offload activations to disk instead of CPU RAM

* don't sort multipack sampler

* increased packing efficiency increases loss

---------

Co-authored-by: Wing Lian <wing@axolotl.ai>

.github/workflows/tests-nightly.yml

@@ -18,9 +18,96 @@ 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

docs/config.qmd

@@ -612,6 +612,7 @@ 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:

examples/llama-4/README.md

@@ -34,3 +34,5 @@ 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.

requirements.txt

@@ -11,6 +11,7 @@ liger-kernel==0.5.9
 
 packaging==23.2
 
+huggingface_hub==0.31.0
 peft==0.15.2
 transformers==4.51.3
 tokenizers>=0.21.1

setup.py

@@ -142,6 +142,7 @@ extras_require = {
         "apollo-torch",
         "lomo-optim==0.1.1",
         "torch-optimi==0.2.1",
+        "came_pytorch==0.1.3",
     ],
     "ray": [
         "ray[train]",

src/axolotl/core/trainer_builder.py

@@ -708,6 +708,20 @@ 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/train.py

@@ -2,6 +2,7 @@
 
 import importlib
 import inspect
+import logging
 import os
 import signal
 import sys
@@ -12,7 +13,6 @@ 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 = get_logger(__name__)
+LOG = logging.getLogger(__name__)
 
 
 def setup_model_and_tokenizer(
@@ -63,7 +63,6 @@ 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)
 

src/axolotl/utils/data/shared.py

@@ -281,6 +281,10 @@ def load_dataset_w_config(
             **load_ds_kwargs,
         )
     if not ds:
-        raise ValueError("unhandled dataset load")
+        raise ValueError(
+            "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

@@ -1,16 +1,59 @@
 """custom checkpointing utils"""
 
+import importlib
 from functools import partial
 
-from axolotl.utils.gradient_checkpointing.unsloth import (
-    Unsloth_Offloaded_Gradient_Checkpointer,
+from packaging import version
+
+from axolotl.utils.gradient_checkpointing.offload_cpu import (
+    CPU_Offloaded_Gradient_Checkpointer,
 )
+from axolotl.utils.gradient_checkpointing.offload_disk import (
+    DiskOffloadedGradientCheckpointer,
+)
+
+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
-    return Unsloth_Offloaded_Gradient_Checkpointer.apply(
+    if uses_gc_layers(decoder_layer):
+        return CPU_Offloaded_Gradient_Checkpointer.apply(
+            decoder_layer,
+            *args,
+        )
+
+    return CPU_Offloaded_Gradient_Checkpointer.apply(
+        (
+            decoder_layer.func.__self__
+            if isinstance(decoder_layer, partial)
+            else decoder_layer.__self__
+        ),
+        *args,
+    )
+
+
+def hf_grad_checkpoint_disk_offload_wrapper(
+    decoder_layer, *args, use_reentrant=None
+):  # pylint: disable=unused-argument
+    if uses_gc_layers(decoder_layer):
+        return DiskOffloadedGradientCheckpointer.apply(
+            decoder_layer,
+            *args,
+        )
+
+    return DiskOffloadedGradientCheckpointer.apply(
         (
             decoder_layer.func.__self__
             if isinstance(decoder_layer, partial)

src/axolotl/utils/gradient_checkpointing/offload_cpu.py

@@ -1,4 +1,4 @@
-"""Unsloth checkpointing"""
+"""CPU offloaded checkpointing"""
 
 # Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
 #
@@ -26,7 +26,7 @@ else:
     torch_cuda_amp_custom_bwd = torch.amp.custom_bwd(device_type="cuda")
 
 
-class Unsloth_Offloaded_Gradient_Checkpointer(  # pylint: disable=invalid-name
+class CPU_Offloaded_Gradient_Checkpointer(  # pylint: disable=invalid-name
     torch.autograd.Function
 ):
     """

