xformers attention with packing (#2619)

* xformers attention with packing

* wire up the patch

* fix xformers + packing validation

* fix warning

* reorder the packing check

* fix fp16 / bf16 reset when using fp16 with bf16 auto

* fix seq lens calc to drop hanging sequences

* handle xformers patch for inference too

* fix batch size setter

* fix xformers inference

* add colab callback to fix inference post train

* PR feedback

docs/config.qmd

@@ -73,11 +73,12 @@ load_in_8bit: true
 load_in_4bit:
 
 # Use CUDA bf16
-bf16: true # bool or 'full' for `bf16_full_eval`. require >=ampere
+bf16: true # bool or 'full' for `bf16_full_eval`, or 'auto' for automatic detection. require >=ampere
 # Use CUDA fp16
 fp16: true
 # Use CUDA tf32
 tf32: true # require >=ampere
+# Note: if bf16 is set to 'auto', and fp16 is set to true, we will prefer the explict fp16 setting
 
 # No AMP (automatic mixed precision)
 bfloat16: true # require >=ampere

src/axolotl/core/trainer_builder.py

@@ -21,6 +21,7 @@ import importlib.util
 import inspect
 import logging
 import math
+import os
 import sys
 from abc import abstractmethod
 from pathlib import Path
@@ -72,6 +73,7 @@ from axolotl.utils.callbacks import (
     SaveBetterTransformerModelCallback,
     bench_eval_callback_factory,
     causal_lm_bench_eval_callback_factory,
+    colab_inference_post_train_callback,
     log_prediction_callback_factory,
 )
 from axolotl.utils.callbacks.lisa import lisa_callback_factory
@@ -293,6 +295,10 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         if self.cfg.lisa_step_interval and self.cfg.lisa_n_layers:
             callbacks.append(lisa_callback_factory(trainer))
 
+        if any("COLAB_" in key for key in os.environ):
+            ColabCallback = colab_inference_post_train_callback(trainer)
+            callbacks.append(ColabCallback(self.cfg))
+
         callbacks.extend(super().get_post_trainer_create_callbacks(trainer=trainer))
         return callbacks
 

src/axolotl/monkeypatch/attention/__init__.py

@@ -0,0 +1,19 @@
+"""
+attention module for attention monkeypatches
+"""
+
+from transformers.integrations.flash_attention import flash_attention_forward
+
+
+def patch_xformers_attn_over_fa2():
+    from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS
+
+    from .xformers import xformers_attention_forward
+
+    ALL_ATTENTION_FUNCTIONS["flash_attention_2"] = xformers_attention_forward
+
+
+def unpatch_xformers_attn_over_fa2():
+    from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS
+
+    ALL_ATTENTION_FUNCTIONS["flash_attention_2"] = flash_attention_forward()

