[PT2E][Quant] Support for embedding op quantization via

ExecuTorchNativeQuantizer (#99106)

ExecuTorchNativeQuantizer

ExecuTorchNativeQuantizer is a terribly name, I admit, however lets fix it once
we align on what the quantized kernel lib within executorch runtime should be called

Differential Revision: [D44986258](https://our.internmc.facebook.com/intern/diff/D44986258/)

**NOTE FOR REVIEWERS**: This PR has internal Meta-specific changes or comments, please review them on [Phabricator](https://our.internmc.facebook.com/intern/diff/D44986258/)!
Pull Request resolved: https://github.com/pytorch/pytorch/pull/99106
Approved by: https://github.com/jerryzh168

test/quantization/pt2e/test_quantize_pt2e.py

@@ -13,6 +13,7 @@ from torch.ao.quantization._pt2e.quantizer import (
 )
 from torch.ao.quantization._quantize_pt2e import convert_pt2e, prepare_pt2e_quantizer
 from torch.ao.quantization.backend_config import get_qnnpack_backend_config
+
 from torch.ao.quantization.qconfig import default_per_channel_symmetric_qnnpack_qconfig
 from torch.ao.quantization.quantize_fx import convert_to_reference_fx, prepare_fx
 from torch.testing._internal.common_quantization import (

torch/ao/quantization/_pt2e/quantizer/qnnpack_quantizer.py

@@ -116,6 +116,8 @@ def _get_act_obs_or_fq_ctr(quantization_config: Optional[QuantizationConfig]):
     if quantization_config is None:
         return None
     assert quantization_config is not None
+    if quantization_config.activation is None:
+        return None
     quantization_spec: QuantizationSpec = quantization_config.activation
     qdtype = _TORCH_DTYPE_TO_QDTYPE[quantization_spec.dtype]
     assert quantization_spec.qscheme in [
@@ -141,6 +143,8 @@ def _get_weight_obs_or_fq_ctr(quantization_config: Optional[QuantizationConfig])
     if quantization_config is None:
         return None
     assert quantization_config is not None
+    if quantization_config.weight is None:
+        return None
     quantization_spec: QuantizationSpec = quantization_config.weight
     qdtype = _TORCH_DTYPE_TO_QDTYPE[quantization_spec.dtype]
     if quantization_spec.qscheme == torch.per_tensor_symmetric:
@@ -169,6 +173,8 @@ def _get_bias_obs_or_fq_ctr(quantization_config: Optional[QuantizationConfig]):
     if quantization_config is None:
         return None
     assert quantization_config is not None
+    if quantization_config.bias is None:
+        return None
     quantization_spec: QuantizationSpec = quantization_config.bias
     assert (
         quantization_spec.dtype == torch.float

torch/ao/quantization/_pt2e/quantizer/quantizer.py

@@ -1,5 +1,5 @@
 from abc import ABC, abstractmethod
-from dataclasses import dataclass
+from dataclasses import asdict, dataclass
 from typing import Callable, List, NamedTuple, Optional
 
 import torch
@@ -18,6 +18,15 @@ SUPPORTED_QSCHEMES = [
     torch.per_channel_affine_float_qparams,
 ]
 
+# TODO: add support for torch dtype in quant code base
+# this includes observers and prepare/convert code
+_TORCH_DTYPE_TO_QDTYPE = {
+    torch.int8: torch.qint8,
+    torch.uint8: torch.quint8,
+    torch.int32: torch.qint32,
+    torch.float16: torch.float16,
+}
+
 
 @dataclass(eq=True, frozen=True)
 class QuantizationSpec:
@@ -53,13 +62,19 @@ class QuantizationSpec:
             raise ValueError("Ch_axis is < 0.")
 
 
+def get_observer_kwargs(quant_spec: QuantizationSpec):
+    kwargs_dict = asdict(quant_spec)
+    kwargs_dict["dtype"] = _TORCH_DTYPE_TO_QDTYPE[quant_spec.dtype]
+    return kwargs_dict
+
+
 # In the absence of better name, just winging it with QuantizationConfig
 QuantizationConfig = NamedTuple(
     "QuantizationConfig",
     [
-        ("activation", QuantizationSpec),
-        ("weight", QuantizationSpec),
-        ("bias", QuantizationSpec),
+        ("activation", Optional[QuantizationSpec]),
+        ("weight", Optional[QuantizationSpec]),
+        ("bias", Optional[QuantizationSpec]),
     ],
 )