[ExecuTorch][WebGPU] Port native_group_norm#21173
Merged
Merged
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/21173
Note: Links to docs will display an error until the docs builds have been completed. ❌ 37 New Failures, 11 Unrelated FailuresAs of commit 2c4b7cf with merge base 4a26c64 ( NEW FAILURES - The following jobs have failed:
FLAKY - The following jobs failed but were likely due to flakiness present on trunk:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This PR needs a
|
meta-cla
Bot
added
the
CLA Signed
This was referenced Jul 22, 2026
psiddh
approved these changes
Jul 23, 2026
psiddh
left a comment
psiddh
left a comment
Contributor
There was a problem hiding this comment.
Approving full WebGPU stack
meta-codesync
Bot
merged commit Jul 24, 2026
c74813f
into
gh/JCNTH/133/base
140 of 197 checks passed
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21173 Ports `aten.native_group_norm.default` as a 2-pass, 3-output group normalization, the second Phase B (vision) op. Pass 1 reduces per group to `mean`/`rstd`; pass 2 applies the per-channel affine. Key changes: - `runtime/ops/native_group_norm/{GroupNorm.cpp, group_norm_reduce.wgsl, group_norm.wgsl}` (+ generated `_wgsl.h`) — reduce: one thread per `(n, group)` serially scans that group's contiguous NCHW block (`base = n*C*HxW + g*(C/G)*HxW`, `group_size = (C/G)*HxW` elements), writing `mean = s/count`, `rstd = inverseSqrt(ss/count - mean^2 + eps)`; normalize: one thread per element, `g = c / (C/G)`, `out = (x - mean[n*G+g]) * rstd[n*G+g] * weight[c] + bias[c]`. A file-local `add_gn_dispatch` helper builds each of the two dispatches from an explicit binding list sharing one params UBO. Registered `aten.native_group_norm.default`. - `CMakeLists.txt` — `runtime/ops/native_group_norm/GroupNorm.cpp` in `WEBGPU_SRCS`. Mirrors Vulkan `impl/GroupNorm.cpp` (2 nodes: reduce -> mean/rstd, then normalize) + `glsl/group_norm_reduce_texture.glsl` (`mean=sum/count`, `variance=sumsq/count - mean^2`, `rstd=1/sqrt(var+eps)`). Buffer NCHW re-derivation (Vulkan is texture/channels-packed), so a serial per-group reduce (a group's channels x HxW are contiguous in NCHW) replaces the cooperative texture reduction. All 3 outputs are real. The inter-pass RAW ordering (normalize reads the reduce's mean/rstd) is guaranteed because `execute()` runs one compute pass per dispatch (`WebGPUGraph.cpp`; proven by `test_dispatch_order.cpp` to 1M elems). A dynamic-shape resize hook recomputes params + both dispatch counts + propagates cur_dims for out/mean/rstd. Fail-loud guards: 4D, `C % group == 0`, fp32, weight/bias len `== C`, mean/rstd size `== N*group`. ghstack-source-id: 406360859 @exported-using-ghexport Differential Revision: [D112257596](https://our.internmc.facebook.com/intern/diff/D112257596/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21173 Ports `aten.native_group_norm.default` as a 2-pass, 3-output group normalization, the second Phase B (vision) op. Pass 1 reduces per group to `mean`/`rstd`; pass 2 applies the per-channel affine. Key changes: - `runtime/ops/native_group_norm/{GroupNorm.cpp, group_norm_reduce.wgsl, group_norm.wgsl}` (+ generated `_wgsl.h`) — reduce: one thread per `(n, group)` serially scans that group's contiguous NCHW block (`base = n*C*HxW + g*(C/G)*HxW`, `group_size = (C/G)*HxW` elements), writing `mean = s/count`, `rstd = inverseSqrt(ss/count - mean^2 + eps)`; normalize: one thread per element, `g = c / (C/G)`, `out = (x - mean[n*G+g]) * rstd[n*G+g] * weight[c] + bias[c]`. A file-local `add_gn_dispatch` helper builds each of the two dispatches from an explicit binding list sharing one params UBO. Registered `aten.native_group_norm.default`. - `CMakeLists.txt` — `runtime/ops/native_group_norm/GroupNorm.cpp` in `WEBGPU_SRCS`. Mirrors Vulkan `impl/GroupNorm.cpp` (2 nodes: reduce -> mean/rstd, then normalize) + `glsl/group_norm_reduce_texture.glsl` (`mean=sum/count`, `variance=sumsq/count - mean^2`, `rstd=1/sqrt(var+eps)`). Buffer NCHW re-derivation (Vulkan is texture/channels-packed), so