Swap Llava export arg order (#14238)

Swaps Llava export arg order, such that forward takes `embeddings,
cache_position`, instead of `cache_position, embeddings`.

Author: jackzhxng

examples/models/llava/export_llava.py

@@ -77,7 +77,7 @@ def __init__(self, llava):
             super().__init__()
             self.text_model = llava.text_model
 
-        def forward(self, input_pos, embeddings):
+        def forward(self, embeddings, input_pos):
             return self.text_model(None, {"input_pos": input_pos}, embeddings)
 
     llava_text_model = LlavaTextModel(llava)
@@ -88,7 +88,7 @@ def forward(self, input_pos, embeddings):
         max_seq_len=llava.text_model_args.max_seq_len,
         dtype=DType.fp32,
         use_kv_cache=True,
-        example_inputs=(torch.tensor([0], dtype=torch.int64), embeddings),
+        example_inputs=(embeddings, torch.tensor([0], dtype=torch.int64)),
         dynamic_shapes=dynamic_shapes,
     )
 

examples/models/llava/model.py

@@ -405,5 +405,5 @@ def _get_image_dynamic_shapes(self):
 
     def _get_prompt_dynamic_shapes(self):
         dim = torch.export.Dim("token_dim", min=2, max=self.max_seq_len)
-        text_model_dynamic_shapes = ({0: 1}, {1: dim})
+        text_model_dynamic_shapes = ({1: dim}, {0: 1})
         return text_model_dynamic_shapes

examples/models/llava/runner/llava_image_prefiller.h

@@ -47,7 +47,7 @@ class ET_EXPERIMENTAL LlavaImagePrefiller {
 
     // Run text model
     auto outputs_res = ET_UNWRAP(module_->execute(
-        kTextModelMethod, {start_pos_tensor, image_encoder_outputs[0]}));
+        kTextModelMethod, {image_encoder_outputs[0], start_pos_tensor}));
     ET_CHECK_MSG(
         outputs_res[0].isTensor(),
         "Non Tensor Output returned from executing image prefill");

examples/models/llava/runner/llava_text_decoder_runner.h

@@ -34,7 +34,7 @@ class ET_EXPERIMENTAL LlavaTextDecoderRunner
         &start_pos, {1}, executorch::aten::ScalarType::Long);
     // run text model
     auto outputs_res = ET_UNWRAP(module_->execute(
-        kTextModelMethod, {start_pos_tensor, token_embedding_outputs[0]}));
+        kTextModelMethod, {token_embedding_outputs[0], start_pos_tensor}));
 
     ET_CHECK_MSG(
         outputs_res.size() == 1,

examples/models/llava/test/test_llava.py

@@ -97,7 +97,7 @@ def test_llava_export(self):
         )[0]
         llava_module.run_method(
             "text_decoder",
-            (torch.tensor([start_pos], dtype=torch.int64), pte_embeds_before_img),
+            (pte_embeds_before_img, torch.tensor([start_pos], dtype=torch.int64)),
         )
 
         # Update the start_pos. start_pos is used in kv cache. The source of truth
@@ -109,8 +109,8 @@ def test_llava_export(self):
         llava_module.run_method(
             "text_decoder",
             (
-                torch.tensor([start_pos], dtype=torch.int64),
                 pte_embeds_img,
+                torch.tensor([start_pos], dtype=torch.int64),
             ),
         )
 
@@ -123,7 +123,7 @@ def test_llava_export(self):
         )[0]
         pte_prefill_after_img = llava_module.run_method(
             "text_decoder",
-            (torch.tensor([start_pos], dtype=torch.int64), pte_embeds_after_img),
+            (pte_embeds_after_img, torch.tensor([start_pos], dtype=torch.int64)),
         )[0]
 
         # Update the logits for each prefill (kv cache) step.
@@ -140,7 +140,7 @@ def test_llava_export(self):
             )[0]
             logits = llava_module.run_method(
                 "text_decoder",
-                (torch.tensor([start_pos + i], dtype=torch.int64), token_embeds),
+                (token_embeds, torch.tensor([start_pos + i], dtype=torch.int64)),
             )[0]
             new_tokens.append(torch.argmax(logits).item())