test v2batch w/ flex attn

fixes silly mistake

src/axolotl/core/trainer_builder.py

@@ -79,6 +79,7 @@ from axolotl.utils.chat_templates import get_chat_template_from_config
 from axolotl.utils.collators import (
     BatchSamplerDataCollatorForSeq2Seq,
     DataCollatorForSeq2Seq,
+    FlexBatchSamplerDataCollatorForSeq2Seq,
     MambaDataCollator,
     V2BatchSamplerDataCollatorForSeq2Seq,
 )
@@ -816,6 +817,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             Union[
                 V2BatchSamplerDataCollatorForSeq2Seq,
                 BatchSamplerDataCollatorForSeq2Seq,
+                FlexBatchSamplerDataCollatorForSeq2Seq,
                 DataCollatorForSeq2Seq,
                 DataCollatorWithFlattening,
                 RewardDataCollatorWithPadding,
@@ -827,7 +829,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             if "max_length" in kwargs:
                 kwargs.pop("max_length")
         elif use_batch_sampler_collator:
-            if self.cfg.model_config_type in SUPPORTED_MULTIPACK_MODEL_TYPES:
+            if self.cfg.flex_attention is True:
+                collator = V2BatchSamplerDataCollatorForSeq2Seq
+            elif self.cfg.model_config_type in SUPPORTED_MULTIPACK_MODEL_TYPES:
                 collator = V2BatchSamplerDataCollatorForSeq2Seq
             elif (
                 self.cfg.model_config_type in ["llama"]

