ],
-                "num_epochs": 1,
-                "max_steps": 100,
-                "micro_batch_size": 4,
-                "gradient_accumulation_steps": 4,
-                "output_dir": temp_dir,
-                "learning_rate": 0.00001,
-                "optimizer": "adamw_torch",
-                "lr_scheduler": "cosine",
-                "flash_attention": True,
-                "deepspeed": str(AXOLOTL_ROOT / "deepspeed_configs/zero3_bf16.json"),
-            }
-        )
-
-        # write cfg to yaml file
-        Path(temp_dir).mkdir(parents=True, exist_ok=True)
-        with open(Path(temp_dir) / "config.yaml", "w", encoding="utf-8") as fout:
-            fout.write(yaml.dump(cfg.to_dict(), Dumper=yaml.Dumper))
-
-        execute_subprocess_async(
-            [
-                "accelerate",
-                "launch",
-                "--num-processes",
-                "2",
-                "-m",
-                "axolotl.cli.train",
-                str(Path(temp_dir) / "config.yaml"),
-            ]
-        )
-
-    @with_temp_dir
-    def test_ds_zero3_qlora_packed(self, temp_dir):
-        # pylint: disable=duplicate-code
-        cfg = DictDefault(
-            {
-                "base_model": "TinyLlama/TinyLlama_v1.1",
-                "tokenizer_type": "LlamaTokenizer",
-                "load_in_4bit": True,
-                "adapter": "qlora",
-                "lora_r": 8,
-                "lora_alpha": 16,
-                "lora_dropout": 0.05,
-                "lora_target_linear": True,
-                "sample_packing": True,
-                "eval_sample_packing": False,
-                "pad_to_sequence_len": True,
-                "sequence_len": 2048,
-                "val_set_size": 0.05,
-                "special_tokens": {
-                    "unk_token": "<unk>",
-                    "bos_token": "<s>",
-                    "eos_token": "</s>",
-                },
-                "datasets": [
-                    {
-                        "path": "tatsu-lab/alpaca",
-                        "type": "alpaca",
-                    },
-                ],
-                "num_epochs": 1,
-                "max_steps": 100,
-                "micro_batch_size": 4,
-                "gradient_accumulation_steps": 4,
-                "output_dir": temp_dir,
-                "learning_rate": 0.0001,
-                "optimizer": "adamw_torch",
-                "lr_scheduler": "cosine",
-                "flash_attention": True,
-                "deepspeed": str(AXOLOTL_ROOT / "deepspeed_configs/zero3_bf16.json"),
-            }
-        )
-
-        # write cfg to yaml file
-        Path(temp_dir).mkdir(parents=True, exist_ok=True)
-        with open(Path(temp_dir) / "config.yaml", "w", encoding="utf-8") as fout:
-            fout.write(yaml.dump(cfg.to_dict(), Dumper=yaml.Dumper))
-
-        execute_subprocess_async(
-            [
-                "accelerate",
-                "launch",
-                "--num-processes",
-                "2",
-                "-m",
-                "axolotl.cli.train",
-                str(Path(temp_dir) / "config.yaml"),
-            ]
-        )

tests/e2e/patched/test_unsloth_integration.py

@@ -1,12 +1,22 @@
 """Test module for checking whether the integration of Unsloth with Hugging Face Transformers is working as expected."""
 import unittest
 
-from axolotl.monkeypatch.unsloth_ import check_self_attn_is_patchable
+from axolotl.monkeypatch.unsloth_ import (
+    check_cel_is_patchable,
+    check_self_attn_is_patchable,
+)
 
 
 class TestUnslothIntegration(unittest.TestCase):
     """Unsloth monkeypatch integration tests."""
 
+    def test_is_cel_patchable(self):
+        # ensures the current version of transformers has loss code that matches our patching code
+        self.assertTrue(
+            check_cel_is_patchable(),
+            "HF transformers loss code has changed and isn't patchable",
+        )
+
     def test_is_self_attn_patchable(self):
         # ensures the current version of transformers has loss code that matches our patching code
         self.assertTrue(

tests/e2e/test_load_model.py

@@ -1,95 +0,0 @@
-"""Module for testing ModelLoader."""
-
-import shutil
-import tempfile
-
-import pytest
-import torch
-
-from axolotl.utils.dict import DictDefault
-from axolotl.utils.models import ModelLoader, load_model, load_tokenizer
-
-
-@pytest.fixture(name="temp_dir")
-def fixture_temp_dir():
-    temp_dir = tempfile.mkdtemp()
-    yield temp_dir
-    shutil.rmtree(temp_dir)
-
-
-class TestLoadModelUtils:
-    """
-    Testing module testing ModelLoader.
-    """
-
-    def setup_method(self):
-        # load config
-        self.cfg = DictDefault(
-            {
-                "base_model": "JackFram/llama-68m",
-                "tokenizer_type": "LlamaTokenizer",
-                "tokenizer_config": "JackFram/llama-68m",
-                "sequence_len": 1024,
-                "load_in_8bit": False,
-                "adapter": "lora",
-                "lora_r": 8,
-                "lora_alpha": 16,
-                "lora_dropout": 0.05,
-                "lora_target_linear": True,
-                "val_set_size": 0.1,
-                "special_tokens": {
-                    "unk_token": "<unk>",
-                    "bos_token": "<s>",
-                    "eos_token": "</s>",
-                },
-                "datasets": [
-                    {
-                        "path": "mhenrichsen/alpaca_2k_test",
-                        "type": "alpaca",
-                    },
-                ],
-                "num_epochs": 1,
-                "micro_batch_size": 8,
-                "gradient_accumulation_steps": 1,
-                "learning_rate": 0.00001,
-                "optimizer": "adamw_torch",
-                "lr_scheduler": "cosine",
-            }
-        )
-        self.model_loader = (  # pylint: disable=attribute-defined-outside-init
-            ModelLoader(
-                cfg=self.cfg,
-                tokenizer="",
-            )
-        )
-
-    @pytest.mark.parametrize("embedding_modules", ["embed_tokens", "lm_head"])
-    @pytest.mark.parametrize(
-        "dist_dtype", [torch.bfloat16, torch.float16, torch.float32]
-    )
-    @pytest.mark.parametrize("before_kbit_train_or_finetune", [True, False])
-    def test_convert_embedding_modules_dtype(
-        self, temp_dir, embedding_modules, dist_dtype, before_kbit_train_or_finetune
-    ):
-        self.cfg.output_dir = temp_dir
-        self.model_loader.tokenizer = load_tokenizer(self.cfg)  # pylint: disable=all
-        self.model_loader.model, _ = load_model(
-            self.cfg,
-            self.model_loader.tokenizer,
-            inference=False,
-            reference_model=True,
-        )
-        self.model_loader.convert_embedding_modules_dtype(
-            embedding_modules, dist_dtype, before_kbit_train_or_finetune
-        )
-        for name, module in self.model_loader.model.named_modules():
-            if (
-                "norm" in name
-                or (before_kbit_train_or_finetune and name.endswith(".gate"))
-                or (
-                    any(m in name for m in embedding_modules)
-                    and hasattr(module, "weight")
-                )
-            ):
-                for _, param in module.named_parameters():
-                    assert param.dtype == dist_dtype

