Merge branch 'main' into mlflow-log-model

llmfoundry/models/hf/hf_causal_lm.py

@@ -65,10 +65,7 @@ def __init__(self, om_model_config: Union[DictConfig,
                                               nn.Module],
                  tokenizer: PreTrainedTokenizerBase):
         # set up training and eval metrics
-        train_metrics = [
-            LanguageCrossEntropy(),
-            LanguagePerplexity(),
-        ]
+        train_metrics = [LanguageCrossEntropy(), LanguagePerplexity()]
         eval_metrics = [
             LanguageCrossEntropy(),
             LanguagePerplexity(),
@@ -92,6 +89,9 @@ def __init__(self, om_model_config: Union[DictConfig,
                     'which is not significantly slower and not compatible with the LLM foundry training code, rather than the code release by MosaicML.'
                 )
 
+            if not om_model_config.get('use_train_metrics', True):
+                train_metrics = []
+
             # load the model config
             trust_remote_code = om_model_config.get('trust_remote_code', True)
             use_auth_token = om_model_config.get('use_auth_token', False)
@@ -109,6 +109,7 @@ def __init__(self, om_model_config: Union[DictConfig,
                     )
 
                 attr = getattr(config, k)
+                # attempt to disallow typos in nested configs
                 if isinstance(attr, Mapping):
                     extra_keys = [
                         _k for _k in v.keys() if _k not in attr.keys()
@@ -120,6 +121,10 @@ def __init__(self, om_model_config: Union[DictConfig,
                             f'Expected (a subset of) keys: {list(attr.keys())}.'
                         )
                     getattr(config, k).update(v)
+                # necessary case to allow for rope_scaling to be overriden in llama config
+                elif attr is None and isinstance(v, Mapping):
+                    setattr(config, k, {})
+                    getattr(config, k).update(v)
                 else:
                     setattr(config, k, v)
 

llmfoundry/models/mpt/modeling_mpt.py

@@ -694,7 +694,9 @@ def __init__(
         hf_config = MPTConfig.from_dict(resolved_om_model_config)
         model = MPTForCausalLM(hf_config)
 
-        train_metrics = [LanguageCrossEntropy(), LanguagePerplexity()]
+        use_train_metrics = om_model_config.get('use_train_metrics', True)
+        train_metrics = [LanguageCrossEntropy(),
+                         LanguagePerplexity()] if use_train_metrics else []
         eval_metrics = [
             LanguageCrossEntropy(),
             LanguagePerplexity(),

mcli/mcli-llama2-finetune.yaml

@@ -2,6 +2,7 @@ integrations:
 - integration_type: git_repo
   git_repo: mosaicml/llm-foundry
   git_branch: v0.3.0
+  # git_commit: # OR use your commit hash
   pip_install: -e .[gpu]
   ssh_clone: false # Should be true if using a private repo
 

mcli/mcli-openai-eval.yaml

@@ -1,8 +1,8 @@
 integrations:
 - integration_type: git_repo
   git_repo: mosaicml/llm-foundry
-  git_branch: # use your branch
-  # git_commit: 29d65cc26853c09f6de7542978056ddb0b07e98c # OR use your commit hash
+  git_branch: v0.3.0
+  # git_commit: # OR use your commit hash
   pip_install: -e ".[gpu,openai]"
   ssh_clone: false # Should be true if using a private repo
 

mcli/mcli-pretokenize-oci-upload.yaml

@@ -14,7 +14,7 @@ integrations:
   - oci-cli==3.23.2
 - integration_type: git_repo
   git_repo: mosaicml/llm-foundry
-  git_branch: v0.2.0
+  git_branch: v0.3.0
   # git_commit: # OR use your commit hash
   pip_install: '.'
   ssh_clone: false # Should be true if using a private repo

tests/test_hf_config.py

@@ -1,6 +1,7 @@
 # Copyright 2022 MosaicML LLM Foundry authors
 # SPDX-License-Identifier: Apache-2.0
 
+import os
 import tempfile
 from copy import deepcopy
 from pathlib import Path
@@ -139,3 +140,30 @@ def test_hf_config_override(
                 assert getattr(hf_model.config, k)[_k] == _v
         else:
             assert getattr(hf_model.config, k) == v
+
+
+@pytest.mark.skipif('HUGGING_FACE_HUB_TOKEN' not in os.environ,
+                    reason='CI does not have access to llama2')
+def test_rope_scaling_override():
+    model_cfg = {
+        'name': 'hf_causal_lm',
+        'pretrained_model_name_or_path': 'meta-llama/Llama-2-7b-hf',
+        'config_overrides': {
+            'num_hidden_layers': 2,
+            'hidden_size': 32,
+            'intermediate_size': 64,
+            'rope_scaling': {
+                'type': 'dynamic',
+                'factor': 0.5
+            }
+        },
+        'use_auth_token': True,
+        'pretrained': False,
+        'init_device': 'cpu',
+    }
+    model_cfg = om.create(model_cfg)
+
+    model = COMPOSER_MODEL_REGISTRY[model_cfg.name](model_cfg, tokenizer=None)
+    # This would error if the config isn't parsed into a proper dictionary
+    model.get_metadata()
+    assert model.config.rope_scaling == {'type': 'dynamic', 'factor': 0.5}