Stack from [ghstack](https://github.com/ezyang/ghstack/tree/0.15.0)
(oldest at bottom):
* #20465
* #20464
* #20463
* #20435
* #20399
* #20398
* #20397
* #20396
* #20395
* #20394
* #20393
* #20392
* #20391
* #20390
* #20363
* __->__ #20362
* #20361
* #20360
* #20359



Adds `aten.select_copy.int` to the WebGPU delegate as a gather: picks a
fixed index along one dim, producing an output of rank (input rank - 1).

Composition (single dispatch):
- `select/Select.cpp` — reads `[self, dim, index, out]` (static `Int`
via `read_scalar`; throws on dynamic `SymInt`), normalizes +
bounds-checks dim/index, builds 2 `TensorMeta` UBOs + a
`SelectParams{dim,index}`, fp32 guard, 1D-dispatch over `numel`,
releases uniforms after the bind group.
- `select/select.wgsl` — seeds the input offset with `index *
in.strides[dim]`, delinearizes the output index, maps each out dim to
its in dim (shifted past the selected dim), relinearizes on input
strides.
@exported-using-ghexport

Differential Revision:
[D108793166](https://our.internmc.facebook.com/intern/diff/D108793166/)

Differential Revision:
[D108793166](https://our.internmc.facebook.com/intern/diff/D108793166)

…ework) (#20363)

Stack from [ghstack](https://github.com/ezyang/ghstack/tree/0.15.0)
(oldest at bottom):
* #20465
* #20464
* #20463
* #20435
* #20399
* #20398
* #20397
* #20396
* #20395
* #20394
* #20393
* #20392
* #20391
* #20390
* __->__ #20363
* #20362
* #20361
* #20360
* #20359



Registers `aten.select_copy.int` in the `cases.py` op-test framework: a
`_select_suite` of 4 configs (leading/middle/last dim + negative index)
that `generate_op_tests` exports and compares to a torch golden on Dawn.
Also adds `test/ops/select/test_select.py` (`SelectModule` + `CONFIGS` +
an export-delegation/eager smoke test) and the `aten.select_copy.int`
partitioner-allowlist entry in `tester.py`.
@exported-using-ghexport

Differential Revision:
[D108793161](https://our.internmc.facebook.com/intern/diff/D108793161/)

Differential Revision:
[D108793161](https://our.internmc.facebook.com/intern/diff/D108793161)

Stack from [ghstack](https://github.com/ezyang/ghstack/tree/0.15.0)
(oldest at bottom):
* #20465
* #20464
* #20463
* #20435
* #20399
* #20398
* #20397
* #20396
* #20395
* #20394
* #20393
* #20392
* #20391
* __->__ #20390
* #20363
* #20362
* #20361
* #20360
* #20359



Adds `aten.sigmoid.default` to the WebGPU delegate: element-wise
`1/(1+exp(-x))` over a flat fp32 buffer. On the Llama critical path
(`F.silu` -> `sigmoid` + `mul`).

Composition (single dispatch):
- `sigmoid/UnaryOp.cpp` — binds input (storage, read-only) + output
(storage) + a `Params{num_elements}` uniform, 1D-dispatches over
`num_elements` with `override wg_size` (clamped to the device limit);
mirrors the `add` op (uniform mapped-at-creation, released after the
bind group).
- `sigmoid/sigmoid.wgsl` — guards `idx >= num_elements` and writes the
logistic of each element.
@exported-using-ghexport

Differential Revision:
[D108793157](https://our.internmc.facebook.com/intern/diff/D108793157/)

Differential Revision:
[D108793157](https://our.internmc.facebook.com/intern/diff/D108793157)

…k) (#20391)

Stack from [ghstack](https://github.com/ezyang/ghstack/tree/0.15.0)
(oldest at bottom):
* #20465
* #20464
* #20463
* #20435
* #20399
* #20398
* #20397
* #20396
* #20395
* #20394
* #20393
* #20392
* __->__ #20391
* #20390
* #20363
* #20362
* #20361
* #20360
* #20359