src/axolotl/monkeypatch/flex_attn.py

@@ -0,0 +1,146 @@
+"""
+Taken from https://github.com/pytorch/torchtune/blob/main/torchtune/modules/attention_utils.py
+"""
+from typing import Union
+
+import torch
+from torch.nn.attention.flex_attention import BlockMask
+from torch.nn.attention.flex_attention import (
+    create_block_mask as create_block_causal_mask_flex,
+)
+
+_MaskType = Union[torch.Tensor, BlockMask]
+
+
+def create_block_causal_mask(
+    seq_lens: list[torch.Tensor], max_seq_len: int
+) -> torch.Tensor:
+    """
+    Given a batch tensor of seq lens defining the lengths of samples in each pack,
+    Construct a 2D block causal mask for each pack in the batch. For example, if
+    a single sample's seq_lens is [3, 2, 1], the mask would be::
+
+        mask = [
+            [1, 0, 0, 0, 0, 0],
+            [1, 1, 0, 0, 0, 0],
+            [1, 1, 1, 0, 0, 0],
+            [0, 0, 0, 1, 0, 0],
+            [0, 0, 0, 1, 1, 0],
+            [0, 0, 0, 0, 0, 1],
+        ]
+
+    Args:
+        seq_lens (List[torch.Tensor]): Sequence lengths of samples in each pack in the batch,
+            shape (batch_size, n), where n is the max number of sequences in a pack and can vary
+            across packs.
+
+
+    Returns:
+        Tensor: Block causal mask of shape (batch_size, max_seq_len, max_seq_len).
+    """
+    batch_block_attn_masks = []
+    batch_size = len(seq_lens)
+    for sample_idx in range(batch_size):
+        block_attn_masks = [
+            torch.trilu( # torch.tril(
+                torch.ones(seq_len, seq_len, dtype=torch.bool, device=seq_len.device)
+            )
+            for seq_len in seq_lens[sample_idx]
+        ]
+
+        """residue_len = max_seq_len - torch.sum(seq_lens[sample_idx])
+        block_attn_masks.append(
+            torch.tril(
+                torch.ones(
+                    residue_len, residue_len, dtype=torch.bool, device=seq_lens[sample_idx].device
+                )
+            )
+        )"""
+
+        batch_block_attn_masks.append(torch.block_diag(*block_attn_masks))
+
+    return torch.stack(batch_block_attn_masks)[:, None, :, :]
+
+
+def _get_document_ids_from_seq_lens(
+    seq_lens: list[torch.Tensor],
+) -> torch.Tensor:
+    """
+    Convert a batch tensor of seq lens into integer IDs denoting sample ownership.
+    For example, seq_lens = [2, 3, 1] would return [0, 0, 1, 1, 1, 2].
+
+    Args:
+        seq_lens (List[torch.Tensor]): Sequence lengths of samples in each pack in the batch,
+            shape (batch_size, n), where n is the max number of sequences in a pack and can vary
+            across packs.
+
+    Returns:
+        Tensor: Document IDs of shape (batch_size, max_seq_len).
+    """
+    batch_size = len(seq_lens)
+    batch_document_ids = []
+    for sample_idx in range(batch_size):
+        # We assume seq lens sum to max seq lens, so document_ids should be of
+        # shape (max_seq_len, )
+        document_ids = torch.cat(
+            [
+                torch.full((seq_len,), i, dtype=torch.long, device=seq_len.device)
+                for i, seq_len in enumerate(seq_lens[sample_idx])
+            ]
+        )
+        batch_document_ids.append(document_ids)
+    batch_document_ids = torch.stack(batch_document_ids)
+    return batch_document_ids
+
+
+def packed_block_causal_mask(
+    seq_lens: list[torch.Tensor], totalseqlens: list[int]
+) -> _MaskType:
+    """
+    Create a block causal document mask for a batch of packed sequences. If
+    flex attention is supported by the current hardware, block causal logic and
+    passing this into :func:`torch.nn.attention.flex_attention.create_block_mask`.
+    The resultant BlockMask is a compressed representation of the full block causal
+    mask. If on an older version, a standard 2D block causal mask is created and returned.
+
+    Args:
+        seq_lens (List[torch.Tensor]): Sequence lengths of samples in each pack in the batch,
+            shape (batch_size, n), where n is the max number of sequences in a pack and can vary
+            across packs.
+
+    Returns:
+        _MaskType: BlockMask or Tensor if torch version < 2.5.0.
+    """
+
+    document_ids = _get_document_ids_from_seq_lens(seq_lens)
+    batch_size , max_seq_len = document_ids.shape
+    document_ids = document_ids.to("cuda")
+    totalseqlens = totalseqlens.to("cuda")
+
+    # Instead of passing a tensor mask, flex attention requires a mask_mod function
+    # that determines which elements of QK^T should be included in the attention
+    # computation prior to the softmax. For sample packing, we need both the
+    # logic for both causal mask and document mask. See PyTorch's official
+    # blog post for more details: https://pytorch.org/blog/flexattention/#mask-mods
+    def mask_mod(b, h, q_idx, kv_idx):
+        """
+        Defines the logic of a block causal mask by combining both a standard causal mask
+        and a block diagonal document mask.
+
+        See :func:`~torchtune.modules.attention_utils.create_block_causal_mask`
+        for an illustration.
+        """
+        causal_mask = q_idx >= kv_idx
+        document_mask = document_ids[b, q_idx] == document_ids[b, kv_idx]
+        finite_mask = q_idx < totalseqlens[b]
+        return causal_mask & document_mask & finite_mask
+
+    return create_block_causal_mask_flex(
+        mask_mod,
+        batch_size,
+        None,
+        max_seq_len,
+        max_seq_len,
+        device="cuda",
+        BLOCK_SIZE=512,
+    )

src/axolotl/monkeypatch/utils.py

@@ -95,6 +95,103 @@ def get_cu_seqlens(attn_mask):
     return torch.stack(results).to(dtype=torch.int32), torch.stack(max_seq_lens)
 
