[Vulkan] Stop clamping per-row reductions and propagate NaN by mergennachin · Pull Request #23244 · pytorch/executorch

mergennachin · 2026-09-29T15:02:34Z

convert.glslh guarded an fp16 clamp with #if T == float16_t, but neither side is a macro, so the preprocessor compares 0 with 0 and the clamp was always on. Per-row buffer sum, mean, amax and amin therefore saturated fp32 results at 65504. The clamp also affected generated int32 shader variants, which are now tested through direct backend lowering; integer reductions remain excluded by the partitioner. The clamp is removed. Single-dimension texture and per-row buffer amax and amin now propagate NaN, argmax and argmin return the index of the first NaN as ATen does, and mean divides in the accumulator type. FP16 texture output uses explicit nearest-even conversion so rounding and overflow agree with ATen across drivers. Exact FP16 texture tests cover even/odd ties, exponent carry, the normal/subnormal and zero/subnormal boundaries, and sums below, at and above the ±65520 overflow threshold. Comparisons are exact against ATen for both signs and include zero signs.

Part 5/15 of the Vulkan transformer and operator-conformance stack. #23243 has landed; review against main. Integration PR: #23254.

Validation: All three focused review tests pass on MoltenVK. The FP16 tests cover 18 signed cases with exact ATen comparisons, actual FP16 texture output, and zero-sign checks. The int32 test lowers amax/amin directly to Vulkan, verifies INT32 buffer inputs and outputs, and matches ATen beyond both ±65504 limits; the same inputs also match portable CPU execution. The previously rebased stack passed 51 native tests with one expected SwiftShader-only skip; the review updates add test coverage. Lintrunner and git diff --check pass. Vulkan CI is rerunning on the updated heads.

Recreates #23205 through ghstack. Prior review discussion remains on that PR.

Authored with OpenAI Codex; split planned with Claude Code.

…ghstack [ghstack-poisoned]

mergennachin · 2026-09-29T15:02:35Z

Stack from ghstack (oldest at bottom):

pytorch-bot · 2026-09-29T15:02:39Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/23244

📄 Preview Python docs built from this PR

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

✅ No Failures

As of commit 1ad69df with merge base 2c85103 ():
💚 Looks good so far! There are no failures yet. 💚

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

@SS-JIA

…ormer blocks The dynamic transformer reproductions now lower to one Vulkan delegate across changing sequence lengths. This final PR makes `expand_copy` resizable and adds the eager-attention and SDPA integration tests, including fully masked rows on devices with 8-bit storage buffers. Fixes #23156. This is part 15 of a linear ghstack series. Review each PR against its selected base branch; this PR contains only the final expansion and transformer changes. The scalar-tensor issue #23158 is owned by part 10. | Part | PR | Change | | --- | --- | --- | | 1 | #23240 | Hardware Vulkan CI | | 2 | #23241 | Scalar cache type and signed-zero keys | | 3 | #23242 | Vulkan-local signed-zero serialization | | 4 | #23243 | GELU modes and view kwargs | | 5 | #23244 | Reduction clamp, NaN, and FP16 rounding | | 6 | #23245 | Reduction dimension guards | | 7 | #23246 | Bool staging and logical_not | | 8 | #23247 | Scalar representability and symbolic guards | | 9 | #23248 | 64-bit dtype and fusion policy | | 10 | #23249 | scalar_tensor with exact integer values | | 11 | #23250 | Typed, resizable full | | 12 | #23251 | Power special values and logical FP16 dtype | | 13 | #23252 | mul.Scalar | | 14 | #23253 | any.dim | | 15 | #23254 | Dynamic expand and transformer integration | The stack retains the FACTO and ATen conformance fixes and their regression tests: typed integer fills, NaN and signed-zero behavior, reduction range and accumulation fixes, safe scalar fallbacks, and the FP16 power contract. FP16 writes use explicit nearest-even conversion, and power keeps the requested tensor dtype when a device emulates FP16 storage with FP32. The test module is now `test_vulkan_dynamic.py`; CI and Buck references follow the rename. Each code patch was linted and tested before its original publication. Prior validation of the operator changes passed 116 FACTO cases against both ATen and portable kernels, plus 144 boundary cases against ATen, with texture and buffer preferences: all 520 Vulkan-configuration executions matched ATen, including CPU fallback where unsupported. The additional 239-case sweep covers all 91 available FACTO specifications for registered ATen overloads and 57 targeted examples. Exact-input replays of every candidate Vulkan failure on a separately built main at `a31838280f9309f04af1b375147375b702d29339` found no new regressions or unresolved comparisons. Validation uses the shared `backends/test/` harness on Apple M1 Pro / MoltenVK with Release portable CPU kernels and Vulkan; optimized kernels and XNNPACK are disabled. Buffer preference uses `texture_limits=(1,1,1)`, with actual storage and fallback recorded separately. Single-texel tensors can still select textures. FACTO revision: `3b8c778c99766a8b4d0d04563ae0b16cbb276829`, seed 0. The complete native regression run on the previously published head `a2543b8a82` passed **50 tests with one expected SwiftShader-only skip** on MoltenVK, covering all 35 dynamic tests, graph-builder and serialization tests, and PT2E quantized linear with downcasting disabled. The subsequent serializer refinement confines `--force-defaults` to Vulkan and restores the shared `exir` FlatBuffers API. That revised source tree passed **26 focused tests**, including shared/Vulkan serialization, signed-zero execution, and power special values. All other code patches are unchanged, and every replacement PR was checked against the tested source trees. Lintrunner and `git diff --check` passed. NVIDIA, SwiftShader, and Windows CI are pending verification on the new ghstack heads. Recreates #23162 through ghstack. Prior review discussion remains on that PR. Authored with OpenAI Codex; split planned with Claude Code. cc @SS-JIA @manuelcandales @digantdesai @cbilgin ghstack-source-id: c6a7e94 ghstack-comment-id: 5892899249 Pull-Request: #23254

