Merge pull request #719 from allenai/shanea/hf-olmo-gradient-checkpoi…

…nting

[HF OLMo] Add flash attention and gradient checkpointing support

CHANGELOG.md

@@ -10,6 +10,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 ### Added
 
 - Added ability to try loading latest checkpoint from save folder using `--try_load_latest_save`.
+- Added support for flash attention and gradient checkpointing to `hf_olmo`.
 
 ## [v0.5.0](https://github.com/allenai/OLMo/releases/tag/v0.5.0) - 2024-08-26
 

hf_olmo/modeling_olmo.py

@@ -1,14 +1,14 @@
 import logging
 from dataclasses import fields
-from typing import List, Optional, Tuple, Union
+from typing import Callable, List, Optional, Tuple, Union
 
 import torch
 from transformers import PreTrainedModel
 from transformers.cache_utils import Cache
 from transformers.modeling_outputs import CausalLMOutputWithPast
 from transformers.models.auto import AutoModelForCausalLM
 
-from olmo.config import ModelConfig
+from olmo.config import ActivationCheckpointingStrategy, ModelConfig
 from olmo.model import OLMo
 
 from .configuration_olmo import OLMoConfig
@@ -26,6 +26,15 @@ def create_model_config_from_pretrained_config(config: OLMoConfig):
         kwargs[field.name] = getattr(config, field.name)
 
     model_config = ModelConfig(**kwargs)
+
+    # Handle flash attention settings
+    if config._attn_implementation == "flash_attention_2":
+        model_config.flash_attention = True
+    elif config._attn_implementation in ("eager", "sdpa"):
+        model_config.flash_attention = False
+    else:
+        raise ValueError(f"Unexpected _attn_implementation {config._attn_implementation}")
+
     return model_config
 
 
@@ -37,10 +46,16 @@ class OLMoForCausalLM(PreTrainedModel):
     config_class = OLMoConfig
     base_model_prefix = "model"
     _no_split_modules = ["OLMoBlock"]
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    supports_gradient_checkpointing = True
 
     def __init__(self, config: OLMoConfig, model: Optional[OLMo] = None, init_params: bool = False):
         super().__init__(config)
 
+        self._gradient_checkpointing_func: Optional[Callable] = None
+        self._gradient_checkpointing = False
+
         if not model:
             model_config = create_model_config_from_pretrained_config(config)
             # Initialize model (always on CPU to start with so we don't run out of GPU memory).
@@ -49,6 +64,23 @@ def __init__(self, config: OLMoConfig, model: Optional[OLMo] = None, init_params
         else:
             self.model = model
 
+    @property
+    def gradient_checkpointing(self) -> bool:
+        return self._gradient_checkpointing
+
+    @gradient_checkpointing.setter
+    def gradient_checkpointing(self, enabled: bool):
+        if self._gradient_checkpointing == enabled:
+            return
+
+        # HF does not specify a way to pass checkpointing strategies, so we pick
+        # whole layer as our strategy. We can make this configurable later if needed.
+        checkpointing_strategy = ActivationCheckpointingStrategy.whole_layer if enabled else None
+        self.model.set_activation_checkpointing(
+            checkpointing_strategy, checkpoint_func=self._gradient_checkpointing_func
+        )
+        self._gradient_checkpointing = enabled
+
     def forward(
         self,
         input_ids: torch.LongTensor = None,

olmo/model.py

@@ -418,7 +418,7 @@ def __init__(self, layer_id: int, config: ModelConfig, cache: BufferCache):
         self.__cache = cache
         assert config.d_model % config.n_heads == 0
 
-        self._activation_checkpoint_fn = None
+        self._activation_checkpoint_fn: Optional[Callable] = None
 
         # Dropout.
         self.dropout = Dropout(config.residual_dropout)
@@ -500,9 +500,11 @@ def reset_parameters(self):
         init_normal(self.attn_out, std=attn_out_std, init_cutoff_factor=cutoff_factor)
         init_normal(self.ff_out, std=ff_out_std, init_cutoff_factor=cutoff_factor)
 
