add colab callback to fix inference post train

fix error with lambda and pickling

make sure things are in float instead of np.float

.github/workflows/tests-nightly.yml

@@ -18,96 +18,9 @@ jobs:
         env:
           SKIP: no-commit-to-branch
 
-  preload-cache:
-    name: Preload HF cache
-    runs-on: ubuntu-latest
-    strategy:
-      fail-fast: false
-      matrix:
-        python_version: ["3.11"]
-        pytorch_version: ["2.6.0"]
-    timeout-minutes: 20
-
-    env:
-      AXOLOTL_IS_CI_CACHE_PRELOAD: "1"
-
-    steps:
-      - name: Check out repository code
-        uses: actions/checkout@v4
-
-      - name: Restore HF cache
-        id: hf-cache-restore
-        uses: actions/cache/restore@v4
-        with:
-          path: |
-            /home/runner/.cache/huggingface/hub/datasets--*
-            /home/runner/.cache/huggingface/hub/models--*
-          key: ${{ runner.os }}-hf-hub-cache-v2
-
-      - name: Setup Python
-        uses: actions/setup-python@v5
-        with:
-          python-version: ${{ matrix.python_version }}
-          cache: 'pip' # caching pip dependencies
-
-      - name: upgrade pip
-        run: |
-          pip3 install --upgrade pip
-          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
-
-      - name: Install PyTorch
-        run: |
-          pip3 install torch==${{ matrix.pytorch_version }}
-
-      - name: Install dependencies
-        run: |
-          pip3 show torch
-          pip3 install --no-build-isolation -U -e .
-          python scripts/unsloth_install.py | sh
-          python scripts/cutcrossentropy_install.py | sh
-          pip3 install -r requirements-dev.txt -r requirements-tests.txt
-
-      - name: Make sure PyTorch version wasn't clobbered
-        run: |
-          python -c "import torch; assert '${{ matrix.pytorch_version }}' in torch.__version__"
-
-      - name: Ensure axolotl CLI was installed
-        run: |
-          axolotl --help
-
-      - name: Pre-Download dataset fixture
-        run: |
-          huggingface-cli download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
-
-      - name: Run tests
-        run: |
-          pytest -v tests/conftest.py
-
-      - name: Upload coverage to Codecov
-        uses: codecov/codecov-action@v5
-        with:
-          token: ${{ secrets.CODECOV_TOKEN }}
-          files: ./coverage.xml
-          flags: unittests,pytorch-${{ matrix.pytorch_version }}
-          fail_ci_if_error: false
-
-      - name: cleanup pip cache
-        run: |
-          find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
-
-      - name: Save HF cache
-        id: hf-cache
-        uses: actions/cache/save@v4
-        with:
-          path: |
-            /home/runner/.cache/huggingface/hub/datasets--*
-            /home/runner/.cache/huggingface/hub/models--*
-          key: ${{ steps.hf-cache-restore.outputs.cache-primary-key }}
-
   pytest:
     name: PyTest
     runs-on: ubuntu-latest
-    needs: [preload-cache]
     strategy:
       fail-fast: false
       max-parallel: 2

.github/workflows/tests.yml

@@ -44,98 +44,12 @@ jobs:
         env:
           SKIP: no-commit-to-branch
 
-  preload-cache:
-    name: Preload HF cache
-    runs-on: ubuntu-latest
-    strategy:
-      fail-fast: false
-      matrix:
-        python_version: ["3.11"]
-        pytorch_version: ["2.6.0"]
-    timeout-minutes: 20
-
-    env:
-      AXOLOTL_IS_CI_CACHE_PRELOAD: "1"
-
-    steps:
-      - name: Check out repository code
-        uses: actions/checkout@v4
-
-      - name: Restore HF cache
-        id: hf-cache-restore
-        uses: actions/cache/restore@v4
-        with:
-          path: |
-            /home/runner/.cache/huggingface/hub/datasets--*
-            /home/runner/.cache/huggingface/hub/models--*
-          key: ${{ runner.os }}-hf-hub-cache-v2
-
-      - name: Setup Python
-        uses: actions/setup-python@v5
-        with:
-          python-version: ${{ matrix.python_version }}
-          cache: 'pip' # caching pip dependencies
-
-      - name: upgrade pip
-        run: |
-          pip3 install --upgrade pip
-          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
-
-      - name: Install PyTorch
-        run: |
-          pip3 install torch==${{ matrix.pytorch_version }}
-
-      - name: Install dependencies
-        run: |
-          pip3 show torch
-          pip3 install --no-build-isolation -U -e .
-          python scripts/unsloth_install.py | sh
-          python scripts/cutcrossentropy_install.py | sh
-          pip3 install -r requirements-dev.txt -r requirements-tests.txt
-
-      - name: Make sure PyTorch version wasn't clobbered
-        run: |
-          python -c "import torch; assert '${{ matrix.pytorch_version }}' in torch.__version__"
-
-      - name: Ensure axolotl CLI was installed
-        run: |
-          axolotl --help
-
-      - name: Pre-Download dataset fixture
-        run: |
-          huggingface-cli download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
-
-      - name: Run tests
-        run: |
-          pytest -v tests/conftest.py
-
-      - name: Upload coverage to Codecov
-        uses: codecov/codecov-action@v5
-        with:
-          token: ${{ secrets.CODECOV_TOKEN }}
-          files: ./coverage.xml
-          flags: unittests,pytorch-${{ matrix.pytorch_version }}
-          fail_ci_if_error: false
-
-      - name: cleanup pip cache
-        run: |
-          find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
-
-      - name: Save HF cache
-        id: hf-cache
-        uses: actions/cache/save@v4
-        with:
-          path: |
-            /home/runner/.cache/huggingface/hub/datasets--*
-            /home/runner/.cache/huggingface/hub/models--*
-          key: ${{ steps.hf-cache-restore.outputs.cache-primary-key }}
-
   pytest:
     name: PyTest
     runs-on: ubuntu-latest
-    needs: [preload-cache]
     strategy:
       fail-fast: false
+      max-parallel: 2
       matrix:
         python_version: ["3.11"]
         pytorch_version: ["2.5.1", "2.6.0", "2.7.0"]
@@ -207,12 +121,21 @@ jobs:
         run: |
           find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
 
+      - name: Save HF cache
+        id: hf-cache
+        uses: actions/cache/save@v4
+        with:
+          path: |
+            /home/runner/.cache/huggingface/hub/datasets--*
+            /home/runner/.cache/huggingface/hub/models--*
+          key: ${{ steps.hf-cache-restore.outputs.cache-primary-key }}
+
   pytest-sdist:
     name: PyTest from Source Dist
     runs-on: ubuntu-latest
-    needs: [preload-cache]
     strategy:
       fail-fast: false
+      max-parallel: 1
       matrix:
         python_version: ["3.11"]
         pytorch_version: ["2.5.1", "2.6.0", "2.7.0"]
@@ -276,6 +199,15 @@ jobs:
         run: |
           find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
 
+      - name: Save HF cache
+        id: hf-cache
+        uses: actions/cache/save@v4
+        with:
+          path: |
+            /home/runner/.cache/huggingface/hub/datasets--*
+            /home/runner/.cache/huggingface/hub/models--*
+          key: ${{ steps.hf-cache-restore.outputs.cache-primary-key }}
+
   docker-e2e-tests-1st:
     if: ${{ ! contains(github.event.commits[0].message, '[skip e2e]') && github.repository_owner == 'axolotl-ai-cloud' }}
     # this job needs to be run on self-hosted GPU runners...

docs/config.qmd

@@ -32,8 +32,6 @@ tokenizer_legacy:
 resize_token_embeddings_to_32x:
 # Optional[bool] Whether to shrink the embeddings to len(tokenizer). By default, we won't shrink.
 shrink_embeddings:
-# Optional[bool] Don't upcast the embeddings to float32 when using PEFT. Useful for low-VRAM GPUs
-embeddings_skip_upcast:
 # Whether to load the model with randomly initialized weights. Useful for
 # pre-training a model from scratch or debugging purposes.
 random_init_weights:
@@ -75,12 +73,11 @@ load_in_8bit: true
 load_in_4bit:
 
 # Use CUDA bf16
-bf16: true # bool or 'full' for `bf16_full_eval`, or 'auto' for automatic detection. require >=ampere
+bf16: true # bool or 'full' for `bf16_full_eval`. require >=ampere
 # Use CUDA fp16
 fp16: true
 # Use CUDA tf32
 tf32: true # require >=ampere
-# Note: if bf16 is set to 'auto', and fp16 is set to true, we will prefer the explict fp16 setting
 
 # No AMP (automatic mixed precision)
 bfloat16: true # require >=ampere
@@ -187,8 +184,8 @@ datasets:
     # adding a system turn with empty content.
     drop_system_message:
 
-    # Optional[bool]. (for Qwen3 template only) Whether to split the assistant content based on a reasoning trace inside delimited tags
+    # Optional[bool]. Whether to split the assistant turn based on a reasoning trace inside delimited tags
-    # See example at `docs/dataset-formats/conversation.qmd`
+    # defaults to False
     split_thinking:
 
     # IMPORTANT: The following fields determine which parts of the conversation to train on.
@@ -550,7 +547,7 @@ gradient_checkpointing: false
 early_stopping_patience: 3
 
 # Specify a scheduler and kwargs to use with the optimizer
-lr_scheduler: # 'one_cycle' | 'rex' | 'log_sweep' | 'linear' | 'cosine_with_restarts' | 'polynomial' | 'constant' | 'constant_with_warmup' | 'inverse_sqrt' | 'reduce_lr_on_plateau' | 'cosine_with_min_lr' | 'warmup_stable_decay' | empty for cosine
+lr_scheduler: # 'one_cycle' | 'rex' | 'log_sweep' | empty for cosine
 lr_scheduler_kwargs:
 cosine_min_lr_ratio: # decay lr to some percentage of the peak lr, e.g. cosine_min_lr_ratio=0.1 for 10% of peak lr
 cosine_constant_lr_ratio: # freeze lr at some percentage of the step, e.g. cosine_constant_lr_ratio=0.8 means start cosine_min_lr at 80% of training step (https://arxiv.org/pdf/2308.04014.pdf)
@@ -612,7 +609,6 @@ lr_div_factor: # Learning rate div factor
 # - optimi_adamw
 # - ao_adamw_8bit
 # - ao_adamw_fp8
-# - came_pytorch
 optimizer:
 # Dictionary of arguments to pass to the optimizer
 optim_args:

docs/dataset-formats/conversation.qmd

@@ -196,34 +196,6 @@ datasets:
 It is not necessary to set both `message_field_training` and `message_field_training_detail` at once.
 :::
 
-8. (For Qwen3 template only) Enable reasoning split, where the reasoning is split from the content and passed as a separate field into the template.
-
-```yaml
-datasets:
-  - path: ...
-    type: chat_template
-    chat_template: qwen3
-    split_thinking: true
-```
-
-For example, a content can look like:
-
-```json
-{
-  "content": "<think>Some thinking outputs</think>Output after thinking."
-}
-```
-
-After split, it will look like:
-
-```json
-{
-  "reasoning_content": "Some thinking outputs",
-  "content": "Output after thinking..."
-}
-```
-
-
 ## sharegpt
 
