Add torchao mps ops

Author: manuelcandales

torchchat/utils/quantize.py

@@ -63,10 +63,10 @@
 def get_named_parameters(func: Callable) -> List[str]:
     # Get the signature of the function
     signature = inspect.signature(func)
-    
+
     # Extract the parameters from the signature
     parameters = signature.parameters
-    
+
     # Filter and return named parameters
     named_params = [
         name for name, param in parameters.items()
@@ -80,8 +80,8 @@ def validate_args(named_params: List[str], q_kwargs: Dict[str, Any], quantizer:
             print(f"Specification for quantizer {quantizer} has extraneous key {key}. Ignoring.")
             del q_kwargs[key]
     return q_kwargs
-    
-            
+
+
 #########################################################################
 ###                  torchchat quantization API                       ###
 
@@ -116,15 +116,18 @@ def quantize_model(
                 if not support_tensor_subclass:
                     unwrap_tensor_subclass(model)
                 continue
-            
+
             if quantizer in ["linear:a8wxdq", "embedding:wx"]:
                 # These quantizers require float32 input weights.  Note that after quantization,
                 # the weights will no longer be float32, but lowbit integers
                 if get_precision() != torch.float32:
                     print(f"Quantizer {quantizer} requires float32 inputs, but received {get_precision()}.  Changing dtype to float32.  Note that after quantization, the weights will be lowbit integers, not float32.")
                     set_precision(torch.float32)
-                
-            # We set global precision from quantize options if it is specified at cli.py:485 
+
+            if quantizer == "linear:fpaxw" and device != "mps":
+                raise RuntimeError("linear:fpaxw quantization can only run on mps device!")
+
+            # We set global precision from quantize options if it is specified at cli.py:485
             # so the precision returned by get_precision() is always the authoritative precision/dtype in torchchat
             precision = get_precision()
 
@@ -915,10 +918,12 @@ def quantized_model(self) -> nn.Module:
     from torchao_experimental_quant_api import (
         Int8DynActIntxWeightLinearQuantizer,
         IntxWeightEmbeddingQuantizer,
+        UIntxWeightOnlyLinearQuantizer,
     )
 
     quantizer_class_dict["linear:a8wxdq"] = Int8DynActIntxWeightLinearQuantizer
     quantizer_class_dict["embedding:wx"] = IntxWeightEmbeddingQuantizer
+    quantizer_class_dict["linear:fpaxw"] = UIntxWeightOnlyLinearQuantizer
 
     # Try loading custom op
     try:
@@ -928,15 +933,14 @@ def quantized_model(self) -> nn.Module:
         libs = list(filter(lambda l: (l.endswith("so") or l.endswith("dylib")), libs))
         torch.ops.load_library(libs[0])
     except Exception as e:
-        print("Failed to torchao ops library with error: ", e)
-        print("Slow fallback kernels will be used.")
+        print("Unabled to load torchao cpu ops library. Slow fallback kernels will be used.")
+
+    try:
+        libname = "libtorchao_ops_mps_linear_fp_act_xbit_weight_aten.dylib"
+        libpath = f"{torchao_build_path}/cmake-out/lib/{libname}"
+        torch.ops.load_library(libpath)
+    except Exception as e:
+        print("Unabled to load torchao mps ops library.")
 
 except Exception as e:
-    class ErrorHandler(QuantHandler):
-        def __init__(self, model: Optional[nn.Module]=None, device="cpu", precision=None):
-            global torchao_experimental_load_error
-            raise Exception(f"Note: Failed to load torchao experimental quantizer with error: {torchao_experimental_load_error}")
-            
-    torchao_experimental_load_error = e
-    quantizer_class_dict["linear:a8wxdq"] = ErrorHandler
-    quantizer_class_dict["embedding:wx"] = ErrorHandler
+    print("Unabled to import torchao experimental quant_api with error: ", e)

torchchat/utils/scripts/build_torchao_ops.sh

@@ -5,12 +5,17 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+device=${1:-cpu}
 
+if [[ "$device" != "cpu" && "$device" != "mps" ]]; then
+  echo "Invalid argument: $device. Valid values are 'cpu' or 'mps'." >&2
+  exit 1
+fi
 
 source "$(dirname "${BASH_SOURCE[0]}")/install_utils.sh"
 
 pushd ${TORCHCHAT_ROOT}
 find_cmake_prefix_path
 clone_torchao
-install_torchao_aten_ops
+install_torchao_aten_ops "$device"
 popd

torchchat/utils/scripts/install_utils.sh

@@ -178,21 +178,35 @@ clone_torchao() {
 
   git clone https://github.com/pytorch/ao.git
   cd ao
-  git checkout $(cat ${TORCHCHAT_ROOT}/install/.pins/torchao-pin.txt)
+  # The next two lines will be removed before landing this PR
+  # Instead, the torcha-pin.txt will be updated once ao PR #1322 lands
+  git fetch origin pull/1322/head:pr-1322
+  git checkout pr-1322
+  # git checkout $(cat ${TORCHCHAT_ROOT}/install/.pins/torchao-pin.txt)
 
   popd
 }
 
 install_torchao_aten_ops() {
-  echo "Building torchao custom ops for ATen"
-  pushd ${TORCHCHAT_ROOT}/torchao-build/src/ao/torchao/experimental
+  local device=${1:-cpu}
+
+  if [[ "$device" == "cpu" ]]; then
+    echo "Building torchao custom ops for ATen"
+    pushd ${TORCHCHAT_ROOT}/torchao-build/src/ao/torchao/experimental
+  elif [[ "$device" == "mps" ]]; then
+    echo "Building torchao mps custom ops for ATen"
+    pushd ${TORCHCHAT_ROOT}/torchao-build/src/ao/torchao/experimental/ops/mps
+  else
+    echo "Invalid argument: $device. Valid values are 'cpu' or 'mps'." >&2
+    return 1
+  fi
 
   CMAKE_OUT_DIR=${TORCHCHAT_ROOT}/torchao-build/cmake-out
   cmake -DCMAKE_PREFIX_PATH=${MY_CMAKE_PREFIX_PATH} \
     -DCMAKE_INSTALL_PREFIX=${CMAKE_OUT_DIR} \
     -DCMAKE_BUILD_TYPE="Release" \
     -S . \
-    -B ${CMAKE_OUT_DIR} -G Ninja
+    -B ${CMAKE_OUT_DIR}
   cmake --build  ${CMAKE_OUT_DIR} --target install --config Release
 
   popd