Qualcomm AI Engine Direct - Reland GA Static QWEN2.5 0.5B (#12582)

### Summary
Previous PR were merged unintentionally:
#12054
On top of previous PR, also had a new commit addressing code review.

### Test plan
[PLEASE REMOVE] How did you test this PR? Please write down any manual
commands you used and note down tests that you have written if
applicable.

Author: winskuo-quic

backends/qualcomm/_passes/layout_transform.py

@@ -103,8 +103,8 @@ class LayoutTransform(ExportPass):
         exir_ops.edge.aten.pow.Tensor_Scalar,
         exir_ops.edge.aten.prelu.default,
         exir_ops.edge.aten.repeat.default,
-        exir_ops.edge.aten.round.default,
         exir_ops.edge.aten.relu.default,
+        exir_ops.edge.aten.round.default,
         exir_ops.edge.aten.sigmoid.default,
         exir_ops.edge.aten.split_with_sizes.default,
         exir_ops.edge.aten.split_with_sizes_copy.default,

backends/qualcomm/quantizer/annotators.py

@@ -278,7 +278,9 @@ def annotate_masked_fill(node: Node, quantization_config: QuantizationConfig) ->
     )
 
 
-@register_annotator([torch.ops.aten.mul, torch.ops.aten.mul.Tensor])
+@register_annotator(
+    [torch.ops.aten.mul, torch.ops.aten.mul.Tensor, torch.ops.aten.mul_.Tensor]
+)
 def annotate_mul(node: Node, quantization_config: QuantizationConfig) -> None:
     annotate_binary(node, quantization_config)
 
@@ -1311,7 +1313,7 @@ def annotate_where(node: Node, quantization_config: QuantizationConfig) -> None:
     )
 
 
-@register_annotator([torch.ops.aten.zeros.default])
+@register_annotator([torch.ops.aten.zeros.default, torch.ops.aten.zeros_like.default])
 def annotate_zeros(node: Node, quantization_config: QuantizationConfig) -> None:
     if _is_annotated([node]) or not _is_float_tensor(node):
         return

backends/qualcomm/quantizer/custom_annotation.py

@@ -153,7 +153,9 @@ def annotate_prefill_kv_output(gm: torch.fx.GraphModule, kv_quant_attrs: dict):
                 )
 
 
-def annotate_matmul_16a8w(gm: torch.fx.GraphModule) -> None:  # noqa: C901
+def annotate_matmul_16a8w(  # noqa: C901
+    gm: torch.fx.GraphModule, annotate_conv=True
+) -> None:
     """
     This function is specific for matmul op 16a8w.
     For k, we will tag such as the below, and
@@ -317,9 +319,10 @@ def annotate_matmul_input1(node: Node):
                 # The arguments of cat op: (the past kv cache, the new kv cache)
                 node = node.args[0][1]
             elif node.target == torch.ops.aten.conv2d.default:
-                annotate_conv2d(
-                    node, quantization_config=quantization_config_8a4w_per_channel
-                )
+                if annotate_conv:
+                    annotate_conv2d(
+                        node, quantization_config=quantization_config_8a4w_per_channel
+                    )
                 break
             elif node.target in [torch.ops.aten.add.Tensor, torch.ops.aten.sub.Tensor]:
                 break

backends/qualcomm/scripts/build.sh

@@ -85,6 +85,7 @@ if [ "$BUILD_AARCH64" = true ]; then
         -DEXECUTORCH_BUILD_EXTENSION_FLAT_TENSOR=ON \
         -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON \
         -DEXECUTORCH_ENABLE_EVENT_TRACER=ON \
+        -DEXECUTORCH_ENABLE_LOGGING=ON \
         -DQNN_SDK_ROOT=$QNN_SDK_ROOT \
         -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK_ROOT/build/cmake/android.toolchain.cmake \
         -DANDROID_ABI='arm64-v8a' \
@@ -104,6 +105,9 @@ if [ "$BUILD_AARCH64" = true ]; then
         -DANDROID_ABI='arm64-v8a' \
         -DANDROID_PLATFORM=android-30 \
         -DCMAKE_PREFIX_PATH=$CMAKE_PREFIX_PATH \
+        -DSUPPORT_REGEX_LOOKAHEAD=ON \
+        -DBUILD_TESTING=OFF \
+        -DEXECUTORCH_ENABLE_LOGGING=ON \
         -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON \
         -DCMAKE_FIND_ROOT_PATH_MODE_PACKAGE=BOTH \
         -DPYTHON_EXECUTABLE=$PYTHON_EXECUTABLE \
@@ -134,6 +138,7 @@ if [ "$BUILD_X86_64" = true ]; then
         -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON \
         -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON \
         -DEXECUTORCH_ENABLE_EVENT_TRACER=ON \
+        -DEXECUTORCH_ENABLE_LOGGING=ON \
         -DPYTHON_EXECUTABLE=$PYTHON_EXECUTABLE \
         -S $PRJ_ROOT \
         -B $BUILD_ROOT \
@@ -157,6 +162,9 @@ if [ "$BUILD_X86_64" = true ]; then
        -DCMAKE_PREFIX_PATH=$CMAKE_PREFIX_PATH \
        -DCMAKE_FIND_ROOT_PATH_MODE_PACKAGE=BOTH \
        -DPYTHON_EXECUTABLE=$PYTHON_EXECUTABLE \
+       -DSUPPORT_REGEX_LOOKAHEAD=ON \
+       -DBUILD_TESTING=OFF \
+       -DEXECUTORCH_ENABLE_LOGGING=ON \
        -B$EXAMPLE_ROOT
 