 ::: {.callout-important}

examples/llama-4/README.md

@@ -34,5 +34,3 @@ We provide a script to delinearize Llama 4 linearized models into regular Huggin
 ```bash
 axolotl delinearize-llama4 --model path/to/model_dir --output path/to/output_dir
 ```
-
-Note: This only works with the non-quantized linearized model. If you have an adapter, merge it with the *non-quantized linearized* model before delinearizing.

examples/orpheus/README.md

@@ -1,341 +0,0 @@
-# Finetuning LLMs to output audio
-
-In this example, we finetune Orpcanopylabs/orpheus-tts-0.1-pretrained (a LLaMA 3.2 3b model) to output audio.
-
-The `finetune.yml` withe current settings will run on any Nvidia GPU with 45GB VRAM or more. If you adjust the batch size it can easily run on any GPU under 24GB.
-
-## Dataset pre-processing for pre-training
-If you are adding another voice in English, please jump ahead to finetuning pre-processing.
-
-For this to work, we need to preprocess our dataset. Since we are expecting to output audio, we will need to add tokens to the tokenizer.
-
-Using this code, it will download the SNAC model and add the correct tokens and upload the final dataset.
-
-```python
-import torch
-from snac import SNAC
-from datasets import load_dataset
-from huggingface_hub import snapshot_download
-from datasets import load_dataset
-import random
-import torchaudio.transforms as T
-from transformers import AutoTokenizer
-import os
-
-my_original_dataset_name = "<huggingface-id-of-dataset-that-we-want-to-preprocess>"
-name_to_push_dataset_to = "<huggingface-id-of-where-to-save-dataset>"
-
-dsn = my_original_dataset_name
-
-snapshot_download(
-    repo_id=dsn,
-    repo_type="dataset",
-    revision="main",
-    max_workers=64,
-)
-
-
-ds = load_dataset(dsn, split="train")
-ds_sample_rate = ds[0]["audio"]["sampling_rate"]
-
-model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz")
-model = model.to("mps")
-
-def tokenise_audio(waveform):
-  waveform = torch.from_numpy(waveform).unsqueeze(0)
-  waveform = waveform.to(dtype=torch.float32)
-  resample_transform = T.Resample(orig_freq=ds_sample_rate, new_freq=24000)
-  waveform = resample_transform(waveform)
-
-  waveform = waveform.unsqueeze(0).to("cuda")
-
-  #generate the codes from snac
-  with torch.inference_mode():
-    codes = model.encode(waveform)
-
-  all_codes = []
-  for i in range(codes[0].shape[1]):
-    all_codes.append(codes[0][0][i].item()+128266)
-    all_codes.append(codes[1][0][2*i].item()+128266+4096)
-    all_codes.append(codes[2][0][4*i].item()+128266+(2*4096))
-    all_codes.append(codes[2][0][(4*i)+1].item()+128266+(3*4096))
-    all_codes.append(codes[1][0][(2*i)+1].item()+128266+(4*4096))
-    all_codes.append(codes[2][0][(4*i)+2].item()+128266+(5*4096))
-    all_codes.append(codes[2][0][(4*i)+3].item()+128266+(6*4096))
-
-
-  return all_codes
-
-def add_codes(example):
-    # Always initialize codes_list to None
-    codes_list = None
-
-    try:
-        answer_audio = example.get("audio")
-        # If there's a valid audio array, tokenise it
-        if answer_audio and "array" in answer_audio:
-            audio_array = answer_audio["array"]
-            codes_list = tokenise_audio(audio_array)
-    except Exception as e:
-        print(f"Skipping row due to error: {e}")
-        # Keep codes_list as None if we fail
-    example["codes_list"] = codes_list
-
-    return example
-
-ds = ds.map(add_codes, remove_columns=["audio"])
-
-#@title Load Tokenizer
-tokeniser_length = 128256
-start_of_text = 128000
-end_of_text = 128009
-
-start_of_speech = tokeniser_length + 1
-end_of_speech = tokeniser_length + 2
-
-start_of_human = tokeniser_length + 3
-end_of_human = tokeniser_length + 4
-
-start_of_ai = tokeniser_length + 5
-end_of_ai =  tokeniser_length + 6
-pad_token = tokeniser_length + 7
-
-audio_tokens_start = tokeniser_length + 10
-
-tokenizer_name = "canopylabs/orpheus-3b-0.1-pretrained"
-
-
-tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
-num_proc = os.cpu_count() - 2
-
-ds = ds.filter(lambda x: x["codes_list"] is not None)
-ds = ds.filter(lambda x: len(x["codes_list"]) > 0)
-
-#@title Create Input Ids
-def remove_duplicate_frames(example):
-    vals = example["codes_list"]
-    if len(vals) % 7 != 0:
-        raise ValueError("Input list length must be divisible by 7")
-
-    result = vals[:7]
-
-    removed_frames = 0
-
-    for i in range(7, len(vals), 7):
-        current_first = vals[i]
-        previous_first = result[-7]
-
-        if current_first != previous_first:
-            result.extend(vals[i:i+7])
-        else:
-            removed_frames += 1
-
-    example["codes_list"] = result
-
-    return example
-
-ds = ds.map(remove_duplicate_frames, num_proc=num_proc)
-
-
-def create_input_ids(example):
-    text_ids = tokenizer.encode({example['text']},  add_special_tokens=True)
-    text_ids.append(end_of_text)
-    example["text_tokens"] = text_ids
-    input_ids = (
-        [start_of_human]
-        + example["text_tokens"]
-        + [end_of_human]
-        + [start_of_ai]
-        + [start_of_speech]
-        + example["codes_list"]
-        + [end_of_speech]
-        + [end_of_ai]
-    )
-    example["input_ids"] = input_ids
-    example["labels"] = input_ids
-    example["attention_mask"] = [1] * len(input_ids)
-
-    return example
-
-ds = ds.map(create_input_ids, num_proc=num_proc, remove_columns=["text", "codes_list"])
-
-#@title Remove unnecessary columns
-columns_to_keep = ["input_ids", "labels", "attention_mask"]
-columns_to_remove = [col for col in ds.column_names if col not in columns_to_keep]
-
-ds = ds.remove_columns(columns_to_remove)
-
-ds.push_to_hub(name_to_push_dataset_to)
-```
-
-
-## Finetune pre-processing
-Use this code to add a new voice.
-
-```python
-import torch
-from snac import SNAC
-from datasets import load_dataset
-from huggingface_hub import snapshot_download
-from datasets import load_dataset
-import random
-import torchaudio.transforms as T
-from transformers import AutoTokenizer
-import os
-
-my_original_dataset_name = "<huggingface-id-of-dataset-that-we-want-to-preprocess>"
-name_to_push_dataset_to = "<huggingface-id-of-where-to-save-dataset>"
-
-dsn = my_original_dataset_name
-
-snapshot_download(
-    repo_id=dsn,
-    repo_type="dataset",
-    revision="main",
-    max_workers=64,
-)
-
-
-ds = load_dataset(dsn, split="train")
-ds_sample_rate = ds[0]["audio"]["sampling_rate"]
-
-model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz")
-model = model.to("mps")
-
-def tokenise_audio(waveform):
-  waveform = torch.from_numpy(waveform).unsqueeze(0)
-  waveform = waveform.to(dtype=torch.float32)
-  resample_transform = T.Resample(orig_freq=ds_sample_rate, new_freq=24000)
-  waveform = resample_transform(waveform)
-
-  waveform = waveform.unsqueeze(0).to("cuda")
-
-  #generate the codes from snac
-  with torch.inference_mode():
-    codes = model.encode(waveform)
-
-  all_codes = []
-  for i in range(codes[0].shape[1]):
-    all_codes.append(codes[0][0][i].item()+128266)
-    all_codes.append(codes[1][0][2*i].item()+128266+4096)
-    all_codes.append(codes[2][0][4*i].item()+128266+(2*4096))
-    all_codes.append(codes[2][0][(4*i)+1].item()+128266+(3*4096))
-    all_codes.append(codes[1][0][(2*i)+1].item()+128266+(4*4096))
-    all_codes.append(codes[2][0][(4*i)+2].item()+128266+(5*4096))
-    all_codes.append(codes[2][0][(4*i)+3].item()+128266+(6*4096))
-
-
-  return all_codes
-
-def add_codes(example):
-    # Always initialize codes_list to None
-    codes_list = None
-
-    try:
-        answer_audio = example.get("audio")
-        # If there's a valid audio array, tokenise it
-        if answer_audio and "array" in answer_audio:
-            audio_array = answer_audio["array"]
-            codes_list = tokenise_audio(audio_array)
-    except Exception as e:
-        print(f"Skipping row due to error: {e}")
-        # Keep codes_list as None if we fail
-    example["codes_list"] = codes_list
-
-    return example
-
-ds = ds.map(add_codes, remove_columns=["audio"])
-
-#@title Load Tokenizer
-tokeniser_length = 128256
-start_of_text = 128000
-end_of_text = 128009
-
-start_of_speech = tokeniser_length + 1
-end_of_speech = tokeniser_length + 2
-
-start_of_human = tokeniser_length + 3
-end_of_human = tokeniser_length + 4
-
-start_of_ai = tokeniser_length + 5
-end_of_ai =  tokeniser_length + 6
-pad_token = tokeniser_length + 7
-
-audio_tokens_start = tokeniser_length + 10
-
-tokenizer_name = "canopylabs/orpheus-3b-0.1-pretrained"
-
-
-tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
-num_proc = os.cpu_count() - 2
-
-ds = ds.filter(lambda x: x["codes_list"] is not None)
-ds = ds.filter(lambda x: len(x["codes_list"]) > 0)
-
-#@title Create Input Ids
-def remove_duplicate_frames(example):
-    vals = example["codes_list"]
-    if len(vals) % 7 != 0:
-        raise ValueError("Input list length must be divisible by 7")
-
-    result = vals[:7]
-
-    removed_frames = 0
-
-    for i in range(7, len(vals), 7):
-        current_first = vals[i]
-        previous_first = result[-7]
-
-        if current_first != previous_first:
-            result.extend(vals[i:i+7])
-        else:
-            removed_frames += 1
-
-    example["codes_list"] = result
-
-    return example
-
-ds = ds.map(remove_duplicate_frames, num_proc=num_proc)
-
-tok_info = '''*** HERE you can modify the text prompt
-i.e. if you wanted a multispeaker model like canopylabs/orpheus-3b-0.1-ft, you can pass:
-f"{example["source"]}:  {example["text"]}", as is passed.
-'''
-print(tok_info)
-
-def create_input_ids(example):
-    text_ids = tokenizer.encode(f"{example['speaker_id']}: {example['text']}",  add_special_tokens=True)
-    text_ids.append(end_of_text)
-    example["text_tokens"] = text_ids
-    input_ids = (
-        [start_of_human]
-        + example["text_tokens"]
-        + [end_of_human]
-        + [start_of_ai]
-        + [start_of_speech]
-        + example["codes_list"]
-        + [end_of_speech]
-        + [end_of_ai]
-    )
-    example["input_ids"] = input_ids
-    example["labels"] = input_ids
-    example["attention_mask"] = [1] * len(input_ids)
-
-    return example
-
-ds = ds.map(create_input_ids, num_proc=num_proc, remove_columns=["text", "codes_list"])
-
-#@title Remove unnecessary columns
-columns_to_keep = ["input_ids", "labels", "attention_mask"]
-columns_to_remove = [col for col in ds.column_names if col not in columns_to_keep]
-
-ds = ds.remove_columns(columns_to_remove)
-
-ds.push_to_hub(name_to_push_dataset_to)
-```
-
-## Training
-After preprocessing is done, fill out the blanks in finetune.yml and simply run `axolotl train finetune.yml`
-
-## Inference
-For inference, please refer to the original [orpheus github](https://github.com/canopyai/Orpheus-TTS/tree/main).