Registers `aten.sigmoid.default` in the `cases.py` op-test framework: a
`_sigmoid_suite` (hard-coded shapes + a saturation case over a
`linspace(-12, 12)` input) that `generate_op_tests` exports and compares
to an fp64 torch golden on Dawn. Also adds
`test/ops/sigmoid/test_sigmoid.py` (`SigmoidModule` + `N` + `_det_input`
+ an export-delegation/eager smoke test) and the `aten.sigmoid.default`
partitioner-allowlist entry in `tester.py`.
@exported-using-ghexport

Differential Revision:
[D108793159](https://our.internmc.facebook.com/intern/diff/D108793159/)

Differential Revision:
[D108793159](https://our.internmc.facebook.com/intern/diff/D108793159)

…20392)

Stack from [ghstack](https://github.com/ezyang/ghstack/tree/0.15.0)
(oldest at bottom):
* #20465
* #20464
* #20463
* #20435
* #20399
* #20398
* #20397
* #20396
* #20395
* #20394
* #20393
* __->__ #20392
* #20391
* #20390
* #20363
* #20362
* #20361
* #20360
* #20359



Adds `aten.squeeze_copy.dims` and `aten.unsqueeze_copy.default` to the
WebGPU delegate. Both are numel-preserving shape ops; on a dense
row-major buffer backend they are the same flat copy as `view_copy` —
only the shape metadata differs (mirrors the Vulkan delegate, which
routes both through `add_view_copy_node`).

Composition (no new kernel):
- `squeeze/Squeeze.cpp` — reads `args = [self, dims, out]`, ignores the
AOT-fixed `dims`, calls `add_flat_copy(graph, in, out)` from
`runtime/ops/view_copy/view_copy.h`.
- `unsqueeze/Unsqueeze.cpp` — reads `args = [self, dim, out]`, ignores
the AOT-fixed `dim`, calls `add_flat_copy(graph, in, out)`.
@exported-using-ghexport

Differential Revision:
[D108793153](https://our.internmc.facebook.com/intern/diff/D108793153/)

Differential Revision:
[D108793153](https://our.internmc.facebook.com/intern/diff/D108793153)

….py op-test framework) (#20393)

Stack from [ghstack](https://github.com/ezyang/ghstack/tree/0.15.0)
(oldest at bottom):
* #20465
* #20464
* #20463
* #20435
* #20399
* #20398
* #20397
* #20396
* #20395
* #20394
* __->__ #20393
* #20392
* #20391
* #20390
* #20363
* #20362
* #20361
* #20360
* #20359



Registers `aten.squeeze_copy.dims` and `aten.unsqueeze_copy.default` in
the `cases.py` op-test framework: a `_squeeze_suite` of 3 configs
(squeeze leading/middle/multiple size-1 dims) and a `_unsqueeze_suite`
of 3 configs (insert dim at front/middle/last) that `generate_op_tests`
exports via `VulkanPartitioner` and compares to a torch golden on Dawn.
Also adds `test/ops/squeeze/test_squeeze.py` (`SqueezeModule` +
`CONFIGS` + `_op_delegated` smoke test),
`test/ops/unsqueeze/test_unsqueeze.py` (`UnsqueezeModule` + `CONFIGS` +
`_op_delegated` smoke test), and the two partitioner-allowlist entries
in `tester.py`.
@exported-using-ghexport

Differential Revision:
[D108793152](https://our.internmc.facebook.com/intern/diff/D108793152/)

Differential Revision:
[D108793152](https://our.internmc.facebook.com/intern/diff/D108793152)

Stack from [ghstack](https://github.com/ezyang/ghstack/tree/0.15.0)
(oldest at bottom):
* #20465
* #20464
* #20463
* #20435
* #20399
* #20398
* #20397
* #20396
* #20395
* __->__ #20394
* #20393
* #20392
* #20391
* #20390
* #20363
* #20362
* #20361
* #20360
* #20359



Adds `aten.slice_copy.Tensor` to the WebGPU delegate as a gather: each
output element is mapped back to its source input element along the
sliced dim via `start + coord * step`.

