Update

[ghstack-poisoned]

Author: swolchok

.ci/scripts/test_model.sh

@@ -100,6 +100,14 @@ test_model() {
       rm "./${MODEL_NAME}.pte"
       return  # Skip running with portable executor runnner since portable doesn't support Qwen's biased linears.
   fi
+  if [[ "${MODEL_NAME}" == "phi4_mini" ]]; then
+      # Install requirements for export_llama
+      bash examples/models/llama/install_requirements.sh
+      # Test export_llama script: python3 -m examples.models.llama.export_llama.
+      "${PYTHON_EXECUTABLE}" -m examples.models.llama.export_llama --model "${MODEL_NAME}" -c examples/models/llama/params/demo_rand_params.pth -p examples/models/phi-4-mini/config.json
+      run_portable_executor_runner
+      rm "./${MODEL_NAME}.pte"
+  fi
 
   # Export a basic .pte and run the model.
   "${PYTHON_EXECUTABLE}" -m examples.portable.scripts.export --model_name="${MODEL_NAME}" "${STRICT}"

backends/arm/scripts/build_executorch_runner.sh

@@ -14,8 +14,9 @@ toolchain_cmake=${et_root_dir}/examples/arm/ethos-u-setup/arm-none-eabi-gcc.cmak
 pte_file=""
 target="ethos-u55-128"
 build_type="Release"
-system_config=""
 bundleio=false
+system_config=""
+memory_mode=""
 build_with_etdump=false
 extra_build_flags=""
 output_folder_set=false
@@ -32,9 +33,12 @@ help() {
     echo "  --pte=<PTE_FILE>                pte file (genrated by the aot_arm_compier from the model to include in the elf"
     echo "  --target=<TARGET>               Target to build and run for Default: ${target}"
     echo "  --build_type=<TYPE>             Build with Release, Debug or RelWithDebInfo, default is ${build_type}"
-    echo "  --system_config=<CONFIG>        System configuration to select from the Vela configuration file (see vela.ini). Default: Ethos_U55_High_End_Embedded for EthosU55 targets, Ethos_U85_SYS_DRAM_Mid for EthosU85 targets."
-    echo "                                     NOTE: If given, this option must match the given target. This option also sets timing adapter values customized for specific hardware, see ./executor_runner/CMakeLists.txt."
     echo "  --bundleio                      Support both pte and Bundle IO bpte using Devtools BundelIO with Input/RefOutput included"
+    echo "  --system_config=<CONFIG>        System configuration to select from the Vela configuration file (see vela.ini). Default: Ethos_U55_High_End_Embedded for EthosU55 targets, Ethos_U85_SYS_DRAM_Mid for EthosU85 targets."
+    echo "                                     NOTE: If given, this option must match the given target. This option along with the memory_mode sets timing adapter values customized for specific hardware, see ./executor_runner/CMakeLists.txt."
+    echo "  --memory_mode=<CONFIG>          Vela memory mode, used for setting the Timing Adapter parameters of the Corstone platforms."
+    echo "                                  Valid values are Shared_Sram(for Ethos-U55, Ethos-U65, Ethos-85), Sram_Only(for Ethos-U55, Ethos-U65, Ethos-U85) or Dedicated_Sram(for Ethos-U65, Ethos-U85)."
+    echo "                                  Default: Shared_Sram for the Ethos-U55 and Sram_Only for the Ethos-U85"
     echo "  --etdump                        Adds Devtools etdump support to track timing, etdump area will be base64 encoded in the log"
     echo "  --extra_build_flags=<FLAGS>     Extra flags to pass to cmake like -DET_ARM_BAREMETAL_METHOD_ALLOCATOR_POOL_SIZE=60000 Default: none "
     echo "  --output=<FOLDER>               Output folder Default: <MODEL>/<MODEL>_<TARGET INFO>.pte"
@@ -49,8 +53,9 @@ for arg in "$@"; do
       --pte=*) pte_file="${arg#*=}";;
       --target=*) target="${arg#*=}";;
       --build_type=*) build_type="${arg#*=}";;
-      --system_config=*) system_config="${arg#*=}";;
       --bundleio) bundleio=true ;;
+      --system_config=*) system_config="${arg#*=}";;
+      --memory_mode=*) memory_mode="${arg#*=}";;
       --etdump) build_with_etdump=true ;;
       --extra_build_flags=*) extra_build_flags="${arg#*=}";;
       --output=*) output_folder="${arg#*=}" ; output_folder_set=true ;;
@@ -83,6 +88,15 @@ then
     fi
 fi
 