a serial per-group reduce (a group's channels x HxW are contiguous in NCHW) replaces the cooperative texture reduction. All 3 outputs are real. The inter-pass RAW ordering (normalize reads the reduce's mean/rstd) is guaranteed because `execute()` runs one compute pass per dispatch (`WebGPUGraph.cpp`; proven by `test_dispatch_order.cpp` to 1M elems). A dynamic-shape resize hook recomputes params + both dispatch counts + propagates cur_dims for out/mean/rstd. Fail-loud guards: 4D, `C % group == 0`, fp32, weight/bias len `== C`, mean/rstd size `== N*group`. ghstack-source-id: 406360859 @exported-using-ghexport Differential Revision: [D112257596](https://our.internmc.facebook.com/intern/diff/D112257596/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21173 Ports `aten.native_group_norm.default` as a 2-pass, 3-output group normalization, the second Phase B (vision) op. Pass 1 reduces per group to `mean`/`rstd`; pass 2 applies the per-channel affine. Key changes: - `runtime/ops/native_group_norm/{GroupNorm.cpp, group_norm_reduce.wgsl, group_norm.wgsl}` (+ generated `_wgsl.h`) — reduce: one thread per `(n, group)` serially scans that group's contiguous NCHW block (`base = n*C*HxW + g*(C/G)*HxW`, `group_size = (C/G)*HxW` elements), writing `mean = s/count`, `rstd = inverseSqrt(ss/count - mean^2 + eps)`; normalize: one thread per element, `g = c / (C/G)`, `out = (x - mean[n*G+g]) * rstd[n*G+g] * weight[c] + bias[c]`. A file-local `add_gn_dispatch` helper builds each of the two dispatches from an explicit binding list sharing one params UBO. Registered `aten.native_group_norm.default`. - `CMakeLists.txt` — `runtime/ops/native_group_norm/GroupNorm.cpp` in `WEBGPU_SRCS`. Mirrors Vulkan `impl/GroupNorm.cpp` (2 nodes: reduce -> mean/rstd, then normalize) + `glsl/group_norm_reduce_texture.glsl` (`mean=sum/count`, `variance=sumsq/count - mean^2`, `rstd=1/sqrt(var+eps)`). Buffer NCHW re-derivation (Vulkan is texture/channels-packed), so a serial per-group reduce (a group's channels x HxW are contiguous in NCHW) replaces the cooperative texture reduction. All 3 outputs are real. The inter-pass RAW ordering (normalize reads the reduce's mean/rstd) is guaranteed because `execute()` runs one compute pass per dispatch (`WebGPUGraph.cpp`; proven by `test_dispatch_order.cpp` to 1M elems). A dynamic-shape resize hook recomputes params + both dispatch counts + propagates cur_dims for out/mean/rstd. Fail-loud guards: 4D, `C % group == 0`, fp32, weight/bias len `== C`, mean/rstd size `== N*group`. ghstack-source-id: 406360859 @exported-using-ghexport Differential Revision: [D112257596](https://our.internmc.facebook.com/intern/diff/D112257596/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21173 Ports `aten.native_group_norm.default` as a 2-pass, 3-output group normalization, the second Phase B (vision) op. Pass 1 reduces per group to `mean`/`rstd`; pass 2 applies the per-channel affine. Key changes: - `runtime/ops/native_group_norm/{GroupNorm.cpp, group_norm_reduce.wgsl, group_norm.wgsl}` (+ generated `_wgsl.h`) — reduce: one thread per `(n, group)` serially scans that group's contiguous NCHW block (`base = n*C*HxW + g*(C/G)*HxW`, `group_size = (C/G)*HxW` elements), writing `mean = s/count`, `rstd = inverseSqrt(ss/count - mean^2 + eps)`; normalize: one thread per element, `g = c / (C/G)`, `out = (x - mean[n*G+g]) * rstd[n*G+g] * weight[c] + bias[c]`. A file-local `add_gn_dispatch` helper builds each of the two dispatches from an explicit binding list sharing one params UBO. Registered `aten.native_group_norm.default`. - `CMakeLists.txt` — `runtime/ops/native_group_norm/GroupNorm.cpp` in `WEBGPU_SRCS`. Mirrors Vulkan `impl/GroupNorm.cpp` (2 nodes: reduce -> mean/rstd, then normalize) + `glsl/group_norm_reduce_texture.glsl` (`mean=sum/count`, `variance=sumsq/count - mean^2`, `rstd=1/sqrt(var+eps)`). Buffer NCHW re-derivation (Vulkan is texture/channels-packed), so a serial per-group reduce (a group's channels x HxW are contiguous in NCHW) replaces the cooperative