@SS-JIA

…ormer blocks The dynamic transformer reproductions now lower to one Vulkan delegate across changing sequence lengths. This final PR makes `expand_copy` resizable and adds the eager-attention and SDPA integration tests, including fully masked rows on devices with 8-bit storage buffers. Fixes #23156. This is part 15 of a linear ghstack series. Review each PR against its selected base branch; this PR contains only the final expansion and transformer changes. The scalar-tensor issue #23158 is owned by part 10. | Part | PR | Change | | --- | --- | --- | | 1 | #23240 | Hardware Vulkan CI | | 2 | #23241 | Scalar cache type and signed-zero keys | | 3 | #23242 | Vulkan-local signed-zero serialization | | 4 | #23243 | GELU modes and view kwargs | | 5 | #23244 | Reduction clamp, NaN, and FP16 rounding | | 6 | #23245 | Reduction dimension guards | | 7 | #23246 | Bool staging and logical_not | | 8 | #23247 | Scalar representability and symbolic guards | | 9 | #23248 | 64-bit dtype and fusion policy | | 10 | #23249 | scalar_tensor with exact integer values | | 11 | #23250 | Typed, resizable full | | 12 | #23251 | Power special values and logical FP16 dtype | | 13 | #23252 | mul.Scalar | | 14 | #23253 | any.dim | | 15 | #23254 | Dynamic expand and transformer integration | The stack retains the FACTO and ATen conformance fixes and their regression tests: typed integer fills, NaN and signed-zero behavior, reduction range and accumulation fixes, safe scalar fallbacks, and the FP16 power contract. FP16 writes use explicit nearest-even conversion, and power keeps the requested tensor dtype when a device emulates FP16 storage with FP32. The test module is now `test_vulkan_dynamic.py`; CI and Buck references follow the rename. Each code patch was linted and tested before its original publication. Prior validation of the operator changes passed 116 FACTO cases against both ATen and portable kernels, plus 144 boundary cases against ATen, with texture and buffer preferences: all 520 Vulkan-configuration executions matched ATen, including CPU fallback where unsupported. The additional 239-case sweep covers all 91 available FACTO specifications for registered ATen overloads and 57 targeted examples. Exact-input replays of every candidate Vulkan failure on a separately built main at `a31838280f9309f04af1b375147375b702d29339` found no new regressions or unresolved comparisons. Validation uses the shared `backends/test/` harness on Apple M1 Pro / MoltenVK with Release portable CPU kernels and Vulkan; optimized kernels and XNNPACK are disabled. Buffer preference uses `texture_limits=(1,1,1)`, with actual storage and fallback recorded separately. Single-texel tensors can still select textures. FACTO revision: `3b8c778c99766a8b4d0d04563ae0b16cbb276829`, seed 0. The complete native regression run on the previously published head `a2543b8a82` passed **50 tests with one expected SwiftShader-only skip** on MoltenVK, covering all 35 dynamic tests, graph-builder and serialization tests, and PT2E quantized linear with downcasting disabled. The subsequent serializer refinement confines `--force-defaults` to Vulkan and restores the shared `exir` FlatBuffers API. That revised source tree passed **26 focused tests**, including shared/Vulkan serialization, signed-zero execution, and power special values. All other code patches are unchanged, and every replacement PR was checked against the tested source trees. Lintrunner and `git diff --check` passed. NVIDIA, SwiftShader, and Windows CI are pending verification on the new ghstack heads. Recreates #23162 through ghstack. Prior review discussion remains on that PR. Authored with OpenAI Codex; split planned with Claude Code. cc @SS-JIA @manuelcandales @digantdesai @cbilgin ghstack-source-id: f8d24d9 ghstack-comment-id: 5892899249 Pull-Request: #23254