tests/e2e/test_packing_loss.py

@@ -1,74 +0,0 @@
-"""
-E2E tests for packed training
-"""
-
-import logging
-import os
-import unittest
-
-from tbparse import SummaryReader
-from transformers.utils import is_torch_bf16_gpu_available
-
-from axolotl.cli import load_datasets
-from axolotl.common.cli import TrainerCliArgs
-from axolotl.train import train
-from axolotl.utils.config import normalize_config
-from axolotl.utils.dict import DictDefault
-
-from .utils import most_recent_subdir, with_temp_dir
-
-LOG = logging.getLogger("axolotl.tests.e2e")
-os.environ["WANDB_DISABLED"] = "true"
-
-
-class TestPackedLlama(unittest.TestCase):
-    """
-    Test case for Packed training of llama models
-    """
-
-    @with_temp_dir
-    def test_loss_packed(self, temp_dir):
-        # pylint: disable=duplicate-code
-        cfg = DictDefault(
-            {
-                "base_model": "HuggingFaceTB/SmolLM-135M",
-                "sequence_len": 1024,
-                "sample_packing": True,
-                "flash_attention": True,
-                "val_set_size": 0.0,
-                "special_tokens": {
-                    "pad_token": "<|endoftext|>",
-                },
-                "datasets": [
-                    {
-                        "path": "vicgalle/alpaca-gpt4",
-                        "type": "alpaca",
-                    },
-                ],
-                "num_epochs": 1,
-                "micro_batch_size": 2,
-                "gradient_accumulation_steps": 4,
-                "output_dir": temp_dir,
-                "learning_rate": 0.00001,
-                "optimizer": "adamw_torch",
-                "lr_scheduler": "cosine",
-                "max_steps": 5,
-                "use_tensorboard": True,
-            }
-        )
-        if is_torch_bf16_gpu_available():
-            cfg.bf16 = True
-        else:
-            cfg.fp16 = True
-        normalize_config(cfg)
-        cli_args = TrainerCliArgs()
-        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
-
-        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
-
-        tb_log_path = most_recent_subdir(temp_dir + "/runs")
-        event_file = os.path.join(tb_log_path, sorted(os.listdir(tb_log_path))[0])
-        reader = SummaryReader(event_file)
-        df = reader.scalars  # pylint: disable=invalid-name
-        df = df[(df.tag == "train/train_loss")]  # pylint: disable=invalid-name
-        assert df.value.values[-1] < 2.0, "Loss is too high"

tests/e2e/test_reward_model_llama.py

@@ -1,74 +0,0 @@
-"""
-E2E tests for reward model lora llama
-"""
-
-import logging
-import os
-import unittest
-from pathlib import Path
-
-from axolotl.cli import load_datasets
-from axolotl.common.cli import TrainerCliArgs
-from axolotl.train import train
-from axolotl.utils.config import normalize_config
-from axolotl.utils.dict import DictDefault
-
-from .utils import with_temp_dir
-
-LOG = logging.getLogger("axolotl.tests.e2e")
-os.environ["WANDB_DISABLED"] = "true"
-
-
-class TestRewardModelLoraLlama(unittest.TestCase):
-    """
-    Test case for Llama reward models using LoRA
-    """
-
-    @with_temp_dir
-    def test_rm_fft(self, temp_dir):
-        # pylint: disable=duplicate-code
-        cfg = DictDefault(
-            {
-                "base_model": "JackFram/llama-68m",
-                "model_type": "AutoModelForSequenceClassification",
-                "tokenizer_type": "LlamaTokenizer",
-                "chat_template": "alpaca",
-                "reward_model": True,
-                "sequence_len": 1024,
-                "pad_to_sequence_len": True,
-                "adapter": "lora",
-                "lora_r": 8,
-                "lora_alpha": 16,
-                "lora_dropout": 0.05,
-                "lora_target_linear": True,
-                "val_set_size": 0.0,
-                "special_tokens": {
-                    "unk_token": "<unk>",
-                    "bos_token": "<s>",
-                    "eos_token": "</s>",
-                },
-                "datasets": [
-                    {
-                        "path": "argilla/distilabel-intel-orca-dpo-pairs",
-                        "type": "bradley_terry.chat_template",
-                    },
-                ],
-                "remove_unused_columns": False,
-                "max_steps": 10,
-                "num_epochs": 1,
-                "micro_batch_size": 4,
-                "gradient_accumulation_steps": 1,
-                "output_dir": temp_dir,
-                "learning_rate": 0.00001,
-                "optimizer": "adamw_bnb_8bit",
-                "lr_scheduler": "cosine",
-                "gradient_checkpointing": True,
-                "warmup_ratio": 0.1,
-            }
-        )
-        normalize_config(cfg)
-        cli_args = TrainerCliArgs()
-        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
-
-        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
-        assert (Path(temp_dir) / "adapter_model.bin").exists()