    cmake --build $EXAMPLE_ROOT -j$BUILD_JOB_NUMBER

backends/qualcomm/tests/test_qnn_delegate.py

@@ -4049,7 +4049,7 @@ def test_llama3_2_1b(self):
             "16a4w",
             "--temperature",
             "0",
-            "--llama_model",
+            "--decoder_model",
             "llama3_2",
             "--model_mode",
             "hybrid",
@@ -4129,7 +4129,7 @@ def test_llama_stories_110m(self):
             "16a4w",
             "--temperature",
             "0",
-            "--llama_model",
+            "--decoder_model",
             "stories110m",
             "--model_mode",
             "hybrid",
@@ -4171,6 +4171,65 @@ def test_llama_stories_110m(self):
                 if not self.compile_only and not self.enable_x86_64:
                     self.assertGreaterEqual(msg["inference_speed"], 220)  # Lanai
 
+    def test_qwen2_5(self):
+        if not self.required_envs():
+            self.skipTest("missing required envs")
+
+        prompt = "My favourite condiment is "
+        cmds = [
+            "python",
+            f"{self.executorch_root}/examples/qualcomm/oss_scripts/llama/llama.py",
+            "--artifact",
+            self.artifact_dir,
+            "--build_folder",
+            self.build_folder,
+            "--model",
+            self.model,
+            "--ip",
+            self.ip,
+            "--port",
+            str(self.port),
+            "--prompt",
+            f"{prompt}",
+            "--ptq",
+            "16a8w",
+            "--decoder_model",
+            "qwen2_5",
+            "--model_mode",
+            "hybrid",
+            "--prefill_ar_len",
+            "32",
+            "--max_seq_len",
+            "128",
+        ]
+        if self.compile_only:
+            cmds.extend(["--compile_only"])
+        elif self.device:
+            cmds.extend(["--device", self.device])
+        if self.host:
+            cmds.extend(["--host", self.host])
+        elif self.enable_x86_64:
+            cmds.extend(["--enable_x86_64"])
+        if self.pre_gen_pte:
+            cmds.extend(["--pre_gen_pte", self.pre_gen_pte])
+
+        # Accuracy is bad for now. Just check user's prompt is returned.
+        golden_start_with = "My favourite condiment is "
+        p = subprocess.Popen(cmds, stdout=subprocess.DEVNULL)
+        with Listener((self.ip, self.port)) as listener:
+            conn = listener.accept()
+            p.communicate()
+            msg = json.loads(conn.recv())
+            if "Error" in msg:
+                self.fail(msg["Error"])
+            else:
+                model_out = msg["result"][0]
+                self.assertTrue(
+                    model_out.startswith(golden_start_with),
+                    f"Expected Output: {golden_start_with}. Actual Output: {model_out}",
+                )
+                self.assertGreaterEqual(msg["inference_speed"], 95)  # Lanai
+
 
 class TestExampleOssScript(TestQNN):
     def test_albert(self):

examples/models/qwen2_5/config/0_5b_config.json

@@ -0,0 +1,14 @@
+{
+  "dim": 896,
+  "ffn_dim_multiplier": 1,
+  "hidden_dim": 4864,
+  "n_heads": 14,
+  "n_kv_heads": 2,
+  "n_layers": 24,
+  "norm_eps": 1e-06,
+  "rope_theta": 1000000.0,
+  "use_scaled_rope": false,
+  "vocab_size": 151936,
+  "use_hf_rope": true,
+  "attention_qkv_bias": true
+}

examples/qualcomm/CMakeLists.txt

@@ -77,8 +77,8 @@ target_include_directories(
 
 # add tokenizers
 add_subdirectory(
-  ${EXECUTORCH_ROOT}/extension/llm/tokenizers
-  ${CMAKE_CURRENT_BINARY_DIR}/../../extension/llm/tokenizers
+  ${EXECUTORCH_ROOT}/extension/llm/runner
+  ${CMAKE_CURRENT_BINARY_DIR}/../../extension/llm/runner
 )
 
 # build qnn_executor_runner

examples/qualcomm/oss_scripts/llama/CMakeLists.txt

@@ -4,6 +4,7 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+
 # model sharding with custom op
 set(CUSTOM_OP_SRCS_FILE
     "${EXECUTORCH_SOURCE_DIR}/extension/llm/custom_ops/op_fallback.cpp"
@@ -63,14 +64,22 @@ target_link_libraries(
   executorch_core
   extension_data_loader
   extension_flat_tensor
+  extension_llm_runner
   extension_module
   extension_tensor
+  tokenizers
   gflags
   custom_ops
   quantized_ops_lib
   quantized_kernels
   tokenizers
 )
+
+target_include_directories(
+  qnn_llama_runner
+  PUBLIC ${EXECUTORCH_ROOT}/extension/llm/tokenizers/include
+)
+
 target_compile_options(qnn_llama_runner PUBLIC ${_common_compile_options})
 set_target_properties(
   qnn_llama_runner PROPERTIES LINK_FLAGS "-Wl,-rpath='$ORIGIN'"

examples/qualcomm/oss_scripts/llama/README.md

@@ -1,10 +1,11 @@
 # Summary
 
 ## Overview
-This file provides you the instructions to run LLAMA model with different parameters via Qualcomm HTP backend. We currently support the following models:
+This file provides you the instructions to run LLM Decoder model with different parameters via Qualcomm HTP backend. We currently support the following models:
  1. LLAMA2 Stories 110M
  2. LLAMA3.2 1B
  3. LLAMA3.2 3B
+ 4. QWEN2.5 0.5B
 
 We offer the following modes to execute the model: