Merge pull request #220 from allenai/hf-integration

Lightweight HF integration

.github/actions/setup-venv/action.yml

@@ -30,7 +30,7 @@ runs:
       id: virtualenv-cache
       with:
         path: .venv
-        key: ${{ inputs.cache-prefix }}-${{ runner.os }}-${{ env.PYTHON_VERSION }}-${{ hashFiles('requirements.txt', 'dev-requirements.txt') }}
+        key: ${{ inputs.cache-prefix }}-${{ runner.os }}-${{ env.PYTHON_VERSION }}-${{ hashFiles('requirements.txt', 'dev-requirements.txt', 'hf_olmo/requirements.txt') }}
 
     - if: steps.virtualenv-cache.outputs.cache-hit != 'true'
       shell: bash
@@ -40,13 +40,15 @@ runs:
         . .venv/bin/activate
         pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cpu
         pip install -e .[dev]
+        pip install -e hf_olmo
 
     - if: steps.virtualenv-cache.outputs.cache-hit == 'true'
       shell: bash
       run: |
         # Set up virtual environment from cache hit.
         . .venv/bin/activate
         pip install --no-deps -e .[dev]
+        pip install --no-deps -e hf_olmo
 
     - shell: bash
       run: |

.github/workflows/main.yml

@@ -152,7 +152,7 @@ jobs:
                     value: ":16:8"
                   - name: TOKENIZERS_PARALLELISM
                     value: "false"
-                command: ["/entrypoint.sh", "pytest", "-v", "-m", "gpu", "tests/"]
+                command: ["/entrypoint.sh", "pytest", "-v", "-m", "gpu", "tests/", "-k", "not hf_olmo"]
                 result:
                   path: /unused
           token: ${{ env.BEAKER_TOKEN }}

conftest.py

@@ -89,3 +89,8 @@ def lorem_ipsum() -> str:
 @pytest.fixture(scope="module")
 def lorem_ipsum_docs() -> List[str]:
     return [text.replace("\n", " ").strip() for text in (LOREM_IPSUM_1, LOREM_IPSUM_2)]
+
+
+@pytest.fixture(scope="function")
+def model_path() -> str:
+    return "test_fixtures/test-olmo-model"

hf_olmo/__init__.py

@@ -0,0 +1,3 @@
+from .configuration_olmo import OLMoConfig
+from .modeling_olmo import OLMoForCausalLM
+from .tokenization_olmo_fast import OLMoTokenizerFast

hf_olmo/add_hf_config_to_olmo_checkpoint.py

@@ -0,0 +1,38 @@
+import argparse
+import logging
+import os
+
+from hf_olmo.configuration_olmo import OLMoConfig
+from olmo import Olmo
+
+logger = logging.getLogger(__name__)
+
+
+def write_config(checkpoint_dir: str):
+    # save config as HF config
+    logger.info(f"Loading checkpoint from {checkpoint_dir}")
+    model = Olmo.from_checkpoint(checkpoint_dir)
+
+    config_kwargs = model.config.asdict()
+    config_kwargs["use_cache"] = True
+    config = OLMoConfig(**config_kwargs)
+
+    logger.info(f"Saving HF-compatible config to {os.path.join(checkpoint_dir, 'config.json')}")
+    config.save_pretrained(checkpoint_dir)
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Adds a config.json to the checkpoint directory, making it easier to load weights as HF models."
+    )
+    parser.add_argument(
+        "--checkpoint-dir",
+        help="Location of OLMo checkpoint.",
+    )
+
+    args = parser.parse_args()
+    write_config(checkpoint_dir=args.checkpoint_dir)
+
+
+if __name__ == "__main__":
+    main()

hf_olmo/configuration_olmo.py

@@ -0,0 +1,26 @@
+"""
+OLMo configuration
+"""
+
+from transformers import AutoConfig, PretrainedConfig
+from transformers.utils import logging
+
+from olmo.config import ModelConfig
+
+logger = logging.get_logger(__name__)
+
+
+class OLMoConfig(PretrainedConfig):
+    model_type = "olmo"
+    keys_to_ignore_at_inference = ["past_key_values"]  # TODO: confirm
+
+    def __init__(self, use_cache: bool = False, **kwargs):
+        model_config = ModelConfig()
+        all_kwargs = model_config.asdict()
+        all_kwargs.update(kwargs)
+        all_kwargs.update({"use_cache": use_cache})
+        super().__init__(**all_kwargs)
+
+
+# Register the config class so that it is available for transformer pipelines, auto-loading etc.
+AutoConfig.register("olmo", OLMoConfig)