tests/prompt_strategies/messages/test_chat.py

@@ -1,62 +0,0 @@
-"""
-tests for chat_template prompt strategy
-"""
-# pylint: disable=duplicate-code
-import logging
-import unittest
-
-from axolotl.prompt_strategies.messages.chat import load
-from axolotl.utils.dict import DictDefault
-
-logging.basicConfig(level=logging.DEBUG)
-LOG = logging.getLogger("axolotl")
-
-
-class TestMessagesChatLlama3:
-    """
-    Test class for assistant style datasets with llama-3 prompts using the messages chat llama3 strategy.
-    """
-
-    def test_llama3_load(self, llama3_tokenizer, assistant_dataset):
-        LOG.info("Loading llama-3 tokenizer with assistant dataset")
-        strategy = load(
-            llama3_tokenizer,
-            DictDefault(
-                {
-                    "train_on_inputs": False,
-                    "sequence_len": 512,
-                }
-            ),
-            DictDefault(
-                {
-                    "chat_template": "llama3",
-                    "message_field_role": "role",
-                    "message_field_content": "content",
-                    "field_messages": "messages",
-                }
-            ),
-        )
-        res = strategy.wrap_dataset(assistant_dataset)
-        input_ids = res[0]["input_ids"]
-        # fmt: off
-        expected_input_ids = [
-            128000,  # bos
-            128006, 882, 128007,  # user header
-            271, 15339, 128009,  # user prompt eot
-            128006, 78191, 128007,  # assistant header
-            271, 15339, 128009,  # assistant response eot
-            128006, 882, 128007,
-            271, 19045, 29474, 128009,
-            128006, 78191, 128007,
-            271, 19045, 29474, 128009,
-        ]
-        # fmt: on
-        LOG.debug(f"Expected input_ids: {expected_input_ids}")
-        LOG.debug(f"Actual input_ids: {input_ids}")
-        assert (
-            input_ids == expected_input_ids
-        ), f"Input IDs mismatch: {input_ids} != {expected_input_ids}"
-
-
-if __name__ == "__main__":
-    unittest.main()

tests/prompt_strategies/test_chat_template_utils.py

@@ -1,125 +0,0 @@
-"""
-Tests for utils in axolotl.utils.chat_templates
-"""
-import unittest
-
-import pytest
-from transformers import AutoTokenizer
-
-from axolotl.utils.chat_templates import (
-    _CHAT_TEMPLATES,
-    extract_chat_template_args,
-    get_chat_template,
-)
-
-
-@pytest.fixture(name="llama3_tokenizer")
-def fixture_llama3_tokenizer():
-    tokenizer = AutoTokenizer.from_pretrained("NousResearch/Meta-Llama-3-8B")
-
-    return tokenizer
-
-
-class TestGetChatTemplateUtils:
-    """
-    Tests the get_chat_template function.
-    """
-
-    def test_known_chat_template(self):
-        chat_template_str = get_chat_template("llama3")
-        assert chat_template_str == _CHAT_TEMPLATES["llama3"]
-
-    def test_invalid_chat_template(self):
-        with pytest.raises(ValueError) as exc:
-            get_chat_template("invalid_template")
-            assert str(exc) == "Template 'invalid_template' not found."
-
-    def test_tokenizer_default_no_tokenizer(self):
-        with pytest.raises(ValueError):
-            get_chat_template("tokenizer_default", tokenizer=None)
-
-    def test_tokenizer_default_no_chat_template_on_tokenizer(self, llama3_tokenizer):
-        with pytest.raises(ValueError):
-            get_chat_template("tokenizer_default", tokenizer=llama3_tokenizer)
-
-    def test_tokenizer_default_with_chat_template_on_tokenizer(self, llama3_tokenizer):
-        llama3_tokenizer.chat_template = "test_template"
-        chat_template_str = get_chat_template(
-            "tokenizer_default", tokenizer=llama3_tokenizer
-        )
-        assert chat_template_str == "test_template"
-
-    def test_tokenizer_default_fallback_no_tokenizer(self):
-        with pytest.raises(ValueError):
-            get_chat_template("tokenizer_default_fallback_test", tokenizer=None)
-
-    def test_tokenizer_default_fallback_no_chat_template_on_tokenizer(
-        self, llama3_tokenizer
-    ):
-        chat_template_str = get_chat_template(
-            "tokenizer_default_fallback_chatml", tokenizer=llama3_tokenizer
-        )
-        assert chat_template_str == get_chat_template("chatml")
-
-    def test_tokenizer_default_fallback_with_chat_template_on_tokenizer(
-        self, llama3_tokenizer
-    ):
-        llama3_tokenizer.chat_template = "test_template"
-        chat_template_str = get_chat_template(
-            "tokenizer_default_fallback_chatml", tokenizer=llama3_tokenizer
-        )
-        assert chat_template_str == "test_template"
-
-    def test_jinja_template_mode(self):
-        jinja_template = "example_jinja_template"
-        chat_template_str = get_chat_template("jinja", jinja_template=jinja_template)
-        assert chat_template_str == jinja_template
-
-    def test_jinja_template_mode_no_jinja_template(self):
-        with pytest.raises(ValueError):
-            get_chat_template("jinja", jinja_template=None)
-
-    def test_extract_chat_template_args(self):
-        # No ds_cfg
-        chat_template_choice, chat_template_jinja = extract_chat_template_args(
-            cfg={"chat_template": "chatml"},
-        )
-        assert chat_template_choice == "chatml"
-        assert chat_template_jinja is None
-
-        # ds_cfg provided
-        chat_template_choice, chat_template_jinja = extract_chat_template_args(
-            cfg={
-                "chat_template": "jinja",
-                "chat_template_jinja": "global_jinja_template",
-            },
-            ds_cfg={"chat_template": "llama3", "chat_template_jinja": None},
-        )
-        assert chat_template_choice == "llama3"
-        assert chat_template_jinja is None
-
-        # ds_cfg provided with jinja template
-        chat_template_choice, chat_template_jinja = extract_chat_template_args(
-            cfg={"chat_template": "chatml", "chat_template_jinja": None},
-            ds_cfg={
-                "chat_template": "jinja",
-                "chat_template_jinja": "ds_jinja_template",
-            },
-        )
-        assert chat_template_choice == "jinja"
-        assert chat_template_jinja == "ds_jinja_template"
-
-        # ds_cfg provided with no chat_template
-        chat_template_choice, chat_template_jinja = extract_chat_template_args(
-            cfg={
-                "chat_template": "jinja",
-                "chat_template_jinja": "global_jinja_template",
-            },
-            ds_cfg={"chat_template": None, "chat_template_jinja": "ds_jinja_template"},
-        )
-        assert chat_template_choice == "jinja"
-        assert chat_template_jinja == "global_jinja_template"
-
-
-if __name__ == "__main__":
-    unittest.main()

