Update fused quant broadcast logic (#20171)

[DrJessop]

Summary:

Unifies QuantParamsStruct (sas_compiler's central quant-params abstraction) onto a single affine-quantization representation and drops the axis argument from every fused-quant op interface.

Core change (ops.py): scale/zero_point are now either a singleton (per-tensor, auto-expanded internally) or a full-rank tensor whose shape encodes the affine block layout — block_size[i] = tensor.shape[i] // scale.shape[i]. This one representation covers per-tensor, per-channel, per-group, and blockwise uniformly. quantize/dequantize delegate to torch.ops.torchao.(de)quantize_affine. The axis field is removed from QuantParamsStruct, all ~60 op-schema fields, _make_qp, and the _lib.define strings. is_per_tensor/is_per_channel/is_per_group and a new channel_axis() helper are now derived from scale shape (channel_axis() returns 0 if all dims are unary, the single non-unary dim if there's
exactly one, else None).

Fusion (fusion_pass.py, fusion_passes/utils.py): the qparams flat block is 6→5 tuple; the per-channel branch inserts an aten.view to make 1-D scales full-rank [1, …, C, …, 1] so their shape encodes the block layout.

Lowering boundary (graph_utils.py, lower_to_turing_linear.py, lower_to_turing_conv_no_nlu_params.py): Helios ParameterExtraction wants the compact scale form ([K] / [K, num_groups]), but the fused op now carries full-rank scales. New compact_scale_node() squeezes size-1 dims at the lowering boundary; the inserted view_copy folds via ConstantPropPass before extraction. Lowering asserts channel_axis() is not None.

Broadcast fix (fuse_mul_into_linear.py): channel_scale is an activation-space [K] vector (out-features is the trailing dim of the mul constant), whereas the weight scale is now full-rank [K, 1] (out-features at dim 0). Reshape the channel factor to [-1, 1, …] so it broadcasts along the weight scale's channel axis instead of producing a [K, K] outer product. The 1-D bias multiply is unchanged (bias is already [K]).

Misc consumers: quant_absorption.py per-tensor check is now out_scale.numel() == 1; BUCK adds the torchao dep.

Reviewed By: ethansfng

Differential Revision: D108065588

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20171

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[meta-codesync]

@DrJessop has exported this pull request. If you are a Meta employee, you can view the originating Diff in D108065588.

[github-actions]

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.