Composition (single compute dispatch):
- `runtime/ops/slice/Slice.cpp` — reads `args = [self, dim, start, end,
step, out]` via `read_scalar` (static `Int`/`Null`-sentinel default;
throws on dynamic `SymInt`); normalizes negative `dim`/`start`, clamps
`start` to `[0, in_size]`; builds two `TensorMeta` UBOs + a
`SliceParams{dim, start, step}` uniform; guards fp32; dispatches over
`compute_1d_workgroup_count(out.numel)` with `override wg_size`;
releases all uniforms after the bind group.
- `runtime/ops/slice/slice.wgsl` — delinearizes the output index over
the contiguous output strides, maps the sliced-dim coordinate back to
the input (`start + coord*step`), relinearizes over the input strides.
@exported-using-ghexport

Differential Revision:
[D108793168](https://our.internmc.facebook.com/intern/diff/D108793168/)

Differential Revision:
[D108793168](https://our.internmc.facebook.com/intern/diff/D108793168)

…work) (#20395)

Stack from [ghstack](https://github.com/ezyang/ghstack/tree/0.15.0)
(oldest at bottom):
* #20465
* #20464
* #20463
* #20435
* #20399
* #20398
* #20397
* #20396
* __->__ #20395
* #20394
* #20393
* #20392
* #20391
* #20390
* #20363
* #20362
* #20361
* #20360
* #20359



Registers `aten.slice_copy.Tensor` in the `cases.py` op-test framework:
a `_slice_suite` of 4 configs (leading-dim slice `[:,1:5]`, last-dim
slice `[...,1:3]`, step-2 `[:,0:8:2]`, negative-end `[:,1:-1]`) that
`generate_op_tests` exports via `VulkanPartitioner` and compares to a
torch golden on Dawn. Also adds `test/ops/slice/test_slice.py`
(`SliceModule` + `CONFIGS` + export-delegation/eager smoke test) and the
`aten.slice_copy.Tensor` partitioner-allowlist entry in `tester.py`.
@exported-using-ghexport

Differential Revision:
[D108793151](https://our.internmc.facebook.com/intern/diff/D108793151/)

Differential Revision:
[D108793151](https://our.internmc.facebook.com/intern/diff/D108793151)

…ermute_copy.default) (#20396)

Stack from [ghstack](https://github.com/ezyang/ghstack/tree/0.15.0)
(oldest at bottom):
* #20465
* #20464
* #20463
* #20435
* #20399
* #20398
* #20397
* __->__ #20396
* #20395
* #20394
* #20393
* #20392
* #20391
* #20390
* #20363
* #20362
* #20361
* #20360
* #20359



Adds `aten.permute_copy.default` (a coordinate-reorder gather) to the
WebGPU delegate, and the `IntList` graph value type it needs to read its
`dims` argument.

Composition:
- `runtime/WebGPUGraph.{h,cpp}` — adds `ValueType::IntList` backed by
`std::vector<std::vector<int64_t>> int_lists_` + `get_int_list(int)`;
`build()` deserializes `vkgraph::GraphTypes::IntList` via
`value_as_IntList()->items()` (int64, matching the FlatBuffer `[long]`);
mirrors the existing scalar value plumbing.
- `runtime/ops/permute/Permute.cpp` — reads the permutation via
`get_int_list`, normalizes negative dims, validates it is a permutation
of `[0, ndim)`, builds two `TensorMeta` UBOs + a `PermuteParams{perm:
vec4<u32>}` uniform, guards fp32 + rank≤4, dispatches over
`compute_1d_workgroup_count(out.numel)` with `override wg_size`;
releases all uniforms after the bind group.
- `runtime/ops/permute/permute.wgsl` — delinearizes the output index
over the contiguous output strides, reads `input` at
`in.strides[perm[d]]` per dim (mirrors Vulkan `permute_buffer.glsl`).
- Registers both `aten.permute_copy.default` and `aten.permute.default`
to the same handler.
@exported-using-ghexport

Differential Revision:
[D108793162](https://our.internmc.facebook.com/intern/diff/D108793162/)

Differential Revision:
[D108793162](https://our.internmc.facebook.com/intern/diff/D108793162)

…mework) (#20397)

Stack from [ghstack](https://github.com/ezyang/ghstack/tree/0.15.0)
(oldest at bottom):
* #20465
* #20464
* #20463
* #20435
* #20399
* #20398
* __->__ #20397
* #20396
* #20395
* #20394
* #20393
* #20392
* #20391
* #20390
* #20363
* #20362
* #20361
* #20360
* #20359



Registers `aten.permute_copy.default` in the `cases.py` op-test
framework: a `_permute_suite` of 4 configs (3D rotation, 4D middle-dim
transpose, 2D transpose, full 4D shuffle) that `generate_op_tests`
exports via `VulkanPartitioner` and compares to a torch golden on Dawn.
Also adds `test/ops/permute/test_permute.py` (`PermuteModule` +
`CONFIGS` + `_op_delegated` smoke test) and the
`aten.permute_copy.default` partitioner-allowlist entry in `tester.py`.
@exported-using-ghexport

Differential Revision:
[D108793156](https://our.internmc.facebook.com/intern/diff/D108793156/)

Differential Revision:
[D108793156](https://our.internmc.facebook.com/intern/diff/D108793156)

…efault) (#20398)

Stack from [ghstack](https://github.com/ezyang/ghstack/tree/0.15.0)
(oldest at bottom):
* #20465
* #20464
* #20463
* #20435
* #20399
* __->__ #20398



Adds `aten.cat.default` to the WebGPU delegate as an index-math scatter,
and the `ValueList` graph value type it needs to read its tensor-list
argument.

Composition (one dispatch per input):
- `runtime/WebGPUGraph.{h,cpp}` — adds `ValueType::ValueList` backed by
`std::vector<std::vector<int>> value_lists_` + `get_value_list(int)`;
`build()` deserializes `vkgraph::GraphTypes::ValueList` via
`value_as_ValueList()->items()`; mirrors the existing scalar value
plumbing.
- `runtime/ops/cat/Cat.cpp` — reads the input-tensor list via
`get_value_list`, reads the static `Int` dim, validates each input rank
+ non-concat dims against the output; builds one shared `out_meta`
`TensorMeta` uniform + a shared bind-group/pipeline layout; per input
builds a fresh pipeline + bind group with `in_meta` +
`CatParams{concat_dim, off_k}`, dispatches over
`compute_1d_workgroup_count(in.numel)`; releases per-input uniforms
after each bind group, the shared `out_meta` after the loop.
- `runtime/ops/cat/cat.wgsl` — delinearizes the input index over the
input strides, shifts the concat-dim coordinate by the host-computed
`off_k` (running sum of prior input sizes), relinearizes over the output
strides to scatter `output[out] = input[in]`.
- The N disjoint-slab writes share one output buffer across separate
dispatches and rely on the existing per-pass `execute()` ordering.
@exported-using-ghexport

Differential Revision:
[D108793165](https://our.internmc.facebook.com/intern/diff/D108793165/)

Differential Revision:
[D108793165](https://our.internmc.facebook.com/intern/diff/D108793165)

…20399)

Stack from [ghstack](https://github.com/ezyang/ghstack/tree/0.15.0)
(oldest at bottom):
* #20465
* #20464
* #20463
* #20435
* __->__ #20399
* #20398