hf_olmo/modeling_olmo.py

@@ -0,0 +1,211 @@
+import os
+
+# import warnings
+from typing import List, Optional, Tuple, Union
+
+import torch
+from transformers import PreTrainedModel
+
+# from transformers.generation.utils import (  # BaseStreamer,
+#     GenerateOutput,
+#     LogitsProcessorList,
+#     StoppingCriteriaList,
+# )
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.models.auto import AutoModelForCausalLM
+
+from olmo.config import ModelConfig
+from olmo.model import Olmo
+
+from .configuration_olmo import OLMoConfig
+
+# from typing import Callable, Sequence
+
+
+def create_model_config_from_pretrained_config(config: OLMoConfig):
+    """
+    Utility function
+    """
+    model_config = ModelConfig(
+        d_model=config.d_model,
+        n_heads=config.n_heads,
+        n_layers=config.n_layers,
+        mlp_ratio=config.mlp_ratio,
+        activation_type=config.activation_type,
+        block_type=config.block_type,
+        alibi=config.alibi,
+        alibi_bias_max=config.alibi_bias_max,
+        rope=config.rope,
+        flash_attention=config.flash_attention,
+        attention_dropout=config.attention_dropout,
+        attention_layer_norm=config.attention_layer_norm,
+        multi_query_attention=config.multi_query_attention,
+        residual_dropout=config.residual_dropout,
+        embedding_dropout=config.embedding_dropout,
+        layer_norm_type=config.layer_norm_type,
+        max_sequence_length=config.max_sequence_length,
+        include_bias=config.include_bias,
+        vocab_size=config.vocab_size,
+        embedding_size=config.embedding_size,
+        eos_token_id=config.eos_token_id,
+        pad_token_id=config.pad_token_id,
+        init_device=config.init_device,
+        init_std=config.init_std,
+        precision=config.precision,
+    )
+    return model_config
+
+
+class OLMoForCausalLM(PreTrainedModel):
+    """
+    Extremely barebones HF model wrapper.
+    """
+
+    config_class = OLMoConfig
+
+    def __init__(self, config: OLMoConfig, model: Optional[Olmo] = None):
+        super().__init__(config)
+
+        if not model:
+            model_config = create_model_config_from_pretrained_config(config)
+            self.model = Olmo(model_config, init_params=True)
+        else:
+            self.model = model
+
+    # def forward(self, *args, **kwargs):
+    #     # use_cache = self.config.use_cache or kwargs.pop("use_cache", False)
+    #     kwargs["use_cache"] = kwargs.pop("use_cache", self.config.use_cache)
+    #     return self.model.forward(*args, **kwargs)
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        if use_cache is None:
+            use_cache = self.config.use_cache
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model.forward(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+        )
+
+        logits = outputs.logits
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = torch.nn.CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.embedding_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.attn_key_values,
+        )
+
+    def can_generate(self) -> bool:
+        return True
+
+    # Note (akshitab): This model does not use OLMo's generate() function as it does not support all the
+    # bells and whistles that HF's generation-compatible models do, such as `StoppingCriteria` or top-p sampling, etc.
+    # Instead, the model sets `can_generate` to True, and relies on HF's default `.generate()`, and implements
+    # supporting functions like `prepare_inputs_for_generation()`. This allows us to use HF's various generation
+    # options.
+
+    # def generate(
+    #     self,
+    #     input_ids: Optional[torch.Tensor] = None,
+    #     max_length: int = 20,
+    #     max_new_tokens: Optional[int] = None,
+    #     logits_processor: Optional[LogitsProcessorList] = None,
+    #     stopping_criteria: Optional[StoppingCriteriaList] = None,
+    #     prefix_allowed_tokens_fn: Optional[Callable[[int, torch.Tensor], List[int]]] = None,
+    #     synced_gpus: Optional[bool] = None,
+    #     assistant_model: Optional["PreTrainedModel"] = None,
+    #     streamer: Optional["BaseStreamer"] = None,
+    #     **kwargs,
+    # ) -> Union[GenerateOutput, torch.LongTensor]:
+    #
+    #     assert input_ids is not None
+    #
+    #     # TODO: use stopping_criteria, since it's being used by instruct-eval
+    #     if stopping_criteria is not None:
+    #         warnings.warn(
+    #             "OLMo's generate() function does not currently support `stopping_criteria`. "
+    #             "This will likely result in worse performance on tasks."
+    #         )
+    #
+    #     max_steps = max_new_tokens or max_length - input_ids.shape[1]
+    #     result = self.model.generate(
+    #         input_ids,
+    #         max_steps=max_steps,
+    #         beam_size=1,
+    #         **kwargs,
+    #     )
+    #
+    #     return torch.cat((input_ids, result.token_ids[:, 0]), dim=1)
+
+    @classmethod
+    def from_pretrained(
+        cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], *model_args, **kwargs
+    ):
+        assert pretrained_model_name_or_path is not None
+        if kwargs.get("device_map", "auto") == "auto":
+            device = "cuda:0" if torch.cuda.is_available() else "cpu"
+        else:
+            device = "cpu"
+        model = Olmo.from_checkpoint(pretrained_model_name_or_path, device=device)
+        try:
+            config = OLMoConfig.from_pretrained(pretrained_model_name_or_path)
+        except FileNotFoundError:
+            config = OLMoConfig(use_cache=True, **model.config.asdict())
+        return cls(config, model)
+
+    def prepare_inputs_for_generation(
+        self, input_ids: torch.LongTensor, past_key_values: Optional[List[Tuple]] = None, **kwargs
+    ):
+        if past_key_values:
+            # This is because we want the model to only process the last generated token.
+            input_ids = input_ids[:, -1:]
+        model_inputs = {"input_ids": input_ids, "past_key_values": past_key_values}
+
+        model_inputs.update(kwargs)
+        model_inputs["use_cache"] = kwargs.pop("use_cache", self.config.use_cache)
+        return model_inputs
+
+    # TODO: these are required to make the implementation complete.
+    # def resize_position_embeddings(self, new_num_position_embeddings: int):
+    #     pass
+    #
+    # def get_position_embeddings(self) -> Union[nn.Embedding, Tuple[nn.Embedding]]:
+    #     pass
+    #
+    # def _reorder_cache(self, past_key_values, beam_idx):
+    #     pass
+
+
+# Register the model so that it is available for transformer pipelines, auto-loading, etc.
+AutoModelForCausalLM.register(OLMoConfig, OLMoForCausalLM)