Add Qwen3 models.

Author: explanare

pyvene/models/intervenable_modelcard.py

@@ -14,6 +14,7 @@
 from .llava.modelings_intervenable_llava import *
 from .olmo.modelings_intervenable_olmo import *
 from .olmo2.modelings_intervenable_olmo2 import *
+from .qwen3.modelings_intervenable_qwen3 import *
 
 #########################################################################
 """
@@ -66,6 +67,8 @@
     hf_models.olmo.modeling_olmo.OlmoForCausalLM: olmo_lm_type_to_module_mapping,
     hf_models.olmo2.modeling_olmo2.Olmo2Model: olmo2_type_to_module_mapping,
     hf_models.olmo2.modeling_olmo2.Olmo2ForCausalLM: olmo2_lm_type_to_module_mapping,
+    hf_models.qwen3.modeling_qwen3.Qwen3Model: qwen3_type_to_module_mapping,
+    hf_models.qwen3.modeling_qwen3.Qwen3ForCausalLM: qwen3_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,
@@ -103,6 +106,8 @@
     hf_models.olmo.modeling_olmo.OlmoForCausalLM: olmo_lm_type_to_dimension_mapping,
     hf_models.olmo2.modeling_olmo2.Olmo2Model: olmo2_type_to_dimension_mapping,
     hf_models.olmo2.modeling_olmo2.Olmo2ForCausalLM: olmo2_lm_type_to_dimension_mapping,
+    hf_models.qwen3.modeling_qwen3.Qwen3Model: qwen3_type_to_dimension_mapping,
+    hf_models.qwen3.modeling_qwen3.Qwen3ForCausalLM: qwen3_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/qwen3/__init__.py

@@ -0,0 +1,77 @@
+"""
+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 ..constants import *
+
+qwen3_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")),
+}
+
+qwen3_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",),
+}
+
+"""qwen3 model with LM head"""
+qwen3_lm_type_to_module_mapping = {}
+for k, v in qwen3_type_to_module_mapping.items():
+    qwen3_lm_type_to_module_mapping[k] = (f"model.{v[0]}", ) + v[1:]
+qwen3_lm_type_to_dimension_mapping = qwen3_type_to_dimension_mapping
+
+"""qwen3 model with classifier head"""
+qwen3_classifier_type_to_module_mapping = {}
+for k, v in qwen3_type_to_module_mapping.items():
+    qwen3_classifier_type_to_module_mapping[k] = (f"model.{v[0]}", ) + v[1:]
+qwen3_classifier_type_to_dimension_mapping = qwen3_type_to_dimension_mapping
+
+def create_qwen3(
+    name="Qwen/Qwen3-8B", cache_dir=None, dtype=torch.bfloat16
+):
+    """Creates a Causal LM model, config, and tokenizer from the given name and revision"""
+    from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
+    
+    config = AutoConfig.from_pretrained(name, cache_dir=cache_dir)
+    tokenizer = AutoTokenizer.from_pretrained(name, cache_dir=cache_dir)
+    model = AutoModelForCausalLM.from_pretrained(
+        name,
+        config=config,
+        cache_dir=cache_dir,
+        torch_dtype=dtype,
+    )
+    print("loaded model")
+    return config, tokenizer, model