texture reduction. All 3 outputs are real. The inter-pass RAW ordering (normalize reads the reduce's mean/rstd) is guaranteed because `execute()` runs one compute pass per dispatch (`WebGPUGraph.cpp`; proven by `test_dispatch_order.cpp` to 1M elems). A dynamic-shape resize hook recomputes params + both dispatch counts + propagates cur_dims for out/mean/rstd. Fail-loud guards: 4D, `C % group == 0`, fp32, weight/bias len `== C`, mean/rstd size `== N*group`. ghstack-source-id: 406360859 @exported-using-ghexport Differential Revision: [D112257596](https://our.internmc.facebook.com/intern/diff/D112257596/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21173 Ports `aten.native_group_norm.default` as a 2-pass, 3-output group normalization, the second Phase B (vision) op. Pass 1 reduces per group to `mean`/`rstd`; pass 2 applies the per-channel affine. Key changes: - `runtime/ops/native_group_norm/{GroupNorm.cpp, group_norm_reduce.wgsl, group_norm.wgsl}` (+ generated `_wgsl.h`) — reduce: one thread per `(n, group)` serially scans that group's contiguous NCHW block (`base = n*C*HxW + g*(C/G)*HxW`, `group_size = (C/G)*HxW` elements), writing `mean = s/count`, `rstd = inverseSqrt(ss/count - mean^2 + eps)`; normalize: one thread per element, `g = c / (C/G)`, `out = (x - mean[n*G+g]) * rstd[n*G+g] * weight[c] + bias[c]`. A file-local `add_gn_dispatch` helper builds each of the two dispatches from an explicit binding list sharing one params UBO. Registered `aten.native_group_norm.default`. - `CMakeLists.txt` — `runtime/ops/native_group_norm/GroupNorm.cpp` in `WEBGPU_SRCS`. Mirrors Vulkan `impl/GroupNorm.cpp` (2 nodes: reduce -> mean/rstd, then normalize) + `glsl/group_norm_reduce_texture.glsl` (`mean=sum/count`, `variance=sumsq/count - mean^2`, `rstd=1/sqrt(var+eps)`). Buffer NCHW re-derivation (Vulkan is texture/channels-packed), so a serial per-group reduce (a group's channels x HxW are contiguous in NCHW) replaces the cooperative texture reduction. All 3 outputs are real. The inter-pass RAW ordering (normalize reads the reduce's mean/rstd) is guaranteed because `execute()` runs one compute pass per dispatch (`WebGPUGraph.cpp`; proven by `test_dispatch_order.cpp` to 1M elems). A dynamic-shape resize hook recomputes params + both dispatch counts + propagates cur_dims for out/mean/rstd. Fail-loud guards: 4D, `C % group == 0`, fp32, weight/bias len `== C`, mean/rstd size `== N*group`. ghstack-source-id: 406360859 @exported-using-ghexport Differential Revision: [D112257596](https://our.internmc.facebook.com/intern/diff/D112257596/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21173 Ports `aten.native_group_norm.default` as a 2-pass, 3-output group normalization, the second Phase B (vision) op. Pass 1 reduces per group to `mean`/`rstd`; pass 2 applies the per-channel affine. Key changes: - `runtime/ops/native_group_norm/{GroupNorm.cpp, group_norm_reduce.wgsl, group_norm.wgsl}` (+ generated `_wgsl.h`) — reduce: one thread per `(n, group)` serially scans that group's contiguous NCHW block (`base = n*C*HxW + g*(C/G)*HxW`, `group_size = (C/G)*HxW` elements), writing `mean = s/count`, `rstd = inverseSqrt(ss/count - mean^2 + eps)`; normalize: one thread per element, `g = c / (C/G)`, `out = (x - mean[n*G+g]) * rstd[n*G+g] * weight[c] + bias[c]`. A file-local `add_gn_dispatch` helper builds each of the two dispatches from an explicit binding list sharing one params UBO. Registered `aten.native_group_norm.default`. - `CMakeLists.txt` — `runtime/ops/native_group_norm/GroupNorm.cpp` in `WEBGPU_SRCS`. Mirrors Vulkan `impl/GroupNorm.cpp` (2 nodes: reduce -> mean/rstd, then normalize) + `glsl/group_norm_reduce_texture.glsl` (`mean=sum/count`, `variance=sumsq/count - mean^2`, `rstd=1/sqrt(var+eps)`). Buffer NCHW re-derivation (Vulkan is texture/channels-packed), so a serial per-group reduce (a group's channels x HxW are contiguous in NCHW) replaces the cooperative texture reduction. All 3 outputs are real. The inter-pass RAW ordering (normalize reads the reduce's mean/rstd) is