Clean checkpoint: fix naming, remove predictor keys, fix tie_word_embeddings

config.json

@@ -15,7 +15,7 @@
   "lm_loss_weight": 0.5,
   "mask_ratio": 0.5,
   "max_target_length": 512,
-  "model_type": "biopan_omni",
+  "model_type": "smb_v1",
   "momentum_tau": 0.996,
   "num_queries": 128,
   "num_resampler_layers": 3,
@@ -48,7 +48,7 @@
     "rope_scaling": null,
     "rope_theta": 1000000,
     "sliding_window": null,
-    "tie_word_embeddings": true,
+    "tie_word_embeddings": false,
     "torch_dtype": "bfloat16",
     "use_cache": true,
     "use_sliding_window": false,
@@ -2120,10 +2120,11 @@
   "vision_model_name_or_path2": "",
   "vocab_size": 151685,
   "architectures": [
-    "BioPANOmniForConditionalGeneration"
+    "SMBV1ForConditionalGeneration"
   ],
   "auto_map": {
-    "AutoConfig": "modeling_biopan_omni.BioPANOmniConfig",
-    "AutoModelForCausalLM": "modeling_biopan_omni.BioPANOmniForConditionalGeneration"
-  }
+    "AutoConfig": "modeling_smb_v1.SMBV1Config",
+    "AutoModelForCausalLM": "modeling_smb_v1.SMBV1ForConditionalGeneration"
+  },
+  "tie_word_embeddings": false
 }

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:fb814b1a4d22b0ad10d5f923e1f1995d3d5316db6cd58f2fc880e6234209f0f8
-size 3658396544
+oid sha256:5dacf6293e6a37868a681daf134df81e6ce51b65d313c017f1512f85ec40332c
+size 4061464080

modeling_smb_v1.py

@@ -402,8 +402,8 @@ def build_connector(config):
 # CONFIGURATION
 # =============================================================================
 
-class BioPANOmniConfig(PretrainedConfig):
-    model_type = "biopan_omni"
+class SMBV1Config(PretrainedConfig):
+    model_type = "smb_v1"
     
     def __init__(
         self,
@@ -475,11 +475,11 @@ class BioPANOmniConfig(PretrainedConfig):
 # MAIN MODEL
 # =============================================================================
 
-class BioPANOmniPreTrainedModel(PreTrainedModel):
-    config_class = BioPANOmniConfig
+class SMBV1PreTrainedModel(PreTrainedModel):
+    config_class = SMBV1Config
     base_model_prefix = "model"
     supports_gradient_checkpointing = True
-    _no_split_modules = ["BioPANOmniVisionAttention"]
+    _no_split_modules = ["SMBV1VisionAttention"]
     _skip_keys_device_placement = "past_key_values"
     _supports_flash_attn_2 = True
 
@@ -493,8 +493,8 @@ class BioPANOmniPreTrainedModel(PreTrainedModel):
             module.weight.data.normal_(mean=0.0, std=std)
 
 
-class BioPANOmniForConditionalGeneration(BioPANOmniPreTrainedModel, GenerationMixin):
-    def __init__(self, config: BioPANOmniConfig):
+class SMBV1ForConditionalGeneration(SMBV1PreTrainedModel, GenerationMixin):
+    def __init__(self, config: SMBV1Config):
         super().__init__(config)
         
         # Detect LLM backend from text_config
@@ -904,5 +904,5 @@ class BioPANOmniForConditionalGeneration(BioPANOmniPreTrainedModel, GenerationMi
 # REGISTER WITH AUTO CLASSES
 # =============================================================================
 
-AutoConfig.register("biopan_omni", BioPANOmniConfig)
-AutoModelForCausalLM.register(BioPANOmniConfig, BioPANOmniForConditionalGeneration)
+AutoConfig.register("smb_v1", SMBV1Config)
+AutoModelForCausalLM.register(SMBV1Config, SMBV1ForConditionalGeneration)