examples/orpheus/finetune.yml

@@ -1,52 +0,0 @@
-base_model: canopylabs/orpheus-3b-0.1-pretrained
-
-hub_model_id: <your-hub-model-id>
-
-plugins:
-  - axolotl.integrations.liger.LigerPlugin
-liger_rope: true
-liger_rms_norm: true
-liger_glu_activation: true
-liger_fused_linear_cross_entropy: true
-
-datasets:
-  - path: <your-hf-dataset-id>
-    type:  # leave empty to load pre-tokenized
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.01
-output_dir: ./outputs/out
-
-sequence_len: 8192
-sample_packing: true
-pad_to_sequence_len: true
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 8
-micro_batch_size: 4
-num_epochs: 3
-optimizer: adamw_torch_fused
-lr_scheduler: cosine
-learning_rate: 2e-5
-
-bf16: auto
-tf32: false
-
-gradient_checkpointing: true
-gradient_checkpointing_kwargs:
-  use_reentrant: false
-resume_from_checkpoint:
-logging_steps: 1
-flash_attention: true
-
-warmup_steps: 20
-evals_per_epoch: 5
-saves_per_epoch: 5
-weight_decay: 0.05
-
-special_tokens:
-  pad_token: <custom_token_7>

requirements.txt

@@ -6,17 +6,16 @@ triton>=3.0.0
 mamba-ssm==1.2.0.post1
 xformers>=0.0.23.post1
 autoawq==0.2.7.post3
-liger-kernel==0.5.9
+liger-kernel==0.5.8
 # END section
 
 packaging==23.2
 
-huggingface_hub==0.31.0
 peft==0.15.2
 transformers==4.51.3
 tokenizers>=0.21.1
 accelerate==1.6.0
-datasets==3.5.1
+datasets==3.5.0
 deepspeed>=0.15.4
 trl==0.17.0
 hf_xet==1.1.0

setup.py

@@ -67,13 +67,13 @@ def parse_requirements(extras_require_map):
             if (major, minor) >= (2, 7):
                 _install_requires.pop(_install_requires.index(xformers_version))
                 # _install_requires.append("xformers==0.0.29.post3")  # xformers seems to be hard pinned to 2.6.0
-                extras_require_map["vllm"] = ["vllm==0.8.5.post1"]
+                extras_require_map["vllm"] = ["vllm==0.8.5"]
             elif (major, minor) >= (2, 6):
                 _install_requires.pop(_install_requires.index(xformers_version))
                 _install_requires.append(
                     "xformers==0.0.29.post2"
                 )  # vllm needs post2 w torch 2.6
-                extras_require_map["vllm"] = ["vllm==0.8.5.post1"]
+                extras_require_map["vllm"] = ["vllm==0.8.5"]
             elif (major, minor) >= (2, 5):
                 _install_requires.pop(_install_requires.index(xformers_version))
                 if patch == 0:
@@ -142,7 +142,6 @@ extras_require = {
         "apollo-torch",
         "lomo-optim==0.1.1",
         "torch-optimi==0.2.1",
-        "came_pytorch==0.1.3",
     ],
     "ray": [
         "ray[train]",

src/axolotl/cli/preprocess.py

@@ -18,7 +18,6 @@ from axolotl.cli.checks import check_accelerate_default_config, check_user_token
 from axolotl.cli.config import load_cfg
 from axolotl.common.const import DEFAULT_DATASET_PREPARED_PATH
 from axolotl.common.datasets import load_datasets, load_preference_datasets
-from axolotl.integrations.base import PluginManager
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.trainer import disable_datasets_caching
 
@@ -48,10 +47,7 @@ def do_preprocess(cfg: DictDefault, cli_args: PreprocessCliArgs) -> None:
         cfg.dataset_prepared_path = DEFAULT_DATASET_PREPARED_PATH
 
     with disable_datasets_caching():
-        plugin_manager = PluginManager.get_instance()
+        if cfg.rl:
-        if plugin_manager.load_datasets(cfg, preprocess=True):
-            pass
-        elif cfg.rl:
             load_preference_datasets(cfg=cfg, cli_args=cli_args)
         else:
             load_datasets(cfg=cfg, cli_args=cli_args)

src/axolotl/cli/train.py

@@ -43,13 +43,10 @@ def do_train(cfg: DictDefault, cli_args: TrainerCliArgs):
     if int(os.getenv("LOCAL_RANK", "0")) == 0:
         check_user_token()
 
-    plugin_manager = PluginManager.get_instance()
+    if cfg.rl:
-    dataset_meta = plugin_manager.load_datasets(cfg, preprocess=False)
+        dataset_meta = load_preference_datasets(cfg=cfg, cli_args=cli_args)
-    if not dataset_meta:
+    else:
-        if cfg.rl:
+        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
-            dataset_meta = load_preference_datasets(cfg=cfg, cli_args=cli_args)
-        else:
-            dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
     model, tokenizer, trainer = train(cfg=cfg, dataset_meta=dataset_meta)
 

src/axolotl/core/trainer_builder.py

@@ -170,9 +170,6 @@ class TrainerBuilderBase(abc.ABC):
                 )
             )
 
-        if self.cfg.gc_steps:
-            callbacks.append(GCCallback(gc_steps=self.cfg.gc_steps))
-
         if self.cfg.use_wandb:
             callbacks.append(
                 SaveAxolotlConfigtoWandBCallback(self.cfg.axolotl_config_path)
@@ -254,6 +251,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         if self.cfg.loss_watchdog_threshold is not None:
             callbacks.append(LossWatchDogCallback(self.cfg))
 
+        if self.cfg.gc_steps:
+            callbacks.append(GCCallback(gc_steps=self.cfg.gc_steps))
+
         return callbacks
 
     def get_post_trainer_create_callbacks(self, trainer):
@@ -708,20 +708,6 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 optimizer_cls = ADOPT
                 adam_kwargs["decouple"] = True
                 optimizer_kwargs.update(adam_kwargs)
-            elif self.cfg.optimizer == "came_pytorch":
-                from came_pytorch import CAME
-
-                optimizer_cls = CAME
-
-                beta1 = training_arguments_kwargs.get("adam_beta1", 0.9)
-                beta2 = training_arguments_kwargs.get("adam_beta2", 0.999)
-                beta3 = training_arguments_kwargs.get("adam_beta2", 0.9999)
-                eps1 = training_arguments_kwargs.get("adam_epsilon", 1e-30)
-                eps2 = training_arguments_kwargs.get("adam_epsilon2", 1e-16)
-                adam_kwargs["betas"] = (beta1, beta2, beta3)
-                adam_kwargs["eps"] = (eps1, eps2)
-
-                optimizer_kwargs.update(adam_kwargs)
 
             # Parse any additional optimizer args from config
             if self.cfg.optim_args:

src/axolotl/core/trainers/base.py

@@ -114,6 +114,8 @@ class AxolotlTrainer(
             packing_efficiency_estimate=self.args.sample_packing_efficiency,
             batch_max_len=batch_max_len,
             batch_size=batch_size,
+            group_size=self.args.sample_packing_group_size,
+            bin_size=self.args.sample_packing_bin_size,
             sequential=self.args.sample_packing_sequentially,
             drop_last=True,
         )

src/axolotl/core/trainers/dpo/trainer.py

@@ -247,9 +247,7 @@ class AxolotlDPOTrainer(RngLoaderMixin, SchedulerMixin, DPOTrainer):
                 )
 
         # Base evaluation
-        initial_output = super(  # pylint: disable=bad-super-call
+        initial_output = super().evaluation_loop(
-            DPOTrainer, self
-        ).evaluation_loop(
             dataloader,
             description,
             prediction_loss_only,

src/axolotl/integrations/base.py

@@ -26,8 +26,6 @@ from typing import OrderedDict
 import torch
 from torch.optim.lr_scheduler import LRScheduler
 
-from axolotl.utils.dict import DictDefault
-
 
 class BasePlugin:
     """
@@ -38,13 +36,11 @@ class BasePlugin:
 
     Methods:
     register(cfg): Registers the plugin with the given configuration.
-    load_datasets(cfg): Loads and preprocesses the dataset for training.
     pre_model_load(cfg): Performs actions before the model is loaded.
     post_model_build(cfg, model): Performs actions after the model is loaded, but before LoRA adapters are applied.
     pre_lora_load(cfg, model): Performs actions before LoRA weights are loaded.
     post_lora_load(cfg, model): Performs actions after LoRA weights are loaded.
     post_model_load(cfg, model): Performs actions after the model is loaded, inclusive of any adapters.
-    post_trainer_create(cfg, trainer): Performs actions after the trainer is created.
     create_optimizer(cfg, trainer): Creates and returns an optimizer for training.
     create_lr_scheduler(cfg, trainer, optimizer, num_training_steps): Creates and returns a learning rate scheduler.
     add_callbacks_pre_trainer(cfg, model): Adds callbacks to the trainer before training.
@@ -67,32 +63,20 @@ class BasePlugin:
         None
         """
 
-    def get_input_args(self) -> str | None:
+    def get_input_args(self):
         """
         Returns a pydantic model for the plugin's input arguments.
         """
 
-    def load_datasets(self, cfg: DictDefault, preprocess: bool = False):
-        """
-        Loads and preprocesses the dataset for training.
-
-        Args:
-            cfg: The configuration for the plugin.
-            preprocess: Whether this is the preprocess step of the datasets.
-
-        Returns:
-            dataset_meta: The metadata for the training dataset.
-        """
-
     def pre_model_load(self, cfg):  # pylint: disable=unused-argument
         """
         Performs actions before the model is loaded.
 
-        Args:
+        Parameters:
-            cfg (dict): The configuration for the plugin.
+        cfg (dict): The configuration for the plugin.
 
         Returns:
-            None
+        None
         """
 
     def post_model_build(self, cfg, model):  # pylint: disable=unused-argument
@@ -107,71 +91,59 @@ class BasePlugin:
         """
         Performs actions after the model is loaded.
 
-        Args:
+        Parameters:
-            cfg (dict): The configuration for the plugin.
+        cfg (dict): The configuration for the plugin.
-            model (object): The loaded model.
+        model (object): The loaded model.
 
         Returns:
-            None
+        None
         """
 
     def pre_lora_load(self, cfg, model):  # pylint: disable=unused-argument
         """
         Performs actions before LoRA weights are loaded.
 
