Pretrain multipack (#2278)

* fix for pretrain with packing

* fix model name and loss expected

* make sure to check with micro batch size for pretraining

* change loss threshholds based on parametrization

* make tests smaller for CI

* fix pretrain packing

* fix pretrain packing test

* address pr feedback

src/axolotl/core/trainer_builder.py

@@ -1880,6 +1880,8 @@ def build_collator(
         if training_args.pretraining:
             if self.cfg.pretraining_sample_concatenation is False:
                 return DataCollatorForSeq2Seq(self.tokenizer, **kwargs)
+            if self.cfg.micro_batch_size > 1:
+                return DataCollatorForSeq2Seq(self.tokenizer, **kwargs)
             return None
 
         if self.cfg.model_config_type == "mamba":

src/axolotl/utils/data/pretraining.py

@@ -191,7 +191,7 @@ def wrap_pretraining_dataset(
             tokenizer,
             return_tensors="pt",
             padding=True,
-            pad_to_multiple_of=max_tokens * batch_size,
+            pad_to_multiple_of=max_tokens,
             multipack_attn=cfg.pretrain_multipack_attn,
         )
         encode = functools.partial(
@@ -201,8 +201,6 @@ def wrap_pretraining_dataset(
             max_seq_length=max_tokens,
             batch_size=batch_size,
             multipack_attn=cfg.pretrain_multipack_attn,
-            group_size=cfg.sample_packing_group_size,
-            bin_size=cfg.sample_packing_bin_size,
         )
         # set this to 1 so downstream data_loader doesn't try to increase the batch again
         cfg.micro_batch_size = 1
@@ -247,9 +245,7 @@ def encode_packed_pretraining(
     examples: Dict[str, List],
     max_seq_length: int = 2048,
     batch_size: int = 4,
-    multipack_attn: Optional[bool] = False,
-    group_size: int = 100000,
-    bin_size: int = 200,
+    multipack_attn: Optional[bool] = True,
 ) -> Dict[str, List]:
     # pylint: disable=duplicate-code
     # tokenize all the examples
@@ -260,15 +256,16 @@ def encode_packed_pretraining(
         train_dataset,
         max_seq_length,
         skip_position_ids=not multipack_attn,
+        # FIXME using attention mask unpad/pad with trainer and packed pretraining is broken atm
+        # workaround by using the position id logic for now in trainer
+        drop_attention_mask=multipack_attn,
     )
 
     sampler = MultipackBatchSampler(
         sampler=RandomSampler(train_dataset),
         lengths=get_dataset_lengths(train_dataset),
         batch_size=1,
         batch_max_len=batch_size * max_seq_length,
-        group_size=group_size,
-        bin_size=bin_size,
         drop_last=True,
     )
 

src/axolotl/utils/trainer.py

@@ -310,19 +310,22 @@ def drop_no_trainable_tokens(sample):
 
 
 def process_pretraining_datasets_for_packing(
-    train_dataset, sequence_len, skip_position_ids=True
+    train_dataset, sequence_len, skip_position_ids=True, drop_attention_mask=False
 ):
     drop_long = partial(drop_long_seq, sequence_len=sequence_len)
 
     train_dataset = train_dataset.filter(
         drop_long,
         desc="Dropping Long Sequences",
+        load_from_cache_file=False,
     )
-    if skip_position_ids:
+    if not skip_position_ids:
         train_dataset = train_dataset.map(
             add_position_ids,
             desc="Add position_id column (Pretraining Sample Packing)",
         )
+    if drop_attention_mask:
+        train_dataset = train_dataset.remove_columns("attention_mask")
 
     return train_dataset
 

tests/e2e/test_llama_pretrain.py

@@ -13,7 +13,7 @@
 from axolotl.utils.config import normalize_config
 from axolotl.utils.dict import DictDefault
 
-from .utils import check_model_output_exists
+from .utils import check_model_output_exists, check_tensorboard
 
 LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"
@@ -28,19 +28,25 @@ class TestPretrainLlama:
         "sample_packing",
         [True, False],
     )
-    def test_pretrain(self, temp_dir, sample_packing):
+    @pytest.mark.parametrize(
+        "pretrain_multipack_attn",
+        [True, False],
+    )
+    def test_pretrain(self, temp_dir, sample_packing, pretrain_multipack_attn):
+        if not sample_packing and pretrain_multipack_attn:
+            return
+
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
-                "tokenizer_type": "LlamaTokenizer",
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
                 "flash_attention": True,
                 "sequence_len": 1024,
                 "sample_packing": sample_packing,
+                "pretrain_multipack_attn": pretrain_multipack_attn,
+                "dataset_processes": 1,
                 "special_tokens": {
-                    "unk_token": "<unk>",
-                    "bos_token": "<s>",
-                    "eos_token": "</s>",
+                    "pad_token": "<|endoftext|>",
                 },
                 "pretraining_dataset": [
                     {
@@ -51,7 +57,7 @@ def test_pretrain(self, temp_dir, sample_packing):
                 ],
                 "max_steps": 5,
                 "num_epochs": 1,
-                "micro_batch_size": 1,
+                "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "val_set_size": 0.0,
                 "output_dir": temp_dir,
@@ -60,6 +66,7 @@ def test_pretrain(self, temp_dir, sample_packing):
                 "lr_scheduler": "cosine",
                 "save_safetensors": True,
                 "bf16": "auto",
+                "use_tensorboard": True,
             }
         )
         normalize_config(cfg)
@@ -68,3 +75,12 @@ def test_pretrain(self, temp_dir, sample_packing):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        loss_threshold = 3.5
+        if sample_packing and not pretrain_multipack_attn:
+            loss_threshold = 6.5
+        check_tensorboard(
+            temp_dir + "/runs",
+            "train/train_loss",
+            loss_threshold,
+            "Train Loss is too high",
+        )

tests/test_packed_pretraining.py

@@ -41,6 +41,7 @@ def test_packing_stream_dataset(self):
                     }
                 ],
                 "sample_packing": True,
+                "pretrain_multipack_attn": True,
                 "pad_to_sequence_len": True,
                 "sequence_len": 2048,
                 "micro_batch_size": 2,
@@ -87,9 +88,11 @@ def test_packing_stream_dataset(self):
             assert data["labels"].shape == torch.Size(
                 [1, original_bsz * cfg.sequence_len]
             )
-            assert data["attention_mask"].shape == torch.Size(
-                [1, original_bsz * cfg.sequence_len]
-            )
+            assert "attention_mask" not in data
+            # FIXME add back once we fix packing unpad/pad with attention mask
+            # assert data["attention_mask"].shape == torch.Size(
+            #     [1, original_bsz * cfg.sequence_len]
+            # )
             idx += 1