Registers `aten.cat.default` in the `cases.py` op-test framework: a
`_cat_suite` of 4 configs (3 inputs along dim 0, 2 inputs along dim 1, 3
inputs along dim 2, uneven split along dim 1) that `generate_op_tests`
exports via `VulkanPartitioner` and compares to a torch golden on Dawn.
Also adds `test/ops/cat/test_cat.py` (`CatModule` + `CONFIGS` +
`_op_delegated` smoke test, with distinct per-input value ranges to
catch cross-slab contamination) and the `aten.cat.default`
partitioner-allowlist entry in `tester.py`.
@exported-using-ghexport

Differential Revision:
[D108793163](https://our.internmc.facebook.com/intern/diff/D108793163/)

Differential Revision:
[D108793163](https://our.internmc.facebook.com/intern/diff/D108793163)

…>.py (#20435)

Stack from [ghstack](https://github.com/ezyang/ghstack/tree/0.15.0)
(oldest at bottom):
* #20465
* #20464
* #20463
* __->__ #20435
* #20399
* #20398



Pure test-file relocation: moves the already-landed ops' tests from
nested `test/ops/<op>/test_<op>.py` to flat `test/ops/test_<op>.py`,
matching the ExecuTorch convention (XNNPACK uses flat
`test/ops/test_<op>.py`; Vulkan uses flat `test/test_*.py`) and
completing the flatten applied to the new ops in the stack below. Drops
the per-op `__init__.py`; the parent `test/ops/__init__.py` is kept.
Ops: `add`, `rms_norm`, `sdpa` (`test_sdpa` + `test_update_cache`),
`dispatch_order`, `quantized_linear`, `embedding_q4gsw`, `rope`,
`prepack`.

No behavior change — the test modules and their export/golden functions
are unchanged; only their path moves. Every reference to the old paths
is updated: the `cases.py` op-test imports (`add`, `rms_norm`),
`test/TARGETS` (`test_add` srcs), `test/ops/test_dispatch_order.py`'s
internal `rms_norm` import, and the build/CI scripts that import the
per-op export functions (`test/test_build_webgpu.sh`,
`scripts/test_webgpu_native_ci.sh`).

Nothing required the per-op subdirectory: the codegen framework imports
only `cases.py`, the one buck target uses a literal path, and the
native-golden scripts import the modules by path — each resolves
identically at the flat path.
@exported-using-ghexport

Differential Revision:
[D109349894](https://our.internmc.facebook.com/intern/diff/D109349894/)

Differential Revision:
[D109349894](https://our.internmc.facebook.com/intern/diff/D109349894)

Stack from [ghstack](https://github.com/ezyang/ghstack/tree/0.15.0)
(oldest at bottom):
* #20465
* __->__ #20464
* #20463
* #20435
* #20399
* #20398



Adds the WebGPU delegate handler for aten.index.Tensor, the 1D-self
advanced-index
gather out[i] = self[index[i]] (output shape == index shape). This is
the form the
VulkanPartitioner delegates -- it requires a 1D self and exactly one
non-None index
(op_registry.py); 2D mask/freqs gathers stay on CPU. It mirrors the
Vulkan delegate's
index_tensor op (IndexTensor.cpp + index_tensor_buffer.glsl) as a single
compute
dispatch over the output elements, each reading the int32 index and
gathering the
corresponding fp32 self element.

The op is composed as:
- index.wgsl: one workgroup-strided pass, out[i] = self[u32(index[i])],
guarded by a
numel bound; buffer-only, fp32 self/out, int32 index, 1D dispatch via
the shared
  WebGPUUtils helpers (clamp workgroup size + 1D count).
- Index.cpp: validates the args (self/out tensors; indices ValueList
with exactly one
index tensor; fp32 self/out; int32 index; out numel == index numel),
failing loud on
any violation, then records the dispatch. row_width is dropped (always 1
for 1D self).
@exported-using-ghexport

Differential Revision:
[D109478967](https://our.internmc.facebook.com/intern/diff/D109478967/)

Differential Revision:
[D109478967](https://our.internmc.facebook.com/intern/diff/D109478967)

…lden) (#20465)

