[ET-VK] Add dynamic-shape resize to q8ta ops

[SS-JIA]

Stack from ghstack (oldest at bottom):

• -> [ET-VK] Add dynamic-shape resize to q8ta ops #20312
• [ET-VK][conv2d_dw] Fix test requesting nonexistent sned b1x1 shader #20343

The q8ta (quantized int8) op DynamicDispatchNodes were constructed with an empty resize-args list and no resize function, so their output tensors were never virtual_resized on trigger_resize(). On a dynamic-shape graph this froze the q8ta outputs at the build-time upper-bound shape — the same failure mode the fp32 ops already avoid. Concretely, in a quantized Vulkan-delegated graph the terminal pointwise conv produces the graph output, so a smaller input (e.g. 238 rows fed into a graph allocated at the 241-row upper bound) left stale rows that propagate downstream, where GroupNorm's global per-group statistics smear them across the whole tensor.

Add resize functions across the q8ta op family, each matching that op's output-shape semantics (mirroring the corresponding fp32 op's resize):

• q8ta_conv2d / q8ta_conv2d_dw: output H/W recomputed from the input via calc_out_sizes_hw.
• q8ta_conv2d_pw: 1x1 conv preserves spatial dims (out H/W == in H/W).
• q8ta_conv2d_transposed: transposed output formula via calc_out_sizes_hw(transposed=true) (threads output_padding through the dispatch, which was previously dropped).
• q8ta im2col scratch: flattened-window K from channels/kernel/groups, H_out/W_out from the current input.
• q8ta_linear: [*input.shape[:-1], out_features].
• q8ta binary: broadcast(in_a, in_b).
• q8ta quantize / dequantize: elementwise, output shape == input shape.

The quantized conv/quant path now honors dynamic input shapes like the fp32 path.

Differential Revision: D108788845

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20312

• 📄 Preview Python docs built from this PR

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

❌ 6 New Failures, 1 Unrelated Failure

As of commit 8530ff7 with merge base 0eb8247 ():

NEW FAILURES - The following jobs have failed:

• pull / test-arm-backend-no-driver (test_pytest_models_tosa) / linux-job (gh)
  RuntimeError: Command docker exec -t e85f66722d07aa845e2c454325b7227954bf6e1fb7653aff6d7d6845e3155da9 /exec failed with exit code 1
• pull / test-llama-runner-qnn-linux (fp32, qnn_16a16w, qnn) / linux-job (gh)
  RuntimeError: Command docker exec -t cf2d9f1a0df301412c1bd70076cd522ab1edf71cbffcd4ac5f4c8a1157b04e60 /exec failed with exit code 137
• pull / test-lora-multimethod-linux / linux-job (gh)
  RuntimeError: Command docker exec -t a0c01b7b7ffa436501995ea11299b943eae23bd981de0a8e52b74bd8d5f152d0 /exec failed with exit code 1
• pull / test-sqnr-static-llm-qnn-linux (smollm2_135m) / linux-job (gh)
  An error occurred trying to start process '/usr/bin/bash' with working directory '/home/ec2-user/actions-runner/_work/executorch/executorch/pytorch/executorch'. No such file or directory
• pull / test-voxtral-realtime-xnnpack-linux / linux-job (gh)
  RuntimeError: Command docker exec -t fb6049a87d70aff1cc915ca3f90573e67bfdce440535c76533530503a03595ef /exec failed with exit code 1
• pull / unittest / windows / windows-job (gh)
  examples/models/test/test_export.py::ExportTest::test_efficient_sam_export_to_executorch

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

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

• pull / unittest-nxp-neutron / linux-job (gh) (trunk failure)
  ##[error]The operation was canceled.

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

[linux-foundation-easycla]

• ❌ The email address for the commit (2dd36aa, 8530ff7, e8c73e8, f1d90f2) is not linked to the GitHub account, preventing the EasyCLA check. Consult this Help Article and GitHub Help to resolve. (To view the commit's email address, add .patch at the end of this PR page's URL.) For further assistance with EasyCLA, please visit our EasyCLA portal and chat with our support bot.

[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.