src/axolotl/monkeypatch/attention/xformers.py

@@ -0,0 +1,160 @@
+"""
+xformers attention implementation for packing
+"""
+
+from typing import Optional
+
+import torch
+import xformers
+import xformers.ops.fmha
+from transformers.modeling_flash_attention_utils import (
+    _upad_input,
+)
+
+from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids
+
+xformers_attention = xformers.ops.fmha.memory_efficient_attention
+
+
+def xformers_attention_forward(
+    module: torch.nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    dropout: float = 0.0,  # pylint: disable=unused-argument
+    scaling: Optional[float] = None,  # pylint: disable=unused-argument
+    sliding_window: Optional[int] = None,  # pylint: disable=unused-argument
+    softcap: Optional[float] = None,  # pylint: disable=unused-argument
+    cu_seq_lens_q: Optional[torch.LongTensor] = None,
+    cu_seq_lens_k: Optional[torch.LongTensor] = None,
+    max_length_q: Optional[int] = None,
+    max_length_k: Optional[int] = None,  # pylint: disable=unused-argument
+    **kwargs,  # pylint: disable=unused-argument
+):
+    # Get dimensions
+    # query: [batch, heads, seq_len, hidden_dim]
+    batch_size = query.size(0)
+    query_length = query.shape[2]
+    key_length = key.shape[2]
+
+    # Default causal mask
+    attn_bias = xformers.ops.LowerTriangularMask()
+
+    # Check if we have sliding window attention
+    has_sliding_window = sliding_window is not None and sliding_window < query_length
+
+    # Transpose dimensions for xformers (Q: [b, h, s, d] -> [b, s, h, d])
+    query = query.transpose(1, 2)
+    key = key.transpose(1, 2)
+    value = value.transpose(1, 2)
+
+    # Get GQA parameters
+    num_attention_heads = module.config.num_attention_heads
+    num_key_value_heads = module.config.num_key_value_heads
+    head_dim = query.size(-1)
+    is_gqa = num_attention_heads != num_key_value_heads
+    n_groups = num_attention_heads // num_key_value_heads if is_gqa else 1
+
+    # If position_ids is provided and check all examples do not contain only 1 sequence, If tensor in increasing
+    # then we probably have one sequence, otherwise it is packed. Additionally check we are in pre-fill/training stage.
+    # Use `flash_attn_varlen_func` to prevent cross-example attention and also allow padding free approach
+    if position_ids is not None and (
+        max_length_q is not None
+        or (query_length != 1 and not (torch.diff(position_ids, dim=-1) >= 0).all())
+    ):
+        if cu_seq_lens_q is None or cu_seq_lens_k is None:
+            cu_seq_lens_q = get_cu_seqlens_from_pos_ids(position_ids)[0]
+            cu_seq_lens_q = cu_seq_lens_q.squeeze()
+            seq_lengths = cu_seq_lens_q[1:] - cu_seq_lens_q[:-1]
+            attn_bias = (
+                xformers.ops.fmha.attn_bias.BlockDiagonalCausalMask.from_seqlens(
+                    q_seqlen=seq_lengths.tolist(),
+                )
+            )
+        else:
+            query = query.reshape(-1, query.size(-2), query.size(-1))
+            key = key.reshape(-1, key.size(-2), key.size(-1))
+            value = value.reshape(-1, value.size(-2), value.size(-1))
+
+        # Handle GQA
+        if is_gqa:
+            key = key.repeat_interleave(n_groups, dim=2)
+            value = value.repeat_interleave(n_groups, dim=2)
+
+    elif attention_mask is not None:
+        query, key, value, _, cu_seq_lens, _ = _upad_input(
+            query, key, value, attention_mask, query_length
+        )
+        cu_seq_lens_q, cu_seq_lens_k = cu_seq_lens
+        seq_lengths = []
+        for i in range(len(cu_seq_lens_q) - 1):
+            seq_lengths.append(cu_seq_lens_q[i + 1] - cu_seq_lens_q[i])
+        attn_bias = xformers.ops.fmha.attn_bias.BlockDiagonalCausalMask.from_seqlens(
+            q_seqlen=seq_lengths,
+            kv_seqlen=seq_lengths,
+        )
+
+        # Handle GQA
+        if is_gqa:
+            key = key.repeat_interleave(n_groups, dim=2)
+            value = value.repeat_interleave(n_groups, dim=2)
+    else:
+        # Handle Group Query Attention (GQA) using view/expand approach from reference
+        key = key.view(batch_size, key_length, num_key_value_heads, 1, head_dim)
+        value = value.view(batch_size, key_length, num_key_value_heads, 1, head_dim)
+        key = key.expand(
+            batch_size, key_length, num_key_value_heads, n_groups, head_dim
+        )
+        value = value.expand(
+            batch_size, key_length, num_key_value_heads, n_groups, head_dim
+        )
+
+        if module.training:
+            key = key.reshape(batch_size, key_length, num_attention_heads, head_dim)
+            value = value.reshape(batch_size, key_length, num_attention_heads, head_dim)
+
+            if has_sliding_window:
+                query = query.view(
+                    1, batch_size * query_length, num_attention_heads, head_dim
+                )
+                key = key.view(
+                    1, batch_size * key_length, num_attention_heads, head_dim
+                )
+                value = value.view(
+                    1, batch_size * key_length, num_attention_heads, head_dim
+                )
+        else:
+            query = query.view(
+                batch_size, query_length, num_key_value_heads, n_groups, head_dim
+            )
+
+            # If we need a sliding window attention
+            if has_sliding_window:
+                query = query.view(
+                    1,
+                    batch_size * query_length,
+                    num_key_value_heads,
+                    n_groups,
+                    head_dim,
+                )
+                key = key.view(
+                    1, batch_size * key_length, num_key_value_heads, n_groups, head_dim
+                )
+                value = value.view(
+                    1, batch_size * key_length, num_key_value_heads, n_groups, head_dim
+                )
+
+    # Run the xformers attention
+    attn_output = xformers_attention(
+        query,
+        key,
+        value,
+        attn_bias=attn_bias,
+    )
+
+    attn_output = attn_output.view(
+        batch_size, -1, attn_output.size(-2), attn_output.size(-1)
+    )
+    return attn_output, None