guaranteed because `execute()` runs one compute pass per dispatch (`WebGPUGraph.cpp`; proven by `test_dispatch_order.cpp` to 1M elems). A dynamic-shape resize hook recomputes params + both dispatch counts + propagates cur_dims for out/mean/rstd. Fail-loud guards: 4D, `C % group == 0`, fp32, weight/bias len `== C`, mean/rstd size `== N*group`. ghstack-source-id: 406360859 @exported-using-ghexport Differential Revision: [D112257596](https://our.internmc.facebook.com/intern/diff/D112257596/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21173 Ports `aten.native_group_norm.default` as a 2-pass, 3-output group normalization, the second Phase B (vision) op. Pass 1 reduces per group to `mean`/`rstd`; pass 2 applies the per-channel affine. Key changes: - `runtime/ops/native_group_norm/{GroupNorm.cpp, group_norm_reduce.wgsl, group_norm.wgsl}` (+ generated `_wgsl.h`) — reduce: one thread per `(n, group)` serially scans that group's contiguous NCHW block (`base = n*C*HxW + g*(C/G)*HxW`, `group_size = (C/G)*HxW` elements), writing `mean = s/count`, `rstd = inverseSqrt(ss/count - mean^2 + eps)`; normalize: one thread per element, `g = c / (C/G)`, `out = (x - mean[n*G+g]) * rstd[n*G+g] * weight[c] + bias[c]`. A file-local `add_gn_dispatch` helper builds each of the two dispatches from an explicit binding list sharing one params UBO. Registered `aten.native_group_norm.default`. - `CMakeLists.txt` — `runtime/ops/native_group_norm/GroupNorm.cpp` in `WEBGPU_SRCS`. Mirrors Vulkan `impl/GroupNorm.cpp` (2 nodes: reduce -> mean/rstd, then normalize) + `glsl/group_norm_reduce_texture.glsl` (`mean=sum/count`, `variance=sumsq/count - mean^2`, `rstd=1/sqrt(var+eps)`). Buffer NCHW re-derivation (Vulkan is texture/channels-packed), so a serial per-group reduce (a group's channels x HxW are contiguous in NCHW) replaces the cooperative texture reduction. All 3 outputs are real. The inter-pass RAW ordering (normalize reads the reduce's mean/rstd) is guaranteed because `execute()` runs one compute pass per dispatch (`WebGPUGraph.cpp`; proven by `test_dispatch_order.cpp` to 1M elems). A dynamic-shape resize hook recomputes params + both dispatch counts + propagates cur_dims for out/mean/rstd. Fail-loud guards: 4D, `C % group == 0`, fp32, weight/bias len `== C`, mean/rstd size `== N*group`. ghstack-source-id: 406360859 @exported-using-ghexport Differential Revision: [D112257596](https://our.internmc.facebook.com/intern/diff/D112257596/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21173 Ports `aten.native_group_norm.default` as a 2-pass, 3-output group normalization, the second Phase B (vision) op. Pass 1 reduces per group to `mean`/`rstd`; pass 2 applies the per-channel affine. Key changes: - `runtime/ops/native_group_norm/{GroupNorm.cpp, group_norm_reduce.wgsl, group_norm.wgsl}` (+ generated `_wgsl.h`) — reduce: one thread per `(n, group)` serially scans that group's contiguous NCHW block (`base = n*C*HxW + g*(C/G)*HxW`, `group_size = (C/G)*HxW` elements), writing `mean = s/count`, `rstd = inverseSqrt(ss/count - mean^2 + eps)`; normalize: one thread per element, `g = c / (C/G)`, `out = (x - mean[n*G+g]) * rstd[n*G+g] * weight[c] + bias[c]`. A file-local `add_gn_dispatch` helper builds each of the two dispatches from an explicit binding list sharing one params UBO. Registered `aten.native_group_norm.default`. - `CMakeLists.txt` — `runtime/ops/native_group_norm/GroupNorm.cpp` in `WEBGPU_SRCS`. Mirrors Vulkan `impl/GroupNorm.cpp` (2 nodes: reduce -> mean/rstd, then normalize) + `glsl/group_norm_reduce_texture.glsl` (`mean=sum/count`, `variance=sumsq/count - mean^2`, `rstd=1/sqrt(var+eps)`). Buffer NCHW re-derivation (Vulkan is texture/channels-packed), so a serial per-group reduce (a group's channels x HxW are contiguous in NCHW) replaces the cooperative texture reduction. All 3 outputs are real. The inter-pass RAW ordering (normalize reads the reduce's mean/rstd) is guaranteed because `execute()` runs one compute pass per dispatch (`WebGPUGraph.cpp`; proven by `test_dispatch_order.cpp` to 