coremltools.optimize.torch.quantization fails to handle w8a8 conv3d quantization properly

[siyryu]

🐞Describing the bug

• The issue arises when using coremltools.optimize.torch.quantization to quantize a model with w8a8 for Conv3d layers. The quantization process does not work properly, causing errors or incorrect behavior when converting the model to Core ML format.

Stack Trace

Traceback (most recent call last):
  File "/Users/silveryu/Developer/lightsvd/diffuserskit/torch_quant.py", line 72, in <module>
    ct.convert(traced_model,
  File "/opt/homebrew/anaconda3/envs/lightsvd/lib/python3.10/site-packages/coremltools/converters/_converters_entry.py", line 635, in convert
    mlmodel = mil_convert(
  File "/opt/homebrew/anaconda3/envs/lightsvd/lib/python3.10/site-packages/coremltools/converters/mil/converter.py", line 188, in mil_convert
    return _mil_convert(model, convert_from, convert_to, ConverterRegistry, MLModel, compute_units, **kwargs)
  File "/opt/homebrew/anaconda3/envs/lightsvd/lib/python3.10/site-packages/coremltools/converters/mil/converter.py", line 212, in _mil_convert
    proto, mil_program = mil_convert_to_proto(
  File "/opt/homebrew/anaconda3/envs/lightsvd/lib/python3.10/site-packages/coremltools/converters/mil/converter.py", line 288, in mil_convert_to_proto
    prog = frontend_converter(model, **kwargs)
  File "/opt/homebrew/anaconda3/envs/lightsvd/lib/python3.10/site-packages/coremltools/converters/mil/converter.py", line 108, in __call__
    return load(*args, **kwargs)
  File "/opt/homebrew/anaconda3/envs/lightsvd/lib/python3.10/site-packages/coremltools/converters/mil/frontend/torch/load.py", line 88, in load
    return _perform_torch_convert(converter, debug)
  File "/opt/homebrew/anaconda3/envs/lightsvd/lib/python3.10/site-packages/coremltools/converters/mil/frontend/torch/load.py", line 151, in _perform_torch_convert
    prog = converter.convert()
  File "/opt/homebrew/anaconda3/envs/lightsvd/lib/python3.10/site-packages/coremltools/converters/mil/frontend/torch/converter.py", line 1383, in convert
    self.convert_const()
  File "/opt/homebrew/anaconda3/envs/lightsvd/lib/python3.10/site-packages/coremltools/converters/mil/frontend/torch/converter.py", line 1251, in convert_const
    self._add_const(name, val)
  File "/opt/homebrew/anaconda3/envs/lightsvd/lib/python3.10/site-packages/coremltools/converters/mil/frontend/torch/converter.py", line 1206, in _add_const
    compression_op = self._construct_compression_op(val.detach().numpy(), name)
  File "/opt/homebrew/anaconda3/envs/lightsvd/lib/python3.10/site-packages/coremltools/converters/mil/frontend/torch/converter.py", line 1189, in _construct_compression_op
    result = self._construct_quantization_op(val, compression_info, param_name, result)
  File "/opt/homebrew/anaconda3/envs/lightsvd/lib/python3.10/site-packages/coremltools/converters/mil/frontend/torch/converter.py", line 953, in _construct_quantization_op
    raise ValueError(
ValueError: In conv1.weight, the `weight` should have same rank as `scale`, but got (1, 1, 3, 3, 3) vs (1, 1)

To Reproduce

import tqdm
import torch
from torch import nn
from torch.utils.data import DataLoader, Dataset
from coremltools.optimize.torch.quantization import (
    LinearQuantizer,
    LinearQuantizerConfig,
    ModuleLinearQuantizerConfig,
)
import coremltools as ct

def torch_quantization(
        model: torch.nn.Module,
        sample_dataloader: DataLoader,
) -> torch.nn.Module:
    config = LinearQuantizerConfig(
        global_config=ModuleLinearQuantizerConfig(
            weight_dtype="qint8",
            activation_dtype="quint8",
            quantization_scheme="symmetric",
            milestones=[0, 1000, 1000, 0],
        )
    )
    quantizer = LinearQuantizer(model, config)

    example_inputs = next(iter(sample_dataloader))
    quantizer.prepare(example_inputs=example_inputs, inplace=True)

    quantizer.step()

    # Do a forward pass through the model with calibration data
    for data in tqdm.tqdm(sample_dataloader, desc="calibrating"):
        with torch.no_grad():
            model(data)

    quantized_model = quantizer.finalize()
    return quantized_model

class SimpleNet(nn.Module):
    def __init__(self):
        super().__init__()
        self.conv1 = nn.Conv3d(1, 1, 3)

    def forward(self, x):
        x = self.conv1(x)
        return x

class sample_dataset(Dataset):
    def __init__(self):
        super().__init__()
        self.len = 10

    def __len__(self):
        return self.len

    def __getitem__(self, item):
        return torch.randn(1, 12, 224, 224, device="mps", dtype=torch.float32)

test_model = SimpleNet().to("mps")
quantized_model = torch_quantization(test_model, DataLoader(sample_dataset()))

with torch.no_grad():
    traced_model = torch.jit.trace(quantized_model.to("cpu"), torch.randn(1, 12, 244, 244, dtype=torch.float32))

    _ = traced_model(torch.randn(1, 12, 224, 224, dtype=torch.float32))

    ct.convert(traced_model,
               inputs=[ct.TensorType(name="input", shape=(3, 244, 244))],
               convert_to="mlprogram",
               minimum_deployment_target=ct.target.iOS18)

• This code triggers the issue when performing quantization on a model with a Conv3d layer using w8a8 quantization.

System environment (please complete the following information):

• Coremltools: 8.1
• OS: macOS 15.3
• PyTorch: 2.4.0
• Python: 3.10.15

Additional context

• The issue appears specifically when attempting to quantize a model with Conv3d. Other layers, like Conv2d, seem to work fine.

[aryanmahawar205]

@siyryu i would like to work on it!

[ankpatr1]

Hi, is this error still exist ? 1)I checked in my local if I used Conv2d and resolved so need confirmation can I use Conv2d or I need to use Conv3d only ? 2)we need to if CoreMLTools fully supports quantized 3D convolutions with the provided quantization settings.

import coremltools as ct

def torch_quantization( model: torch.nn.Module, sample_dataloader: DataLoader, ) -> torch.nn.Module: config = LinearQuantizerConfig( global_config=ModuleLinearQuantizerConfig( weight_dtype="qint8", activation_dtype="quint8", quantization_scheme="symmetric", -------------------- we resolve this issue to change "symmetric" to "asymmetric" milestones=[0, 1000, 1000, 0], ) ) let me know is this correct approach or not so that I ll raise PR .