fix: Correct embedding dimension logic in LoRA dummy creation

Fixed incorrect fallback logic for embedding layers where dimensions were reversed.

## Problem
For embedding layers with shape [vocab_size, embedding_dim]:
- input_dim should be vocab_size (shape[0])
- output_dim should be embedding_dim (shape[1])
- embeddings_tensor_dim should be embedding_dim (shape[1])

Previous code had:
- input_dim fallback: shape[1] ❌ (was getting embedding_dim instead of vocab_size)
- output_dim fallback: shape[0] ❌ (was getting vocab_size instead of embedding_dim)
- embeddings_tensor_dim: Used input_size instead of output_size ❌

## Fix
Corrected all fallback paths to use proper dimensions for embedding layers:
- input_dim: shape[0] (vocab_size)
- output_dim: shape[1] (embedding_dim)
- embeddings_tensor_dim: shape[1] (embedding_dim)

Also fixed elif chain to check output_size instead of input_size for embeddings_tensor_dim.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>
Signed-off-by: sheikheddy <[email protected]>

Author: sheikheddy

vllm/lora/models.py

@@ -624,28 +624,34 @@ def create_dummy_lora(
                         input_dim = module.base_layer.input_size
                     elif hasattr(module.base_layer, "weight_shape"):
                         # Compressed tensors: weight_shape stores [output, input]
-                        input_dim = module.base_layer.weight_shape[1].item()
+                        # For embeddings: [vocab_size, embedding_dim]
+                        input_dim = module.base_layer.weight_shape[0].item()
                     else:
-                        input_dim = module.weight.shape[1]
+                        # For embeddings: weight.shape = [vocab_size, embedding_dim]
+                        input_dim = module.weight.shape[0]
 
                     if hasattr(module.base_layer, "embedding_dim"):
                         output_dim = module.base_layer.embedding_dim
                     elif hasattr(module.base_layer, "output_size"):
                         output_dim = module.base_layer.output_size
                     elif hasattr(module.base_layer, "weight_shape"):
                         # Compressed tensors: weight_shape stores [output, input]
-                        output_dim = module.base_layer.weight_shape[0].item()
+                        # For embeddings: [vocab_size, embedding_dim]
+                        output_dim = module.base_layer.weight_shape[1].item()
                     else:
-                        output_dim = module.weight.shape[0]
+                        # For embeddings: weight.shape = [vocab_size, embedding_dim]
+                        output_dim = module.weight.shape[1]
 
                     if hasattr(module.base_layer, "embedding_dim"):
                         embeddings_tensor_dim = module.base_layer.embedding_dim
-                    elif hasattr(module.base_layer, "input_size"):
-                        embeddings_tensor_dim = module.base_layer.input_size
+                    elif hasattr(module.base_layer, "output_size"):
+                        embeddings_tensor_dim = module.base_layer.output_size
                     elif hasattr(module.base_layer, "weight_shape"):
                         # Compressed tensors: weight_shape stores [output, input]
+                        # For embeddings: [vocab_size, embedding_dim]
                         embeddings_tensor_dim = module.base_layer.weight_shape[1].item()
                     else:
+                        # For embeddings: weight.shape = [vocab_size, embedding_dim]
                         embeddings_tensor_dim = module.weight.shape[1]
                     lora = LoRALayerWeights.create_dummy_lora_weights(
                         module_name,