tests/prompt_strategies/test_chat_templates.py

@@ -11,7 +11,7 @@ from axolotl.prompt_strategies.chat_template import (
     load,
 )
 from axolotl.prompters import IGNORE_TOKEN_ID
-from axolotl.utils.chat_templates import get_chat_template
+from axolotl.utils.chat_templates import chat_templates
 from axolotl.utils.dict import DictDefault
 
 logging.basicConfig(level=logging.DEBUG)
@@ -73,7 +73,7 @@ class TestAssistantChatTemplateLlama3:
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
                 llama3_tokenizer,
-                chat_template=get_chat_template("llama3"),
+                chat_template=chat_templates("llama3"),
                 message_field_role="role",
                 message_field_content="content",
                 roles={
@@ -113,7 +113,7 @@ class TestAssistantChatTemplateLlama3:
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
                 phi35_tokenizer,
-                chat_template=get_chat_template("phi_35"),
+                chat_template=chat_templates("phi_35"),
                 message_field_role="role",
                 message_field_content="content",
                 roles={
@@ -171,7 +171,7 @@ class TestAssistantChatTemplateLlama3:
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
                 llama3_tokenizer,
-                chat_template=get_chat_template("llama3"),
+                chat_template=chat_templates("llama3"),
                 message_field_role="role",
                 message_field_content="content",
                 message_field_training="training",
@@ -230,7 +230,7 @@ class TestSharegptChatTemplateLlama3:
         # pylint: disable=duplicate-code
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
-                llama3_tokenizer, chat_template=get_chat_template("llama3")
+                llama3_tokenizer, chat_template=chat_templates("llama3")
             ),
             tokenizer=llama3_tokenizer,
             train_on_inputs=False,
@@ -283,7 +283,7 @@ class TestSharegptChatTemplateLlama3:
         # pylint: disable=duplicate-code
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
-                llama3_tokenizer, chat_template=get_chat_template("llama3")
+                llama3_tokenizer, chat_template=chat_templates("llama3")
             ),
             tokenizer=llama3_tokenizer,
             train_on_inputs=False,
@@ -336,7 +336,7 @@ class TestSharegptChatTemplateLlama3:
         # pylint: disable=duplicate-code
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
-                llama3_tokenizer, chat_template=get_chat_template("llama3")
+                llama3_tokenizer, chat_template=chat_templates("llama3")
             ),
             tokenizer=llama3_tokenizer,
             train_on_inputs=False,

tests/prompt_strategies/test_chat_templates_advanced.py

@@ -12,7 +12,7 @@ from axolotl.prompt_strategies.chat_template import (
     ChatTemplateStrategy,
 )
 from axolotl.prompters import IGNORE_TOKEN_ID
-from axolotl.utils.chat_templates import get_chat_template
+from axolotl.utils.chat_templates import chat_templates
 