1M elems). A dynamic-shape resize hook recomputes params + both dispatch counts + propagates cur_dims for out/mean/rstd. Fail-loud guards: 4D, `C % group == 0`, fp32, weight/bias len `== C`, mean/rstd size `== N*group`. ghstack-source-id: 406360859 @exported-using-ghexport Differential Revision: [D112257596](https://our.internmc.facebook.com/intern/diff/D112257596/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21173 Ports `aten.native_group_norm.default` as a 2-pass, 3-output group normalization, the second Phase B (vision) op. Pass 1 reduces per group to `mean`/`rstd`; pass 2 applies the per-channel affine. Key changes: - `runtime/ops/native_group_norm/{GroupNorm.cpp, group_norm_reduce.wgsl, group_norm.wgsl}` (+ generated `_wgsl.h`) — reduce: one thread per `(n, group)` serially scans that group's contiguous NCHW block (`base = n*C*HxW + g*(C/G)*HxW`, `group_size = (C/G)*HxW` elements), writing `mean = s/count`, `rstd = inverseSqrt(ss/count - mean^2 + eps)`; normalize: one thread per element, `g = c / (C/G)`, `out = (x - mean[n*G+g]) * rstd[n*G+g] * weight[c] + bias[c]`. A file-local `add_gn_dispatch` helper builds each of the two dispatches from an explicit binding list sharing one params UBO. Registered `aten.native_group_norm.default`. - `CMakeLists.txt` — `runtime/ops/native_group_norm/GroupNorm.cpp` in `WEBGPU_SRCS`. Mirrors Vulkan `impl/GroupNorm.cpp` (2 nodes: reduce -> mean/rstd, then normalize) + `glsl/group_norm_reduce_texture.glsl` (`mean=sum/count`, `variance=sumsq/count - mean^2`, `rstd=1/sqrt(var+eps)`). Buffer NCHW re-derivation (Vulkan is texture/channels-packed), so a serial per-group reduce (a group's channels x HxW are contiguous in NCHW) replaces the cooperative texture reduction. All 3 outputs are real. The inter-pass RAW ordering (normalize reads the reduce's mean/rstd) is guaranteed because `execute()` runs one compute pass per dispatch (`WebGPUGraph.cpp`; proven by `test_dispatch_order.cpp` to 1M elems). A dynamic-shape resize hook recomputes params + both dispatch counts + propagates cur_dims for out/mean/rstd. Fail-loud guards: 4D, `C % group == 0`, fp32, weight/bias len `== C`, mean/rstd size `== N*group`. ghstack-source-id: 406360859 @exported-using-ghexport Differential Revision: [D112257596](https://our.internmc.facebook.com/intern/diff/D112257596/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21173 Ports `aten.native_group_norm.default` as a 2-pass, 3-output group normalization, the second Phase B (vision) op. Pass 1 reduces per group to `mean`/`rstd`; pass 2 applies the per-channel affine. Key changes: - `runtime/ops/native_group_norm/{GroupNorm.cpp, group_norm_reduce.wgsl, group_norm.wgsl}` (+ generated `_wgsl.h`) — reduce: one thread per `(n, group)` serially scans that group's contiguous NCHW block (`base = n*C*HxW + g*(C/G)*HxW`, `group_size = (C/G)*HxW` elements), writing `mean = s/count`, `rstd = inverseSqrt(ss/count - mean^2 + eps)`; normalize: one thread per element, `g = c / (C/G)`, `out = (x - mean[n*G+g]) * rstd[n*G+g] * weight[c] + bias[c]`. A file-local `add_gn_dispatch` helper builds each of the two dispatches from an explicit binding list sharing one params UBO. Registered `aten.native_group_norm.default`. - `CMakeLists.txt` — `runtime/ops/native_group_norm/GroupNorm.cpp` in `WEBGPU_SRCS`. Mirrors Vulkan `impl/GroupNorm.cpp` (2 nodes: reduce -> mean/rstd, then normalize) + `glsl/group_norm_reduce_texture.glsl` (`mean=sum/count`, `variance=sumsq/count - mean^2`, `rstd=1/sqrt(var+eps)`). Buffer NCHW re-derivation (Vulkan is texture/channels-packed), so a serial per-group reduce (a group's channels x HxW are contiguous in NCHW) replaces the cooperative texture reduction. All 3 outputs are real. The inter-pass RAW ordering (normalize reads the reduce's mean/rstd) is guaranteed because `execute()` runs one compute pass per dispatch (`WebGPUGraph.cpp`; proven by `test_dispatch_order.cpp` to 1M elems). A dynamic-shape resize hook recomputes params + both dispatch counts + propagates cur_dims