bump transformers and update attention class map name

.github/workflows/tests-docker.yml

@@ -0,0 +1,62 @@
+name: e2e-docker-tests
+
+on:
+  pull_request:
+    paths:
+      - '**.py'
+      - 'requirements.txt'
+  workflow_dispatch:
+
+jobs:
+  build-axolotl:
+    if: github.repository_owner == 'OpenAccess-AI-Collective'
+    # this job needs to be run on self-hosted GPU runners...
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - cuda: 118
+            cuda_version: 11.8.0
+            python_version: "3.10"
+            pytorch: 2.0.1
+            axolotl_extras:
+            is_latest: true
+          - cuda: 121
+            cuda_version: 12.1.0
+            python_version: "3.10"
+            pytorch: 2.1.1
+            axolotl_extras:
+    runs-on: [self-hosted, gpu, docker]
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+      - name: Docker metadata
+        id: metadata
+        uses: docker/metadata-action@v5
+        with:
+          images: winglian/axolotl
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_TOKEN }}
+      # guidance for testing before pushing: https://docs.docker.com/build/ci/github-actions/test-before-push/
+      - name: Build and export to Docker
+        uses: docker/build-push-action@v5
+        with:
+          context: .
+          load: true
+          build-args: |
+            BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}
+            CUDA=${{ matrix.cuda }}
+            PYTORCH_VERSION=${{ matrix.pytorch }}
+          file: ./docker/Dockerfile
+          tags: |
+            ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
+            ${{ (matrix.is_latest) && format('{0}-latest', steps.metadata.outputs.tags) || '' }}
+          labels: ${{ steps.metadata.outputs.labels }}
+      - name: Unit Tests
+        run: |
+          docker run --rm ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }} pytest --ignore=tests/e2e/ /workspace/axolotl/tests/

requirements.txt

@@ -2,7 +2,7 @@
 auto-gptq==0.5.1
 packaging
 peft==0.6.0
-transformers==4.36.2
+transformers @ git+https://github.com/huggingface/transformers.git@3cefac1d974db5e2825a0cb2b842883a628be7a0
 tokenizers==0.15.0
 bitsandbytes>=0.41.1
 accelerate==0.24.1

src/axolotl/monkeypatch/mixtral/__init__.py

@@ -17,6 +17,6 @@ def replace_mixtral_attn_with_multipack_flash_attn():
     transformers.models.mixtral.modeling_mixtral.MixtralModel.forward = (
         mixtral_model_forward
     )
-    transformers.models.mixtral.modeling_mixtral.MISTRAL_ATTENTION_CLASSES[
+    transformers.models.mixtral.modeling_mixtral.MIXTRAL_ATTENTION_CLASSES[
         "flash_attention_2"
     ] = MixtralMultipackFlashAttention2

src/axolotl/monkeypatch/mixtral/modeling_mixtral.py

@@ -261,7 +261,11 @@ def mixtral_model_forward(
     if inputs_embeds is None:
         inputs_embeds = self.embed_tokens(input_ids)
 
-    if attention_mask is not None and self._use_flash_attention_2 and use_cache:
+    if (
+        attention_mask is not None
+        and self._attn_implementation == "flash_attention_2"
+        and use_cache
+    ):
         is_padding_right = attention_mask[:, -1].sum().item() != batch_size
         if is_padding_right:
             raise ValueError(
@@ -270,7 +274,7 @@ def mixtral_model_forward(
                 " call `tokenizer.padding_side  = 'left'` before tokenizing the input. "
             )
 
-    if self._use_flash_attention_2:
+    if self._attn_implementation == "flash_attention_2":
         # 2d mask is passed through the layers
         attention_mask = (
             attention_mask

src/axolotl/utils/models.py

@@ -332,15 +332,18 @@ def load_model(
                 or cfg.is_mistral_derived_model
                 or model_config.model_type == "mixtral"
             ):
+                model_kwargs["attn_implementation"] = "flash_attention_2"
                 model_config._attn_implementation = (  # pylint: disable=protected-access
                     "flash_attention_2"
                 )
         else:
             if model_config.model_type == "mixtral":
+                model_kwargs["attn_implementation"] = "flash_attention_2"
                 model_config._attn_implementation = (  # pylint: disable=protected-access
                     "flash_attention_2"
                 )
             else:
+                model_kwargs["attn_implementation"] = "eager"
                 model_config._attn_implementation = (  # pylint: disable=protected-access
                     "eager"
                 )