-    def set_activation_checkpointing(self, strategy: Optional[ActivationCheckpointingStrategy]):
+    def set_activation_checkpointing(
+        self, strategy: Optional[ActivationCheckpointingStrategy], checkpoint_func: Optional[Callable] = None
+    ):
         if strategy == ActivationCheckpointingStrategy.fine_grained:
-            self._activation_checkpoint_fn = activation_checkpoint_function(self.config)
+            self._activation_checkpoint_fn = checkpoint_func or activation_checkpoint_function(self.config)
         else:
             self._activation_checkpoint_fn = None
 
@@ -980,7 +982,7 @@ class OLMoOutput(NamedTuple):
     Attention keys and values from each block.
     """
 
-    hidden_states: Optional[Tuple[torch.Tensor]]
+    hidden_states: Optional[Tuple[torch.Tensor, ...]]
     """
     Hidden states from each block.
     """
@@ -1050,10 +1052,12 @@ def reset_parameters(self):
         for block in self:
             block.reset_parameters()
 
-    def set_activation_checkpointing(self, strategy: Optional[ActivationCheckpointingStrategy]):
+    def set_activation_checkpointing(
+        self, strategy: Optional[ActivationCheckpointingStrategy], checkpoint_func: Optional[Callable] = None
+    ):
         self.activation_checkpointing_strategy = strategy
         for block in self:
-            block.set_activation_checkpointing(strategy)
+            block.set_activation_checkpointing(strategy, checkpoint_func=checkpoint_func)
 
 
 class OLMo(nn.Module):
@@ -1140,14 +1144,16 @@ def __init__(self, config: ModelConfig, init_params: bool = True):
             get_causal_attention_bias(self.__cache, config.max_sequence_length, _non_meta_init_device(config))
             self.get_alibi_attention_bias(config.max_sequence_length, _non_meta_init_device(config))
 
-    def set_activation_checkpointing(self, strategy: Optional[ActivationCheckpointingStrategy]):
+    def set_activation_checkpointing(
+        self, strategy: Optional[ActivationCheckpointingStrategy], checkpoint_func: Optional[Callable] = None
+    ):
         self.activation_checkpointing_strategy = strategy
         if self.config.block_group_size != 1:
             for block_group in self.transformer.block_groups:
-                block_group.set_activation_checkpointing(strategy)
+                block_group.set_activation_checkpointing(strategy, checkpoint_func=checkpoint_func)
         else:
             for block in self.transformer.blocks:
-                block.set_activation_checkpointing(strategy)
+                block.set_activation_checkpointing(strategy, checkpoint_func=checkpoint_func)
 
     @property
     def device(self) -> torch.device:
@@ -1445,7 +1451,11 @@ def forward(
         if self.config.scale_logits:
             logits.mul_(1 / math.sqrt(self.config.d_model))
 
-        return OLMoOutput(logits=logits, attn_key_values=attn_key_values, hidden_states=tuple(all_hidden_states) if output_hidden_states else None)  # type: ignore[arg-type]
+        return OLMoOutput(
+            logits=logits,
+            attn_key_values=attn_key_values,
+            hidden_states=tuple(all_hidden_states) if output_hidden_states else None,
+        )
 
     def get_fsdp_wrap_policy(self, wrap_strategy: Optional[FSDPWrapStrategy] = None):
         if wrap_strategy is None:

test_fixtures/test-olmo-model/config.json

@@ -13,7 +13,9 @@
   "block_type": "sequential",
   "clip_qkv": null,
   "d_model": 32,
+  "emb_init_std": null,
   "embedding_dropout": 0.1,
+  "embedding_layer_norm": false,
   "embedding_size": 50304,
   "eos_token_id": 50256,
   "flash_attention": false,
@@ -22,6 +24,7 @@
   "init_device": null,
   "init_fn": "normal",
   "init_std": 0.02,
+  "layer_norm_eps": 1e-05,
   "layer_norm_type": "default",
   "layer_norm_with_affine": true,
   "max_sequence_length": 1024,
@@ -32,13 +35,16 @@
   "n_heads": 1,
   "n_kv_heads": null,
   "n_layers": 1,
+  "norm_after": false,
   "pad_token_id": 50256,
   "precision": null,
   "residual_dropout": 0.1,
   "rope": false,
   "rope_full_precision": true,
+  "rope_theta": 10000,
+  "scale_emb_init": false,
   "scale_logits": false,
-  "transformers_version": "4.40.2",
+  "transformers_version": "4.44.2",
   "use_cache": true,
   "vocab_size": 50257,
   "weight_tying": true

tests/hf_olmo/modeling_olmo_test.py

@@ -21,7 +21,86 @@ def test_olmo_model(model_path: str):
     output = model(input_tensor)
     hf_output = hf_model(input_tensor)
 