 logging.basicConfig(level=logging.DEBUG)
 LOG = logging.getLogger("axolotl")
@@ -35,7 +35,7 @@ class TestChatTemplateConfigurations:
         LOG.info("Testing with train_on_inputs=True")
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
-                llama3_tokenizer, chat_template=get_chat_template("llama3")
+                llama3_tokenizer, chat_template=chat_templates("llama3")
             ),
             tokenizer=llama3_tokenizer,
             train_on_inputs=True,
@@ -80,7 +80,7 @@ class TestChatTemplateConfigurations:
         LOG.info("Testing with train_on_inputs=False")
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
-                llama3_tokenizer, chat_template=get_chat_template("llama3")
+                llama3_tokenizer, chat_template=chat_templates("llama3")
             ),
             tokenizer=llama3_tokenizer,
             train_on_inputs=False,
@@ -123,7 +123,7 @@ class TestChatTemplateConfigurations:
         LOG.info("Testing roles_to_train with assistant only")
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
-                llama3_tokenizer, chat_template=get_chat_template("llama3")
+                llama3_tokenizer, chat_template=chat_templates("llama3")
             ),
             tokenizer=llama3_tokenizer,
             train_on_inputs=False,
@@ -151,7 +151,7 @@ class TestChatTemplateConfigurations:
         LOG.info("Testing roles_to_train with all roles")
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
-                llama3_tokenizer, chat_template=get_chat_template("llama3")
+                llama3_tokenizer, chat_template=chat_templates("llama3")
             ),
             tokenizer=llama3_tokenizer,
             train_on_inputs=True,
@@ -184,7 +184,7 @@ class TestChatTemplateConfigurations:
         LOG.info("Testing with empty roles_to_train")
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
-                llama3_tokenizer, chat_template=get_chat_template("llama3")
+                llama3_tokenizer, chat_template=chat_templates("llama3")
             ),
             tokenizer=llama3_tokenizer,
             train_on_inputs=False,
@@ -205,7 +205,7 @@ class TestChatTemplateConfigurations:
         LOG.info("Testing with train_on_eos='all'")
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
-                llama3_tokenizer, chat_template=get_chat_template("llama3")
+                llama3_tokenizer, chat_template=chat_templates("llama3")
             ),
             tokenizer=llama3_tokenizer,
             train_on_inputs=False,
@@ -232,7 +232,7 @@ class TestChatTemplateConfigurations:
         LOG.info("Testing with train_on_eos='turn'")
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
-                llama3_tokenizer, chat_template=get_chat_template("llama3")
+                llama3_tokenizer, chat_template=chat_templates("llama3")
             ),
             tokenizer=llama3_tokenizer,
             train_on_inputs=False,
@@ -282,7 +282,7 @@ class TestChatTemplateConfigurations:
         LOG.info("Testing with train_on_eos='last'")
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
-                llama3_tokenizer, chat_template=get_chat_template("llama3")
+                llama3_tokenizer, chat_template=chat_templates("llama3")
             ),
             tokenizer=llama3_tokenizer,
             train_on_inputs=False,
@@ -315,7 +315,7 @@ class TestChatTemplateConfigurations:
         LOG.info("Testing with train_on_eos='none'")
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
-                llama3_tokenizer, chat_template=get_chat_template("llama3")
+                llama3_tokenizer, chat_template=chat_templates("llama3")
             ),
             tokenizer=llama3_tokenizer,
             train_on_inputs=False,
@@ -343,7 +343,7 @@ class TestChatTemplateConfigurations:
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
                 llama3_tokenizer,
-                chat_template=get_chat_template("llama3"),
+                chat_template=chat_templates("llama3"),
                 drop_system_message=True,
             ),
             tokenizer=llama3_tokenizer,
@@ -371,7 +371,7 @@ class TestChatTemplateConfigurations:
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
                 llama3_tokenizer,
-                chat_template=get_chat_template("llama3"),
+                chat_template=chat_templates("llama3"),
                 roles=custom_roles,
             ),
             tokenizer=llama3_tokenizer,
@@ -424,7 +424,7 @@ class TestChatTemplateConfigurations:
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
                 llama3_tokenizer,
-                chat_template=get_chat_template("llama3"),
+                chat_template=chat_templates("llama3"),
                 message_field_training="train",
                 message_field_training_detail="train_detail",
             ),

tests/prompt_strategies/test_dpo_chat_templates.py

@@ -86,20 +86,6 @@ def fixture_llama3_tokenizer():
     return tokenizer
 
 
-@pytest.fixture(name="phi3_tokenizer")
-def fixture_phi3_tokenizer():
-    tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3-medium-128k-instruct")
-
-    return tokenizer
-
-
-@pytest.fixture(name="gemma_tokenizer")
-def fixture_gemma_tokenizer():
-    tokenizer = AutoTokenizer.from_pretrained("unsloth/gemma-2b-it", revision="703fb4a")
-
-    return tokenizer
-
-
 class TestAssistantDPOChatTemplateLlama3:
     """
     Test class for assistant style datasets with llama-3 prompts using the chat_template strategy.
@@ -113,7 +99,7 @@ class TestAssistantDPOChatTemplateLlama3:
                     "chat_template": "llama3",
                     "datasets": [
                         {
-                            "type": "chat_template",
+                            "chat_template": "llama3",
                         }
                     ],
                 }
@@ -138,7 +124,7 @@ class TestAssistantDPOChatTemplateLlama3:
                     "chat_template": "llama3",
                     "datasets": [
                         {
-                            "type": "chat_template",
+                            "chat_template": "llama3",
                             "field_messages": "conversation",
                             "field_chosen": "better",
                             "field_rejected": "worse",
@@ -166,65 +152,5 @@ class TestAssistantDPOChatTemplateLlama3:
         assert result["rejected"] == "party on<|eot_id|>"
 
 
-class TestAssistantDPOChatTemplatePhi3:
-    """
-    Test class for assistant style datasets with phi-3 prompts using the tokenizer's chat_template strategy.
-    """
-
-    def test_phi3_defaults(self, phi3_tokenizer, assistant_dataset):
-        # pylint: disable=duplicate-code
-        transform_fn = default(
-            DictDefault(
-                {
-                    "chat_template": "tokenizer_default",
-                    "datasets": [
-                        {
-                            "type": "chat_template",
-                        }
-                    ],
-                }
-            )
-        )
-        result = transform_fn(assistant_dataset[0], tokenizer=phi3_tokenizer)
-        assert result["prompt"] == (
-            "<|user|>\nhello<|end|>\n"
-            + "<|assistant|>\nhello<|end|>\n"
-            + "<|user|>\ngoodbye<|end|>\n"
-            + "<|assistant|>\n"
-        )
-        assert result["chosen"] == "goodbye<|end|>"
-        assert result["rejected"] == "party on<|end|>"
-
-
-class TestAssistantDPOChatTemplateGemma:
-    """
-    Test class for assistant style datasets with gemma prompts using the tokenizer's chat_template strategy.
-    """
-
-    def test_gemma_defaults(self, gemma_tokenizer, assistant_dataset):
-        # pylint: disable=duplicate-code
-        transform_fn = default(
-            DictDefault(
-                {
-                    "chat_template": "tokenizer_default",
-                    "datasets": [
-                        {
-                            "type": "chat_template",
-                        }
-                    ],
-                }
-            )
-        )
-        result = transform_fn(assistant_dataset[0], tokenizer=gemma_tokenizer)
-        assert result["prompt"] == (
-            "<bos><start_of_turn>user\nhello<end_of_turn>\n"
-            + "<start_of_turn>model\nhello<end_of_turn>\n"
-            + "<start_of_turn>user\ngoodbye<end_of_turn>\n"
-            + "<start_of_turn>model\n"
-        )
-        assert result["chosen"] == "goodbye<end_of_turn>"
-        assert result["rejected"] == "party on<end_of_turn>"
-
-
 if __name__ == "__main__":
     unittest.main()