-        Args:
+        Parameters:
-            cfg (dict): The configuration for the plugin.
+        cfg (dict): The configuration for the plugin.
-            model (object): The loaded model.
+        model (object): The loaded model.
 
         Returns:
-            None
+        None
         """
 
     def post_lora_load(self, cfg, model):  # pylint: disable=unused-argument
         """
         Performs actions after LoRA weights are loaded.
 
-        Args:
+        Parameters:
-            cfg (dict): The configuration for the plugin.
+        cfg (dict): The configuration for the plugin.
-            model (object): The loaded model.
+        model (object): The loaded model.
 
         Returns:
-            None
+        None
         """
 
     def get_trainer_cls(self, cfg):  # pylint: disable=unused-argument):
         """
         Returns a custom class for the trainer.
 
-        Args:
+        Parameters:
-            cfg (dict): The global axolotl configuration.
+        cfg (dict): The global axolotl configuration.
 
         Returns:
-            class: The class for the trainer.
+        class: The class for the trainer.
-        """
-
-    def post_trainer_create(self, cfg, trainer):  # pylint: disable=unused-argument
-        """
-        Performs actions after the trainer is created.
-
-        Args:
-            cfg (dict): The configuration for the plugin.
-            trainer (object): The trainer object for training.
-
-        Returns:
-            None
         """
 
     def create_optimizer(self, cfg, trainer):  # pylint: disable=unused-argument
         """
         Creates and returns an optimizer for training.
 
-        Args:
+        Parameters:
-            cfg (dict): The configuration for the plugin.
+        cfg (dict): The configuration for the plugin.
-            trainer (object): The trainer object for training.
+        trainer (object): The trainer object for training.
 
         Returns:
-            object: The created optimizer.
+        object: The created optimizer.
         """
 
     def create_lr_scheduler(
@@ -180,26 +152,26 @@ class BasePlugin:
         """
         Creates and returns a learning rate scheduler.
 
-        Args:
+        Parameters:
-            cfg (dict): The configuration for the plugin.
+        cfg (dict): The configuration for the plugin.
-            trainer (object): The trainer object for training.
+        trainer (object): The trainer object for training.
-            optimizer (object): The optimizer for training.
+        optimizer (object): The optimizer for training.
-            num_training_steps (int): Total number of training steps
+        num_training_steps (int): Total number of training steps
 
         Returns:
-            object (LRScheduler): The created learning rate scheduler.
+        object (LRScheduler): The created learning rate scheduler.
         """
 
     def add_callbacks_pre_trainer(self, cfg, model):  # pylint: disable=unused-argument
         """
         setup callbacks before creating the trainer.
 
-        Args:
+        Parameters:
-            cfg (dict): The configuration for the plugin.
+        cfg (dict): The configuration for the plugin.
-            model (object): The loaded model.
+        model (object): The loaded model.
 
         Returns:
-            List[callable]: A list of callback functions to be added to the TrainingArgs
+        List[callable]: A list of callback functions to be added to the TrainingArgs
         """
         return []
 
@@ -210,12 +182,12 @@ class BasePlugin:
         Adds callbacks to the trainer after creating the trainer.
         This is useful for callbacks that require access to the model or trainer.
 
-        Args:
+        Parameters:
-            cfg (dict): The configuration for the plugin.
+        cfg (dict): The configuration for the plugin.
-            trainer (object): The trainer object for training.
+        trainer (object): The trainer object for training.
 
         Returns:
-            List[callable]: A list of callback functions to be added
+        List[callable]: A list of callback functions to be added
         """
         return []
 
@@ -223,23 +195,23 @@ class BasePlugin:
         """
         Performs actions after training is complete.
 
-        Args:
+        Parameters:
-            cfg (dict): The axolotl configuration
+        cfg (dict): The axolotl configuration
-            model (object): The loaded model.
+        model (object): The loaded model.
 
         Returns:
-            None
+        None
         """
 
     def post_train_unload(self, cfg):  # pylint: disable=unused-argument
         """
         Performs actions after training is complete and the model is unloaded.
 
-        Args:
+        Parameters:
-            cfg (dict): The configuration for the plugin.
+        cfg (dict): The configuration for the plugin.
 
         Returns:
-            None
+        None
         """
 
 
@@ -366,27 +338,6 @@ class PluginManager:
                 input_args.append(input_args_from_plugin)
         return input_args
 
-    def load_datasets(self, cfg, preprocess: bool = False):
-        """
-        Calls the load_datasets method of each registered plugin.
-
-        Args:
-            cfg: The configuration for the plugins.
-            preprocess : Whether this is preprocess step of the datasets.
-
-        Returns:
-            dataset_meta: The dataset metadata loaded from all registered plugins.
-        """
-        return_ds_meta = None
-        for plugin in self.plugins.values():
-            dataset_meta = plugin.load_datasets(cfg, preprocess)
-            if dataset_meta is not None:
-                if return_ds_meta is None:
-                    return_ds_meta = dataset_meta
-                else:
-                    raise RuntimeError("Multiple plugins loaded datasets")
-        return return_ds_meta
-
     def pre_model_load(self, cfg):
         """
         Calls the pre_model_load method of all registered plugins.
@@ -471,20 +422,6 @@ class PluginManager:
                 return trainer_cls
         return None
 
-    def post_trainer_create(self, cfg, trainer):
-        """
-        Calls the post_trainer_create method of all registered plugins.
-
-        Parameters:
-        cfg (dict): The configuration for the plugins.
-        trainer (object): The trainer object for training.
-
-        Returns:
-        None
-        """
-        for plugin in self.plugins.values():
-            plugin.post_trainer_create(cfg, trainer)
-
     def create_optimizer(self, trainer):
         """
         Calls the create_optimizer method of all registered plugins and returns the first non-None optimizer.

src/axolotl/integrations/cut_cross_entropy/__init__.py

@@ -72,7 +72,7 @@ class CutCrossEntropyPlugin(BasePlugin):
         if cfg.cut_cross_entropy:
             self._check_requirements()
 
-            from axolotl.integrations.cut_cross_entropy.monkeypatch.patch import (
+            from .monkeypatch.patch import (
                 cce_patch,
             )
 

src/axolotl/monkeypatch/loss/chunked.py

@@ -0,0 +1,134 @@
+"""
+chunked ce loss
+"""
+
+from typing import List, Optional
+
+import torch
+import torch.nn.functional as F
+
+
+# copied and modified from torchtune.modules.loss.CEWithChunkedOutputLoss
+class CEWithChunkedOutputLoss(torch.nn.Module):
+    """
+    Cross-entropy with chunked outputs that saves memory by only upcasting one chunk at a time.
+
+    For more details, please refer to: https://github.com/pytorch/torchtune/pull/1390
+    """
+
+    def __init__(self, num_output_chunks: int = 8, ignore_index: int = -100):
+        super().__init__()
+        self.num_output_chunks = num_output_chunks
+        self.ignore_index = ignore_index
+
+    def compute_cross_entropy(
+        self,
+        logits: torch.Tensor,
+        labels: torch.Tensor,
+        normalize: bool = True,  # pylint: disable=unused-argument
+    ) -> torch.Tensor:
+        """
+        Upcast logits to fp32 and compute cross entropy loss.
+        """
+        return F.cross_entropy(
+            logits.float(), labels, ignore_index=self.ignore_index, reduction="sum"
+        )
+
+    def forward(
+        self, logits: List[torch.Tensor], labels: torch.Tensor, reduction="sum"
+    ) -> torch.Tensor:
+        """
+        Args:
+            logits (List[torch.Tensor]): List of chunked logits of length
+                ``self.num_output_chunks``, where each chunk has shape
+                ``(batch_size, num_tokens / num_output_chunks, vocab_size)``.
+            labels (torch.Tensor): Ground truth labels of shape ``(batch_size, num_tokens)``.
+            reduction (str): The reduction to apply to the output.
+
+        Returns:
+            torch.Tensor: Cross entropy loss of shape (1,).
+        """
+
+        total_elements = (labels != self.ignore_index).sum()
+
+        # chunk and reshape labels (bsz, num_tokens, vocab) -> [(bsz*num_tokens/num_chunks, vocab)]
+        labels = [
+            target_chunk.reshape(-1)
+            for target_chunk in labels.chunk(self.num_output_chunks, dim=1)
+        ]
+        # reshape logits [(bsz, num_tokens/num_chunks, vocab)] -> [(bsz*num_tokens/num_chunks, vocab)]
+        logits = [
+            logit_chunk.reshape(-1, logit_chunk.size(-1)) for logit_chunk in logits
+        ]
+
+        # compute one chunk at a time
+        total_loss = 0.0
+        for logits_chunk, labels_chunk in zip(logits, labels):
+            total_loss += self.compute_cross_entropy(logits_chunk, labels_chunk)
+
+        if reduction == "sum":
+            return total_loss
+        return total_loss / total_elements
+
+
+def _build_chunked_ce_loss_fn(num_output_chunks: int = 8, ignore_index: int = -100):
+    loss_fn_ce = CEWithChunkedOutputLoss(num_output_chunks, ignore_index)
+    loss_fn_ce.compute_cross_entropy = torch.compile(
+        loss_fn_ce.compute_cross_entropy, backend="inductor"
+    )
+    return loss_fn_ce
+
+
+def get_causal_lm_loss(num_output_chunks: int = 8, ignore_index: int = -100):
+    loss_fn_ce = _build_chunked_ce_loss_fn(num_output_chunks, ignore_index)
+
+    def chunked_fix_cross_entropy(
+        source,
+        target,
+        num_items_in_batch: int = None,
+        ignore_index: int = -100,
+        **kwargs,
+    ):  # pylint: disable=unused-argument
+        reduction = "sum" if num_items_in_batch is not None else "mean"
+        logit_chunks = [  # pylint: disable=unnecessary-comprehension
+            chunk for chunk in source.chunk(loss_fn_ce.num_output_chunks, dim=1)
+        ]
+        loss = loss_fn_ce(logit_chunks, target, reduction=reduction)
+        if reduction == "sum":
+            loss = loss / num_items_in_batch
+        return loss
+
+    def for_causal_lm_chunked_loss(
+        logits,
+        labels,
+        vocab_size: int = None,  # pylint: disable=unused-argument
+        num_items_in_batch: Optional[int] = None,
+        ignore_index: int = -100,
+        shift_labels: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> torch.Tensor:
+        # skip the upcast to float since we handle that in the chunking loss
+        if shift_labels is None:
+            # Shift so that tokens < n predict n
+            labels = F.pad(labels, (0, 1), value=ignore_index)
+            shift_labels = labels[..., 1:].contiguous()
+
+        # Skip Flattening the tokens
+        # Enable model parallelism
+        shift_labels = shift_labels.to(logits.device)
+        loss = chunked_fix_cross_entropy(
+            logits, shift_labels, num_items_in_batch, ignore_index, **kwargs
+        )
+        return loss
+
+    return for_causal_lm_chunked_loss
+
+
+def patch_chunked_ce_loss_fn(num_output_chunks: int = 8, ignore_index: int = -100):
+    import transformers.loss.loss_utils
+
+    for_causal_lm_chunked_loss = get_causal_lm_loss(num_output_chunks, ignore_index)
+    transformers.loss.loss_utils.ForCausalLMLoss = for_causal_lm_chunked_loss
+    transformers.loss.loss_utils.LOSS_MAPPING["ForCausalLM"] = (
+        for_causal_lm_chunked_loss
+    )

src/axolotl/monkeypatch/peft/utils.py

@@ -24,7 +24,7 @@ PATCHED_PREPARE_CODE = """
         for name, param in model.named_parameters():
             if (
                 (param.dtype == torch.float16) or (param.dtype == torch.bfloat16)
-            ) and param.__class__.__name__ != "Params4bit" and all(embed_name not in name for embed_name in ["embed_tokens", "lm_head"]):
+            ) and param.__class__.__name__ != "Params4bit" and "norm" in name:
                 param.data = param.data.to(torch.float32)
 """
 

src/axolotl/train.py

@@ -2,7 +2,6 @@
 
 import importlib
 import inspect
-import logging
 import os
 import signal
 import sys
@@ -13,6 +12,7 @@ from typing import Any, Dict
 
 import torch
 import transformers.modelcard
+from accelerate.logging import get_logger
 from accelerate.utils import save_fsdp_model
 from datasets import Dataset
 from huggingface_hub.errors import OfflineModeIsEnabled
@@ -42,7 +42,7 @@ try:
 except ImportError:
     BetterTransformer = None
 