-    torch.testing.assert_allclose(output.logits, hf_output.logits)
+    torch.testing.assert_close(hf_output.logits, output.logits)
+
+
+@pytest.mark.gpu
+@pytest.mark.skipif(torch.cuda.device_count() < 1, reason="Requires CUDA devices")
+def test_flash_attention_2(model_path: str):
+    from transformers import AutoModelForCausalLM, AutoTokenizer
+
+    import hf_olmo  # noqa: F401
+
+    hf_model = AutoModelForCausalLM.from_pretrained(model_path)
+    hf_model_flash_attn = AutoModelForCausalLM.from_pretrained(model_path, attn_implementation="flash_attention_2")
+
+    tokenizer = AutoTokenizer.from_pretrained(model_path)
+    encoded_input = tokenizer.encode("My name is OLMo!")
+    input_tensor = torch.tensor(encoded_input).unsqueeze(0)
+
+    hf_output = hf_model(input_tensor)
+    hf_output_flash_attn = hf_model_flash_attn(input_tensor)
+
+    torch.testing.assert_close(hf_output_flash_attn.logits, hf_output.logits)
+
+
+def test_sdpa(model_path: str):
+    from transformers import AutoModelForCausalLM, AutoTokenizer
+
+    import hf_olmo  # noqa: F401
+
+    hf_model = AutoModelForCausalLM.from_pretrained(model_path)
+    hf_model_sdpa = AutoModelForCausalLM.from_pretrained(model_path, attn_implementation="sdpa")
+
+    tokenizer = AutoTokenizer.from_pretrained(model_path)
+    encoded_input = tokenizer.encode("My name is OLMo!")
+    input_tensor = torch.tensor(encoded_input).unsqueeze(0)
+
+    hf_output = hf_model(input_tensor)
+    hf_output_sdpa = hf_model_sdpa(input_tensor)
+
+    torch.testing.assert_close(hf_output_sdpa.logits, hf_output.logits)
+
+
+def test_gradient_checkpointing(model_path: str):
+    from transformers import AutoModelForCausalLM, AutoTokenizer, PreTrainedModel
+
+    import hf_olmo  # noqa: F401
+
+    hf_model: PreTrainedModel = AutoModelForCausalLM.from_pretrained(model_path)
+
+    tokenizer = AutoTokenizer.from_pretrained(model_path)
+    encoded_input = tokenizer.encode("My name is OLMo!")
+    input_tensor = torch.tensor(encoded_input).unsqueeze(0)
+
+    hf_output_no_checkpointing = hf_model(input_tensor)
+
+    hf_model.gradient_checkpointing_enable()
+
+    hf_output_checkpointing = hf_model(input_tensor)
+
+    torch.testing.assert_close(hf_output_checkpointing.logits, hf_output_no_checkpointing.logits)
+
+
+def test_gradient_checkpointing_disable(model_path: str):
+    from transformers import AutoModelForCausalLM, AutoTokenizer, PreTrainedModel
+
+    import hf_olmo  # noqa: F401
+
+    hf_model: PreTrainedModel = AutoModelForCausalLM.from_pretrained(model_path)
+
+    tokenizer = AutoTokenizer.from_pretrained(model_path)
+    encoded_input = tokenizer.encode("My name is OLMo!")
+    input_tensor = torch.tensor(encoded_input).unsqueeze(0)
+
+    hf_output = hf_model(input_tensor)
+
+    hf_model.gradient_checkpointing_enable()
+    hf_model.gradient_checkpointing_disable()
+
+    hf_output_after_disable = hf_model(input_tensor)
+
+    torch.testing.assert_close(hf_output_after_disable.logits, hf_output.logits)
 
 
 def test_save_pretrained(model_path: str):