src/axolotl/utils/gradient_checkpointing/offload_disk.py

@@ -0,0 +1,93 @@
+"""Disk offloaded checkpointing"""
+
+import os
+import tempfile
+import uuid
+
+import torch
+
+torch_cuda_amp_custom_fwd = torch.amp.custom_fwd(device_type="cuda")
+torch_cuda_amp_custom_bwd = torch.amp.custom_bwd(device_type="cuda")
+
+
+class DiskOffloadedGradientCheckpointer(torch.autograd.Function):
+    """
+    Saves both VRAM and RAM by offloading activations to disk.
+    Greater hit to performance than RAM offloading, but useful for extremely memory-constrained environments.
+    """
+
+    # Create a temporary directory for storing tensors
+    _temp_dir = tempfile.mkdtemp(prefix="disk_checkpoint_")
+
+    @staticmethod
+    def _get_temp_file_path():
+        """Generate a unique file path for tensor storage"""
+        return os.path.join(
+            DiskOffloadedGradientCheckpointer._temp_dir, f"{uuid.uuid4()}.pt"
+        )
+
+    @staticmethod
+    @torch_cuda_amp_custom_fwd
+    def forward(ctx, forward_function, hidden_states, *args):
+        # Generate a unique file path for this tensor
+        file_path = DiskOffloadedGradientCheckpointer._get_temp_file_path()
+
+        # Save tensor to disk in a non-blocking way (detached from compute)
+        # First move to CPU, then save
+        cpu_hidden_states = hidden_states.detach().cpu()
+        torch.save(cpu_hidden_states, file_path)
+
+        # Free CPU memory
+        del cpu_hidden_states
+
+        # Run forward pass
+        with torch.no_grad():
+            output = forward_function(hidden_states, *args)
+
+        # Store the path instead of the tensor
+        ctx.save_for_backward(torch.tensor([0]))  # Dummy tensor
+        ctx.file_path = file_path
+        ctx.forward_function = forward_function
+        ctx.args = args
+        return output
+
+    @staticmethod
+    @torch_cuda_amp_custom_bwd
+    def backward(ctx, dY):  # pylint: disable=invalid-name
+        # Load the hidden states from disk
+        hidden_states = torch.load(ctx.file_path, weights_only=True)
+
+        # Move to CUDA and prepare for gradient computation
+        hidden_states = hidden_states.to("cuda", non_blocking=True).detach()
+        hidden_states.requires_grad = True
+
+        # Clean up the temporary file
+        try:
+            os.remove(ctx.file_path)
+        except FileNotFoundError:
+            pass  # Ignore errors in file deletion
+
+        # Compute gradients
+        with torch.enable_grad():
+            output = ctx.forward_function(hidden_states, *ctx.args)
+        # pylint: disable=duplicate-code
+        torch.autograd.backward(output, dY)
+
+        return (
+            None,
+            hidden_states.grad,
+        ) + (
+            None,
+        ) * len(ctx.args)
+
+    @staticmethod
+    def cleanup():
+        """Clean up the temporary directory when done"""
+        import shutil
+
+        try:
+            shutil.rmtree(
+                DiskOffloadedGradientCheckpointer._temp_dir
+            )  # pylint: disable=protected-access
+        except FileNotFoundError:
+            pass

src/axolotl/utils/models.py

@@ -70,7 +70,10 @@ from axolotl.utils.distributed import (
     is_local_main_process,
     is_main_process,
 )
-from axolotl.utils.gradient_checkpointing import hf_grad_checkpoint_offload_wrapper
+from axolotl.utils.gradient_checkpointing import (
+    hf_grad_checkpoint_disk_offload_wrapper,
+    hf_grad_checkpoint_offload_wrapper,
+)
 from axolotl.utils.lora_embeddings import get_linear_embedding_layers
 from axolotl.utils.model_shard_quant import load_sharded_model, load_sharded_model_quant
 
@@ -619,6 +622,10 @@ class ModelLoader:
 
         if self.cfg.gradient_checkpointing in ["unsloth", "offload"]:
             transformers.modeling_utils.checkpoint = hf_grad_checkpoint_offload_wrapper
+        if self.cfg.gradient_checkpointing == "offload_disk":
+            transformers.modeling_utils.checkpoint = (
+                hf_grad_checkpoint_disk_offload_wrapper
+            )
 
         if self.cfg.flash_attention:
             self.patch_attention()

src/axolotl/utils/samplers/multipack.py

@@ -78,15 +78,11 @@ def pack_group(
     Returns:
         List of bins, where each bin contains indices of sequences assigned to it
     """
-    # Get sorting indices and sort lengths in descending order
-    indices = np.argsort(sequence_lengths)[::-1]
-    sorted_lengths = sequence_lengths[indices]
-
     bins_remaining_space: list = []  # Tracks remaining capacity in each bin
     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
+    for seq_id, size in enumerate(sequence_lengths):
+        global_idx = seq_id + group_offset
 
