Pretrain multipack v2

requirements.txt

@@ -40,3 +40,4 @@ gcsfs
 # adlfs
 
 trl @ git+https://github.com/huggingface/trl.git@0ee349dcd43b0f4b3169449f16751c38ac4a609f
+zstandard==0.22.0

src/axolotl/utils/collators.py

@@ -217,13 +217,24 @@ class PretrainingBatchSamplerDataCollatorForSeq2Seq(DataCollatorForSeq2Seq):
     Collator for multipack specific to the using the BatchSampler
     """
 
+    def __init__(self, *args, multipack_attn=True, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.multipack_attn = multipack_attn
+
     def __call__(self, features, return_tensors=None):
         chunked_data = {}
         for feature in features.keys():
             if feature == "length":
                 continue
             if feature == "attention_mask":
-                arrays = [(1) * np.array(item) for item in features[feature]]
+                if self.multipack_attn:
+                    arrays = [
+                        (i + 1) * np.array(item[feature])
+                        for i, item in enumerate(features[feature])
+                        if feature in item
+                    ]
+                else:
+                    arrays = [(1) * np.array(item) for item in features[feature]]
                 chunked_data[feature] = np.concatenate(arrays)
             else:
                 arrays = [np.array(item) for item in features[feature]]

src/axolotl/utils/config/models/input/v0_4_1/__init__.py

@@ -489,6 +489,14 @@ class Config:
     eval_sample_packing: Optional[bool] = None
     pad_to_sequence_len: Optional[bool] = None
 
+    pretrain_multipack_buffer_size: Optional[int] = 10_000
+    pretrain_multipack_attn: Optional[bool] = Field(
+        default=True,
+        metadata={
+            "help": "whether to prevent cross attention for packed sequences during pretraining",
+        },
+    )
+
     xformers_attention: Optional[bool] = None
     sdp_attention: Optional[bool] = None
     s2_attention: Optional[bool] = None

src/axolotl/utils/data.py

@@ -105,6 +105,7 @@ def prepare_dataset(cfg, tokenizer):
             max_tokens=cfg.sequence_len,
             batch_size=cfg.micro_batch_size,
             seed=cfg.seed or 42,
+            buffer_size=cfg.pretrain_multipack_buffer_size or 10_000,
         )
         # https://discuss.huggingface.co/t/how-to-use-huggingface-trainer-streaming-datasets-without-wrapping-it-with-torchdatas-iterablewrapper/25230
         train_dataset = train_dataset.with_format("torch")
@@ -813,13 +814,15 @@ def wrap_pretraining_dataset(
             return_tensors="pt",
             padding=True,
             pad_to_multiple_of=max_tokens * batch_size,
+            multipack_attn=cfg.pretrain_multipack_attn,
         )
         encode = functools.partial(
             encode_packed_pretraining,
             collate_fn,
             ds_wrapper_fn,
             max_seq_length=max_tokens,
             batch_size=batch_size,
+            multipack_attn=cfg.pretrain_multipack_attn,
         )
         # set this to 1 so downstream data_loader doesn't try to increase the batch again
         cfg.micro_batch_size = 1
@@ -849,14 +852,17 @@ def encode_packed_pretraining(
     examples: Dict[str, List],
     max_seq_length: int = 2048,
     batch_size: int = 4,
+    multipack_attn: Optional[bool] = False,
 ) -> Dict[str, List]:
     # pylint: disable=duplicate-code
     # tokenize all the examples
     # rows get split with stride (overlap)
     train_dataset = ds_wrapper(Dataset.from_dict(examples))[0]
 
     train_dataset = process_pretraining_datasets_for_packing(
-        train_dataset, max_seq_length
+        train_dataset,
+        max_seq_length,
+        skip_position_ids=not multipack_attn,
     )
 
     sampler = MultipackBatchSampler(

src/axolotl/utils/trainer.py

@@ -172,17 +172,21 @@ def process_datasets_for_packing(cfg, train_dataset, eval_dataset):
     return train_dataset, eval_dataset
 
 
-def process_pretraining_datasets_for_packing(train_dataset, sequence_len):
+def process_pretraining_datasets_for_packing(
+    train_dataset, sequence_len, skip_position_ids=True
+):
     drop_long = partial(drop_long_seq, sequence_len=sequence_len)
 
     train_dataset = train_dataset.filter(
         drop_long,
         desc="Dropping Long Sequences",
     )
-    train_dataset = train_dataset.map(
-        add_position_ids,
-        desc="Add position_id column (Pretraining Sample Packing)",
-    )
+    if skip_position_ids:
+        train_dataset = train_dataset.map(
+            add_position_ids,
+            desc="Add position_id column (Pretraining Sample Packing)",
+        )
+
     return train_dataset