Stack from [ghstack](https://github.com/ezyang/ghstack/tree/0.15.0)
(oldest at bottom):
* __->__ #20465
* #20464
* #20463
* #20435
* #20399
* #20398



Adds the test suite for the aten.index.Tensor op (stacked on the op
diff):
- test/ops/index/test_index.py: exports a module computing x[idx]
through
VulkanPartitioner for four configs (reorder/repeat indices over distinct
self
values, so a wrong-gather is visible), asserts a VulkanBackend delegate
with
index.Tensor absorbed (not a CPU fallback), and writes per-config .pte +
  .self/.idx/.golden.bin.
- test/native/test_index.cpp: a standalone Dawn binary that loads each
.pte, feeds
self (fp32) + index (int64 at the program boundary, narrowed to the
int32 buffer)
and compares the gather against the torch golden at 1e-3, with a
single-output
  shape guard.
- Wired into CMake (webgpu_index_test), test/TARGETS (python_unittest
test_index),
  and the Dawn native CI script.
@exported-using-ghexport

Differential Revision:
[D109479000](https://our.internmc.facebook.com/intern/diff/D109479000/)

Differential Revision:
[D109479000](https://our.internmc.facebook.com/intern/diff/D109479000)

Pull Request resolved: #20463

`aten.clone.default` is a pure flat copy on the buffer-only WebGPU backend, identical to `view_copy`: `clone_impl` reuses the existing `add_flat_copy` helper (`output[i] = input[i]`) and registers a handler under `aten.clone.default`. No new shader, generated WGSL header, or CMake source — it shares the `view_copy` flat-copy compute pipeline. Required for end-to-end Llama 3.2 1B (4-bit, KV cache): the exported model serializes 2 `aten.clone.default` ops into its runtime operator chain (the RoPE-frequency clones reused across all 16 transformer layers), so without a handler the partition graph-breaks at those nodes. Mirrors the Vulkan delegate, which registers the same op and routes a buffer clone to a flat view-copy.
ghstack-source-id: 397534700
@exported-using-ghexport

@diff-train-skip-merge

Differential Revision: [D109477717](https://our.internmc.facebook.com/intern/diff/D109477717/)

Pull Request resolved: #20463

`aten.clone.default` is a pure flat copy on the buffer-only WebGPU backend, identical to `view_copy`: `clone_impl` reuses the existing `add_flat_copy` helper (`output[i] = input[i]`) and registers a handler under `aten.clone.default`. No new shader, generated WGSL header, or CMake source — it shares the `view_copy` flat-copy compute pipeline. Required for end-to-end Llama 3.2 1B (4-bit, KV cache): the exported model serializes 2 `aten.clone.default` ops into its runtime operator chain (the RoPE-frequency clones reused across all 16 transformer layers), so without a handler the partition graph-breaks at those nodes. Mirrors the Vulkan delegate, which registers the same op and routes a buffer clone to a flat view-copy.
ghstack-source-id: 397534700
@exported-using-ghexport

@diff-train-skip-merge

Differential Revision: [D109477717](https://our.internmc.facebook.com/intern/diff/D109477717/)

Pull Request resolved: #20463

`aten.clone.default` is a pure flat copy on the buffer-only WebGPU backend, identical to `view_copy`: `clone_impl` reuses the existing `add_flat_copy` helper (`output[i] = input[i]`) and registers a handler under `aten.clone.default`. No new shader, generated WGSL header, or CMake source — it shares the `view_copy` flat-copy compute pipeline. Required for end-to-end Llama 3.2 1B (4-bit, KV cache): the exported model serializes 2 `aten.clone.default` ops into its runtime operator chain (the RoPE-frequency clones reused across all 16 transformer layers), so without a handler the partition graph-breaks at those nodes. Mirrors the Vulkan delegate, which registers the same op and routes a buffer clone to a flat view-copy.
ghstack-source-id: 397534700
@exported-using-ghexport

@diff-train-skip-merge

Differential Revision: [D109477717](https://our.internmc.facebook.com/intern/diff/D109477717/)