-LOG = logging.getLogger(__name__)
+LOG = get_logger(__name__)
 
 
 def setup_model_and_tokenizer(
@@ -63,6 +63,7 @@ def setup_model_and_tokenizer(
     # Load tokenizer
     LOG.debug(
         f"loading tokenizer... {cfg.tokenizer_config or cfg.base_model_config}",
+        main_process_only=True,
     )
     tokenizer = load_tokenizer(cfg)
 
@@ -527,9 +528,6 @@ def train(
         processor,
     ) = setup_model_and_trainer(cfg, dataset_meta)
 
-    plugin_manager = PluginManager.get_instance()
-    plugin_manager.post_trainer_create(cfg, trainer)
-
     # Handle untrained tokens if configured
     safe_serialization = cfg.save_safetensors is True
     train_dataset = dataset_meta.train_dataset
@@ -552,6 +550,7 @@ def train(
     if not cfg.use_ray:
         cleanup_distributed()
 
+    plugin_manager = PluginManager.get_instance()
     plugin_manager.post_train(cfg, model)
 
     return model, tokenizer, trainer

src/axolotl/utils/callbacks/__init__.py

@@ -885,9 +885,10 @@ def colab_inference_post_train_callback(trainer: Trainer):
             handle T4 gpu, we need to convert attention to eager for inference
             """
             if "Tesla T4" in self.gpu_name and self.cfg.xformers_attention:
-                trainer.model.config._attn_implementation = (  # pylint: disable=protected-access
+                trainer.model.eval()
-                    "eager"
+            trainer.model.config._attn_implementation = (  # pylint: disable=protected-access
-                )
+                "eager"
+            )
             trainer.model.gradient_checkpointing_disable()
             trainer.model.config.use_cache = True
             trainer.model.eval()

src/axolotl/utils/data/shared.py

@@ -281,10 +281,6 @@ def load_dataset_w_config(
             **load_ds_kwargs,
         )
     if not ds:
-        raise ValueError(
+        raise ValueError("unhandled dataset load")
-            "The dataset could not be loaded. This could be due to a misconfigured dataset path "
-            f"({config_dataset.path}). Try double-check your path / name / data_files. "
-            "This is not caused by the dataset type."
-        )
 
     return ds

src/axolotl/utils/gradient_checkpointing/__init__.py

@@ -1,36 +1,15 @@
 """custom checkpointing utils"""
 
-import importlib
 from functools import partial
 
-from packaging import version
-
 from axolotl.utils.gradient_checkpointing.unsloth import (
     Unsloth_Offloaded_Gradient_Checkpointer,
 )
 
-transformers_version = version.parse(importlib.metadata.version("transformers"))
-if transformers_version > version.parse("4.51.3"):
-    from transformers.modeling_layers import GradientCheckpointingLayer
-
-    def uses_gc_layers(decoder_layer):
-        return isinstance(decoder_layer.func.__self__, GradientCheckpointingLayer)
-
-else:
-
-    def uses_gc_layers(_):
-        return False
-
 
 def hf_grad_checkpoint_offload_wrapper(
     decoder_layer, *args, use_reentrant=None
 ):  # pylint: disable=unused-argument
-    if uses_gc_layers(decoder_layer):
-        return Unsloth_Offloaded_Gradient_Checkpointer.apply(
-            decoder_layer,
-            *args,
-        )
-
     return Unsloth_Offloaded_Gradient_Checkpointer.apply(
         (
             decoder_layer.func.__self__

src/axolotl/utils/models.py

@@ -561,12 +561,21 @@ class ModelLoader:
 
             patch_xformers_attn_over_fa2()
             self.cfg.flash_attention = True
+
+        if self.cfg.chunked_cross_entropy:
+            from axolotl.monkeypatch.loss.chunked import patch_chunked_ce_loss_fn
+
+            if self.cfg.chunked_cross_entropy_num_chunks:
+                patch_chunked_ce_loss_fn(self.cfg.chunked_cross_entropy_num_chunks)
+            else:
+                patch_chunked_ce_loss_fn()
+
         if self.cfg.fsdp_config and str(self.cfg.fsdp_config.fsdp_version) == "2":
             from axolotl.monkeypatch.accelerate.fsdp2 import patch_accelerate_fsdp_utils
 
             patch_accelerate_fsdp_utils()
 
-        if self.cfg.adapter and self.cfg.embeddings_skip_upcast:
+        if self.cfg.adapter:
             from axolotl.monkeypatch.peft.utils import patch_peft_prep_code
 
             patch_peft_prep_code()
@@ -1319,11 +1328,8 @@ class ModelLoader:
 
         # make sure these are fp32 per Ramesh et al. (2021)
         embedding_modules = get_linear_embedding_layers(self.cfg.model_config_type)
-        if not self.cfg.fsdp:
+        if self.cfg.fsdp:
-            # we don't run this during FSDP because this will leave mixed
+            # FSDP doesn't like mixed Float and BFloat16
-            # float and bfloat16 dtypes in the model which FSDP doesn't like
-            if self.cfg.load_in_4bit and self.cfg.embeddings_skip_upcast:
-                embedding_modules = []
             self.convert_embedding_modules_dtype(
                 embedding_modules,
                 dist_dtype=torch.float32,

src/axolotl/utils/samplers/multipack.py

@@ -1,10 +1,13 @@
-# pylint: skip-file
 """
-Multipack Batch Sampler
+Multipack Batch Sampler - An efficient batch sampler for packing variable-length sequences
+into fixed-capacity batches to optimize memory usage and training throughput.
 """
+
 import logging
 import math
-from typing import Any, Iterable, List, Union
+from concurrent.futures import ProcessPoolExecutor
+from multiprocessing import cpu_count
+from typing import Iterable, List, Union
 
 import numba
 import numpy as np
@@ -13,26 +16,39 @@ from torch.utils.data import BatchSampler, Sampler, SequentialSampler
 from axolotl.utils.distributed import reduce_and_broadcast
 
 LOG = logging.getLogger(__name__)
-
 LOG.setLevel(logging.INFO)
 
 
 @numba.njit
-def ffd_check(a: np.ndarray, c: int, n: int):
+def ffd_check(sequence_lengths: np.ndarray, bin_capacity: int, num_bins: int):
-    # First-fit-decreasing bin packing
+    """
-    # Check if a[] could fit in n bins with capacity c
+    First-fit-decreasing bin packing algorithm check
-    # https://en.wikipedia.org/wiki/First-fit-decreasing_bin_packing
 
-    a = np.sort(a)[::-1]
+    Checks if sequences with the given lengths could fit in the specified number of bins
-    bins = np.full((n,), c, dtype=a.dtype)
+
-    for size in a:
+    Args:
+        sequence_lengths: Array of sequence lengths
+        bin_capacity: Maximum capacity of each bin
+        num_bins: Number of bins available
+
+    Returns:
+        True if all sequences can be packed, False otherwise
+    """
+    # Sort sequence lengths in descending order for optimal packing
+    sequence_lengths = np.sort(sequence_lengths)[::-1]
+    # Initialize all bins with full capacity
+    bins = np.full((num_bins,), bin_capacity, dtype=sequence_lengths.dtype)
+
+    # Try to place each sequence in the first bin it fits
+    for size in sequence_lengths:
         not_found = True
-        for idx in range(n):
+        for idx in range(num_bins):
             if bins[idx] >= size:
                 bins[idx] -= size
                 not_found = False
                 break
 
+        # If no bin could fit this sequence, packing failed
         if not_found:
             return False
 
@@ -40,240 +56,380 @@ def ffd_check(a: np.ndarray, c: int, n: int):
 
 
 @numba.njit
-def ffd_with_result(a: np.ndarray, c: int, start_index: int):
+def pack_group(
-    # First-fit-decreasing bin packing (with result return)
+    sequence_lengths: np.ndarray,
+    group_offset: int,
+    bin_capacity: int,
+    max_bins: int,
+    bin_size: int,
+    safe_mode: bool = True,
+):
+    """
+    Pack a group of sequences into bins using First-Fit Decreasing algorithm
 
-    indices = np.argsort(a)[::-1]
+    Args:
-    a = a[indices]
+        sequence_lengths: Array of sequence lengths
+        group_offset: Offset to apply to indices when returning results
+        bin_capacity: Maximum capacity of each bin
+        max_bins: Maximum number of bins to use
+        bin_size: Maximum number of sequences per bin
+        safe_mode: If True, use a more conservative packing approach
 
-    bins: List[Any] = []
+    Returns:
-    bins_result: List[Any] = []
+        List of bins, where each bin contains indices of sequences assigned to it
-    for a_id, size in enumerate(a):
+    """
-        add_new = True
+    # Get sorting indices and sort lengths in descending order
-        for idx in range(len(bins)):
+    indices = np.argsort(sequence_lengths)[::-1]
-            if bins[idx] >= size:
+    sorted_lengths = sequence_lengths[indices]
-                bins[idx] -= size
+
-                bins_result[idx].append(indices[a_id] + start_index)
+    bins_remaining_space: list = []  # Tracks remaining capacity in each bin
-                add_new = False
+    bins_assigned_sequences: list = []  # Tracks sequence indices assigned to each bin
+
+    for seq_id, size in enumerate(sorted_lengths):
+        global_idx = indices[seq_id] + group_offset
+
+        # Try to place sequence in existing bins
+        add_new_bin = True
+        for bin_idx, _ in enumerate(bins_remaining_space):
+            if (
+                bins_remaining_space[bin_idx] >= size
+                and len(bins_assigned_sequences[bin_idx]) < bin_size
+            ):
+                bins_remaining_space[bin_idx] -= size
+                bins_assigned_sequences[bin_idx].append(global_idx)
+                add_new_bin = False
                 break
 