@SS-JIA

…ormer blocks The dynamic transformer reproductions now lower to one Vulkan delegate across changing sequence lengths. This final PR makes `expand_copy` resizable and adds the eager-attention and SDPA integration tests, including fully masked rows on devices with 8-bit storage buffers. Fixes #23156. This is part 15 of a linear ghstack series. Review each PR against its selected base branch; this PR contains only the final expansion and transformer changes. The scalar-tensor issue #23158 is owned by part 10. | Part | PR | Change | | --- | --- | --- | | 1 | #23240 | Hardware Vulkan CI | | 2 | #23241 | Scalar cache type and signed-zero keys | | 3 | #23242 | Vulkan-local signed-zero serialization | | 4 | #23243 | GELU modes and view kwargs | | 5 | #23244 | Reduction clamp, NaN, and FP16 rounding | | 6 | #23245 | Reduction dimension guards | | 7 | #23246 | Bool staging and logical_not | | 8 | #23247 | Scalar representability and symbolic guards | | 9 | #23248 | 64-bit dtype and fusion policy | | 10 | #23249 | scalar_tensor with exact integer values | | 11 | #23250 | Typed, resizable full | | 12 | #23251 | Power special values and logical FP16 dtype | | 13 | #23252 | mul.Scalar | | 14 | #23253 | any.dim | | 15 | #23254 | Dynamic expand and transformer integration | The stack retains the FACTO and ATen conformance fixes and their regression tests: typed integer fills, NaN and signed-zero behavior, reduction range and accumulation fixes, safe scalar fallbacks, and the FP16 power contract. FP16 writes use explicit nearest-even conversion, and power keeps the requested tensor dtype when a device emulates FP16 storage with FP32. The test module is now `test_vulkan_dynamic.py`; CI and Buck references follow the rename. Each code patch was linted and tested before its original publication. Prior validation of the operator changes passed 116 FACTO cases against both ATen and portable kernels, plus 144 boundary cases against ATen, with texture and buffer preferences: all 520 Vulkan-configuration executions matched ATen, including CPU fallback where unsupported. The additional 239-case sweep covers all 91 available FACTO specifications for registered ATen overloads and 57 targeted examples. Exact-input replays of every candidate Vulkan failure on a separately built main at `a31838280f9309f04af1b375147375b702d29339` found no new regressions or unresolved comparisons. Validation uses the shared `backends/test/` harness on Apple M1 Pro / MoltenVK with Release portable CPU kernels and Vulkan; optimized kernels and XNNPACK are disabled. Buffer preference uses `texture_limits=(1,1,1)`, with actual storage and fallback recorded separately. Single-texel tensors can still select textures. FACTO revision: `3b8c778c99766a8b4d0d04563ae0b16cbb276829`, seed 0. The complete native regression run on the previously published head `a2543b8a82` passed **50 tests with one expected SwiftShader-only skip** on MoltenVK, covering all 35 dynamic tests, graph-builder and serialization tests, and PT2E quantized linear with downcasting disabled. The subsequent serializer refinement confines `--force-defaults` to Vulkan and restores the shared `exir` FlatBuffers API. That revised source tree passed **26 focused tests**, including shared/Vulkan serialization, signed-zero execution, and power special values. All other code patches are unchanged, and every replacement PR was checked against the tested source trees. Lintrunner and `git diff --check` passed. NVIDIA, SwiftShader, and Windows CI are pending verification on the new ghstack heads. Recreates #23162 through ghstack. Prior review discussion remains on that PR. Authored with OpenAI Codex; split planned with Claude Code. cc @SS-JIA @manuelcandales @digantdesai @cbilgin ghstack-source-id: 2d9c5b5 ghstack-comment-id: 5892899249 Pull-Request: #23254

…ension kernel; rebase after PR 23243 landed [ghstack-poisoned]

