Use pretrained generation config if possible

llmfoundry/models/hf/hf_causal_lm.py

@@ -22,6 +22,7 @@
 from transformers import (
     AutoConfig,
     AutoModelForCausalLM,
+    GenerationConfig,
     PreTrainedModel,
     PreTrainedTokenizerBase,
 )
@@ -337,6 +338,17 @@ def build_inner_model(
         if dist.get_local_rank() == 0:
             os.remove(signal_file_path)
 
+        # Use the pretrained generation config for the model if it exists.
+        try:
+            model.generation_config = GenerationConfig.from_pretrained(
+                pretrained_model_name_or_path,
+                use_auth_token=use_auth_token,
+            )
+        except OSError:
+            log.warning(
+                f'No existing generation config found for the model with name or path {pretrained_model_name_or_path}. Using default generation config.',
+            )
+
         # Hugging Face's weight tying does not succeed if the model is inited on meta device
         # so we manually apply the weight tying here
         if model.config.tie_word_embeddings and resolved_init_device == 'meta':

tests/models/hf/test_hf_config.py

@@ -3,19 +3,28 @@
 
 import os
 from copy import deepcopy
+from pathlib import Path
 from typing import Any, Dict, Mapping
 from unittest.mock import Mock, patch
 
 import pytest
 import torch
 from omegaconf import OmegaConf as om
-from transformers import PretrainedConfig
+from transformers import (
+    AutoConfig,
+    AutoModelForCausalLM,
+    PretrainedConfig,
+    PreTrainedModel,
+)
 
 from llmfoundry.models.hf.hf_fsdp import rgetattr
 from llmfoundry.models.mpt import MPTConfig, MPTForCausalLM
 from llmfoundry.utils import build_tokenizer
 from llmfoundry.utils.builders import build_composer_model
-from llmfoundry.utils.config_utils import to_dict_container
+from llmfoundry.utils.config_utils import (
+    set_config_overrides,
+    to_dict_container,
+)
 
 
 def test_remote_code_false_mpt(
@@ -279,3 +288,57 @@ def test_use_flash():
 
     # Make sure that HF has not cast the parameters to bf16
     assert next(model.parameters()).dtype == torch.float32
+
+
+def test_generation_config(tmp_path: Path):
+    # Create a small llama model to edit and save.
+    config = AutoConfig.from_pretrained('meta-llama/Llama-2-7b-hf')
+    set_config_overrides(
+        config,
+        config_overrides={
+            'num_hidden_layers': 2,
+            'hidden_size': 32,
+            'intermediate_size': 64,
+        },
+    )
+    model = AutoModelForCausalLM.from_config(config)
+
+    assert isinstance(model, PreTrainedModel)
+    assert model.generation_config is not None
+
+    new_bos_token_id = 100
+
+    # Set the bos_token_id to something else
+    model.generation_config.bos_token_id = new_bos_token_id
+
+    # Generation config and model config no longer match
+    assert model.generation_config.bos_token_id != model.config.bos_token_id
+
+    save_dir = tmp_path / 'model'
+
+    # Save the model.
+    model.save_pretrained(save_dir)
+
+    # Now load the model from the save directory and check that the bos_token_id is the same as what we set.
+    model_cfg = {
+        'name': 'hf_causal_lm',
+        'pretrained_model_name_or_path': str(save_dir),
+        'use_auth_token': True,
+        'pretrained': False,
+        'init_device': 'cpu',
+    }
+
+    name = model_cfg.pop('name')
+    model = build_composer_model(
+        name=name,
+        cfg=model_cfg,
+        tokenizer=None,  # type: ignore
+    )
+
+    inner_model = model.model
+
+    assert isinstance(inner_model, PreTrainedModel)
+    assert inner_model.generation_config is not None
+
+    # save_pretrained and reloading with hf_causal_lm should use the bos_token_id we set from earlier.
+    assert inner_model.generation_config.bos_token_id == new_bos_token_id