Qualcomm AI Engine Direct - Streaming Mimi Enablement

Author: winskuo-quic

backends/qualcomm/_passes/lift_constant_scalar_operands.py

@@ -40,10 +40,13 @@ class TensorOpInfo:
     aten.ne.Scalar: TensorOpInfo(aten.ne.Tensor, False, False),
     aten.add.Scalar: TensorOpInfo(aten.add.Tensor, False, False),
     aten.add_.Scalar: TensorOpInfo(aten.add_.Tensor, False, False),
+    # For below cases, refer to LiftAddTensor Model in UT for sample
+    aten.add.Tensor: TensorOpInfo(aten.add.Tensor, False, False),
     aten.div.Scalar: TensorOpInfo(aten.div.Tensor, False, False),
     aten.mul.Scalar: TensorOpInfo(aten.mul.Tensor, False, False),
     aten.rsub.Scalar: TensorOpInfo(aten.rsub.Tensor, False, False),
     aten.sub.Scalar: TensorOpInfo(aten.sub.Tensor, False, False),
+    aten.sub.Tensor: TensorOpInfo(aten.sub.Tensor, False, False),
     aten.pow.Tensor_Scalar: TensorOpInfo(aten.pow.Tensor_Tensor, False, False),
     # The scalar number arg[1] is missing when using default. Result in a corner case to deal
     aten.leaky_relu.default: TensorOpInfo(aten.prelu.default, True, False),

backends/qualcomm/_passes/remove_redundancy.py

@@ -22,6 +22,7 @@ def __init__(self):
             exir_ops.edge.aten.clone.default: self._default_condition,
             torch.ops.aten.alias.default: self._default_condition,
             exir_ops.edge.aten.alias.default: self._default_condition,
+            exir_ops.edge.aten.alias_copy.default: self._default_condition,
             exir_ops.edge.aten.lift_fresh_copy.default: self._default_condition,
             # remove this target if '_skip_dim_order' is set to False
             exir_ops.edge.dim_order_ops._to_dim_order_copy.default: self._dim_order_op_condition,

backends/qualcomm/tests/models.py

@@ -939,6 +939,15 @@ def forward(self, x):
         return self.constant < x
 
 
+class LiftAddTensor(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x):
+        N = 2 - 1
+        return x + N
+
+
 class Linear(torch.nn.Module):
     def __init__(self, use_bias: bool = True):
         super().__init__()

backends/qualcomm/tests/test_qnn_delegate.py

@@ -980,6 +980,12 @@ def test_qnn_backend_einsum_outer_product_relu(self):
         )
         self.lower_module_and_test_output(module, sample_input)
 
+    # TODO: Create a new UT class for passes specific checks
+    def test_qnn_backend_lift_add_tensor(self):
+        module = LiftAddTensor()  # noqa: F405
+        sample_input = (torch.Tensor([1, 2, 3, 4]).to(torch.int32),)
+        self.lower_module_and_test_output(module, sample_input)
+
     @unittest.skip("Fail because of bad accuracy")
     def test_qnn_backend_moe_feed_forward(self):
         args = ModelArgs()

examples/qualcomm/CMakeLists.txt

@@ -90,6 +90,9 @@ add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/executor_runner)
 # build qnn_llama_runner for llama
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/oss_scripts/llama)
 
+# build qnn_mimi_decoder_runner
+add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/oss_scripts/moshi)
+
 # build qaihub_llama2_7b_runner and qaihub_llama3_8b_runner
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/qaihub_scripts/llama)
 

examples/qualcomm/oss_scripts/moshi/CMakeLists.txt

@@ -0,0 +1,33 @@
+# Copyright (c) Qualcomm Innovation Center, Inc.
+# All rights reserved
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+set(_qnn_mimi_decoder_runner__srcs
+    ${CMAKE_CURRENT_LIST_DIR}/qnn_mimi_decoder_runner.cpp
+    ${CMAKE_CURRENT_LIST_DIR}/runner/runner.cpp
+  ${CMAKE_CURRENT_LIST_DIR}/runner/runner.h
+)
+
+# build mimi decoder runner
+add_executable(qnn_mimi_decoder_runner ${_qnn_mimi_decoder_runner__srcs})
+target_include_directories(
+    qnn_mimi_decoder_runner PUBLIC ${_common_include_directories}
+)
+target_link_libraries(
+    qnn_mimi_decoder_runner
+    qnn_executorch_backend
+    executorch_core
+    extension_module
+    extension_data_loader
+    gflags
+)
+
+target_compile_options(
+  qnn_llama_runner PUBLIC ${_common_compile_options}
+)
+
+set_target_properties(
+    qnn_mimi_decoder_runner PROPERTIES LINK_FLAGS "-Wl,-rpath='$ORIGIN'"
+)