+if [[ ${memory_mode} == "" ]]
+then
+    memory_mode="Shared_Sram"
+    if [[ ${target} =~ "ethos-u85" ]]
+    then
+        memory_mode="Sram_Only"
+    fi
+fi
+
 output_folder=$(realpath ${output_folder})
 
 if [[ ${target} == *"ethos-u55"*  ]]; then
@@ -91,7 +105,7 @@ else
     target_cpu=cortex-m85
 fi
 echo "--------------------------------------------------------------------------------"
-echo "Build Arm Baremetal executor_runner for ${target} with ${pte_file} using ${system_config} ${extra_build_flags} to '${output_folder}/cmake-out'"
+echo "Build Arm Baremetal executor_runner for ${target} with ${pte_file} using ${system_config} ${memory_mode} ${extra_build_flags} to '${output_folder}/cmake-out'"
 echo "--------------------------------------------------------------------------------"
 
 cd ${et_root_dir}/examples/arm/executor_runner
@@ -120,6 +134,7 @@ cmake \
     ${build_with_etdump_flags}                  \
     -DPYTHON_EXECUTABLE=$(which python3)        \
     -DSYSTEM_CONFIG=${system_config}            \
+    -DMEMORY_MODE=${memory_mode}                \
     ${extra_build_flags}                        \
     -B ${output_folder}/cmake-out
 

backends/cadence/aot/functions_hifi.yaml

@@ -20,7 +20,7 @@
 - op: _softmax.out
   kernels:
     - arg_meta: null
-      kernel_name: cadence::impl::HiFi::softmax_out
+      kernel_name: cadence::impl::HiFi::_softmax_out
 
 - op: atan2.out
   kernels:
@@ -100,7 +100,7 @@
 - op: mean.out
   kernels:
     - arg_meta: null
-      kernel_name: cadence::impl::HiFi::mean_dim_out
+      kernel_name: cadence::impl::HiFi::mean_out   
 
 - op: minimum.out
   kernels:
@@ -175,7 +175,7 @@
 - op: where.self_out
   kernels:
     - arg_meta: null
-      kernel_name: cadence::impl::HiFi::where_out
+      kernel_name: cadence::impl::HiFi::where_self_out
 
 # custom ops
 - func: cadence::quantize_per_tensor.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)
@@ -189,6 +189,11 @@
   kernels:
     - arg_meta: null
       kernel_name: cadence::impl::HiFi::dequantize_per_tensor_out
+      
+- func: cadence::quantized_conv.out(Tensor input, Tensor weight, Tensor bias, int[] stride, SymInt[] padding, int[] dilation, int groups, int input_zero_point, Tensor weight_zero_point, Tensor bias_scale, float out_scale, int out_zero_point, Tensor out_multiplier, Tensor out_shift, bool channel_last=False, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: cadence::impl::HiFi::quantized_conv_out      
 
 - func: cadence::quantized_layer_norm.out(Tensor input, Tensor in_scale, Tensor in_zero_point, int[] normalized_shape, Tensor weight, Tensor bias, float eps, float output_scale, int output_zero_point, *, Tensor(a!) out) -> Tensor(a!)
   kernels:
@@ -209,6 +214,11 @@
     - arg_meta: null
       kernel_name: cadence::impl::HiFi::quantized_linear_per_tensor_out
 
+- func: cadence::quantized_relu_per_tensor.out(Tensor X, Tensor X_zero_point, int out_zero_point, Tensor out_multiplier, Tensor out_shift, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: cadence::impl::HiFi::quantized_relu_per_tensor_out
+
 - func: cadence::quantized_relu.out(Tensor X, Tensor X_zero_point, int out_zero_point, Tensor out_multiplier, Tensor out_shift, *, Tensor(a!) out) -> Tensor(a!)
   kernels:
     - arg_meta: null

backends/cadence/hifi/operators/CMakeLists.txt

@@ -78,7 +78,8 @@ target_include_directories(
 # Custom ops that are needed to run the test model.
 add_library(
   custom_ops "op_quantized_linear_out.cpp" "op_quantized_layer_norm.cpp"
-             "op_quantize_per_tensor.cpp" "op_quantized_relu_out.cpp" "op_dequantize_per_tensor.cpp" "op_quantized_fully_connected_out"
+             "op_quantize_per_tensor.cpp" "op_quantized_relu_out.cpp" "op_dequantize_per_tensor.cpp"
+             "op_quantized_conv_out.cpp" "op_quantized_fully_connected_out"
 )
 target_include_directories(
   custom_ops PUBLIC ${ROOT_DIR}/.. ${CMAKE_BINARY_DIR}