Can not export Gemma3 model with int4 weight only

[Gasoonjia]

i;m trying to quantize Gemma3 model with int4 weight only but failed:

(optimum-test) [gasoonjia@devvm40058.cln0 ~/optimum-executorch (main)]$ optimum-cli export executorch             --model "google/gemma-3-4b-it"             --task "multimodal-text-to-text"             --recipe "cuda"             --dtype bfloat16             --device cuda             --max_seq_len 64             --output_dir int4-2 --qlinear_encoder 4w
Loading checkpoint shards: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:09<00:00,  4.58s/it]
Traceback (most recent call last):
  File "/home/gasoonjia/.conda/envs/optimum-test/bin/optimum-cli", line 7, in <module>
    sys.exit(main())
  File "/home/gasoonjia/.conda/envs/optimum-test/lib/python3.10/site-packages/optimum/commands/optimum_cli.py", line 208, in main
    service.run()
  File "/home/gasoonjia/.conda/envs/optimum-test/lib/python3.10/site-packages/optimum/commands/export/executorch.py", line 230, in run
    main_export(
  File "/home/gasoonjia/.conda/envs/optimum-test/lib/python3.10/site-packages/optimum/exporters/executorch/__main__.py", line 138, in main_export
    model = task_func(model_name_or_path, **kwargs)
  File "/home/gasoonjia/.conda/envs/optimum-test/lib/python3.10/site-packages/optimum/exporters/executorch/tasks/multimodal_text_to_text.py", line 226, in load_multimodal_text_to_text_model
    quantize_model_(**quantize_encoder_kwargs)
  File "/home/gasoonjia/.conda/envs/optimum-test/lib/python3.10/site-packages/optimum/exporters/executorch/quantization.py", line 108, in quantize_model_
    quantize_(
  File "/home/gasoonjia/.conda/envs/optimum-test/lib/python3.10/site-packages/torchao/quantization/quant_api.py", line 541, in quantize_
    _replace_with_custom_fn_if_matches_filter(
  File "/home/gasoonjia/.conda/envs/optimum-test/lib/python3.10/site-packages/torchao/quantization/quant_api.py", line 217, in _replace_with_custom_fn_if_matches_filter
    new_child = _replace_with_custom_fn_if_matches_filter(
  File "/home/gasoonjia/.conda/envs/optimum-test/lib/python3.10/site-packages/torchao/quantization/quant_api.py", line 217, in _replace_with_custom_fn_if_matches_filter
    new_child = _replace_with_custom_fn_if_matches_filter(
  File "/home/gasoonjia/.conda/envs/optimum-test/lib/python3.10/site-packages/torchao/quantization/quant_api.py", line 217, in _replace_with_custom_fn_if_matches_filter
    new_child = _replace_with_custom_fn_if_matches_filter(
  [Previous line repeated 3 more times]
  File "/home/gasoonjia/.conda/envs/optimum-test/lib/python3.10/site-packages/torchao/quantization/quant_api.py", line 212, in _replace_with_custom_fn_if_matches_filter
    model = replacement_fn(model, *extra_args)
  File "/home/gasoonjia/.conda/envs/optimum-test/lib/python3.10/site-packages/torchao/quantization/quant_api.py", line 2233, in _intx_weight_only_transform
    new_weight = _intx_weight_only_quantize_tensor(module.weight, config)
  File "/home/gasoonjia/.conda/envs/optimum-test/lib/python3.10/site-packages/torchao/quantization/quant_api.py", line 2188, in _intx_weight_only_quantize_tensor
    new_weight = IntxUnpackedToInt8Tensor.from_hp(
  File "/home/gasoonjia/.conda/envs/optimum-test/lib/python3.10/site-packages/torchao/quantization/quantize_/workflows/intx/intx_unpacked_to_int8_tensor.py", line 185, in from_hp
    scale, zero_point = choose_qparams_affine(
  File "/home/gasoonjia/.conda/envs/optimum-test/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 124, in decorate_context
    return func(*args, **kwargs)
  File "/home/gasoonjia/.conda/envs/optimum-test/lib/python3.10/site-packages/torchao/quantization/quant_primitives.py", line 1238, in choose_qparams_affine
    return _choose_qparams_affine(
  File "/home/gasoonjia/.conda/envs/optimum-test/lib/python3.10/site-packages/torch/_ops.py", line 1251, in __call__
    return self._op(*args, **kwargs)
  File "/home/gasoonjia/.conda/envs/optimum-test/lib/python3.10/site-packages/torchao/quantization/quant_primitives.py", line 1546, in _choose_qparams_affine
    shape_for_reduction, reduction_dims = _get_reduction_params(
  File "/home/gasoonjia/.conda/envs/optimum-test/lib/python3.10/site-packages/torchao/quantization/quant_primitives.py", line 300, in _get_reduction_params
    assert input_size[i] % block_size[i] == 0, (
AssertionError: Expecting input size at 1 dimension: 4304 to be divisible by block_size at 1 dimension: 32

i've tried to patch these two PRs:

patch 2 PRs:

#178

#165

but the error message is still the same.