[quantization] Introduce wrapper for Qwen3VLVisionPatchEmbed

[dvsav]

This change introduces QuantQwen3VLVisionPatchEmbed wrapper to support post-training quantization of Qwen3VLVisionPatchEmbed module.

Why?

Qwen3VLVisionPatchEmbed module is used in the image encoder part of Qwen model.
Trying to quantize Qwen3VLVisionPatchEmbed via PTQ generates exception PTQQuantizer: no quantization wrapper for Qwen3VLVisionPatchEmbed.

What

This change introduces:

• Class QuantQwen3VLVisionPatchEmbed (tico/quantization/wrapq/wrappers/qwen_vl/quant_vision_patch_embed.py).
• Unit tests: class TestQuantQwen3VLTextAttention (test/quantization/wrapq/wrappers/qwen_vl/test_quant_vision_patch_embed.py) - skipped if transformers package is not installed.
• New entry quant_vision_patch_embed in _CORE_MODULES (tico/quantization/wrapq/wrappers/registry.py).
• Example of Qwen3VLVisionPatchEmbed quantization and conversion to Circle (tico/quantization/wrapq/examples/qwen/quantize_qwen_vision_patch_embed.py).

Unit Tests

Unit tests results with coverage information:

$ coverage run -m pytest test/quantization/wrapq/wrappers/qwen_vl/test_quant_vision_patch_embed.py -v
======================================================================================= test session starts ========================================================================================
platform linux -- Python 3.10.12, pytest-8.4.0, pluggy-1.6.0 -- /home/d.savchenkov/myenv/bin/python3
cachedir: .pytest_cache
rootdir: /home/d.savchenkov/TICO
configfile: pyproject.toml
plugins: anyio-4.12.0, mock-3.15.1, xdist-3.7.0, cov-6.2.1
collected 9 items                                                                                                                                                                                  

test/quantization/wrapq/wrappers/qwen_vl/test_quant_vision_patch_embed.py::TestQuantQwen3VLVisionPatchEmbed::test_activation_stats_collected PASSED                                          [ 11%]
test/quantization/wrapq/wrappers/qwen_vl/test_quant_vision_patch_embed.py::TestQuantQwen3VLVisionPatchEmbed::test_different_batch_sizes      PASSED                                          [ 22%]
test/quantization/wrapq/wrappers/qwen_vl/test_quant_vision_patch_embed.py::TestQuantQwen3VLVisionPatchEmbed::test_forward_diff               PASSED                                          [ 33%]
test/quantization/wrapq/wrappers/qwen_vl/test_quant_vision_patch_embed.py::TestQuantQwen3VLVisionPatchEmbed::test_mode_transitions           PASSED                                          [ 44%]
test/quantization/wrapq/wrappers/qwen_vl/test_quant_vision_patch_embed.py::TestQuantQwen3VLVisionPatchEmbed::test_multiple_calibration_steps PASSED                                          [ 55%]
test/quantization/wrapq/wrappers/qwen_vl/test_quant_vision_patch_embed.py::TestQuantQwen3VLVisionPatchEmbed::test_observer_count             PASSED                                          [ 66%]
test/quantization/wrapq/wrappers/qwen_vl/test_quant_vision_patch_embed.py::TestQuantQwen3VLVisionPatchEmbed::test_output_shape               PASSED                                          [ 77%]
test/quantization/wrapq/wrappers/qwen_vl/test_quant_vision_patch_embed.py::TestQuantQwen3VLVisionPatchEmbed::test_proj_override              PASSED                                          [ 88%]
test/quantization/wrapq/wrappers/qwen_vl/test_quant_vision_patch_embed.py::TestQuantQwen3VLVisionPatchEmbed::test_registration_in_registry   PASSED                                          [100%]

================================================================================== 9 passed, 2 warnings in 6.50s ===================================================================================

Coverage info (irrelevant files skipped):

$ coverage report -m
Name                                                                   Stmts   Miss  Cover   Missing
----------------------------------------------------------------------------------------------------
...
tico/quantization/wrapq/wrappers/nn/quant_conv3d.py                       30      0   100%
tico/quantization/wrapq/wrappers/qwen_vl/quant_vision_patch_embed.py      30      0   100%
...
----------------------------------------------------------------------------------------------------
TOTAL                                                                  10170   6520    36%

[dayo09]

Suggested change

	# Copyright (c) 2025 Samsung Electronics Co., Ltd. All Rights Reserved
	# Copyright (c) 2026 Samsung Electronics Co., Ltd. All Rights Reserved

[dvsav]

👍 done

[dayo09]

(Just to note) Oh...? This model looks a bit different from my vision patch embed. Maybe because spacial_merge_size ..

[mhs4670go]

@dayo09 How different is it?

[dayo09]

@mhs4670go I cannot attach image files here, see here

[dayo09]

@dvsav

As below are our target configuration for this layer, could you use this?

Qwen3VLVisionPatchEmbed(
  (proj): Conv3d(3, 1024, kernel_size=(2, 16, 16), stride=(2, 16, 16))
)

'args': ('Tensor(shape=[468, 1536], dtype=torch.float32)',)

The reason why is that, your current example generates some float32 ADD operator remains. (See #489 for details)
We are planning to lower above specific Conv3d operator into Conv2d+Reshape (@llFreetimell is working on it). Above specifics are derived from a use case scenario (which is not 100% fixed for now, though).
Thus, it would be good to provide quantization example with above version.

(+ Do you have any specific reason to decide your configuration of this Qwen3VLVisionPatchEmbed?)

[dvsav]

As below are our target configuration for this layer, could you use this?

@dayo09 👍 Thanks for noticing this! I've changed the code of example and added assertions checking that Conv3d has the right configuration.

[dayo09]

@dvsav Well, after applying the config, the graph remains the same. (I am sorry that I cannot show you the image. I am not yet permitted to upload image, I will process that soon to alleviate your inconvenience)

Convolution's weight is lifted up as constant input and not constant-folded. I believe constant folding after quantization is required in this case. 😅

[dayo09]

LGTM

[mhs4670go]

LGTM