[SS-JIA]

@claude Review this PR

[claude]

Claude finished @SS-JIA's task in 8m 59s —— View job

---

PR Review: Add dynamic-shape resize to q8ta ops

• Read the full diff against base
• Review each q8ta op's resize function for correctness
• Check resize-args / shape-semantics consistency with fp32 ops
• Review test utility additions
• Post review feedback

---

Reviewed against origin/gh/SS-JIA/558/base. This is a solid, well-reasoned fix. The root cause (q8ta DynamicDispatchNodes built with empty resize-args and nullptr resize fn → outputs frozen at build-time upper bound) is clearly identified, and each resize function correctly mirrors its op's output-shape semantics. The commentary is excellent and the force_resize test mechanism is a clever way to exercise every resize formula across the swept shapes.

I verified the supporting infra and ArgGroup layouts, and the resize logic checks out:

• Binary (Q8taBinary.cpp:127): ArgGroup is {{out, kWrite}, {{in_a, in_b}, kRead}}, so args.at(1).refs.at(0/1) correctly reaches in_a/in_b. Matches fp32 resize_binary_op_node. ✅
• Linear (Q8taLinear.cpp:150-156): args.at(1).refs.at(0) is packed_int8_input; keeps last dim (out_features), takes leading dims from input. ✅
• Quant/dequant (Q8taQuantizeDequantize.cpp:87): args.at(1).refs.at(0) is the fp/int8 input; elementwise copy is correct. ✅
• Conv / DW / PW-im2col / transposed: calc_out_sizes_hw H/W recompute from current input, tail dims overwritten, N/C preserved — consistent with the fp32 resize_conv2d_node. The transposed path correctly threads the previously-dropped output_padding into the transposed=true formula. ✅
• PW standalone vs PW im2col split via conv_input == kDummyValueRef is a clean dispatch; the standalone path's stride-1/pad-0 assumption matches the op contract (Q8taConv2dPW.cpp:372). ✅
• im2col scratch (Q8taConv2dIm2Col.cpp:137-149): K/W recompute exactly mirrors calculate_q8ta_im2col_sizes, and recomputing from the original input (not the conv output, which may still be frozen during resize) is the right call and well-documented. ✅

A few minor, non-blocking notes:

1. The {stride, padding, dilation, dilation} filler arg is obscure. In the non-transposed conv/DW/im2col/PW resize calls, the 4th slot of calc_out_sizes_hw is the ceil_mode position, and dilation (a non-bool IntList) is passed purely to resolve val_is_bool(...) → false. The comment at Q8taConv2d.cpp:249-251 explains it, but reusing dilation reads as if dilation matters twice. If calc_out_sizes_hw's val_is_bool check handles kDummyValueRef gracefully, passing kDummyValueRef in that slot would express intent more clearly. Worth a quick check rather than a required change.

2. Resize-function duplication. resize_q8ta_conv2d_node, resize_q8ta_conv2d_dw_node, and resize_q8ta_conv2d_pw_im2col_node are nearly identical (same calc_out_sizes_hw call + tail-dim overwrite). Per the repo's "no duplicated logic" preference, these could collapse into one shared helper taking the resize_args. That said, the fp32 side has similar per-op resize fns, so matching that pattern is defensible — flagging only for consideration.

3. force_resize default-on is a repo-wide test behavior change. TestCase::reset() defaults force_resize_ = true (utils.h:667), so all ~24 custom_ops test files now run every DynamicDispatchNode's resize fn on each execute. This is good (it surfaces latent resize bugs), but it means a pre-existing buggy/missing resize in an unrelated op could now show up as a failure here. Worth confirming the full custom_ops suite is green, not just the q8ta tests. The opt-out path (set_force_resize(false)) is documented well for the intentionally-non-shape-preserving cases.

4. CI failures appear unrelated. The Dr. CI failures (arm-tosa, qnn-llama, lora-multimethod, voxtral) are infra/docker exit-code failures and SQNR jobs on other backends — none touch Vulkan q8ta. Confirm the Vulkan custom_ops job specifically passed.

No correctness issues found. Nice fix.
· branch gh/SS-JIA/558/head