 
+def get_packed_mask_from_pos_ids(position_ids):
+    if len(position_ids.shape) == 1:
+        position_ids = position_ids.unsqueeze(0)
+
+    device = position_ids.device
+    results = []
+
+    for i, row in enumerate(position_ids):
+        # Count the number of consecutive zeros from the right side
+        padding_length = (row == 0).int().flip(dims=[0]).cumprod(dim=0).sum().item()
+
+        # Adjust the row to exclude padding
+        adjusted_row = row[:-padding_length] if padding_length else row.clone()
+
+        # Find where the position resets to 0 (indicating a new sequence)
+        seq_starts = torch.cat(
+            [
+                torch.tensor([True], dtype=torch.bool, device=device),
+                adjusted_row[1:] == 0,
+            ]
+        )
+        # Get the indices where the sequence starts
+        start_indices = torch.cat(
+            [
+                torch.nonzero(seq_starts).unbind(dim=1)[0],
+                torch.tensor([len(adjusted_row)], dtype=torch.int32, device=device),
+            ]
+        )
+        # Calculate the sequence lengths
+        seq_lengths = start_indices[1:] - start_indices[:-1]
+        # Append the padding length to the sequence lengths
+        doc_mask = torch.ones(len(row), dtype=torch.int32, device=device)
+        for i, seq_len in enumerate(seq_lengths):
+            start_id = start_indices[i]
+            doc_mask[start_id : start_id + seq_len] = (
+                (i+1) * doc_mask[start_id : start_id + seq_len]
+            )
+        if padding_length:
+            doc_mask[len(adjusted_row) :] = 0 * doc_mask[len(adjusted_row) :]
+
+        results.append(doc_mask)
+
+    return torch.stack(results)
+
+
+def get_seqlens_from_pos_ids(position_ids):
+    """generate a sequence length set using pos ids for doc mask creation in flex attention"""
+    if len(position_ids.shape) == 1:
+        position_ids = position_ids.unsqueeze(0)
+    max_seq_len = position_ids.shape[1]
+
+    device = position_ids.device
+    results = []
+    totalseqlens = []
+
+    for row in position_ids:
+        # Count the number of consecutive zeros from the right side
+        padding_length = (row == 0).int().flip(dims=[0]).cumprod(dim=0).sum().item()
+
+        # Adjust the row to exclude padding
+        adjusted_row = row[:-padding_length] if padding_length else row.clone()
+
+        # Find where the position resets to 0 (indicating a new sequence)
+        seq_starts = torch.cat(
+            [
+                torch.tensor([True], dtype=torch.bool, device=device),
+                adjusted_row[1:] == 0,
+            ]
+        )
+        # Get the indices where the sequence starts
+        start_indices = torch.cat(
+            [
+                torch.nonzero(seq_starts).unbind(dim=1)[0],
+                torch.tensor([len(adjusted_row)], dtype=torch.int32, device=device),
+            ]
+        )
+        # Calculate the sequence lengths
+        seq_lengths = start_indices[1:] - start_indices[:-1]
+        # Append the padding length to the sequence lengths
+        if padding_length:
+            seq_lengths = torch.cat(
+                [
+                    seq_lengths,
+                    torch.tensor(
+                        [len(row) - torch.sum(seq_lengths)],
+                        dtype=torch.int32,
+                        device=device,
+                    ),
+                ]
+            )
+
+        results.append(seq_lengths)
+        totalseqlens.append(len(adjusted_row))
+
+    return results, torch.tensor(totalseqlens, dtype=torch.int32, device=device)
+
+
 def get_cu_seqlens_from_pos_ids(position_ids):
     """generate a cumulative sequence length mask for flash attention using pos ids"""
     if len(position_ids.shape) == 1:
@@ -176,7 +273,10 @@ def mask_2d_to_4d(
     when they attend to each other within that sequence.
     This expansion transforms the mask to lower triangular form to prevent future peeking.
     """
-    bsz, src_len = mask.size()
+
+    if len(mask.size()) == 4:
+        return mask
+    bsz, src_len = int(mask.size()[0]), int(mask.size()[1])
     tgt_len = tgt_len if tgt_len is not None else src_len
 
     mask = mask.unsqueeze(1).unsqueeze(2)

src/axolotl/utils/collators/__init__.py

@@ -4,6 +4,7 @@ shared axolotl collators for multipack, mamba, multimodal
 from .batching import (  # noqa: F401
     BatchSamplerDataCollatorForSeq2Seq,
     DataCollatorForSeq2Seq,
+    FlexBatchSamplerDataCollatorForSeq2Seq,
     PretrainingBatchSamplerDataCollatorForSeq2Seq,
     V2BatchSamplerDataCollatorForSeq2Seq,
 )

src/axolotl/utils/collators/batching.py

@@ -3,12 +3,21 @@ DataCollator for axolotl to pad labels and position_ids for packed sequences
 """
 
 from dataclasses import dataclass
-from typing import Any, Optional, Union
+from typing import Any, List, Optional, Union
 
 import numpy as np
+import torch
 from transformers import PreTrainedTokenizerBase
 from transformers.utils import PaddingStrategy
 
+from axolotl.monkeypatch.flex_attn import (
+    create_block_causal_mask,
+    packed_block_causal_mask,
+)
+from axolotl.monkeypatch.utils import (
+    get_packed_mask_from_pos_ids,
+)
+
 
 @dataclass
 class DataCollatorForSeq2Seq:
@@ -151,6 +160,42 @@ class BatchSamplerDataCollatorForSeq2Seq(DataCollatorForSeq2Seq):
         return super().__call__(out_features, return_tensors=return_tensors)
 