-        if add_new:
+        # Create a new bin if needed and if we haven't reached the limit
-            bins.append(c - size)
+        if add_new_bin:
-            bins_result.append([indices[a_id] + start_index])
+            if len(bins_remaining_space) >= max_bins and safe_mode:
+                # In safe mode, skip items that would exceed max_bins
+                continue
+            bins_remaining_space.append(bin_capacity - size)
+            bins_assigned_sequences.append([global_idx])
 
-    return bins_result
+            # Safety check to avoid infinite bins
+            if len(bins_remaining_space) > len(sequence_lengths):
+                break
+
+    return bins_assigned_sequences
 
 
-@numba.njit
+# Define a standalone function for multiprocessing
-def allocate(
+def _process_group(args):
-    lengths: np.ndarray, lengths_cumsum: np.ndarray, rank: int, c: int, n: int
+    group_lengths, start_idx, bin_capacity, max_bins, bin_size, safe_mode = args
+    return pack_group(
+        group_lengths, start_idx, bin_capacity, max_bins, bin_size, safe_mode
+    )
+
+
+def pack_parallel(
+    sequence_lengths: np.ndarray,
+    bin_capacity: int,
+    group_size: int,
+    bin_size: int,
+    num_processes: int | None = None,
+    safe_mode: bool = True,
 ):
-    # Dynamic batch allocator, similar to Multifit
+    """
-    # https://en.wikipedia.org/wiki/Multifit_algorithm
+    Pack sequences into bins using parallel processing
-    # ~99.5% efficiency on OpenChat training set (12 * 2048 ctx len)
 
-    s = 0
+    Args:
-    start_index = 0
+        sequence_lengths: Array of sequence lengths
-    result = []
+        bin_capacity: Maximum capacity of each bin as total number of tokens
+        group_size: Number of sequences to process in each group
+        bin_size: Maximum number of bins to use
+        num_processes: Number of parallel processes to use
+        safe_mode: If True, use a more conservative packing approach
 
-    while True:
+    Returns:
-        # binary search [l, r)
+        List of bins, where each bin contains indices of sequences assigned to it
-        left = 1
+    """
-        right = 1 + np.searchsorted(lengths_cumsum[start_index:], s + c * n, "right")
+    num_items = len(sequence_lengths)
+    if num_processes is None:
+        num_processes = max(1, min(num_items // group_size, cpu_count()))
 
-        while right - left > 1:
+    # Create tasks for parallel processing
-            mid = (left + right) // 2
+    tasks = []
-            if ffd_check(lengths[start_index : start_index + mid], c, n):
+    for i in range(0, num_items, group_size):
-                left = mid
+        group_lengths = sequence_lengths[i : i + group_size]
-            else:
+        max_bins = len(group_lengths)  # Allow as many bins as items in the group
-                right = mid
+        tasks.append((group_lengths, i, bin_capacity, max_bins, bin_size, safe_mode))
 
-        # use length l
+    # Process groups in parallel
-        batch = ffd_with_result(
+    all_bins = []
-            lengths[start_index : start_index + left], c, start_index
+    with ProcessPoolExecutor(max_workers=num_processes) as executor:
-        )
+        for group_bins in executor.map(_process_group, tasks):
-        assert len(batch) <= n
+            all_bins.extend(group_bins)
-        if len(batch) < n:
-            break
 
-        start_index += left
+    return all_bins
-        s = lengths_cumsum[start_index - 1]
-
-        # add local rank
-        result.append(batch[rank])
-
-    return result, s, len(result) * c * n
 
 
 @numba.njit
-def allocate_sequentially(lengths: np.ndarray, rank: int, c: int, n: int):
+def allocate_sequentially(
+    sequence_lengths: np.ndarray, rank: int, bin_capacity: int, num_ranks: int
+):
     """
     Sequential allocator that preserves example order
 
     Parameters:
-    - lengths: The lengths of all examples
+        sequence_lengths: The lengths of all examples
-    - rank: The current rank (for distributed training)
+        rank: The current rank (for distributed training)
-    - c: The capacity of each bin (maximum sequence length)
+        bin_capacity: The capacity of each bin (maximum sequence length)
-    - n: Number of ranks
+        num_ranks: Number of ranks (processes/GPUs)
 
     Returns:
-    - result: List of batches for the current rank
+        rank_batches: List of batches for the current rank
-    - total_used: Number of actual example tokens
+        total_tokens_used: Number of actual example tokens
-    - total_slots: Maximum theoretical number of example tokens (number of bins * bin capacity)
+        total_token_slots: Maximum theoretical number of example tokens (number of bins * bin capacity)
     """
-    result = []
+    rank_batches = []
-    total_used = 0
+    total_tokens_used = 0
 
     # First, do sequential packing into bins
     all_bins = []
-    current_bin = [0 for i in range(0)]  # numba hint
+    current_bin = []
-    remaining_capacity = c
+    remaining_capacity = bin_capacity
 
-    for idx, size in enumerate(lengths):
+    # Process each sequence in order
+    for idx, size in enumerate(sequence_lengths):
         if size <= remaining_capacity:
             # Example fits in current bin
             current_bin.append(idx)
             remaining_capacity -= size
-            total_used += size
+            total_tokens_used += size
         else:
             # Example doesn't fit, start a new bin
             if current_bin:  # Add non-empty bin to all_bins
                 all_bins.append(current_bin)
             current_bin = [idx]
-            remaining_capacity = c - size
+            remaining_capacity = bin_capacity - size
-            total_used += size
+            total_tokens_used += size
 
     # Add the last bin if not empty
     if current_bin:
         all_bins.append(current_bin)
 
-    # Assign bins to ranks - each rank gets every n-th bin
+    # Assign bins to ranks - each rank gets every num_ranks-th bin
-    for bin_idx in range(rank, len(all_bins), n):
+    for bin_idx in range(rank, len(all_bins), num_ranks):
-        result.append(all_bins[bin_idx])
+        rank_batches.append(all_bins[bin_idx])
 
-    return result, total_used, len(all_bins) * c
+    return rank_batches, total_tokens_used, len(all_bins) * bin_capacity
 
 
 class MultipackBatchSampler(BatchSampler):
-    """Batch sampler class for multipack"""
+    """
+    Batch sampler class for efficient packing of variable-length sequences
+
+    This sampler packs sequences into fixed-capacity bins (batches) to maximize
+    GPU memory utilization and training throughput by reducing padding.
+
+    It supports both parallel packing (using FFD algorithm) and
+    sequential packing (preserving original sequence order).
+    """
 
     def __init__(
         self,
         sampler: Union[Sampler[int], Iterable[int]],
-        batch_size: int,
+        batch_size: int,  # Number of bins per batch
-        batch_max_len: int,
+        batch_max_len: int,  # Maximum sequence length (bin capacity)
-        lengths: np.ndarray,
+        lengths: np.ndarray,  # Sequence lengths
-        packing_efficiency_estimate: float = 1.0,
+        packing_efficiency_estimate: float = 1.0,  # Initial efficiency estimate
-        drop_last: bool = False,
+        drop_last: bool = False,  # Whether to drop incomplete batches
-        num_count_samples: int = 16,
+        num_count_samples: int = 16,  # Number of samples to estimate batch count
-        sequential: bool = False,
+        sequential: bool = False,  # Whether to use sequential packing
-        **kwargs,
+        group_size: int = 100_000,  # Size of groups for parallel packing
+        bin_size: int = 200,  # The max number of samples that can be packed in a single bin
+        num_processes: int | None = None,  # Number of processes for parallel packing
+        safe_mode: bool = True,  # Conservative packing to prevent training instability
+        **kwargs,  # pylint: disable=unused-argument
     ):
         super().__init__(sampler, batch_size, drop_last)
         self.batch_size = batch_size
         self.batch_max_len = batch_max_len
-        self.lengths: np.ndarray = lengths
+        self.lengths = np.array(lengths, dtype=np.int32)
         self.packing_efficiency_estimate = packing_efficiency_estimate or 1.0
         self.sequential = sequential
+        self.group_size = group_size
+        self.bin_size = bin_size
+        self.num_processes = num_processes
+        self.safe_mode = safe_mode
 
         assert isinstance(self.lengths, np.ndarray)
 
         self.epoch = 0
 
-        # statistics
+        # Efficiency statistics tracking
-        self.eff_total_used = 0
+        self.total_tokens_used = 0
-        self.eff_total_slots = 0
+        self.total_token_slots = 0
 
-        # The number of times to calculate the batches to determine the minimum packed dataset length for the local rank
+        # The number of times to calculate batches to determine minimum packed dataset length
         self.num_count_samples = num_count_samples
-        # the minimum packed dataset length across all ranks determined by a gather/broadcast
+        # Minimum packed dataset length across all ranks (determined by gather/broadcast)
         self.len_across_ranks = None
 
+        # Cache for batches
+        self._batches = None
+
         if self.sequential and not isinstance(sampler, SequentialSampler):
             LOG.warning(
                 "using sequential sample packing with non-sequential sampler, did you want to also enable curriculum_sampling?"
             )
 
     def set_epoch(self, epoch: int):
+        """Set the epoch number, used for reproducible shuffling across epochs"""
         self.epoch = epoch
+        self._batches = None  # Invalidate batch cache
 
     def generate_batches(self, set_stats=False):
-        indices = [idx for idx in self.sampler]
+        """
+        Generate packed batches for training
 
-        lengths = self.lengths[indices]
+        Args:
-        lengths_cumsum = np.cumsum(lengths)
+            set_stats: Whether to update efficiency statistics
 
-        if self.sequential:
+        Returns:
-            batches, total_used, total_slots = allocate_sequentially(
+            List of batches, where each batch contains multiple bins,
-                lengths=lengths,
+            and each bin contains multiple sequence indices
-                rank=0,
+        """
-                c=self.batch_max_len,
+        if self._batches is not None:
-                n=1,
+            return self._batches
-            )
-        else:
-            batches, total_used, total_slots = allocate(
-                lengths=lengths,
-                lengths_cumsum=lengths_cumsum,
-                rank=0,
-                c=self.batch_max_len,
-                n=1,
-            )
 
-        batches = [
+        # Get indices from the sampler
-            [
+        indices = [  # pylint: disable=unnecessary-comprehension
-                [indices[b_idx] for b_idx in batch]
+            idx for idx in self.sampler
-                for batch in batches[i : i + self.batch_size]
-            ]
-            for i in range(0, len(batches), self.batch_size)
         ]
 
-        # statistics
+        # Get lengths of the selected sequences
-        if set_stats:
+        lengths = self.lengths[indices]
-            self.eff_total_used += total_used
-            self.eff_total_slots += total_slots
 
