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

model/model.py

@@ -148,10 +148,17 @@ class IvmeCoderV1ForCausalLM(PreTrainedModel, GenerationMixin):
         self.head = nn.Linear(config.d_model, config.vocab_size, bias=False)
         self.head.weight = self.tok_emb.weight  # tied embeddings
 
-        d_head = config.d_model // config.n_head
-        cos, sin = precompute_rope(d_head, config.context_len, theta=config.rope_theta)
-        self.register_buffer("rope_cos", cos, persistent=False)
-        self.register_buffer("rope_sin", sin, persistent=False)
+        # NOTE: RoPE cos/sin tables are intentionally NOT stored as a persistent=False
+        # buffer here. transformers v5's from_pretrained() has a known bug where
+        # persistent=False buffers get overwritten with uninitialized/garbage memory
+        # after loading (https://github.com/huggingface/transformers/issues/44534),
+        # even though they're computed correctly at __init__ time. That garbage then
+        # silently produces huge (but finite) values through the RoPE rotation, which
+        # overflow to NaN inside scaled_dot_product_attention. Recomputing the tables
+        # fresh on every forward() call sidesteps this entirely - it's cheap (a cos/sin
+        # over d_head/2 * context_len elements) relative to the rest of the forward pass.
+        self.d_head = config.d_model // config.n_head
+        self.rope_theta = config.rope_theta
 
         self.post_init()
 
@@ -179,9 +186,14 @@ class IvmeCoderV1ForCausalLM(PreTrainedModel, GenerationMixin):
         # generate()'s kwarg validation, but intentionally unused: this model only
         # supports left-padding for batched generation (see model card), and single-
         # sequence causal generation (the common case) needs no mask at all.
+        # Recompute RoPE tables fresh each call - see note in __init__ for why this
+        # isn't cached in a buffer.
+        rope_cos, rope_sin = precompute_rope(
+            self.d_head, self.context_len, theta=self.rope_theta, device=input_ids.device
+        )
         x = self.tok_emb(input_ids)
         for block in self.blocks:
-            x = block(x, self.rope_cos, self.rope_sin)
+            x = block(x, rope_cos, rope_sin)
         x = self.ln_f(x)
         logits = self.head(x)
         loss = None

modeling_ivme_coder.py

@@ -148,10 +148,17 @@ class IvmeCoderV1ForCausalLM(PreTrainedModel, GenerationMixin):
         self.head = nn.Linear(config.d_model, config.vocab_size, bias=False)
         self.head.weight = self.tok_emb.weight  # tied embeddings
 
-        d_head = config.d_model // config.n_head
-        cos, sin = precompute_rope(d_head, config.context_len, theta=config.rope_theta)
-        self.register_buffer("rope_cos", cos, persistent=False)
-        self.register_buffer("rope_sin", sin, persistent=False)
+        # NOTE: RoPE cos/sin tables are intentionally NOT stored as a persistent=False
+        # buffer here. transformers v5's from_pretrained() has a known bug where
+        # persistent=False buffers get overwritten with uninitialized/garbage memory
+        # after loading (https://github.com/huggingface/transformers/issues/44534),
+        # even though they're computed correctly at __init__ time. That garbage then
+        # silently produces huge (but finite) values through the RoPE rotation, which
+        # overflow to NaN inside scaled_dot_product_attention. Recomputing the tables
+        # fresh on every forward() call sidesteps this entirely - it's cheap (a cos/sin
+        # over d_head/2 * context_len elements) relative to the rest of the forward pass.
+        self.d_head = config.d_model // config.n_head
+        self.rope_theta = config.rope_theta
 
         self.post_init()
 
@@ -179,9 +186,14 @@ class IvmeCoderV1ForCausalLM(PreTrainedModel, GenerationMixin):
         # generate()'s kwarg validation, but intentionally unused: this model only
         # supports left-padding for batched generation (see model card), and single-
         # sequence causal generation (the common case) needs no mask at all.
+        # Recompute RoPE tables fresh each call - see note in __init__ for why this
+        # isn't cached in a buffer.
+        rope_cos, rope_sin = precompute_rope(
+            self.d_head, self.context_len, theta=self.rope_theta, device=input_ids.device
+        )
         x = self.tok_emb(input_ids)
         for block in self.blocks:
-            x = block(x, self.rope_cos, self.rope_sin)
+            x = block(x, rope_cos, rope_sin)
         x = self.ln_f(x)
         logits = self.head(x)
         loss = None