[ExecuTorch][WebGPU] Add dequantize_per_tensor op (int8 buffer path)

[JCNTH]

Stack from ghstack (oldest at bottom):

• [ExecuTorch][WebGPU] Broadcast add.Tensor (mirror mul) + un-skip 6 models #21233
• [ExecuTorch][WebGPU] Add dynamic-shape resize hook to cat (output mismatch) #21232
• [ExecuTorch][WebGPU] Raise TensorMeta rank cap 4->8 to fix rank>4 ops (0x30) #21231
• [ExecuTorch][WebGPU] Add bitwise_or + logical_or ops (bool) #21230
• [ExecuTorch][WebGPU] Add linear_dq8ca_q4gsw + choose_qparams_affine (8da4w) #21229
• [ExecuTorch][WebGPU] Enable + golden int8 quantized ops (q8ta family + qcs4w/q8ta_q8csw linear) #21228
• [ExecuTorch][WebGPU] Enable + golden granular ops (elementwise/bool, shape/movement, pooling, conv/sampling) #21227
• [ExecuTorch][WebGPU] Add creation/cast ops (full family, scalar_tensor, _to_copy) + enable unary activation family #21226
• [ExecuTorch][WebGPU] Optimize reduction-family kernels to cooperative shared-memory #21225
• [ExecuTorch][WebGPU] Op-tests for bitwise_and + bitwise_not #21224
• [ExecuTorch][WebGPU] Add bitwise_and + bitwise_not ops (bool) #21223
• [ExecuTorch][WebGPU] Op-tests for logical_and #21222
• [ExecuTorch][WebGPU] Add logical_and op (aten.logical_and.default) #21221
• [ExecuTorch][WebGPU] Op-tests for comparisons + bool-output golden harness #21220
• [ExecuTorch][WebGPU] Add comparison ops (aten.eq/lt/le/gt/ge.Tensor -> bool) #21219
• [ExecuTorch][WebGPU] Op-tests for conv_with_clamp #21218
• [ExecuTorch][WebGPU] Add conv_with_clamp op (et_vk.conv_with_clamp) #21217
• [ExecuTorch][WebGPU] Op-tests for grid_priors #21216
• [ExecuTorch][WebGPU] Add grid_priors op (et_vk.grid_priors) #21215
• [ExecuTorch][WebGPU] Op-tests for linear_q8ta_q8csw #21214
• [ExecuTorch][WebGPU] Add linear_q8ta_q8csw op (et_vk.linear_q8ta_q8csw) #21213
• [ExecuTorch][WebGPU] Op-tests for linear_qcs4w #21212
• [ExecuTorch][WebGPU] Add linear_qcs4w op (et_vk.linear_qcs4w) #21211
• [ExecuTorch][WebGPU] Op-tests for argmax/argmin + int64-golden harness #21210
• [ExecuTorch][WebGPU] Add argmax/argmin ops + int64-output path #21209
• [ExecuTorch][WebGPU] Op-tests for floor_divide #21208
• [ExecuTorch][WebGPU] Add floor_divide op (aten.div.Tensor_mode) #21207
• [ExecuTorch][WebGPU] Op-tests for q8ta_conv2d_transposed #21206
• [ExecuTorch][WebGPU] Add q8ta_conv2d_transposed op (int8 transposed conv) #21205
• [ExecuTorch][WebGPU] Op-tests for q8ta_conv2d #21204
• [ExecuTorch][WebGPU] Add q8ta_conv2d op (int8 general conv) #21203
• [ExecuTorch][WebGPU] Op-tests for q8ta_conv2d_dw #21202
• [ExecuTorch][WebGPU] Add q8ta_conv2d_dw op (int8 depthwise conv) #21201
• [ExecuTorch][WebGPU] Op-tests for q8ta_conv2d_pw #21200
• [ExecuTorch][WebGPU] Add q8ta_conv2d_pw op (int8 pointwise conv) #21199
• [ExecuTorch][WebGPU] Op-tests for q8ta_linear #21198
• [ExecuTorch][WebGPU] Add q8ta_linear op (int8 quantized linear) #21197
• [ExecuTorch][WebGPU] Op-tests for q8ta_pixel_shuffle #21196
• [ExecuTorch][WebGPU] Add q8ta_pixel_shuffle op (int8 pixel_shuffle) #21195
• [ExecuTorch][WebGPU] Op-tests for q8ta_relu #21194
• [ExecuTorch][WebGPU] Add q8ta_relu op (int8 relu) #21193
• [ExecuTorch][WebGPU] Op-tests for q8ta_add #21192
• [ExecuTorch][WebGPU] Add q8ta_add op (int8 elementwise add) #21191
• [ExecuTorch][WebGPU] Op-tests for quantize/dequantize_per_tensor #21190
• [ExecuTorch][WebGPU] Op-test harness: int8-output golden support #21189
• -> [ExecuTorch][WebGPU] Add dequantize_per_tensor op (int8 buffer path) #21188
• [ExecuTorch][WebGPU] Add quantize_per_tensor op (int8 buffer path) #21187
• [ExecuTorch][WebGPU] Test apply_rotary_emb_interleaved #21186
• [ExecuTorch][WebGPU] Add apply_rotary_emb_interleaved op #21185
• [ExecuTorch][WebGPU] Op-tests for conv1d pointwise #21184
• [ExecuTorch][WebGPU] Port conv1d pointwise (aten.convolution) #21183
• [ExecuTorch][WebGPU] Op-tests for conv1d_dw (depthwise conv1d) #21182
• [ExecuTorch][WebGPU] Port aten.convolution (depthwise conv1d) #21181
• [ExecuTorch][WebGPU] Op-tests for grid_sampler_2d #21180
• [ExecuTorch][WebGPU] Port grid_sampler_2d #21179
• [ExecuTorch][WebGPU] Op-tests for pixel_shuffle #21178
• [ExecuTorch][WebGPU] Port pixel_shuffle #21177
• [ExecuTorch][WebGPU] Op-tests for avg_pool2d #21176
• [ExecuTorch][WebGPU] Port avg_pool2d #21175
• [ExecuTorch][WebGPU] Op-tests for native_group_norm #21174
• [ExecuTorch][WebGPU] Port native_group_norm #21173
• [ExecuTorch][WebGPU] Op-tests for index_select #21172
• [ExecuTorch][WebGPU] Port index_select #21171
• [ExecuTorch][WebGPU] Op-test framework: multi-output goldens #21170
• [ExecuTorch][WebGPU] Op-tests for repeat #21169
• [ExecuTorch][WebGPU] Port repeat #21168
• [ExecuTorch][WebGPU] Op-tests for flip #21167
• [ExecuTorch][WebGPU] Port flip #21166
• [ExecuTorch][WebGPU] Op-tests for amax + amin #21165
• [ExecuTorch][WebGPU] Port amax + amin (last-dim reductions) #21164
• [ExecuTorch][WebGPU] Op-test for pow.Tensor_Scalar #21163
• [ExecuTorch][WebGPU] Port pow.Tensor_Scalar #21162
• [ExecuTorch][WebGPU] Op-tests for pow #21161
• [ExecuTorch][WebGPU] Port pow (aten.pow.Tensor_Tensor) #21160
• [ExecuTorch][WebGPU] Op-test for minimum #21159
• [ExecuTorch][WebGPU] Port minimum (first A2 binary op) #21158
• [ExecuTorch][WebGPU] Op-tests for clamp + hardtanh #21157
• [ExecuTorch][WebGPU] Port clamp + hardtanh (clamp-family unary ops) #21156
• [ExecuTorch][WebGPU] Op-tests for the 10 unary activations #21155
• [ExecuTorch][WebGPU] Port 10 unary activations to the WebGPU backend #21154
• [ExecuTorch][WebGPU] Promote unary op plumbing to a shared 2D-dispatch helper #21153

