Merge branch 'main' into android-thinking

Author: kirklandsign

backends/arm/_passes/arm_pass_manager.py

@@ -59,6 +59,9 @@
 )
 
 from executorch.backends.arm.tosa_specification import Tosa_0_80, TosaSpecification
+from executorch.backends.transforms.decompose_sdpa import (
+    DecomposeScaledDotProductAttention,
+)
 from executorch.backends.transforms.fuse_view_copy import FuseViewCopyTransform
 from executorch.backends.xnnpack._passes.remove_getitem_op import RemoveGetItemPass
 from executorch.exir import ExportedProgram
@@ -194,6 +197,7 @@ def transform_to_backend_pipeline(self, exported_program: ExportedProgram):
             )
 
     def transform_for_annotation_pipeline(self, graph_module: GraphModule):
+        self.add_pass(DecomposeScaledDotProductAttention())
         self.add_pass(ReplaceScalarWithTensorArgPassTOSABI())
         self.add_pass(ScalarsToAttributePass())
         self.add_pass(DecomposeLayerNormPass())

backends/arm/_passes/decompose_softmax_pass.py

@@ -8,7 +8,11 @@
 from executorch.exir.pass_base import ExportPass
 
 # For BI case
-torch_softmax = (torch.ops.aten.softmax.int, torch.ops.aten.log_softmax.int)
+torch_softmax = (
+    torch.ops.aten.softmax.int,
+    torch.ops.aten._safe_softmax.default,
+    torch.ops.aten.log_softmax.int,
+)
 # For MI case
 edge_softmax = (
     exir_ops.edge.aten._softmax.default,

backends/arm/test/models/test_conformer.py

@@ -83,7 +83,6 @@ def test_conformer_tosa_BI(self):
             )
         )
 
-    @unittest.expectedFailure  # TODO(MLETORCH-635)
     def test_conformer_u55_BI(self):
         tester = (
             ArmTester(
@@ -97,13 +96,20 @@ def test_conformer_u55_BI(self):
             .to_executorch()
             .serialize()
         )
+
         if conftest.is_option_enabled("corstone_fvp"):
-            tester.run_method_and_compare_outputs(
-                qtol=1.0,
-                rtol=1.0,
-                atol=5.0,
-                inputs=get_test_inputs(self.dim, self.lengths, self.num_examples),
-            )
+            try:
+                tester.run_method_and_compare_outputs(
+                    qtol=1.0,
+                    rtol=1.0,
+                    atol=5.0,
+                    inputs=get_test_inputs(self.dim, self.lengths, self.num_examples),
+                )
+                self.fail(
+                    "TODO(MLETORCH-635): Expected failure under FVP option, but test passed."
+                )
+            except Exception:
+                pass
 
     @unittest.expectedFailure  # TODO(MLETORCH-635)
     def test_conformer_u85_BI(self):

backends/arm/test/ops/test_sdpa.py

@@ -0,0 +1,45 @@
+# Copyright 2025 Arm Limited and/or its affiliates.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+from typing import Tuple
+
+import torch
+
+from executorch.backends.arm.test.tester.test_pipeline import (
+    TosaPipelineBI,
+    TosaPipelineMI,
+)
+
+
+class SDPA(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, query, key, value):
+        return torch.nn.functional.scaled_dot_product_attention(
+            query, key, value, attn_mask=None, dropout_p=0.0, is_causal=False
+        )
+
+
+input_t = Tuple[torch.Tensor, torch.Tensor, torch.Tensor]
+
+
+def test_sdpa_MI():
+    test_input = tuple(torch.randn(1, 3, 197, 64) for x in range(3))
+    pipeline = TosaPipelineMI[input_t](SDPA(), test_input, [], [])
+    pipeline.pop_stage("check_count.exir")
+    pipeline.run()
+
+
+def test_sdpa_BI():
+    test_input = tuple(torch.randn(1, 3, 197, 64) for x in range(3))
+    pipeline = TosaPipelineBI[input_t](SDPA(), test_input, [], [])
+    pipeline.pop_stage("check.quant_nodes")
+    pipeline.pop_stage("check_count.exir")
+    pipeline.pop_stage(
+        "run_method_and_compare_outputs"
+    )  # TODO: reference is not quantized
+    pipeline.run()