+        # Pack sequences into bins using either sequential or parallel packing
+        if self.sequential:
+            bins, total_used, total_slots = allocate_sequentially(
+                lengths,
+                rank=0,
+                bin_capacity=self.batch_max_len,
+                num_ranks=1,
+            )
+        else:
+            # Use parallel packing
+            all_bins = pack_parallel(
+                lengths,
+                bin_capacity=self.batch_max_len,
+                group_size=self.group_size,
+                bin_size=self.bin_size,
+                num_processes=self.num_processes,
+                safe_mode=self.safe_mode,
+            )
+
+            # Map bin indices back to original indices
+            bins = [
+                [indices[b_idx] for b_idx in bin_indices] for bin_indices in all_bins
+            ]
+
+            # Calculate efficiency statistics
+            total_used = lengths.sum()
+            total_slots = len(all_bins) * self.batch_max_len
+
+        # Group bins into batches (each batch contains batch_size bins)
+        batches = [
+            bins[i : i + self.batch_size] for i in range(0, len(bins), self.batch_size)
+        ]
+
+        # Drop last batch if requested and it's incomplete
+        if self.drop_last and len(batches[-1]) < self.batch_size:
+            batches = batches[:-1]
+            # Adjust total_slots if we dropped a batch
+            if not self.sequential:
+                total_slots -= (self.batch_size - len(batches[-1])) * self.batch_max_len
+
+        # Update statistics if requested
+        if set_stats:
+            self.total_tokens_used += total_used
+            self.total_token_slots += total_slots
+
+        self._batches = batches
         return batches
 
     def __iter__(self):
+        """
+        Return an iterator over batches
+
+        The batches are truncated to match the minimum number of batches across all ranks
+        to ensure distributed training balance
+        """
         batches = self.generate_batches(set_stats=True)
         if self.len_across_ranks:
-            # make sure the batches we iterate over is truncated to the same min length across all ranks
+            # Truncate batches to ensure all ranks have the same number of batches
             batches = batches[: self.len_across_ranks]
         return iter(batches)
 
-    def num_batches(self):
-        batches = self.generate_batches(set_stats=True)
-        return len(batches)
-
     def efficiency(self):
-        return self.eff_total_used / self.eff_total_slots
+        """
+        Calculate the packing efficiency (ratio of tokens used to total token slots)
+        Higher is better - 1.0 would mean perfect packing with no wasted space
+        """
+        if self.total_token_slots == 0:
+            self.generate_batches(set_stats=True)
+        if self.total_token_slots == 0:
+            return 0.0
+        # Return a Python float instead of potentially a numpy float
+        return float(self.total_tokens_used / self.total_token_slots)
 
     def gather_efficiency(self):
+        """
+        Gather and synchronize packing efficiency estimates across all distributed ranks
+        Returns a conservative efficiency estimate based on the measurements
+        """
+
         def calc_sample_packing_eff_est(estimates: List[float]):
             LOG.debug(f"sample_packing_eff_est across ranks: {repr(estimates)}")
-            return math.floor(0.997 * max(estimates))
+            # Use 99.7% of max observed efficiency as a safe estimate
+            max_eff = max(float(eff) for eff in estimates)
+            return math.floor(0.997 * max_eff)
 
+        # Gather efficiency from all ranks and apply the calculation function
         sample_packing_actual_eff_all = reduce_and_broadcast(
-            lambda: self.efficiency(),  # pylint: disable=unnecessary-lambda
+            lambda: float(self.efficiency()),  # pylint: disable=unnecessary-lambda
             calc_sample_packing_eff_est,
         )
+
+        # Quantize to 0.5% intervals for stability
         sample_packing_eff_est = (
             math.ceil(sample_packing_actual_eff_all * 200.0) / 200.0
         )
         return sample_packing_eff_est
 
     def gather_len_batches(self, num):
+        """
+        Gather and synchronize batch counts across all distributed ranks
+        Returns the minimum number of batches available on any rank
+        """
+
         def calc_min_len(estimates: list[(int, float)]):
             LOG.info(f"gather_len_batches: {repr(estimates)}")
             return math.floor(min(estimates))
 
+        # Find minimum batch count across ranks to ensure balance
         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 = min(
+        Return the total number of batches that will be yielded by this sampler
-                [self.num_batches() for _ in range(self.num_count_samples)]
+
+        This is calculated as the minimum number of batches available on any rank
+        to ensure balanced distributed training
+        """
+        if self._batches is None:
+            self._batches = self.generate_batches(set_stats=True)
+
+        if self.len_across_ranks is None:
+            # Sample multiple times to get stable estimate
+            len_batches = min(  # pylint: disable=consider-using-generator
+                [len(self._batches) for _ in range(self.num_count_samples)]
             )
+            # Gather minimum across all ranks
             self.len_across_ranks = self.gather_len_batches(len_batches)
+
         return self.len_across_ranks

src/axolotl/utils/schemas/config.py

@@ -82,7 +82,6 @@ class AxolotlInputConfig(
     mean_resizing_embeddings: bool | None = False
     # optionally shrink the embeddings when the tokenizer vocab size is smaller
     shrink_embeddings: bool | None = None
-    embeddings_skip_upcast: bool | None = None
 
     rl: RLType | None = None
     trl: TRLConfig | None = Field(
@@ -243,6 +242,9 @@ class AxolotlInputConfig(
     unsloth_rms_norm: bool | None = None
     unsloth_rope: bool | None = None
 
+    chunked_cross_entropy: bool | None = None
+    chunked_cross_entropy_num_chunks: int | None = None
+
     lora_mlp_kernel: bool | None = None
     lora_qkv_kernel: bool | None = None
     lora_o_kernel: bool | None = None
@@ -462,10 +464,9 @@ class AxolotlInputConfig(
             and not data.get("flash_attention")
             and not data.get("sdp_attention")
             and not data.get("flex_attention")
-            and not data.get("xformers_attention")
         ):
             LOG.warning(
-                "sample_packing without flash, sdp, xformers or flex attention does not handle cross sample decontamination."
+                "sample_packing without flash, sdp or flex attention does not handle cross sample decontamination."
             )
 
         return data

src/axolotl/utils/schemas/enums.py

@@ -53,5 +53,4 @@ class CustomSupportedOptimizers(str, Enum):
     ao_adamw_8bit = "ao_adamw_8bit"  # pylint: disable=invalid-name
     ao_adamw_fp8 = "ao_adamw_fp8"  # pylint: disable=invalid-name
     adopt_adamw = "adopt_adamw"  # pylint: disable=invalid-name
-    came_pytorch = "came_pytorch"  # pylint: disable=invalid-name
     muon = "muon"  # pylint: disable=invalid-name

src/axolotl/utils/schemas/training.py

@@ -75,10 +75,8 @@ class HyperparametersConfig(BaseModel):
     lr_groups: list[LrGroup] | None = None
 
     adam_epsilon: float | None = None
-    adam_epsilon2: float | None = None
     adam_beta1: float | None = None
     adam_beta2: float | None = None
-    adam_beta3: float | None = None
     max_grad_norm: float | None = None
     num_epochs: float = Field(default=1.0)
 

tests/conftest.py

@@ -4,7 +4,6 @@ shared pytest fixtures
 
 import functools
 import importlib
-import os
 import shutil
 import sys
 import tempfile
@@ -530,32 +529,31 @@ def dataset_fozziethebeat_alpaca_messages_2k_dpo_test_rev_ea82cff(
 
 
 # # pylint: disable=redefined-outer-name,unused-argument
-@pytest.mark.skipif(
+# def test_load_fixtures(
-    os.environ.get("AXOLOTL_IS_CI_CACHE_PRELOAD", "-1") != "1",
+#     download_smollm2_135m_model,
-    reason="Not running in CI cache preload",
+#     download_llama_68m_random_model,
-)
+#     download_qwen_2_5_half_billion_model,
-def test_load_fixtures(
+#     download_tatsu_lab_alpaca_dataset,
-    download_smollm2_135m_model,
+#     download_mhenrichsen_alpaca_2k_dataset,
-    download_qwen_2_5_half_billion_model,
+#     download_mhenrichsen_alpaca_2k_w_revision_dataset,
-    download_tatsu_lab_alpaca_dataset,
+#     download_mlabonne_finetome_100k_dataset,
-    download_mhenrichsen_alpaca_2k_dataset,
+#     download_argilla_distilabel_capybara_dpo_7k_binarized_dataset,
-    download_mhenrichsen_alpaca_2k_w_revision_dataset,
+#     download_argilla_ultrafeedback_binarized_preferences_cleaned_dataset,
-    download_mlabonne_finetome_100k_dataset,
+#     download_fozzie_alpaca_dpo_dataset,
-    download_argilla_distilabel_capybara_dpo_7k_binarized_dataset,
+#     download_arcee_ai_distilabel_intel_orca_dpo_pairs_dataset,
-    download_arcee_ai_distilabel_intel_orca_dpo_pairs_dataset,
+#     download_argilla_dpo_pairs_dataset,
-    download_argilla_dpo_pairs_dataset,
+#     download_tiny_shakespeare_dataset,
-    download_tiny_shakespeare_dataset,
+#     download_deepseek_model_fixture,
-    download_deepseek_model_fixture,
+#     download_huggyllama_model_fixture,
-    download_huggyllama_model_fixture,
+#     download_llama_1b_model_fixture,
-    download_llama_1b_model_fixture,
+#     download_llama3_8b_model_fixture,
-    download_llama3_8b_model_fixture,
+#     download_llama3_8b_instruct_model_fixture,
-    download_llama3_8b_instruct_model_fixture,
+#     download_phi_35_mini_model_fixture,
-    download_phi_35_mini_model_fixture,
+#     download_phi_3_medium_model_fixture,
-    download_phi_3_medium_model_fixture,
+#     download_mistral_7b_model_fixture,
-    download_mistral_7b_model_fixture,
+#     download_gemma_2b_model_fixture,
-    download_gemma_2b_model_fixture,
+#     download_gemma2_9b_model_fixture,
-    download_gemma2_9b_model_fixture,
+#     download_mlx_mistral_7b_model_fixture,
-    download_mlx_mistral_7b_model_fixture,
+#     download_llama2_model_fixture,
-    download_llama2_model_fixture,
+# ):
-):
+#     pass
-    pass

tests/e2e/integrations/test_hooks.py

@@ -29,12 +29,6 @@ class LogHooksPlugin(BasePlugin):
         except FileNotFoundError:
             pass
 