tests/test_datasets.py

@@ -12,7 +12,6 @@ from huggingface_hub import snapshot_download
 from transformers import AutoTokenizer
 
 from axolotl.utils.data import load_tokenized_prepared_datasets
-from axolotl.utils.data.rl import load_prepare_dpo_datasets
 from axolotl.utils.dict import DictDefault
 
 
@@ -268,40 +267,6 @@ class TestDatasetPreparation(unittest.TestCase):
             assert "attention_mask" in dataset.features
             assert "labels" in dataset.features
 
-    def test_load_hub_with_dpo(self):
-        """Verify that processing dpo data from the hub works"""
-
-        cfg = DictDefault(
-            {
-                "tokenizer_config": "huggyllama/llama-7b",
-                "sequence_len": 1024,
-                "rl": "dpo",
-                "chat_template": "llama3",
-                "datasets": [
-                    {
-                        "path": "fozziethebeat/alpaca_messages_2k_dpo_test",
-                        "type": "chat_template.default",
-                        "chat_template": "llama3",
-                        "field_messages": "conversation",
-                        "field_chosen": "chosen",
-                        "field_rejected": "rejected",
-                        "message_field_role": "role",
-                        "message_field_content": "content",
-                        "roles": {
-                            "system": ["system"],
-                            "user": ["user"],
-                            "assistant": ["assistant"],
-                        },
-                    }
-                ],
-            }
-        )
-
-        train_dataset, _ = load_prepare_dpo_datasets(cfg)
-
-        assert len(train_dataset) == 1800
-        assert "conversation" in train_dataset.features
-
     def test_load_hub_with_revision(self):
         """Verify that processing data from the hub works with a specific revision"""
         with tempfile.TemporaryDirectory() as tmp_dir:
@@ -329,41 +294,6 @@ class TestDatasetPreparation(unittest.TestCase):
             assert "attention_mask" in dataset.features
             assert "labels" in dataset.features
 
-    def test_load_hub_with_revision_with_dpo(self):
-        """Verify that processing dpo data from the hub works with a specific revision"""
-
-        cfg = DictDefault(
-            {
-                "tokenizer_config": "huggyllama/llama-7b",
-                "sequence_len": 1024,
-                "rl": "dpo",
-                "chat_template": "llama3",
-                "datasets": [
-                    {
-                        "path": "fozziethebeat/alpaca_messages_2k_dpo_test",
-                        "type": "chat_template.default",
-                        "chat_template": "llama3",
-                        "revision": "ea82cff",
-                        "field_messages": "conversation",
-                        "field_chosen": "chosen",
-                        "field_rejected": "rejected",
-                        "message_field_role": "role",
-                        "message_field_content": "content",
-                        "roles": {
-                            "system": ["system"],
-                            "user": ["user"],
-                            "assistant": ["assistant"],
-                        },
-                    }
-                ],
-            }
-        )
-
-        train_dataset, _ = load_prepare_dpo_datasets(cfg)
-
-        assert len(train_dataset) == 1800
-        assert "conversation" in train_dataset.features
-
     def test_load_local_hub_with_revision(self):
         """Verify that a local copy of a hub dataset can be loaded with a specific revision"""
         with tempfile.TemporaryDirectory() as tmp_dir:

tests/test_validation.py

@@ -13,7 +13,6 @@ from axolotl.utils import is_comet_available
 from axolotl.utils.config import validate_config
 from axolotl.utils.config.models.input.v0_4_1 import AxolotlConfigWCapabilities
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.mlflow_ import setup_mlflow_env_vars
 from axolotl.utils.models import check_model_config
 from axolotl.utils.wandb_ import setup_wandb_env_vars
 
@@ -1433,58 +1432,3 @@ class TestValidationComet(BaseValidation):
 
         for key in comet_env.keys():
             os.environ.pop(key, None)
-
-
-class TestValidationMLflow(BaseValidation):
-    """
-    Validation test for MLflow
-    """
-
-    def test_hf_mlflow_artifacts_config_sets_env(self, minimal_cfg):
-        cfg = (
-            DictDefault(
-                {
-                    "hf_mlflow_log_artifacts": True,
-                }
-            )
-            | minimal_cfg
-        )
-
-        new_cfg = validate_config(cfg)
-
-        assert new_cfg.hf_mlflow_log_artifacts is True
-
-        # Check it's not already present in env
-        assert "HF_MLFLOW_LOG_ARTIFACTS" not in os.environ
-
-        setup_mlflow_env_vars(new_cfg)
-
-        assert os.environ.get("HF_MLFLOW_LOG_ARTIFACTS") == "true"
-
-        os.environ.pop("HF_MLFLOW_LOG_ARTIFACTS", None)
-
-    def test_mlflow_not_used_by_default(self, minimal_cfg):
-        cfg = DictDefault({}) | minimal_cfg
-
-        new_cfg = validate_config(cfg)
-
-        setup_mlflow_env_vars(new_cfg)
-
-        assert cfg.use_mlflow is not True
-
-        cfg = (
-            DictDefault(
-                {
-                    "mlflow_experiment_name": "foo",
-                }
-            )
-            | minimal_cfg
-        )
-
-        new_cfg = validate_config(cfg)
-
-        setup_mlflow_env_vars(new_cfg)
-
-        assert new_cfg.use_mlflow is True
-
-        os.environ.pop("MLFLOW_EXPERIMENT_NAME", None)

