Upload Ivme-Coder-v1 (Otter 1): safetensors + custom modeling code

model/model.py

@@ -129,6 +129,13 @@ class Block(nn.Module):
 
 class IvmeCoderV1ForCausalLM(PreTrainedModel, GenerationMixin):
     config_class = IvmeCoderConfig
+    # Required so from_pretrained's tie-recovery logic knows how to reconnect
+    # head.weight to tok_emb.weight when head.weight is absent from the checkpoint
+    # (it's deliberately excluded from the safetensors file, since it's tied storage,
+    # not distinct data). Without this, HF's loader treats the missing key as needing
+    # fresh random initialization instead of re-tying it - which silently produces a
+    # working-looking model with a completely untrained output head.
+    _tied_weights_keys = {"head.weight": "tok_emb.weight"}
 
     def __init__(self, config):
         super().__init__(config)
@@ -160,6 +167,12 @@ class IvmeCoderV1ForCausalLM(PreTrainedModel, GenerationMixin):
     def set_input_embeddings(self, value):
         self.tok_emb = value
 
+    def get_output_embeddings(self):
+        return self.head
+
+    def set_output_embeddings(self, value):
+        self.head = value
+
     def forward(self, input_ids, labels=None, use_cache=False, past_key_values=None,
                 attention_mask=None, **kwargs):
         # attention_mask is accepted for API compatibility with tokenizer output /

modeling_ivme_coder.py

@@ -129,6 +129,13 @@ class Block(nn.Module):
 
 class IvmeCoderV1ForCausalLM(PreTrainedModel, GenerationMixin):
     config_class = IvmeCoderConfig
+    # Required so from_pretrained's tie-recovery logic knows how to reconnect
+    # head.weight to tok_emb.weight when head.weight is absent from the checkpoint
+    # (it's deliberately excluded from the safetensors file, since it's tied storage,
+    # not distinct data). Without this, HF's loader treats the missing key as needing
+    # fresh random initialization instead of re-tying it - which silently produces a
+    # working-looking model with a completely untrained output head.
+    _tied_weights_keys = {"head.weight": "tok_emb.weight"}
 
     def __init__(self, config):
         super().__init__(config)
@@ -160,6 +167,12 @@ class IvmeCoderV1ForCausalLM(PreTrainedModel, GenerationMixin):
     def set_input_embeddings(self, value):
         self.tok_emb = value
 
+    def get_output_embeddings(self):
+        return self.head
+
+    def set_output_embeddings(self, value):
+        self.head = value
+
     def forward(self, input_ids, labels=None, use_cache=False, past_key_values=None,
                 attention_mask=None, **kwargs):
         # attention_mask is accepted for API compatibility with tokenizer output /