tests/e2e/test_mixtral.py

@@ -0,0 +1,109 @@
+"""
+E2E tests for mixtral
+"""
+
+import logging
+import os
+import unittest
+from pathlib import Path
+
+from transformers.utils import is_torch_bf16_gpu_available
+
+from axolotl.cli import load_datasets
+from axolotl.common.cli import TrainerCliArgs
+from axolotl.train import train
+from axolotl.utils.config import normalize_config
+from axolotl.utils.dict import DictDefault
+
+from .utils import with_temp_dir
+
+LOG = logging.getLogger("axolotl.tests.e2e")
+os.environ["WANDB_DISABLED"] = "true"
+
+
+class TestMixtral(unittest.TestCase):
+    """
+    Test case for Llama models using LoRA
+    """
+
+    @with_temp_dir
+    def test_qlora(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "hf-internal-testing/Mixtral-tiny",
+                "tokenizer_config": "mistralai/Mixtral-8x7B-v0.1",
+                "flash_attention": True,
+                "sequence_len": 1024,
+                "load_in_4bit": True,
+                "adapter": "qlora",
+                "lora_r": 16,
+                "lora_alpha": 32,
+                "lora_dropout": 0.1,
+                "lora_target_linear": True,
+                "val_set_size": 0.1,
+                "special_tokens": {},
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 2,
+                "micro_batch_size": 2,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_bnb_8bit",
+                "lr_scheduler": "cosine",
+                "max_steps": 20,
+                "save_steps": 10,
+                "eval_steps": 10,
+            }
+        )
+        normalize_config(cfg)
+        cli_args = TrainerCliArgs()
+        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
+
+        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
+        assert (Path(temp_dir) / "adapter_model.bin").exists()
+
+    @with_temp_dir
+    def test_ft(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "hf-internal-testing/Mixtral-tiny",
+                "tokenizer_config": "mistralai/Mixtral-8x7B-v0.1",
+                "flash_attention": True,
+                "sequence_len": 1024,
+                "val_set_size": 0.1,
+                "special_tokens": {},
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 2,
+                "micro_batch_size": 2,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_bnb_8bit",
+                "lr_scheduler": "cosine",
+                "max_steps": 20,
+                "save_steps": 10,
+                "eval_steps": 10,
+            }
+        )
+        if is_torch_bf16_gpu_available():
+            cfg.bf16 = True
+        else:
+            cfg.fp16 = True
+        normalize_config(cfg)
+        cli_args = TrainerCliArgs()
+        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
+
+        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
+        assert (Path(temp_dir) / "pytorch_model.bin").exists()

tests/e2e/test_mixtral_samplepack.py

@@ -0,0 +1,123 @@
+"""
+E2E tests for mixtral
+"""
+
+import logging
+import os
+import unittest
+from pathlib import Path
+
+from transformers.utils import is_torch_bf16_gpu_available
+
+from axolotl.cli import load_datasets
+from axolotl.common.cli import TrainerCliArgs
+from axolotl.train import train
+from axolotl.utils.config import normalize_config
+from axolotl.utils.dict import DictDefault
+
+from .utils import with_temp_dir
+
+LOG = logging.getLogger("axolotl.tests.e2e")
+os.environ["WANDB_DISABLED"] = "true"
+
+
+class TestMixtral(unittest.TestCase):
+    """
+    Test case for Llama models using LoRA
+    """
+
+    @with_temp_dir
+    def test_qlora(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "hf-internal-testing/Mixtral-tiny",
+                "tokenizer_config": "mistralai/Mixtral-8x7B-v0.1",
+                "flash_attention": True,
+                "sequence_len": 2048,
+                "load_in_4bit": True,
+                "adapter": "qlora",
+                "lora_r": 16,
+                "lora_alpha": 32,
+                "lora_dropout": 0.1,
+                "lora_target_linear": True,
+                "val_set_size": 0.1,
+                "special_tokens": {},
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 2,
+                "micro_batch_size": 2,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_bnb_8bit",
+                "lr_scheduler": "cosine",
+                "max_steps": 20,
+                "save_steps": 10,
+                "eval_steps": 10,
+                "sample_packing": True,
+            }
+        )
+        if is_torch_bf16_gpu_available():
+            cfg.bf16 = True
+        else:
+            cfg.fp16 = True
+        normalize_config(cfg)
+        cli_args = TrainerCliArgs()
+        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
+
+        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
+        assert (Path(temp_dir) / "adapter_model.bin").exists()
+
+    @with_temp_dir
+    def test_ft(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "hf-internal-testing/Mixtral-tiny",
+                "tokenizer_config": "mistralai/Mixtral-8x7B-v0.1",
+                "flash_attention": True,
+                "sequence_len": 2048,
+                "val_set_size": 0.1,
+                "special_tokens": {},
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 2,
+                "micro_batch_size": 2,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_bnb_8bit",
+                "lr_scheduler": "cosine",
+                "max_steps": 20,
+                "save_steps": 10,
+                "eval_steps": 10,
+                "sample_packing": True,
+            }
+        )
+        if is_torch_bf16_gpu_available():
+            cfg.bf16 = True
+        else:
+            cfg.fp16 = True
+        normalize_config(cfg)
+        cli_args = TrainerCliArgs()
+        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
+
+        model, _ = train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
+        assert (
+            "axolotl.monkeypatch.mixtral.modeling_mixtral"
+            in model.model.layers[0].self_attn.__class__.__module__
+        )
+        assert (
+            "MixtralMultipackFlashAttention2"
+            in model.model.layers[0].self_attn.__class__.__name__
+        )
+        assert (Path(temp_dir) / "pytorch_model.bin").exists()