Revert "Fix sample packing producing longer sequences than specified by `sequ…"

This reverts commit 8dfadc2b3c.

docs/lora_optims.qmd

@@ -12,6 +12,7 @@ to leverage operator fusion and tensor re-use in order to improve speed and redu
 memory usage during the forward and backward passes of these calculations.
 
 We currently support several common model architectures, including (but not limited to):
+
 - `llama`
 - `mistral`
 - `qwen2`
@@ -82,7 +83,7 @@ lora_o_kernel: true
 ## Requirements
 
 - One or more NVIDIA or AMD GPUs (in order to use the Triton kernels)
-    - AMD can be used with experimental Triton support by setting the environment variable `TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1`
+    - Note: Set `TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1` to enable [memory-efficient attention on AMD GPUs](https://github.com/ROCm/aotriton/issues/16#issuecomment-2346675491)
 - Targeted LoRA adapters cannot use Dropout
     - This may limit model expressivity / cause overfitting
 - Targeted LoRA adapters cannot have bias terms

src/axolotl/__init__.py

@@ -4,4 +4,4 @@ import pkgutil
 
 __path__ = pkgutil.extend_path(__path__, __name__)  # Make this a namespace package
 
-__version__ = "0.7.0"
+__version__ = "0.8.0.dev0"

tests/prompt_strategies/conftest.py

@@ -125,6 +125,12 @@ def fixture_llama3_tokenizer():
     return tokenizer
 
 
+@pytest.fixture(name="smollm2_tokenizer", scope="session", autouse=True)
+def fixture_smollm2_tokenizer():
+    tokenizer = AutoTokenizer.from_pretrained("HuggingFaceTB/SmolLM2-135M")
+    return tokenizer
+
+
 @pytest.fixture(name="mistralv03_tokenizer", scope="session", autouse=True)
 def fixture_mistralv03_tokenizer():
     tokenizer = AutoTokenizer.from_pretrained(

tests/prompt_strategies/test_dpo_chatml.py

@@ -0,0 +1,61 @@
+"""
+Tests for loading DPO preference datasets with chatml formatting
+"""
+import unittest
+
+import pytest
+
+from axolotl.prompt_strategies.dpo import load as load_dpo
+from axolotl.utils.data.rl import load_prepare_preference_datasets
+from axolotl.utils.dict import DictDefault
+
+
+@pytest.fixture(name="minimal_dpo_cfg")
+def fixture_cfg():
+    return DictDefault(
+        {
+            "base_model": "HuggingFaceTB/SmolLM2-135M",
+            "tokenizer_config": "HuggingFaceTB/SmolLM2-135M",
+            "rl": "dpo",
+            "learning_rate": 0.000001,
+            "micro_batch_size": 1,
+            "gradient_accumulation_steps": 1,
+            "special_tokens": {
+                "pad_token": "<|endoftext|>",
+            },
+            "sequence_len": 2048,
+        }
+    )
+
+
+class TestDPOChatml:
+    """
+    Test loading DPO preference datasets with chatml formatting
+    """
+
+    def test_default(self, minimal_dpo_cfg):
+        cfg = DictDefault(
+            {
+                "datasets": [
+                    {
+                        "path": "argilla/distilabel-intel-orca-dpo-pairs",
+                        "type": "chatml",
+                        "split": "train[:1%]",
+                    }
+                ]
+            }
+            | minimal_dpo_cfg
+        )
+
+        # test that dpo.load works
+        load_dpo("chatml", cfg)
+        # now actually load the datasets with the strategy
+        train_ds, _ = load_prepare_preference_datasets(cfg)
+        assert train_ds[0]["prompt"].startswith("<|im_start|>")
+        assert train_ds[0]["prompt"].endswith("<|im_start|>assistant\n")
+        assert "chosen" in train_ds[0]
+        assert "rejected" in train_ds[0]
+
+
+if __name__ == "__main__":
+    unittest.main()