Merge pull request #182 from stanfordnlp/olmo

[P2] Add OLMo models.

pyvene/models/intervenable_modelcard.py

@@ -11,7 +11,7 @@
 from .blip.modelings_intervenable_blip_itm import *
 from .backpack_gpt2.modelings_intervenable_backpack_gpt2 import *
 from .llava.modelings_intervenable_llava import *
-
+from .olmo.modelings_intervenable_olmo import *
 
 #########################################################################
 """
@@ -58,6 +58,8 @@
     hf_models.gemma.modeling_gemma.GemmaModel: gemma_type_to_module_mapping,
     hf_models.gemma.modeling_gemma.GemmaForCausalLM: gemma_lm_type_to_module_mapping,
     hf_models.gemma.modeling_gemma.GemmaForSequenceClassification: gemma_classifier_type_to_module_mapping,
+    hf_models.olmo.modeling_olmo.OlmoModel: olmo_type_to_module_mapping,
+    hf_models.olmo.modeling_olmo.OlmoForCausalLM: olmo_lm_type_to_module_mapping,  
     hf_models.blip.modeling_blip.BlipForQuestionAnswering: blip_type_to_module_mapping,
     hf_models.blip.modeling_blip.BlipForImageTextRetrieval: blip_itm_type_to_module_mapping,
     MLPModel: mlp_type_to_module_mapping,
@@ -89,6 +91,8 @@
     hf_models.gemma.modeling_gemma.GemmaModel: gemma_type_to_dimension_mapping,
     hf_models.gemma.modeling_gemma.GemmaForCausalLM: gemma_lm_type_to_dimension_mapping,
     hf_models.gemma.modeling_gemma.GemmaForSequenceClassification: gemma_classifier_type_to_dimension_mapping,
+    hf_models.olmo.modeling_olmo.OlmoModel: olmo_type_to_dimension_mapping,
+    hf_models.olmo.modeling_olmo.OlmoForCausalLM: olmo_lm_type_to_dimension_mapping, 
     hf_models.blip.modeling_blip.BlipForQuestionAnswering: blip_type_to_dimension_mapping,
     hf_models.blip.modeling_blip.BlipForImageTextRetrieval: blip_itm_type_to_dimension_mapping,
     MLPModel: mlp_type_to_dimension_mapping,

pyvene/models/olmo/modelings_intervenable_olmo.py

@@ -0,0 +1,93 @@
+"""
+Each modeling file in this library is a mapping between
+abstract naming of intervention anchor points and actual
+model module defined in the huggingface library.
+
+We also want to let the intervention library know how to
+config the dimensions of intervention based on model config
+defined in the huggingface library.
+"""
+
+
+import torch
+from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
+from ..constants import *
+
+
+olmo_type_to_module_mapping = {
+    "block_input": ("layers[%s]", CONST_INPUT_HOOK),
+    "block_output": ("layers[%s]", CONST_OUTPUT_HOOK),
+    "mlp_activation": ("layers[%s].mlp.act_fn", CONST_OUTPUT_HOOK),
+    "mlp_output": ("layers[%s].mlp", CONST_OUTPUT_HOOK),
+    "mlp_input": ("layers[%s].mlp", CONST_INPUT_HOOK),
+    "attention_value_output": ("layers[%s].self_attn.o_proj", CONST_INPUT_HOOK),
+    "head_attention_value_output": ("layers[%s].self_attn.o_proj", CONST_INPUT_HOOK, (split_head_and_permute, "n_head")),
+    "attention_output": ("layers[%s].self_attn", CONST_OUTPUT_HOOK),
+    "attention_input": ("layers[%s].self_attn", CONST_INPUT_HOOK),
+    "query_output": ("layers[%s].self_attn.q_proj", CONST_OUTPUT_HOOK),
+    "key_output": ("layers[%s].self_attn.k_proj", CONST_OUTPUT_HOOK),
+    "value_output": ("layers[%s].self_attn.v_proj", CONST_OUTPUT_HOOK),
+    "head_query_output": ("layers[%s].self_attn.q_proj", CONST_OUTPUT_HOOK, (split_head_and_permute, "n_head")),
+    "head_key_output": ("layers[%s].self_attn.k_proj", CONST_OUTPUT_HOOK, (split_head_and_permute, "n_kv_head")),
+    "head_value_output": ("layers[%s].self_attn.v_proj", CONST_OUTPUT_HOOK, (split_head_and_permute, "n_kv_head")),
+}
+
+
+olmo_type_to_dimension_mapping = {
+    "n_head": ("num_attention_heads",),
+    "n_kv_head": ("num_key_value_heads",),
+    "block_input": ("hidden_size",),
+    "block_output": ("hidden_size",),
+    "mlp_activation": ("intermediate_size",),
+    "mlp_output": ("hidden_size",),
+    "mlp_input": ("hidden_size",),
+    "attention_value_output": ("hidden_size",),
+    "head_attention_value_output": ("hidden_size/num_attention_heads",),
+    "attention_output": ("hidden_size",),
+    "attention_input": ("hidden_size",),
+    "query_output": ("hidden_size",),
+    "key_output": ("hidden_size",),
+    "value_output": ("hidden_size",),
+    "head_query_output": ("hidden_size/num_attention_heads",),
+    "head_key_output": ("hidden_size/num_attention_heads",),
+    "head_value_output": ("hidden_size/num_attention_heads",),
+}
+
+
+"""olmo model with LM head"""
+olmo_lm_type_to_module_mapping = {}
+for k, v in olmo_type_to_module_mapping.items():
+    olmo_lm_type_to_module_mapping[k] = (f"model.{v[0]}", ) + v[1:]
+
+
+olmo_lm_type_to_dimension_mapping = olmo_type_to_dimension_mapping
+
+
+"""olmo model with classifier head"""
+olmo_classifier_type_to_module_mapping = {}
+for k, v in olmo_type_to_module_mapping.items():
+    olmo_classifier_type_to_module_mapping[k] = (f"model.{v[0]}", ) + v[1:]
+
+
+olmo_classifier_type_to_dimension_mapping = olmo_type_to_dimension_mapping
+
+
+def create_olmo(
+    name="allenai/OLMo-7B-0424-hf", cache_dir=None, dtype=torch.bfloat16, config=None,
+    revision='main'
+):
+    """Creates a OLMo Causal LM model, config, and tokenizer from the given name and revision"""
+    if config is None:
+        config = AutoConfig.from_pretrained(name, cache_dir=cache_dir)
+        olmo = AutoModelForCausalLM.from_pretrained(
+            name,
+            config=config,
+            cache_dir=cache_dir,
+            torch_dtype=dtype,
+        )
+        tokenizer = AutoTokenizer.from_pretrained(name, cache_dir=cache_dir)
+    else:
+        olmo = AutoModelForCausalLM(config, cache_dir=cache_dir, revision=revision)
+        tokenizer = AutoTokenizer.from_pretrained(name, cache_dir=cache_dir)
+    print("loaded model")
+    return config, tokenizer, olmo