tests/test_validation_dataset.py

@@ -1,238 +0,0 @@
-"""Module for testing the validation module for the dataset config"""
-
-import warnings
-from typing import Optional
-
-import pytest
-
-from axolotl.utils.config import validate_config
-from axolotl.utils.config.models.input.v0_4_1 import ChatTemplate
-from axolotl.utils.dict import DictDefault
-
-warnings.filterwarnings("error")
-
-
-@pytest.fixture(name="minimal_cfg")
-def fixture_cfg():
-    return DictDefault(
-        {
-            "base_model": "TinyLlama/TinyLlama-1.1B-Chat-v0.6",
-            "learning_rate": 0.000001,
-            "micro_batch_size": 1,
-            "gradient_accumulation_steps": 1,
-        }
-    )
-
-
-# pylint: disable=too-many-public-methods (duplicate-code)
-class BaseValidation:
-    """
-    Base validation module to setup the log capture
-    """
-
-    _caplog: Optional[pytest.LogCaptureFixture] = None
-
-    @pytest.fixture(autouse=True)
-    def inject_fixtures(self, caplog):
-        self._caplog = caplog
-
-
-class TestValidationCheckDatasetConfig(BaseValidation):
-    """
-    Test the validation for the dataset config to ensure no correct parameters are dropped
-    """
-
-    def test_dataset_config_no_drop_param(self, minimal_cfg):
-        cfg = DictDefault(
-            minimal_cfg
-            | {
-                "datasets": [
-                    {
-                        "path": "LDJnr/Puffin",
-                        "type": "sharegpt",
-                        "conversation": "chatml",
-                        "shards": 10,
-                    }
-                ]
-            }
-        )
-
-        checked_cfg = validate_config(cfg)
-
-        def _check_config():
-            assert checked_cfg.datasets[0].path == cfg.datasets[0].path
-            assert checked_cfg.datasets[0].type == cfg.datasets[0].type
-            assert checked_cfg.datasets[0].conversation == cfg.datasets[0].conversation
-            assert checked_cfg.datasets[0].shards == cfg.datasets[0].shards
-
-        _check_config()
-
-        checked_cfg = validate_config(
-            cfg,
-            capabilities={
-                "bf16": "false",
-                "n_gpu": 1,
-                "compute_capability": "8.0",
-            },
-        )
-
-        _check_config()
-
-    def test_dataset_default_chat_template_no_drop_param(self, minimal_cfg):
-        cfg = DictDefault(
-            minimal_cfg
-            | {
-                "datasets": [
-                    {
-                        "path": "LDJnr/Puffin",
-                        "type": "chat_template",
-                        "field_messages": "conversations",
-                        "shards": 10,
-                        "message_field_role": "from",
-                        "message_field_content": "value",
-                    }
-                ],
-            }
-        )
-
-        checked_cfg = validate_config(cfg)
-
-        def _check_config():
-            assert checked_cfg.datasets[0].path == cfg.datasets[0].path
-            assert checked_cfg.datasets[0].type == cfg.datasets[0].type
-            assert checked_cfg.chat_template is None
-            assert (
-                checked_cfg.datasets[0].chat_template == ChatTemplate.tokenizer_default
-            )
-            assert (
-                checked_cfg.datasets[0].field_messages == cfg.datasets[0].field_messages
-            )
-            assert checked_cfg.datasets[0].shards == cfg.datasets[0].shards
-            assert (
-                checked_cfg.datasets[0].message_field_role
-                == cfg.datasets[0].message_field_role
-            )
-            assert (
-                checked_cfg.datasets[0].message_field_content
-                == cfg.datasets[0].message_field_content
-            )
-
-        _check_config()
-
-        checked_cfg = validate_config(
-            cfg,
-            capabilities={
-                "bf16": "false",
-                "n_gpu": 1,
-                "compute_capability": "8.0",
-            },
-        )
-
-        _check_config()
-
-    def test_dataset_partial_default_chat_template_no_drop_param(self, minimal_cfg):
-        cfg = DictDefault(
-            minimal_cfg
-            | {
-                "chat_template": "chatml",
-                "datasets": [
-                    {
-                        "path": "LDJnr/Puffin",
-                        "type": "chat_template",
-                        "field_messages": "conversations",
-                        "shards": 10,
-                        "message_field_role": "from",
-                        "message_field_content": "value",
-                    }
-                ],
-            }
-        )
-
-        checked_cfg = validate_config(cfg)
-
-        def _check_config():
-            assert checked_cfg.datasets[0].path == cfg.datasets[0].path
-            assert checked_cfg.datasets[0].type == cfg.datasets[0].type
-            assert checked_cfg.chat_template == ChatTemplate.chatml
-            assert (
-                checked_cfg.datasets[0].chat_template == ChatTemplate.tokenizer_default
-            )
-            assert (
-                checked_cfg.datasets[0].field_messages == cfg.datasets[0].field_messages
-            )
-            assert checked_cfg.datasets[0].shards == cfg.datasets[0].shards
-            assert (
-                checked_cfg.datasets[0].message_field_role
-                == cfg.datasets[0].message_field_role
-            )
-            assert (
-                checked_cfg.datasets[0].message_field_content
-                == cfg.datasets[0].message_field_content
-            )
-
-        _check_config()
-
-        checked_cfg = validate_config(
-            cfg,
-            capabilities={
-                "bf16": "false",
-                "n_gpu": 1,
-                "compute_capability": "8.0",
-            },
-        )
-
-        _check_config()
-
-    def test_dataset_chatml_chat_template_no_drop_param(self, minimal_cfg):
-        cfg = DictDefault(
-            minimal_cfg
-            | {
-                "chat_template": "chatml",
-                "datasets": [
-                    {
-                        "path": "LDJnr/Puffin",
-                        "type": "chat_template",
-                        "chat_template": "gemma",
-                        "field_messages": "conversations",
-                        "shards": 10,
-                        "message_field_role": "from",
-                        "message_field_content": "value",
-                    }
-                ],
-            }
-        )
-
-        checked_cfg = validate_config(cfg)
-
-        def _check_config():
-            assert checked_cfg.datasets[0].path == cfg.datasets[0].path
-            assert checked_cfg.datasets[0].type == cfg.datasets[0].type
-            assert checked_cfg.chat_template == cfg.chat_template
-            assert (
-                checked_cfg.datasets[0].chat_template == cfg.datasets[0].chat_template
-            )
-            assert (
-                checked_cfg.datasets[0].field_messages == cfg.datasets[0].field_messages
-            )
-            assert checked_cfg.datasets[0].shards == cfg.datasets[0].shards
-            assert (
-                checked_cfg.datasets[0].message_field_role
-                == cfg.datasets[0].message_field_role
-            )
-            assert (
-                checked_cfg.datasets[0].message_field_content
-                == cfg.datasets[0].message_field_content
-            )
-
-        _check_config()
-
-        checked_cfg = validate_config(
-            cfg,
-            capabilities={
-                "bf16": "false",
-                "n_gpu": 1,
-                "compute_capability": "8.0",
-            },
-        )
-
-        _check_config()