src/axolotl/utils/callbacks/__init__.py

@@ -868,3 +868,28 @@ class GCCallback(TrainerCallback):
     ):
         torch.cuda.empty_cache()
         gc.collect()
+
+
+def colab_inference_post_train_callback(trainer: Trainer):
+    class ColabCallback(TrainerCallback):
+        """Callback to prep model for inference on Google Colab"""
+
+        def __init__(self, cfg):
+            self.gpu_name = torch.cuda.get_device_name(0)
+            self.cfg = cfg
+
+        def on_train_end(
+            self, args, state, control, **kwargs
+        ):  # pylint: disable=unused-argument
+            """
+            handle T4 gpu, we need to convert attention to eager for inference
+            """
+            if "Tesla T4" in self.gpu_name and self.cfg.xformers_attention:
+                trainer.model.config._attn_implementation = (  # pylint: disable=protected-access
+                    "eager"
+                )
+            trainer.model.gradient_checkpointing_disable()
+            trainer.model.config.use_cache = True
+            trainer.model.eval()
+
+    return ColabCallback

src/axolotl/utils/config/__init__.py

@@ -70,6 +70,9 @@ def resolve_dtype(cfg):
             if cfg.fp16 is None and not cfg.float16:
                 cfg.fp16 = True
 
+    if cfg.fp16 and cfg.bf16 == "auto":
+        cfg.bf16 = False
+
     if cfg.device == "mps":
         cfg.load_in_8bit = False
         cfg.tf32 = False

src/axolotl/utils/models.py

@@ -540,6 +540,11 @@ class ModelLoader:
         self.auto_model_loader = AutoModelForCausalLM  # pylint: disable=invalid-name
 
     def apply_patches(self) -> None:
+        if self.cfg.xformers_attention and self.cfg.sample_packing:
+            from axolotl.monkeypatch.attention import patch_xformers_attn_over_fa2
+
+            patch_xformers_attn_over_fa2()
+            self.cfg.flash_attention = True
         if self.cfg.fsdp_config and str(self.cfg.fsdp_config.fsdp_version) == "2":
             from axolotl.monkeypatch.accelerate.fsdp2 import patch_accelerate_fsdp_utils
 

src/axolotl/utils/schemas/config.py

@@ -435,16 +435,6 @@ class AxolotlInputConfig(
             )
         return data
 
-    @model_validator(mode="before")
-    @classmethod
-    def check_sample_packing_w_xformers(cls, data):
-        if data.get("sample_packing") and data.get("xformers_attention"):
-            raise ValueError(
-                "sample_packing not compatible with xformers_attention. Use flash_attention"
-            )
-
-        return data
-
     @model_validator(mode="before")
     @classmethod
     # pylint: disable=duplicate-code
@@ -471,9 +461,10 @@ class AxolotlInputConfig(
             and not data.get("flash_attention")
             and not data.get("sdp_attention")
             and not data.get("flex_attention")
+            and not data.get("xformers_attention")
         ):
             LOG.warning(
-                "sample_packing without flash, sdp or flex attention does not handle cross sample decontamination."
+                "sample_packing without flash, sdp, xformers or flex attention does not handle cross sample decontamination."
             )
 
         return data