Problem: The WebGPU delegate has no per-tensor dequantization — the int8 -> fp32 half of the C0 quant gate, needed at the fp32 boundaries of a quantized graph.

Solution: Port quantized_decomposed.dequantize_per_tensor.default (int8 -> fp32). The int8 input is read as array<u32> (4 elements per word), each byte sign-extended, matching the layout quantize_per_tensor writes.

Implementation: DequantizePerTensor.cpp registers only .default, takes out = args.back() (robust to the extra out_dtype kwarg this overload carries), and guards int8 input (is_int8) / numel % 4 == 0 / fp32 output / non-null buffers, fail-loud. dequantize_per_tensor.wgsl unpacks each byte, sign-extends via (b ^ 0x80) - 128, and computes (q - zero_point) * scale. Mirrors Vulkan runtime/graph/ops/glsl/q8ta_dequantize.glsl + impl/QuantizeDequantize.cpp.

Differential Revision: D112257669

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/21188

• 📄 Preview Python docs built from this PR

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

❗ 1 Active SEVs

There are 1 currently active SEVs. If your PR is affected, please view them below:

• Unrelated MacOS MPS test jobs time out (30 min) due to stale cpp_extension lock file poisoning self-hosted runners

❌ 1 New Failure, 1 Unrelated Failure

As of commit bfdc3e2 with merge base 266e0dc ():

NEW FAILURE - The following job has failed:

• Build Presets / windows (windows) / build (gh)
  Can't find 'action.yml', 'action.yaml' or 'Dockerfile' under 'C:\actions-runner\_work\executorch\executorch\test-infra\.github\actions\teardown-windows'. Did you forget to run actions/checkout before running your local action?

BROKEN TRUNK - The following job failed but were present on the merge base:

👉 Rebase onto the `viable/strict` branch to avoid these failures

• pull / test-sqnr-static-llm-qnn-linux (smollm2_135m) / linux-job (gh) (trunk failure)
  test_static_llm_model

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

[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.

[psiddh]

Approving full WebGPU stack