[digantdesai]

Stack from ghstack (oldest at bottom):

• -> [executorch][native] Record dynamic-shape linears as chained bundled samples #24405
• [executorch][native] Run the ggml CPU act_quant sample test in CI #24373
• [executorch][vulkan] Round f16 image stores to nearest even #24372
• [executorch][native-vk] Note the int4 weight-only N % 4 and rank-4 linear gaps #24371
• [executorch][vulkan] Fix dq8ca q4gsw weight sums stride when N % 8 is 1 to 4 #24370
• [executorch][vulkan] Accept 1-D activations in 8da4w linear #24369
• [executorch][native] Cover negative, rounded and subnormal GGUF super-block scales in the op samples #24368
• [executorch][native] Decode quantized weights in float32 for precise samples #24367
• [executorch][native] Add GGUF q4_k and q6_k packages to the op samples #24366
• [executorch][native] Add FACTO-drawn linear cases to the op samples #24365
• [executorch][native] Suggest rounding to nearest when engines narrow values #24364
• [executorch][native-ggml] Compute cuBLAS GEMMs in f32 in the CUDA sample tests #24363
• [executorch][native] Judge activation-quantized engines against an int8-activation golden #24362
• [executorch][native-ggml] Reject weight-only quantized linears unless activation quant is allowed #24361
• [executorch][native] Run engines on bundled op samples with gtest #24360
• [executorch][native] Judge engine outputs against golden and precise samples #24359
• [executorch][native] Pin eager kernels when generating op samples #24358
• [executorch][native] Record op packages with bundled golden and precise samples #24357
• [executorch][native] Split engine host factories out of the PTN runners #24356
• [executorch][test-suite] Feed the Linear tests inputs with the model's 67 input features #24355

Engines resize dynamic inputs between inferences, so the op samples now cover a shape changing mid-run. A case built with chain is exported with its Dims and recorded as one sample of four calls. Each Dim starts mid-range, where export traces it, visits both bounds and returns, every other Dim in the opposite order. The harness already runs a sample's calls in order on one executable from one context, resizing inputs between them, so only its header comment changes.

The new linear cases are the XNNPACK delegate's dynamic linear tests, each also followed by a ReLU: one and two batch dims, one Dim on two dims or two inputs, and sequential and parallel linears, with batch dims down to 1, which tracing mid-range keeps export from specializing. The cases every quantization scheme fits take 256 input features over up to 64 tokens: dynamic_tokens in 2-D and, with a static batch of one as in an LLM, in 3-D, two batch dims sharing one Dim of up to 16, and the sequential and parallel linears.

Authored with Claude Code.

Differential Revision: D128770754

NOTE FOR REVIEWERS: This PR has internal Meta-specific changes or comments, please review them on Phabricator!

Differential Revision: D128770754

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/24405

• 📄 Preview Python docs built from this PR

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

❌ 1 New Failure

As of commit 43d5ffb with merge base 0f8e096 ():

NEW FAILURE - The following job has failed:

• pull / nxp-mcuxpresso-test / linux-job (gh)
  undefined reference to SYS_INFO_PrintAll'`

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.