@SS-JIA

…ormer blocks The dynamic transformer reproductions now lower to one Vulkan delegate across changing sequence lengths. This final PR makes `expand_copy` resizable and adds the eager-attention and SDPA integration tests, including fully masked rows on devices with 8-bit storage buffers. Fixes #23156. This is part 15 of a linear ghstack series. Review each PR against its selected base branch; this PR contains only the final expansion and transformer changes. The scalar-tensor issue #23158 is owned by part 10. | Part | PR | Change | | --- | --- | --- | | 1 | #23240 | Hardware Vulkan CI | | 2 | #23241 | Scalar cache type and signed-zero keys | | 3 | #23242 | Vulkan-local signed-zero serialization | | 4 | #23243 | GELU modes and view kwargs | | 5 | #23244 | Reduction clamp, NaN, and FP16 rounding | | 6 | #23245 | Reduction dimension guards | | 7 | #23246 | Bool staging and logical_not | | 8 | #23247 | Scalar representability and symbolic guards | | 9 | #23248 | 64-bit dtype and fusion policy | | 10 | #23249 | scalar_tensor with exact integer values | | 11 | #23250 | Typed, resizable full | | 12 | #23251 | Power special values and logical FP16 dtype | | 13 | #23252 | mul.Scalar | | 14 | #23253 | any.dim | | 15 | #23254 | Dynamic expand and transformer integration | The stack retains the FACTO and ATen conformance fixes and their regression tests: typed integer fills, NaN and signed-zero behavior, reduction range and accumulation fixes, safe scalar fallbacks, and the FP16 power contract. FP16 writes use explicit nearest-even conversion, and power keeps the requested tensor dtype when a device emulates FP16 storage with FP32. The test module is now `test_vulkan_dynamic.py`; CI and Buck references follow the rename. Each code patch was linted and tested before its original publication. Prior validation of the operator changes passed 116 FACTO cases against both ATen and portable kernels, plus 144 boundary cases against ATen, with texture and buffer preferences: all 520 Vulkan-configuration executions matched ATen, including CPU fallback where unsupported. The additional 239-case sweep covers all 91 available FACTO specifications for registered ATen overloads and 57 targeted examples. Exact-input replays of every candidate Vulkan failure on a separately built main at `a31838280f9309f04af1b375147375b702d29339` found no new regressions or unresolved comparisons. Validation uses the shared `backends/test/` harness on Apple M1 Pro / MoltenVK with Release portable CPU kernels and Vulkan; optimized kernels and XNNPACK are disabled. Buffer preference uses `texture_limits=(1,1,1)`, with actual storage and fallback recorded separately. Single-texel tensors can still select textures. FACTO revision: `3b8c778c99766a8b4d0d04563ae0b16cbb276829`, seed 0. The complete native regression run on the previously published head `a2543b8a82` passed **50 tests with one expected SwiftShader-only skip** on MoltenVK, covering all 35 dynamic tests, graph-builder and serialization tests, and PT2E quantized linear with downcasting disabled. The subsequent serializer refinement confines `--force-defaults` to Vulkan and restores the shared `exir` FlatBuffers API. That revised source tree passed **26 focused tests**, including shared/Vulkan serialization, signed-zero execution, and power special values. All other code patches are unchanged, and every replacement PR was checked against the tested source trees. Lintrunner and `git diff --check` passed. NVIDIA, SwiftShader, and Windows CI are pending verification on the new ghstack heads. Recreates #23162 through ghstack. Prior review discussion remains on that PR. Authored with OpenAI Codex; split planned with Claude Code. cc @SS-JIA @manuelcandales @digantdesai @cbilgin ghstack-source-id: 26d2d76 ghstack-comment-id: 5892899249 Pull-Request: #23254

github-actions · 2026-10-03T03:10:34Z

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.

Copilot

Copilot review overview

🟡 Changes recommended

Regression coverage omits integer clamping and key FP16 underflow and overflow boundaries.

Review effort: Balanced
Findings: 2

Open (2)

Add int32 range regression test above half-float limits · New
Test zero/subnormal and max-finite/infinity rounding ties · New

What changed in this PR

Updates Vulkan reductions to match ATen range, NaN, index-selection, and FP16 rounding semantics.

Changes:

Removes erroneous output clamping and fixes mean accumulation.
Propagates NaNs and selects the first NaN for arg-reductions.
Adds exact reduction regression tests.

File	Description
`backends/vulkan/test/test_vulkan_dynamic.py`	Adds reduction conformance tests.
`backends/vulkan/runtime/graph/ops/glsl/reduce.yaml`	Uses NaN-propagating extrema.
`backends/vulkan/runtime/graph/ops/glsl/reduce.glsl`	Adds NaN helpers and FP16 rounding.
`backends/vulkan/runtime/graph/ops/glsl/reduce_op_defs.glslh`	Fixes mean and first-NaN handling.
`backends/vulkan/runtime/graph/ops/glsl/convert.glslh`	Removes clamping and implements nearest-even FP16 conversion.