 
+@dataclass
+class FlexBatchSamplerDataCollatorForSeq2Seq(DataCollatorForSeq2Seq):
+    """
+    Collator for multipack specific to Flex Attention using the BatchSampler
+    """
+
+    def __call__(self, features, return_tensors=None):
+        if not isinstance(features[0], list):
+            features = [features]
+        out_features = [{} for _ in features]
+        for i, features_ in enumerate(features):
+            for feature in features_[0].keys():
+                if feature == "length":
+                    continue
+                elif feature == "attention_mask":
+                    """arrays = [
+                        i * np.array(item[feature])
+                        for i, item in enumerate(features_)
+                        if feature in item
+                    ]
+                    out_features[i][feature] = np.concatenate(arrays)"""
+                    continue
+                else:
+                    arrays = [
+                        np.array(item[feature]) for item in features_ if feature in item
+                    ]
+                    out_features[i][feature] = np.concatenate(arrays)
+        out = super().__call__(out_features, return_tensors=return_tensors)
+
+        # collated_seq_lens, totalseqlens = get_seqlens_from_pos_ids(out["position_ids"])
+        # out["attention_mask"] = packed_block_causal_mask(collated_seq_lens, totalseqlens)
+        out["attention_mask"] = get_packed_mask_from_pos_ids(out["position_ids"])
+        # out["attention_mask"] = create_block_causal_mask(collated_seq_lens, max_seq_len)
+        return out
+
+
 @dataclass
 class V2BatchSamplerDataCollatorForSeq2Seq(DataCollatorForSeq2Seq):
     """

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

@@ -784,6 +784,7 @@ class AxolotlInputConfig(
     xformers_attention: Optional[bool] = None
     sdp_attention: Optional[bool] = None
     s2_attention: Optional[bool] = None
+    flex_attention: Optional[bool] = None
     flash_attention: Optional[bool] = None
     flash_attn_cross_entropy: Optional[bool] = None
     flash_attn_rms_norm: Optional[bool] = None
@@ -1679,6 +1680,26 @@ class AxolotlConfigWCapabilities(AxolotlInputConfig):
                 )
         return data
 
+    @model_validator(mode="before")
+    @classmethod
+    def check_flex_torch_version(cls, data):
+        if (data.get("flex_attention") is not None) and (
+            data.get("flex_attention") is True
+        ):
+            env_capabilities = data.get("env_capabilities", {})
+            torch_version = env_capabilities.get("torch_version")
+
+            if torch_version is None:
+                import torch
+
+                torch_version = str(torch.__version__).split("+", maxsplit=1)[0]
+
+            if version.parse(torch_version) < version.parse("2.5.1"):
+                raise ValueError(
+                    "Flex attention is not supported on torch version < 2.5.1"
+                )
+        return data
+
     @model_validator(mode="before")
     @classmethod
     def check_torch_compile_auto(cls, data):

src/axolotl/utils/models.py

@@ -403,7 +403,7 @@ class ModelLoader:
 
         if (
             self.cfg.model_config_type in SUPPORTED_MULTIPACK_MODEL_TYPES
-            and self.cfg.flash_attention
+            and (self.cfg.flash_attention or self.cfg.flex_attention)
             and self.cfg.sample_packing
         ):
             if "auto_map" in self.model_config:
@@ -708,7 +708,13 @@ class ModelLoader:
         """
         sample packing uses custom FA2 patch
         """
-        if self.cfg.flash_attention:
+
+        if self.cfg.flex_attention:
+            self.model_kwargs["attn_implementation"] = "flex_attention"
+            self.model_config._attn_implementation = (  # pylint: disable=protected-access
+                "flex_attention"
+            )
+        elif self.cfg.flash_attention:
             if not self.cfg.sample_packing and self.cfg.s2_attention:
                 pass
             self.model_kwargs["attn_implementation"] = "flash_attention_2"
@@ -1100,7 +1106,7 @@ class ModelLoader:
         should_convert = (
             # LlamaRMSNorm layers are in fp32 after kbit_training or full finetune, so we need to
             # convert them back to fp16/bf16 for flash-attn compatibility.
-            ((needs_fa2_dtype or self.cfg.flash_attention) and not qlora_fsdp)
+            ((needs_fa2_dtype or self.cfg.flash_attention or self.cfg.flex_attention) and not qlora_fsdp)
             or self.cfg.cut_cross_entropy  # Cut cross entropy requires embedding layers to be in fp16/bf16 for backward pass
         )