disable torchao kernels by default

Author: metascroy

.github/workflows/trunk.yml

@@ -387,7 +387,7 @@ jobs:
         eval "$(conda shell.bash hook)"
 
         # Install requirements
-        ${CONDA_RUN} python install_executorch.py
+        ${CONDA_RUN} USE_CPP=1 python install_executorch.py
         ${CONDA_RUN} sh examples/models/llama/install_requirements.sh
 
         # Run test

examples/models/llama/README.md

@@ -339,7 +339,12 @@ Please refer to [this tutorial](https://pytorch.org/executorch/main/llm/llama-de
 
 ## Running with low-bit kernels
 
-We now give instructions for quantizating and running your model with low-bit kernels.  These are still experimental, and require you do development on an Arm-based Mac.  Also note that low-bit quantization often requires QAT (quantization-aware training) to give good quality results.
+We now give instructions for quantizating and running your model with low-bit kernels.  These are still experimental, and require you do development on an Arm-based Mac, and install executorch from source with the environment variable USE_CPP=1 defined:
+```
+USE_CPP=1 python install_executorch.py
+```
+
+Also note that low-bit quantization often requires QAT (quantization-aware training) to give good quality results.
 
 First export your model for lowbit quantization (step 2 above):
 

install_requirements.py

@@ -118,7 +118,8 @@ def install_requirements(use_pytorch_nightly):
     # Install packages directly from local copy instead of pypi.
     # This is usually not recommended.
     new_env = os.environ.copy()
-    new_env["USE_CPP"] = "1"  # install torchao kernels
+    if "USE_CPP" not in new_env:
+        new_env["USE_CPP"] = "0"
     subprocess.run(
         [
             sys.executable,