💡 Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[ghstack-poisoned]

@SS-JIA

…ormer blocks The dynamic transformer reproductions now lower to one Vulkan delegate across changing sequence lengths. This final PR makes `expand_copy` resizable and adds the eager-attention and SDPA integration tests, including fully masked rows on devices with 8-bit storage buffers. Fixes #23156. This is part 15 of a linear ghstack series. Review each PR against its selected base branch; this PR contains only the final expansion and transformer changes. The scalar-tensor issue #23158 is owned by part 10. | Part | PR | Change | | --- | --- | --- | | 1 | #23240 | Hardware Vulkan CI | | 2 | #23241 | Scalar cache type and signed-zero keys | | 3 | #23242 | Vulkan-local signed-zero serialization | | 4 | #23243 | GELU modes and view kwargs | | 5 | #23244 | Reduction clamp, NaN, and FP16 rounding | | 6 | #23245 | Reduction dimension guards | | 7 | #23246 | Bool staging and logical_not | | 8 | #23247 | Scalar representability and symbolic guards | | 9 | #23248 | 64-bit dtype and fusion policy | | 10 | #23249 | scalar_tensor with exact integer values | | 11 | #23250 | Typed, resizable full | | 12 | #23251 | Power special values and logical FP16 dtype | | 13 | #23252 | mul.Scalar | | 14 | #23253 | any.dim | | 15 | #23254 | Dynamic expand and transformer integration | The stack retains the FACTO and ATen conformance fixes and their regression tests: typed integer fills, NaN and signed-zero behavior, reduction range and accumulation fixes, safe scalar fallbacks, and the FP16 power contract. FP16 writes use explicit nearest-even conversion, and power keeps the requested tensor dtype when a device emulates FP16 storage with FP32. The test module is now `test_vulkan_dynamic.py`; CI and Buck references follow the rename. Each code patch was linted and tested before its original publication. Prior validation of the operator changes passed 116 FACTO cases against both ATen and portable kernels, plus 144 boundary cases against ATen, with texture and buffer preferences: all 520 Vulkan-configuration executions matched ATen, including CPU fallback where unsupported. The additional 239-case sweep covers all 91 available FACTO specifications for registered ATen overloads and 57 targeted examples. Exact-input replays of every candidate Vulkan failure on a separately built main at `a31838280f9309f04af1b375147375b702d29339` found no new regressions or unresolved comparisons. Validation uses the shared `backends/test/` harness on Apple M1 Pro / MoltenVK with Release portable CPU kernels and Vulkan; optimized kernels and XNNPACK are disabled. Buffer preference uses `texture_limits=(1,1,1)`, with actual storage and fallback recorded separately. Single-texel tensors can still select textures. FACTO revision: `3b8c778c99766a8b4d0d04563ae0b16cbb276829`, seed 0. The complete native regression run on the previously published head `a2543b8a82` passed **50 tests with one expected SwiftShader-only skip** on MoltenVK, covering all 35 dynamic tests, graph-builder and serialization tests, and PT2E quantized linear with downcasting disabled. The subsequent serializer refinement confines `--force-defaults` to Vulkan and restores the shared `exir` FlatBuffers API. That revised source tree passed **26 focused tests**, including shared/Vulkan serialization, signed-zero execution, and power special values. All other code patches are unchanged, and every replacement PR was checked against the tested source trees. Lintrunner and `git diff --check` passed. NVIDIA, SwiftShader, and Windows CI are pending verification on the new ghstack heads. Recreates #23162 through ghstack. Prior review discussion remains on that PR. Authored with OpenAI Codex; split planned with Claude Code. cc @SS-JIA @manuelcandales @digantdesai @cbilgin ghstack-source-id: 3a6574c ghstack-comment-id: 5892899249 Pull-Request: #23254

Copilot

Copilot review overview

🔵 Needs a closer look

The claimed int32 clamping fix lacks regression coverage.

