Feat: Allow usage of native Mistral FA when no sample_packing (#669)

* Allow usage of native Mistral FA when no sample_packing

* fix: do not apply custom patch when sample_pack off

* chore: lint

* chore: pin transformer to v4.35.0.dev0

* fix: split sample_packing to separate test

requirements.txt

@@ -4,7 +4,7 @@ torch==2.0.1
 auto-gptq
 packaging
 peft @ git+https://github.com/huggingface/peft.git
-transformers @ git+https://github.com/huggingface/transformers.git@5e11d72d4d0939138fbabfebe9a69d2061519547
+transformers @ git+https://github.com/huggingface/transformers.git@bd6205919aad4d3a2300a39a98a642f1cc3a5348
 bitsandbytes>=0.41.1
 accelerate @ git+https://github.com/huggingface/accelerate@80da9cfb09bb3cc9f1b385cb55d6b90d025a5fd9
 deepspeed

src/axolotl/utils/models.py

@@ -149,7 +149,7 @@ def load_model(
         # Note: This might overwrite previous additional_special_tokens
         tokenizer.add_special_tokens({"additional_special_tokens": [MEM_TOKEN]})
 
-    if cfg.is_mistral_derived_model and cfg.flash_attention:
+    if cfg.is_mistral_derived_model and cfg.flash_attention and cfg.sample_packing:
         from axolotl.monkeypatch.mistral_attn_hijack_flash import (
             replace_mistral_attn_with_flash_attn,
         )
@@ -200,7 +200,11 @@ def load_model(
         )
     # sample packing uses custom FA2 patch
     if cfg.flash_attention and not cfg.sample_packing:
-        if cfg.is_llama_derived_model or cfg.is_falcon_derived_model:
+        if (
+            cfg.is_llama_derived_model
+            or cfg.is_falcon_derived_model
+            or cfg.is_mistral_derived_model
+        ):
             model_kwargs["use_flash_attention_2"] = True
     try:
         if cfg.is_llama_derived_model and not cfg.trust_remote_code and not cfg.gptq:

tests/e2e/test_mistral.py

@@ -71,53 +71,6 @@ class TestMistral(unittest.TestCase):
         train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
         assert (Path(output_dir) / "adapter_model.bin").exists()
 
-    def test_lora_packing(self):
-        # pylint: disable=duplicate-code
-        output_dir = tempfile.mkdtemp()
-        cfg = DictDefault(
-            {
-                "base_model": "openaccess-ai-collective/tiny-mistral",
-                "base_model_config": "openaccess-ai-collective/tiny-mistral",
-                "flash_attention": True,
-                "sample_packing": True,
-                "sequence_len": 1024,
-                "load_in_8bit": True,
-                "adapter": "lora",
-                "lora_r": 32,
-                "lora_alpha": 64,
-                "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": 2,
-                "micro_batch_size": 2,
-                "gradient_accumulation_steps": 1,
-                "output_dir": output_dir,
-                "learning_rate": 0.00001,
-                "optimizer": "adamw_torch",
-                "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(output_dir) / "adapter_model.bin").exists()
-
     def test_ft(self):
         # pylint: disable=duplicate-code
         output_dir = tempfile.mkdtemp()
@@ -161,48 +114,3 @@ class TestMistral(unittest.TestCase):
 
         train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
         assert (Path(output_dir) / "pytorch_model.bin").exists()
-
-    def test_ft_packing(self):
-        # pylint: disable=duplicate-code
-        output_dir = tempfile.mkdtemp()
-        cfg = DictDefault(
-            {
-                "base_model": "openaccess-ai-collective/tiny-mistral",
-                "base_model_config": "openaccess-ai-collective/tiny-mistral",
-                "flash_attention": True,
-                "sample_packing": True,
-                "sequence_len": 1024,
-                "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": 2,
-                "micro_batch_size": 2,
-                "gradient_accumulation_steps": 1,
-                "output_dir": output_dir,
-                "learning_rate": 0.00001,
-                "optimizer": "adamw_torch",
-                "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(output_dir) / "pytorch_model.bin").exists()

tests/e2e/test_mistral_samplepack.py

@@ -0,0 +1,118 @@
+"""
+E2E tests for lora llama
+"""
+
+import logging
+import os
+import tempfile
+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
+
+LOG = logging.getLogger("axolotl.tests.e2e")
+os.environ["WANDB_DISABLED"] = "true"
+
+
+class TestMistral(unittest.TestCase):
+    """
+    Test case for Llama models using LoRA
+    """
+
+    def test_lora_packing(self):
+        # pylint: disable=duplicate-code
+        output_dir = tempfile.mkdtemp()
+        cfg = DictDefault(
+            {
+                "base_model": "openaccess-ai-collective/tiny-mistral",
+                "base_model_config": "openaccess-ai-collective/tiny-mistral",
+                "flash_attention": True,
+                "sample_packing": True,
+                "sequence_len": 1024,
+                "load_in_8bit": True,
+                "adapter": "lora",
+                "lora_r": 32,
+                "lora_alpha": 64,
+                "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": 2,
+                "micro_batch_size": 2,
+                "gradient_accumulation_steps": 1,
+                "output_dir": output_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_torch",
+                "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(output_dir) / "adapter_model.bin").exists()
+
+    def test_ft_packing(self):
+        # pylint: disable=duplicate-code
+        output_dir = tempfile.mkdtemp()
+        cfg = DictDefault(
+            {
+                "base_model": "openaccess-ai-collective/tiny-mistral",
+                "base_model_config": "openaccess-ai-collective/tiny-mistral",
+                "flash_attention": True,
+                "sample_packing": True,
+                "sequence_len": 1024,
+                "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": 2,
+                "micro_batch_size": 2,
+                "gradient_accumulation_steps": 1,
+                "output_dir": output_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_torch",
+                "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(output_dir) / "pytorch_model.bin").exists()