-    def post_trainer_create(self, cfg, trainer):  # pylint: disable=unused-argument
-        with open(
-            self.base_dir.joinpath("plugin_hooks.log"), "a", encoding="utf-8"
-        ) as f:
-            f.write("post_trainer_create\n")
-
     def pre_model_load(self, cfg):  # pylint: disable=unused-argument
         with open(
             self.base_dir.joinpath("plugin_hooks.log"), "a", encoding="utf-8"
@@ -171,7 +165,6 @@ class TestPluginHooks:
         ) as f:
             file_contents = f.readlines()
             file_contents = "\n".join(file_contents)
-            assert "post_trainer_create" in file_contents
             assert "pre_model_load" in file_contents
             assert "post_model_build" in file_contents
             assert "pre_lora_load" in file_contents

tests/e2e/multigpu/test_llama.py

@@ -479,7 +479,7 @@ class TestMultiGPULlama:
                 "sample_packing": True,
                 "pad_to_sequence_len": True,
                 "sequence_len": 2048,
-                "val_set_size": 0.1,
+                "val_set_size": 0.05,
                 "special_tokens": {
                     "pad_token": "<|endoftext|>",
                 },

tests/e2e/patched/lora_kernels/test_lora_kernel_patching.py

@@ -29,12 +29,12 @@ from axolotl.utils.dict import DictDefault
 
 MODEL_CONFIGS = [
     {
-        "name": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
+        "name": "openaccess-ai-collective/tiny-mistral",
         "expected_activation": apply_lora_mlp_swiglu,
         "dtype": torch.float16,
     },
     {
-        "name": "trl-internal-testing/tiny-Qwen2ForCausalLM-2.5",
+        "name": "Qwen/Qwen2-7B",
         "expected_activation": apply_lora_mlp_swiglu,
         "dtype": torch.float16,
     },
@@ -44,7 +44,7 @@ MODEL_CONFIGS = [
         "dtype": torch.float32,
     },
     {
-        "name": "trl-internal-testing/tiny-Gemma2ForCausalLM",
+        "name": "mhenrichsen/gemma-2b",
         "expected_activation": apply_lora_mlp_geglu,
         "dtype": torch.float16,
     },
@@ -156,9 +156,7 @@ def test_swiglu_mlp_integration(small_llama_model):
 def test_geglu_model_integration():
     """Test GeGLU activation with Gemma model."""
     model = AutoModelForCausalLM.from_pretrained(
-        "trl-internal-testing/tiny-Gemma2ForCausalLM",
+        "mhenrichsen/gemma-2b", torch_dtype=torch.float16, device_map="cuda:0"
-        torch_dtype=torch.float16,
-        device_map="cuda:0",
     )
     peft_config = get_peft_config(
         {

tests/e2e/patched/test_falcon_samplepack.py

@@ -6,8 +6,6 @@ import logging
 import os
 import unittest
 
-import pytest
-
 from axolotl.cli.args import TrainerCliArgs
 from axolotl.common.datasets import load_datasets
 from axolotl.train import train
@@ -25,7 +23,6 @@ class TestFalconPatched(unittest.TestCase):
     Test case for Falcon models
     """
 
-    @pytest.mark.skip(reason="no tiny models for testing with safetensors")
     @with_temp_dir
     def test_qlora(self, temp_dir):
         # pylint: disable=duplicate-code
@@ -74,7 +71,6 @@ class TestFalconPatched(unittest.TestCase):
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
 
-    @pytest.mark.skip(reason="no tiny models for testing with safetensors")
     @with_temp_dir
     def test_ft(self, temp_dir):
         # pylint: disable=duplicate-code

tests/e2e/patched/test_mistral_samplepack.py

@@ -28,7 +28,7 @@ class TestMistral(unittest.TestCase):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
+                "base_model": "openaccess-ai-collective/tiny-mistral",
                 "flash_attention": True,
                 "sample_packing": True,
                 "sequence_len": 1024,
@@ -76,7 +76,7 @@ class TestMistral(unittest.TestCase):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
+                "base_model": "openaccess-ai-collective/tiny-mistral",
                 "flash_attention": True,
                 "sample_packing": True,
                 "sequence_len": 1024,

tests/e2e/patched/test_model_patches.py

@@ -56,7 +56,7 @@ class TestModelPatches(unittest.TestCase):
     def test_mistral_multipack(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
+                "base_model": "openaccess-ai-collective/tiny-mistral",
                 "flash_attention": True,
                 "sample_packing": True,
                 "sequence_len": 2048,

tests/e2e/patched/test_peft_embeddings.py

@@ -1,63 +0,0 @@
-"""
-Test case for handling embeddings when using peft
-"""
-
-import torch
-
-from axolotl.train import setup_model_and_tokenizer
-from axolotl.utils.config import normalize_config, validate_config
-from axolotl.utils.dict import DictDefault
-
-
-class TestLlamaPeftEmbeddings:
-    """
-    test class for handling embeddings when using peft
-    """
-
-    def test_peft_embeddings_upcast(self, temp_dir):
-        # pylint: disable=duplicate-code
-        cfg = DictDefault(
-            {
-                "base_model": "HuggingFaceTB/SmolLM2-135M",
-                "load_in_4bit": True,
-                "adapter": "qlora",
-                "lora_r": 8,
-                "lora_alpha": 16,
-                "lora_target_linear": True,
-                "trust_remote_code": True,
-                "sequence_len": 512,
-                "val_set_size": 0.01,
-                "special_tokens": {
-                    "pad_token": "<|endoftext|>",
-                },
-                "datasets": [
-                    {
-                        "path": "mhenrichsen/alpaca_2k_test",
-                        "type": "alpaca",
-                    },
-                ],
-                "num_epochs": 1,
-                "max_steps": 2,
-                "micro_batch_size": 1,
-                "gradient_accumulation_steps": 1,
-                "output_dir": temp_dir,
-                "learning_rate": 0.00001,
-                "optimizer": "adamw_8bit",
-                "lr_scheduler": "cosine",
-                "flash_attention": True,
-                "sample_packing": False,
-                "bf16": "auto",
-                "save_safetensors": True,
-                "embeddings_skip_upcast": True,
-            }
-        )
-
-        cfg = validate_config(cfg)
-        normalize_config(cfg)
-
-        model, _, _, _ = setup_model_and_tokenizer(cfg)
-
-        # Check if the embeddings are upcast correctly
-        # only embed_tokens is a parameter that may be upcast
-        assert model.base_model.model.model.embed_tokens.weight.dtype == torch.bfloat16
-        assert model.base_model.model.lm_head.weight.dtype == torch.bfloat16

tests/e2e/patched/test_resume.py

@@ -15,7 +15,7 @@ from axolotl.train import train
 from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 
-from ..utils import check_model_output_exists, most_recent_subdir, require_torch_2_6_0
+from ..utils import check_model_output_exists, most_recent_subdir
 
 LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"
@@ -26,7 +26,6 @@ class TestResumeLlama:
     Test case for resuming training of llama models
     """
 
-    @require_torch_2_6_0
     def test_resume_lora_packed(self, temp_dir):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
@@ -63,7 +62,6 @@ class TestResumeLlama:
                 "save_total_limit": 5,
                 "max_steps": 15,
                 "use_tensorboard": True,
-                "save_safetensors": True,
             }
         )
         if is_torch_bf16_gpu_available():

tests/e2e/test_evaluate.py

@@ -19,11 +19,14 @@ class TestE2eEvaluate:
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "base_model": "JackFram/llama-68m",
+                "tokenizer_type": "LlamaTokenizer",
                 "sequence_len": 1024,
                 "val_set_size": 0.02,
                 "special_tokens": {
-                    "pad_token": "<|endoftext|>",
+                    "unk_token": "<unk>",
+                    "bos_token": "<s>",
+                    "eos_token": "</s>",
                 },
                 "datasets": [
                     {

tests/e2e/test_falcon.py

@@ -6,8 +6,6 @@ import logging
 import os
 import unittest
 
-import pytest
-
 from axolotl.cli.args import TrainerCliArgs
 from axolotl.common.datasets import load_datasets
 from axolotl.train import train
@@ -25,7 +23,6 @@ class TestFalcon(unittest.TestCase):
     Test case for falcon
     """
 
-    @pytest.mark.skip(reason="no tiny models for testing with safetensors")
     @with_temp_dir
     def test_lora(self, temp_dir):
         # pylint: disable=duplicate-code
@@ -77,7 +74,6 @@ class TestFalcon(unittest.TestCase):
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
 
-    @pytest.mark.skip(reason="no tiny models for testing with safetensors")
     @with_temp_dir
     def test_lora_added_vocab(self, temp_dir):
         # pylint: disable=duplicate-code
@@ -133,7 +129,6 @@ class TestFalcon(unittest.TestCase):
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
 
-    @pytest.mark.skip(reason="no tiny models for testing with safetensors")
     @with_temp_dir
     def test_ft(self, temp_dir):
         # pylint: disable=duplicate-code

tests/e2e/test_mistral.py

@@ -30,7 +30,7 @@ class TestMistral(unittest.TestCase):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
+                "base_model": "openaccess-ai-collective/tiny-mistral",
                 "flash_attention": True,
                 "sequence_len": 1024,
                 "load_in_8bit": True,
@@ -77,7 +77,7 @@ class TestMistral(unittest.TestCase):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
+                "base_model": "openaccess-ai-collective/tiny-mistral",
                 "flash_attention": True,
                 "sequence_len": 1024,
                 "val_set_size": 0.02,

tests/e2e/test_optimizers.py

@@ -199,50 +199,3 @@ class TestCustomOptimizers(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
-
-    @with_temp_dir
-    def test_came_pytorch(self, temp_dir):
-        # pylint: disable=duplicate-code
-        cfg = DictDefault(
-            {
-                "base_model": "JackFram/llama-68m",
-                "tokenizer_type": "LlamaTokenizer",
-                "sequence_len": 1024,
-                "load_in_8bit": True,
-                "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,
-                "output_dir": temp_dir,
-                "learning_rate": 0.00001,
-                "optimizer": "came_pytorch",
-                "adam_beta3": 0.9999,
-                "adam_epsilon2": 1e-16,
-                "max_steps": 5,
-                "lr_scheduler": "cosine",
-            }
-        )
-
-        cfg = validate_config(cfg)
-        normalize_config(cfg)
-        cli_args = TrainerCliArgs()
-        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
-
-        train(cfg=cfg, dataset_meta=dataset_meta)
-        check_model_output_exists(temp_dir, cfg)

tests/test_chunked_xentropy.py

@@ -0,0 +1,40 @@
+"""
+test suite for chunked cross entropy
+"""
+
+import pytest
+import torch
+from torch import nn
+
+from axolotl.monkeypatch.loss.chunked import get_causal_lm_loss
+
+
+@pytest.fixture
+def chunked_fixtures():
+    model_dim = 512
+    vocab_size = 1024 * 256
+    seq_len = 2048
+    batch_size = 1
+
+    lm_head = nn.Linear(model_dim, vocab_size)
+    hidden_state = torch.randn(batch_size, seq_len, model_dim)
+    labels = torch.randint(low=0, high=vocab_size, size=(batch_size, seq_len))
+    return lm_head, hidden_state, labels, vocab_size
+
+
+def test_chunked_forward(chunked_fixtures):  # pylint: disable=redefined-outer-name
+    lm_head, hidden_state, labels, vocab_size = chunked_fixtures
+    lm_loss = get_causal_lm_loss()
+
+    logits = lm_head(hidden_state)
+
+    chunked_lm_loss = lm_loss(logits, labels)
+
+    logits_flattened = logits.view(-1, vocab_size)
+    labels_flattened = labels.view(-1)
+
+    loss = nn.functional.cross_entropy(
+        logits_flattened.float(), labels_flattened, reduction="mean"
+    )
+
+    assert torch.allclose(chunked_lm_loss, loss, atol=1e-2, rtol=1e-2)

tests/test_datasets.py

@@ -414,6 +414,7 @@ class TestDatasetPreparation:
             snapshot_path = snapshot_download(
                 repo_id="mhenrichsen/alpaca_2k_test",
                 repo_type="dataset",
+                local_dir=tmp_ds_path,
             )
             shutil.copytree(snapshot_path, tmp_ds_path, dirs_exist_ok=True)