Review effort: Balanced
Findings: 1

Open (1)

Add int32 range regression test above half-float limits

Resolved since last review (1)

Test zero/subnormal and max-finite/infinity rounding ties

… clamp [ghstack-poisoned]

@SS-JIA

…ormer blocks The dynamic transformer reproductions now lower to one Vulkan delegate across changing sequence lengths. This final PR makes `expand_copy` resizable and adds the eager-attention and SDPA integration tests, including fully masked rows on devices with 8-bit storage buffers. Fixes #23156. This is part 15 of a linear ghstack series. Review each PR against its selected base branch; this PR contains only the final expansion and transformer changes. The scalar-tensor issue #23158 is owned by part 10. | Part | PR | Change | | --- | --- | --- | | 1 | #23240 | Hardware Vulkan CI | | 2 | #23241 | Scalar cache type and signed-zero keys | | 3 | #23242 | Vulkan-local signed-zero serialization | | 4 | #23243 | GELU modes and view kwargs | | 5 | #23244 | Reduction clamp, NaN, and FP16 rounding | | 6 | #23245 | Reduction dimension guards | | 7 | #23246 | Bool staging and logical_not | | 8 | #23247 | Scalar representability and symbolic guards | | 9 | #23248 | 64-bit dtype and fusion policy | | 10 | #23249 | scalar_tensor with exact integer values | | 11 | #23250 | Typed, resizable full | | 12 | #23251 | Power special values and logical FP16 dtype | | 13 | #23252 | mul.Scalar | | 14 | #23253 | any.dim | | 15 | #23254 | Dynamic expand and transformer integration | The stack retains the FACTO and ATen conformance fixes and their regression tests: typed integer fills, NaN and signed-zero behavior, reduction range and accumulation fixes, safe scalar fallbacks, and the FP16 power contract. FP16 writes use explicit nearest-even conversion, and power keeps the requested tensor dtype when a device emulates FP16 storage with FP32. The test module is now `test_vulkan_dynamic.py`; CI and Buck references follow the rename. Each code patch was linted and tested before its original publication. Prior validation of the operator changes passed 116 FACTO cases against both ATen and portable kernels, plus 144 boundary cases against ATen, with texture and buffer preferences: all 520 Vulkan-configuration executions matched ATen, including CPU fallback where unsupported. The additional 239-case sweep covers all 91 available FACTO specifications for registered ATen overloads and 57 targeted examples. Exact-input replays of every candidate Vulkan failure on a separately built main at `a31838280f9309f04af1b375147375b702d29339` found no new regressions or unresolved comparisons. Validation uses the shared `backends/test/` harness on Apple M1 Pro / MoltenVK with Release portable CPU kernels and Vulkan; optimized kernels and XNNPACK are disabled. Buffer preference uses `texture_limits=(1,1,1)`, with actual storage and fallback recorded separately. Single-texel tensors can still select textures. FACTO revision: `3b8c778c99766a8b4d0d04563ae0b16cbb276829`, seed 0. The complete native regression run on the previously published head `a2543b8a82` passed **50 tests with one expected SwiftShader-only skip** on MoltenVK, covering all 35 dynamic tests, graph-builder and serialization tests, and PT2E quantized linear with downcasting disabled. The subsequent serializer refinement confines `--force-defaults` to Vulkan and restores the shared `exir` FlatBuffers API. That revised source tree passed **26 focused tests**, including shared/Vulkan serialization, signed-zero execution, and power special values. All other code patches are unchanged, and every replacement PR was checked against the tested source trees. Lintrunner and `git diff --check` passed. NVIDIA, SwiftShader, and Windows CI are pending verification on the new ghstack heads. Recreates #23162 through ghstack. Prior review discussion remains on that PR. Authored with OpenAI Codex; split planned with Claude Code. cc @SS-JIA @manuelcandales @digantdesai @cbilgin ghstack-source-id: c9b5269 ghstack-comment-id: 5892899249 Pull-Request: #23254

Copilot

Copilot review overview

🔵 Needs a closer look

Cross-driver IEEE-754 shader behavior warrants final human review despite comprehensive targeted tests.

Review effort: Balanced
Findings: None

Resolved since last review (2)

Test zero/subnormal and max-finite/infinity rounding ties
Add int32 range regression test above half-float limits

