add variant

Author: mht-sharma

optimum/exporters/onnx/model_configs.py

@@ -1890,15 +1890,15 @@ class LlavaOnnxConfig(OnnxConfigWithPast):
 
     VARIANTS = {
         "default": "The export follows the Transformers implementation of forward in LlavaModelForConditionalGeneration, with the following components exported:\n\t - "
-        "model.onnx: corresponds to the vision encoder + projection + decoder in a single file without past key value support in https://github.com/huggingface/transformers/blob/v4.33.2/src/transformers/models/speecht5/modeling_speecht5.py#L2544-L2556.\n\t - "
-        "decoder_model.onnx: corresponds to the decoder part in with past_key_values input https://github.com/huggingface/transformers/blob/v4.33.2/src/transformers/models/speecht5/modeling_speecht5.py#L2572-L2602.",
+        "model.onnx: corresponds to the vision encoder + projection + decoder in a single file without past key value support in https://github.com/huggingface/transformers/blob/v4.39.3/src/transformers/models/llava/modeling_llava.py#L360-L519.\n\t - "
+        "decoder_model.onnx: corresponds to the decoder part in with past_key_values input https://github.com/huggingface/transformers/blob/v4.39.3/src/transformers/models/llava/modeling_llava.py#L449-L489.",
         "optimized": "The export follows the memory optimized implementation of Transformers forward. This is a recommended export as decoder is exported only once`. It has the following components exported:\n\t - "
-        "encoder_model.onnx: corresponds to the vision encoder + projection + decoder in https://github.com/huggingface/transformers/blob/v4.33.2/src/transformers/models/speecht5/modeling_speecht5.py#L2544-L2556.\n\t - "
-        "decoder_model.onnx: corresponds to the decoder part in https://github.com/huggingface/transformers/blob/v4.33.2/src/transformers/models/speecht5/modeling_speecht5.py#L2572-L2602.\n\t - "
-        "attention_position_id_generator.onnx: corresponds to attention_mask and position_ids generation when past_key_values is provided in https://github.com/huggingface/transformers/blob/v4.33.2/src/transformers/models/speecht5/modeling_speecht5.py#L2572-L2602.",
+        "encoder_model.onnx: corresponds to the vision encoder + projection + decoder in https://github.com/huggingface/transformers/blob/v4.39.3/src/transformers/models/llava/modeling_llava.py#L421-L445.\n\t - "
+        "decoder_model.onnx: corresponds to the decoder part in https://github.com/huggingface/transformers/blob/v4.39.3/src/transformers/models/llava/modeling_llava.py#L480-L489.\n\t - "
+        "attention_position_id_generator.onnx: corresponds to attention_mask and position_ids generation when past_key_values is provided in https://github.com/huggingface/transformers/blob/v4.39.3/src/transformers/models/llava/modeling_llava.py#L421-L478.",
     }
 
-    DEFAULT_VARIANT = "default"
+    DEFAULT_VARIANT = "optimized"
 
     def __init__(
         self,