         # Try to place sequence in existing bins
         add_new_bin = True

src/axolotl/utils/schemas/config.py

@@ -178,9 +178,9 @@ class AxolotlInputConfig(
 
     # torch_dtype: torch.dtype | None
 
-    gradient_checkpointing: Literal["unsloth", "offload"] | bool | None = Field(
-        default=False
-    )
+    gradient_checkpointing: (
+        Literal["unsloth", "offload", "offload_disk"] | bool | None
+    ) = Field(default=False)
     gradient_checkpointing_kwargs: dict[str, Any] | None = None
 
     unfrozen_parameters: list[str] | None = None

src/axolotl/utils/schemas/enums.py

@@ -53,4 +53,5 @@ 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

@@ -75,8 +75,10 @@ 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/e2e/integrations/test_kd.py

@@ -90,7 +90,7 @@ class TestKnowledgeDistillation:
         train(cfg=cfg, dataset_meta=dataset_meta)
         assert (Path(temp_dir) / "model.safetensors").exists()
         check_tensorboard(
-            temp_dir + "/runs", "train/loss", 1.0, "Train Loss is too high"
+            temp_dir + "/runs", "train/loss", 1.2, "Train Loss (%s) is too high"
         )
 
     @pytest.mark.parametrize(
@@ -121,5 +121,5 @@ class TestKnowledgeDistillation:
         train(cfg=cfg, dataset_meta=dataset_meta)
         assert (Path(temp_dir) / "adapter_model.safetensors").exists()
         check_tensorboard(
-            temp_dir + "/runs", "train/loss", 1.0, "Train Loss is too high"
+            temp_dir + "/runs", "train/loss", 1.2, "Train Loss (%s) is too high"
         )

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": 0.05,
+                "val_set_size": 0.1,
                 "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": "openaccess-ai-collective/tiny-mistral",
+        "name": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
         "expected_activation": apply_lora_mlp_swiglu,
         "dtype": torch.float16,
     },
     {
-        "name": "Qwen/Qwen2-7B",
+        "name": "trl-internal-testing/tiny-Qwen2ForCausalLM-2.5",
         "expected_activation": apply_lora_mlp_swiglu,
         "dtype": torch.float16,
     },
@@ -44,7 +44,7 @@ MODEL_CONFIGS = [
         "dtype": torch.float32,
     },
     {
-        "name": "mhenrichsen/gemma-2b",
+        "name": "trl-internal-testing/tiny-Gemma2ForCausalLM",
         "expected_activation": apply_lora_mlp_geglu,
         "dtype": torch.float16,
     },
@@ -156,7 +156,9 @@ def test_swiglu_mlp_integration(small_llama_model):
 def test_geglu_model_integration():
     """Test GeGLU activation with Gemma model."""
     model = AutoModelForCausalLM.from_pretrained(
-        "mhenrichsen/gemma-2b", torch_dtype=torch.float16, device_map="cuda:0"
+        "trl-internal-testing/tiny-Gemma2ForCausalLM",
+        torch_dtype=torch.float16,
+        device_map="cuda:0",
     )
     peft_config = get_peft_config(
         {

tests/e2e/patched/test_falcon_samplepack.py

@@ -6,6 +6,8 @@ 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
@@ -23,6 +25,7 @@ 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
@@ -71,6 +74,7 @@ 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": "openaccess-ai-collective/tiny-mistral",
+                "base_model": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
                 "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": "openaccess-ai-collective/tiny-mistral",
+                "base_model": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
                 "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": "openaccess-ai-collective/tiny-mistral",
+                "base_model": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
                 "flash_attention": True,
                 "sample_packing": True,
                 "sequence_len": 2048,

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
+from ..utils import check_model_output_exists, most_recent_subdir, require_torch_2_6_0
 
 LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"
@@ -26,6 +26,7 @@ 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(
@@ -62,6 +63,7 @@ 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

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

tests/e2e/test_falcon.py

@@ -6,6 +6,8 @@ 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
@@ -23,6 +25,7 @@ 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
@@ -74,6 +77,7 @@ 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
@@ -129,6 +133,7 @@ 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": "openaccess-ai-collective/tiny-mistral",
+                "base_model": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
                 "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": "openaccess-ai-collective/tiny-mistral",
+                "base_model": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
                 "flash_attention": True,
                 "sequence_len": 1024,
                 "val_set_size": 0.02,

tests/e2e/test_optimizers.py

@@ -199,3 +199,50 @@ 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_datasets.py

@@ -414,7 +414,6 @@ 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)
 

tests/test_packed_batch_sampler.py

@@ -106,3 +106,4 @@ class TestBatchedSamplerPacking:
 
         original_idxs = set(range(len(train_dataset)))
         assert original_idxs == set(batch_idxs)
+        assert len(batch_idxs) == len(set(batch_idxs))