@SS-JIA

…ormer blocks The dynamic transformer reproductions now lower to one Vulkan delegate across changing sequence lengths. This final PR makes `expand_copy` resizable and adds the eager-attention and SDPA integration tests, including fully masked rows on devices with 8-bit storage buffers. Fixes #23156. This is part 15 of the original ghstack series. Parts 1–4 (#23240 through #23243) have landed in main; the remaining eleven PRs run from #23244 through #23254. Review each PR against its selected base branch. This PR contains the final expansion and transformer changes; part 10 owns the scalar-tensor issue #23158. | Part | PR | Change | | --- | --- | --- | | 1 | #23240 (landed) | Hardware Vulkan CI | | 2 | #23241 (landed) | Scalar cache type and signed-zero keys | | 3 | #23242 (landed) | Vulkan-local signed-zero serialization | | 4 | #23243 (landed) | GELU modes and view kwargs | | 5 | #23244 | Reduction clamp, NaN, and FP16 rounding | | 6 | #23245 | Reduction and arg-reduction dimension/storage guards | | 7 | #23246 | Bool staging and logical_not | | 8 | #23247 | Scalar representability and symbolic guards | | 9 | #23248 | 64-bit dtype and fusion policy | | 10 | #23249 | scalar_tensor with exact integer values | | 11 | #23250 | Typed, resizable full | | 12 | #23251 | Power special values and logical FP16 dtype | | 13 | #23252 | mul.Scalar | | 14 | #23253 | any.dim with keepdim=True | | 15 | #23254 | Dynamic expand and transformer integration | The stack retains the FACTO and ATen conformance fixes and their regression tests: typed integer fills, NaN and signed-zero behavior, reduction range and accumulation fixes, safe scalar fallbacks, and the FP16 power contract. On native FP16 devices, the scalar-tensor, full texture, single-dimension texture reduction, and binary scalar shaders round their FP16 outputs to nearest-even. The binary scalar shaders also preserve the requested FP16 dtype when storage is emulated with FP32. Other operators retain their existing FP32 intermediate behavior on those devices. PR4 introduces `test_vulkan_dynamic.py` and its CI/Buck references. Each code patch was linted and tested before its original publication. Prior validation of the operator changes passed 116 FACTO cases against both ATen and portable kernels, plus 144 boundary cases against ATen, with texture and buffer preferences: all 520 Vulkan-configuration executions matched ATen, including CPU fallback where unsupported. The additional 239-case sweep covers all 91 available FACTO specifications for registered ATen overloads and 57 targeted examples. Exact-input replays of every candidate Vulkan failure on a separately built main at `a31838280f9309f04af1b375147375b702d29339` found no new regressions or unresolved comparisons. Validation uses the shared `backends/test/` harness on Apple M1 Pro / MoltenVK with Release portable CPU kernels and Vulkan; optimized kernels and XNNPACK are disabled. Buffer preference uses `texture_limits=(1,1,1)`, with actual storage and fallback recorded separately. Single-texel tensors can still select textures. FACTO revision: `3b8c778c99766a8b4d0d04563ae0b16cbb276829`, seed 0. Validation: Rebased onto main at `2c8510343e47` after parts 1–4 landed. `git range-diff` confirms all eleven remaining patches are unchanged by this rebase. On the previous base (`0b3d26d8c1f9`), the Release Vulkan/portable runtime rebuilt successfully and the native suite passed **51 tests with one expected SwiftShader-only skip**, before subsequent review tests were added. Coverage includes the dynamic tests, graph builder, serialization, and PT2E quantized linear with downcasting disabled. The two FP16 texture rounding tests pass 18 exact ATen cases covering both signs, signed zero, even/odd ties, exponent carry, the normal/subnormal and zero/subnormal boundaries, and values around the ±65520 overflow threshold. A separate test directly lowers int32 amax/amin to Vulkan buffers and checks exact ATen agreement beyond ±65504; the same inputs also match portable CPU execution. The arg-reduction guard update then passed four focused native tests, including 28 argmax/argmin cases checking Vulkan buffer execution and portable CPU fallback against ATen. The final rebase preserved these tested Vulkan sources; only upstream CI configuration and documentation changed. Lintrunner and `git diff --check` passed. Delegation of `any.dim` requires `keepdim=True`. Omitting `keepdim` or setting it to `False` keeps the reduction on CPU, avoiding a new bool-buffer requirement for surrounding texture delegates. The default/explicit-false regression test verifies texture storage and exact ATen results across changing shapes. Five focused native tests pass on MoltenVK, including existing `keepdim=True` execution and the full dynamic transformer tests; lintrunner and `git diff --check` also pass. SwiftShader CI will verify the updated heads. Vulkan CI will run again on the rebased ghstack heads. Recreates #23162 through ghstack. Prior review discussion remains on that PR. Regression coverage includes scalar exponents 2.0001 and 2049 and a two-node FP16 mul.Scalar chain. The power and chained-multiplication tests compare with ATen in both texture and buffer storage. Authored with OpenAI Codex; split planned with Claude Code. cc @SS-JIA @manuelcandales @digantdesai @cbilgin ghstack-source-id: afcd265 ghstack-comment-id: 5892899249 Pull-Request: #23254

mergennachin added 5 commits September 29, 2026 11:02

[INITIAL] Recreate the Vulkan transformer and conformance stack with …

aae8159

…ghstack [ghstack-poisoned]