tests/utils/test_models.py

@@ -1,64 +1,18 @@
 """Module for testing models utils file."""
 
-from unittest.mock import MagicMock, patch
+
+import unittest
+from unittest.mock import patch
 
 import pytest
-from transformers import BitsAndBytesConfig, PreTrainedTokenizerBase
-from transformers.integrations.deepspeed import is_deepspeed_zero3_enabled
-from transformers.utils.import_utils import is_torch_mps_available
 
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.models import ModelLoader, load_model
+from axolotl.utils.models import load_model
 
 
-class TestModelsUtils:
+class ModelsUtilsTest(unittest.TestCase):
     """Testing module for models utils."""
 
-    def setup_method(self) -> None:
-        # load config
-        self.cfg = DictDefault(  # pylint: disable=attribute-defined-outside-init
-            {
-                "base_model": "JackFram/llama-68m",
-                "model_type": "LlamaForCausalLM",
-                "tokenizer_type": "LlamaTokenizer",
-                "load_in_8bit": True,
-                "load_in_4bit": False,
-                "adapter": "lora",
-                "flash_attention": False,
-                "sample_packing": True,
-                "device_map": "auto",
-            }
-        )
-        self.tokenizer = MagicMock(  # pylint: disable=attribute-defined-outside-init
-            spec=PreTrainedTokenizerBase
-        )
-        self.inference = False  # pylint: disable=attribute-defined-outside-init
-        self.reference_model = True  # pylint: disable=attribute-defined-outside-init
-
-        # init ModelLoader
-        self.model_loader = (  # pylint: disable=attribute-defined-outside-init
-            ModelLoader(
-                cfg=self.cfg,
-                tokenizer=self.tokenizer,
-                inference=self.inference,
-                reference_model=self.reference_model,
-            )
-        )
-
-    def test_set_device_map_config(self):
-        # check device_map
-        device_map = self.cfg.device_map
-        if is_torch_mps_available():
-            device_map = "mps"
-        self.model_loader.set_device_map_config()
-        if is_deepspeed_zero3_enabled():
-            assert "device_map" not in self.model_loader.model_kwargs
-        else:
-            assert device_map in self.model_loader.model_kwargs["device_map"]
-
-        # check torch_dtype
-        assert self.cfg.torch_dtype == self.model_loader.model_kwargs["torch_dtype"]
-
     def test_cfg_throws_error_with_s2_attention_and_sample_packing(self):
         cfg = DictDefault(
             {
@@ -81,38 +35,3 @@ class TestModelsUtils:
                 "shifted-sparse attention does not currently support sample packing"
                 in str(exc.value)
             )
-
-    @pytest.mark.parametrize("adapter", ["lora", "qlora", None])
-    @pytest.mark.parametrize("load_in_8bit", [True, False])
-    @pytest.mark.parametrize("load_in_4bit", [True, False])
-    @pytest.mark.parametrize("gptq", [True, False])
-    def test_set_quantization_config(
-        self,
-        adapter,
-        load_in_8bit,
-        load_in_4bit,
-        gptq,
-    ):
-        # init cfg as args
-        self.cfg.load_in_8bit = load_in_8bit
-        self.cfg.load_in_4bit = load_in_4bit
-        self.cfg.gptq = gptq
-        self.cfg.adapter = adapter
-
-        self.model_loader.set_quantization_config()
-        if "quantization_config" in self.model_loader.model_kwargs or self.cfg.gptq:
-            assert not (
-                hasattr(self.model_loader.model_kwargs, "load_in_8bit")
-                and hasattr(self.model_loader.model_kwargs, "load_in_4bit")
-            )
-        elif load_in_8bit and self.cfg.adapter is not None:
-            assert self.model_loader.model_kwargs["load_in_8bit"]
-        elif load_in_4bit and self.cfg.adapter is not None:
-            assert self.model_loader.model_kwargs["load_in_4bit"]
-
-        if (self.cfg.adapter == "qlora" and load_in_4bit) or (
-            self.cfg.adapter == "lora" and load_in_8bit
-        ):
-            assert self.model_loader.model_kwargs.get(
-                "quantization_config", BitsAndBytesConfig
-            )