docs/source/_static/img/swiftpm_xcode.mp4

@@ -25,4 +25,6 @@ You are a helpful assistant.
 """
 
     public static let llama3PromptTemplate = "<|begin_of_text|><|start_header_id|>user<|end_header_id|>%@<|eot_id|><|start_header_id|>assistant<|end_header_id|>"
+    
+public static let phi4PromptTemplate = "<|user|>%@<|end|><|assistant|>"
 }

docs/source/_static/img/swiftpm_xcode.png

@@ -86,6 +86,7 @@ struct ContentView: View {
     case llama
     case llava
     case qwen3
+    case phi4
 
     static func fromPath(_ path: String) -> ModelType {
       let filename = (path as NSString).lastPathComponent.lowercased()
@@ -95,8 +96,10 @@ struct ContentView: View {
         return .llava
       } else if filename.hasPrefix("qwen3") {
         return .qwen3
+      } else if filename.hasPrefix("phi4") {
+        return .phi4
       }
-      print("Unknown model type in path: \(path). Model filename should start with one of: llama, llava, or qwen3")
+      print("Unknown model type in path: \(path). Model filename should start with one of: llama, llava, qwen3, or phi4")
       exit(1)
     }
   }
@@ -343,15 +346,15 @@ struct ContentView: View {
       }
 
       switch modelType {
-      case .llama, .qwen3:
+      case .llama, .qwen3, .phi4:
         runnerHolder.runner = runnerHolder.runner ?? Runner(modelPath: modelPath, tokenizerPath: tokenizerPath)
       case .llava:
         runnerHolder.llavaRunner = runnerHolder.llavaRunner ?? LLaVARunner(modelPath: modelPath, tokenizerPath: tokenizerPath)
       }
 
       guard !shouldStopGenerating else { return }
       switch modelType {
-      case .llama, .qwen3:
+      case .llama, .qwen3, .phi4:
         if let runner = runnerHolder.runner, !runner.isLoaded() {
           var error: Error?
           let startLoadTime = Date()
@@ -474,12 +477,14 @@ struct ContentView: View {
             prompt = String(format: Constants.llama3PromptTemplate, text)
           case .llava:
             prompt = String(format: Constants.llama3PromptTemplate, text)
+          case .phi4:
+              prompt = String(format: Constants.phi4PromptTemplate, text)
           }
 
           try runnerHolder.runner?.generate(prompt, sequenceLength: seq_len) { token in
 
             if token != prompt {
-              if token == "<|eot_id|>" {
+                if token == "<|eot_id|>" {
                 // hack to fix the issue that extension/llm/runner/text_token_generator.h
                 // keeps generating after <|eot_id|>
                 shouldStopShowingToken = true

examples/demo-apps/apple_ios/LLaMA/LLaMA/Application/Constants.swift

@@ -32,12 +32,12 @@ Download already exported LLaMA/LLaVA models along with tokenizers from [Hugging
     ```bash
     open examples/demo-apps/apple_ios/LLaMA/LLaMA.xcodeproj
     ```
-    
+
 3. Click the Play button to launch the app in the Simulator.
 
 4. To run on a device, ensure you have it set up for development and a provisioning profile with the `increased-memory-limit` entitlement. Update the app's bundle identifier to match your provisioning profile with the required capability.
 
-5. After successfully launching the app, copy the exported ExecuTorch model (`.pte`) and tokenizer (`.model`) files to the iLLaMA folder. Three models are currently supported at the moment - Llama, Qwen3, and Llava multimodal. Please ensure that your model `.pte` file starts with `llama`, `qwen3`, or `llava` so that the app selects the correct model type.
+5. After successfully launching the app, copy the exported ExecuTorch model (`.pte`) and tokenizer (`.model`) files to the iLLaMA folder. Four models are currently supported at the moment - Llama, Qwen3, Phi4-mini, and Llava multimodal. Please ensure that your model `.pte` file starts with `llama`, `qwen3`, `phi4` or `llava` so that the app selects the correct model type.
 
     - **For the Simulator:** Drag and drop both files onto the Simulator window and save them in the `On My iPhone > iLLaMA` folder.
     - **For a Device:** Open a separate Finder window, navigate to the Files tab, drag and drop both files into the iLLaMA folder, and wait for the copying to finish.