[INITIAL] Recreate the Vulkan transformer and conformance stack with …

bcab9d0

…ghstack [ghstack-poisoned]

[INITIAL] Recreate the Vulkan transformer and conformance stack with …

d6db604

…ghstack [ghstack-poisoned]

[INITIAL] Recreate the Vulkan transformer and conformance stack with …

446a5d6

…ghstack [ghstack-poisoned]

[INITIAL] Recreate the Vulkan transformer and conformance stack with …

fb905cf

…ghstack [ghstack-poisoned]

mergennachin added the module: vulkan Issues related to the Vulkan delegate and code under backends/vulkan/ label Sep 29, 2026

mergennachin deployed to cadence September 29, 2026 15:02 — with GitHub Actions Active

meta-cla Bot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Sep 29, 2026

mergennachin mentioned this pull request Sep 29, 2026

[Vulkan] Stop clamping per-row reductions and propagate NaN #23205

Closed

mergennachin deployed to cadence September 29, 2026 15:21 — with GitHub Actions Active

SS-JIA reviewed Oct 2, 2026

View reviewed changes

Comment thread backends/vulkan/runtime/graph/ops/glsl/convert.glslh

mergennachin deployed to cadence October 3, 2026 02:55 — with GitHub Actions Active

Base automatically changed from gh/mergennachin/27/head to main October 3, 2026 02:58

[UPDATE] Match arg-reduction partitioning to the buffer-only last-dim…

a231002

…ension kernel; rebase after PR 23243 landed [ghstack-poisoned]

mergennachin deployed to cadence October 3, 2026 03:10 — with GitHub Actions Active

mergennachin marked this pull request as ready for review October 3, 2026 03:15

Copilot AI balanced review requested due to automatic review settings October 3, 2026 03:15

Copilot started reviewing on behalf of mergennachin October 3, 2026 03:15 View session

Copilot AI reviewed Oct 3, 2026

View reviewed changes

Comment thread backends/vulkan/test/test_vulkan_dynamic.py

Comment thread backends/vulkan/test/test_vulkan_dynamic.py

mergennachin deployed to cadence October 3, 2026 03:28 — with GitHub Actions Active

[UPDATE] Add exact FP16 zero/subnormal and overflow-threshold coverage

7baf82c

[ghstack-poisoned]

Copilot AI balanced review requested due to automatic review settings October 3, 2026 03:46

mergennachin deployed to cadence October 3, 2026 03:47 — with GitHub Actions Active

Copilot started reviewing on behalf of mergennachin October 3, 2026 03:47 View session

Copilot AI reviewed Oct 3, 2026

View reviewed changes

[UPDATE] Exercise int32 reduction shaders directly beyond the removed…

1ad69df

… clamp [ghstack-poisoned]

Copilot AI balanced review requested due to automatic review settings October 3, 2026 03:59

Copilot started reviewing on behalf of mergennachin October 3, 2026 03:59 View session

mergennachin deployed to cadence October 3, 2026 04:00 — with GitHub Actions Active

Copilot AI reviewed Oct 3, 2026

View reviewed changes

mergennachin deployed to cadence October 3, 2026 04:19 — with GitHub Actions Active

Navigation Menu

[Vulkan] Stop clamping per-row reductions and propagate NaN - #23244

[Vulkan] Stop clamping per-row reductions and propagate NaN#23244
mergennachin wants to merge 11 commits into
mainfrom
gh/mergennachin/28/head

mergennachin commented Sep 29, 2026 •

edited

Loading

Uh oh!

mergennachin commented Sep 29, 2026 •

edited

Loading

Uh oh!

pytorch-bot Bot commented Sep 29, 2026 •

edited

Loading

Uh oh!

Uh oh!

github-actions Bot commented Oct 3, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI left a comment

Uh oh!

This branch was successfully deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

mergennachin commented Sep 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

mergennachin commented Sep 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

pytorch-bot Bot commented Sep 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/23244

✅ No Failures

Uh oh!

Uh oh!

github-actions Bot commented Oct 3, 2026

This PR needs a release notes: label

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Copilot review overview

🟡 Changes recommended

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Copilot review overview

🔵 Needs a closer look

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Copilot review overview

🔵 Needs a closer look

Uh oh!

This branch was successfully deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

mergennachin commented Sep 29, 2026 •

edited

Loading

mergennachin commented Sep 29, 2026 •

edited

Loading

pytorch-bot Bot commented Sep 29, 2026 •

edited

Loading

This PR needs a `release notes:` label