for out/mean/rstd. Fail-loud guards: 4D, `C % group == 0`, fp32, weight/bias len `== C`, mean/rstd size `== N*group`. ghstack-source-id: 406360859 @exported-using-ghexport Differential Revision: [D112257596](https://our.internmc.facebook.com/intern/diff/D112257596/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21173 Ports `aten.native_group_norm.default` as a 2-pass, 3-output group normalization, the second Phase B (vision) op. Pass 1 reduces per group to `mean`/`rstd`; pass 2 applies the per-channel affine. Key changes: - `runtime/ops/native_group_norm/{GroupNorm.cpp, group_norm_reduce.wgsl, group_norm.wgsl}` (+ generated `_wgsl.h`) — reduce: one thread per `(n, group)` serially scans that group's contiguous NCHW block (`base = n*C*HxW + g*(C/G)*HxW`, `group_size = (C/G)*HxW` elements), writing `mean = s/count`, `rstd = inverseSqrt(ss/count - mean^2 + eps)`; normalize: one thread per element, `g = c / (C/G)`, `out = (x - mean[n*G+g]) * rstd[n*G+g] * weight[c] + bias[c]`. A file-local `add_gn_dispatch` helper builds each of the two dispatches from an explicit binding list sharing one params UBO. Registered `aten.native_group_norm.default`. - `CMakeLists.txt` — `runtime/ops/native_group_norm/GroupNorm.cpp` in `WEBGPU_SRCS`. Mirrors Vulkan `impl/GroupNorm.cpp` (2 nodes: reduce -> mean/rstd, then normalize) + `glsl/group_norm_reduce_texture.glsl` (`mean=sum/count`, `variance=sumsq/count - mean^2`, `rstd=1/sqrt(var+eps)`). Buffer NCHW re-derivation (Vulkan is texture/channels-packed), so a serial per-group reduce (a group's channels x HxW are contiguous in NCHW) replaces the cooperative texture reduction. All 3 outputs are real. The inter-pass RAW ordering (normalize reads the reduce's mean/rstd) is guaranteed because `execute()` runs one compute pass per dispatch (`WebGPUGraph.cpp`; proven by `test_dispatch_order.cpp` to 1M elems). A dynamic-shape resize hook recomputes params + both dispatch counts + propagates cur_dims for out/mean/rstd. Fail-loud guards: 4D, `C % group == 0`, fp32, weight/bias len `== C`, mean/rstd size `== N*group`. ghstack-source-id: 406360859 @exported-using-ghexport Differential Revision: [D112257596](https://our.internmc.facebook.com/intern/diff/D112257596/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21173 Ports `aten.native_group_norm.default` as a 2-pass, 3-output group normalization, the second Phase B (vision) op. Pass 1 reduces per group to `mean`/`rstd`; pass 2 applies the per-channel affine. Key changes: - `runtime/ops/native_group_norm/{GroupNorm.cpp, group_norm_reduce.wgsl, group_norm.wgsl}` (+ generated `_wgsl.h`) — reduce: one thread per `(n, group)` serially scans that group's contiguous NCHW block (`base = n*C*HxW + g*(C/G)*HxW`, `group_size = (C/G)*HxW` elements), writing `mean = s/count`, `rstd = inverseSqrt(ss/count - mean^2 + eps)`; normalize: one thread per element, `g = c / (C/G)`, `out = (x - mean[n*G+g]) * rstd[n*G+g] * weight[c] + bias[c]`. A file-local `add_gn_dispatch` helper builds each of the two dispatches from an explicit binding list sharing one params UBO. Registered `aten.native_group_norm.default`. - `CMakeLists.txt` — `runtime/ops/native_group_norm/GroupNorm.cpp` in `WEBGPU_SRCS`. Mirrors Vulkan `impl/GroupNorm.cpp` (2 nodes: reduce -> mean/rstd, then normalize) + `glsl/group_norm_reduce_texture.glsl` (`mean=sum/count`, `variance=sumsq/count - mean^2`, `rstd=1/sqrt(var+eps)`). Buffer NCHW re-derivation (Vulkan is texture/channels-packed), so a serial per-group reduce (a group's channels x HxW are contiguous in NCHW) replaces the cooperative texture reduction. All 3 outputs are real. The inter-pass RAW ordering (normalize reads the reduce's mean/rstd) is guaranteed because `execute()` runs one compute pass per dispatch (`WebGPUGraph.cpp`; proven by `test_dispatch_order.cpp` to 1M elems). A dynamic-shape resize hook recomputes params + both dispatch counts + propagates cur_dims for out/mean/rstd. Fail-loud guards: 4D, `C % group == 0`, fp32, weight/bias len `== C`, mean/rstd size `== N*group`. ghstack-source-id: 406360859 @exported-using-ghexport Differential Revision: [D112257596](https://our.internmc.facebook.com/intern/diff/D112257596/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21173 Ports `aten.native_group_norm.default` as a 2-pass, 3-output group normalization, the second Phase B (vision) op. Pass 1 reduces per group to `mean`/`rstd`; pass 2 applies the per-channel affine. Key changes: - `runtime/ops/native_group_norm/{GroupNorm.cpp, group_norm_reduce.wgsl, group_norm.wgsl}` (+ generated `_wgsl.h`) — reduce: one thread per `(n, group)` serially scans that group's contiguous NCHW block (`base = n*C*HxW + g*(C/G)*HxW`, `group_size = (C/G)*HxW` elements), writing `mean = s/count`, `rstd = inverseSqrt(ss/count - mean^2 + eps)`; normalize: one thread per element, `g = c / (C/G)`, `out = (x - mean[n*G+g]) * rstd[n*G+g] * weight[c] + bias[c]`. A file-local `add_gn_dispatch` helper builds each of the two dispatches from an explicit binding list sharing one params UBO. Registered `aten.native_group_norm.default`. - `CMakeLists.txt` — `runtime/ops/native_group_norm/GroupNorm.cpp` in `WEBGPU_SRCS`. Mirrors Vulkan `impl/GroupNorm.cpp` (2 nodes: reduce -> mean/rstd, then normalize) + `glsl/group_norm_reduce_texture.glsl` (`mean=sum/count`, `variance=sumsq/count - mean^2`, `rstd=1/sqrt(var+eps)`). Buffer NCHW re-derivation (Vulkan is texture/channels-packed), so a serial per-group reduce (a group's channels x HxW are contiguous in NCHW) replaces the cooperative texture reduction. All 3 outputs are real. The inter-pass RAW ordering (normalize reads the reduce's mean/rstd) is guaranteed because `execute()` runs one compute pass per dispatch (`WebGPUGraph.cpp`; proven by `test_dispatch_order.cpp` to 1M elems). A dynamic-shape resize hook recomputes params + both dispatch counts + propagates cur_dims for out/mean/rstd. Fail-loud guards: 4D, `C % group == 0`, fp32, weight/bias len `== C`, mean/rstd size `== N*group`. ghstack-source-id: 406360859 @exported-using-ghexport Differential Revision: [D112257596](https://our.internmc.facebook.com/intern/diff/D112257596/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21173 Ports `aten.native_group_norm.default` as a 2-pass, 3-output group normalization, the second Phase B (vision) op. Pass 1 reduces per group to `mean`/`rstd`; pass 2 applies the per-channel affine. Key changes: - `runtime/ops/native_group_norm/{GroupNorm.cpp, group_norm_reduce.wgsl, group_norm.wgsl}` (+ generated `_wgsl.h`) — reduce: one thread per `(n, group)` serially scans that group's contiguous NCHW block (`base = n*C*HxW + g*(C/G)*HxW`, `group_size = (C/G)*HxW` elements), writing `mean = s/count`, `rstd = inverseSqrt(ss/count - mean^2 + eps)`; normalize: one thread per element, `g = c / (C/G)`, `out = (x - mean[n*G+g]) * rstd[n*G+g] * weight[c] + bias[c]`. A file-local `add_gn_dispatch` helper builds each of the two dispatches from an explicit binding list sharing one params UBO. Registered `aten.native_group_norm.default`. - `CMakeLists.txt` — `runtime/ops/native_group_norm/GroupNorm.cpp` in `WEBGPU_SRCS`. Mirrors Vulkan `impl/GroupNorm.cpp` (2 nodes: reduce -> mean/rstd, then normalize) + `glsl/group_norm_reduce_texture.glsl` (`mean=sum/count`, `variance=sumsq/count - mean^2`, `rstd=1/sqrt(var+eps)`). Buffer NCHW re-derivation (Vulkan is texture/channels-packed), so a serial per-group reduce (a group's channels x HxW are contiguous in NCHW) replaces the cooperative texture reduction. All 3 outputs are real. The inter-pass RAW ordering (normalize reads the reduce's mean/rstd) is guaranteed because `execute()` runs one compute pass per dispatch (`WebGPUGraph.cpp`; proven by `test_dispatch_order.cpp` to 1M elems). A dynamic-shape resize hook recomputes params + both dispatch counts + propagates cur_dims for out/mean/rstd. Fail-loud guards: 4D, `C % group == 0`, fp32, weight/bias len `== C`, mean/rstd size `== N*group`. ghstack-source-id: 406360859 @exported-using-ghexport Differential Revision: [D112257596](https://our.internmc.facebook.com/intern/diff/D112257596/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21173 Ports `aten.native_group_norm.default` as a 2-pass, 3-output group normalization, the second Phase B (vision) op. Pass 1 reduces per group to `mean`/`rstd`; pass 2 applies the per-channel affine. Key changes: - `runtime/ops/native_group_norm/{GroupNorm.cpp, group_norm_reduce.wgsl, group_norm.wgsl}` (+ generated `_wgsl.h`) — reduce: one thread per `(n, group)` serially scans that group's contiguous NCHW block (`base = n*C*HxW + g*(C/G)*HxW`, `group_size = (C/G)*HxW` elements), writing `mean = s/count`, `rstd = inverseSqrt(ss/count - mean^2 + eps)`; normalize: one thread per element, `g = c / (C/G)`, `out = (x - mean[n*G+g]) * rstd[n*G+g] * weight[c] + bias[c]`. A file-local `add_gn_dispatch` helper builds each of the two dispatches from an explicit binding list sharing one params UBO. Registered `aten.native_group_norm.default`. - `CMakeLists.txt` — `runtime/ops/native_group_norm/GroupNorm.cpp` in `WEBGPU_SRCS`. Mirrors Vulkan `impl/GroupNorm.cpp` (2 nodes: reduce -> mean/rstd, then normalize) + `glsl/group_norm_reduce_texture.glsl` (`mean=sum/count`, `variance=sumsq/count - mean^2`, `rstd=1/sqrt(var+eps)`). Buffer NCHW re-derivation (Vulkan is texture/channels-packed), so a serial per-group reduce (a group's channels x HxW are contiguous in NCHW) replaces the cooperative texture reduction. All 3 outputs are real. The inter-pass RAW ordering (normalize reads the reduce's mean/rstd) is guaranteed because `execute()` runs one compute pass per dispatch (`WebGPUGraph.cpp`; proven by `test_dispatch_order.cpp` to 1M elems). A dynamic-shape resize hook recomputes params + both dispatch counts + propagates cur_dims for out/mean/rstd. Fail-loud guards: 4D, `C % group == 0`, fp32, weight/bias len `== C`, mean/rstd size `== N*group`. ghstack-source-id: 406360859 @exported-using-ghexport Differential Revision: [D112257596](https://our.internmc.facebook.com/intern/diff/D112257596/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21173 Ports `aten.native_group_norm.default` as a 2-pass, 3-output group normalization, the second Phase B (vision) op. Pass 1 reduces per group to `mean`/`rstd`; pass 2 applies the per-channel affine. Key changes: - `runtime/ops/native_group_norm/{GroupNorm.cpp, group_norm_reduce.wgsl, group_norm.wgsl}` (+ generated `_wgsl.h`) — reduce: one thread per `(n, group)` serially scans that group's contiguous NCHW block (`base = n*C*HxW + g*(C/G)*HxW`, `group_size = (C/G)*HxW` elements), writing `mean = s/count`, `rstd = inverseSqrt(ss/count - mean^2 + eps)`; normalize: one thread per element, `g = c / (C/G)`, `out = (x - mean[n*G+g]) * rstd[n*G+g] * weight[c] + bias[c]`. A file-local `add_gn_dispatch` helper builds each of the two dispatches from an explicit binding list sharing one params UBO. Registered `aten.native_group_norm.default`. - `CMakeLists.txt` — `runtime/ops/native_group_norm/GroupNorm.cpp` in `WEBGPU_SRCS`. Mirrors Vulkan `impl/GroupNorm.cpp` (2 nodes: reduce -> mean/rstd, then normalize) + `glsl/group_norm_reduce_texture.glsl` (`mean=sum/count`, `variance=sumsq/count - mean^2`, `rstd=1/sqrt(var+eps)`). Buffer NCHW re-derivation (Vulkan is texture/channels-packed), so a serial per-group reduce (a group's channels x HxW are contiguous in NCHW) replaces the cooperative texture reduction. All 3 outputs are real. The inter-pass RAW ordering (normalize reads the reduce's mean/rstd) is guaranteed because `execute()` runs one compute pass per dispatch (`WebGPUGraph.cpp`; proven by `test_dispatch_order.cpp` to 1M elems). A dynamic-shape resize hook recomputes params + both dispatch counts + propagates cur_dims for out/mean/rstd. Fail-loud guards: 4D, `C % group == 0`, fp32, weight/bias len `== C`, mean/rstd size `== N*group`. ghstack-source-id: 406360859 @exported-using-ghexport Differential Revision: [D112257596](https://our.internmc.facebook.com/intern/diff/D112257596/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21173 Ports `aten.native_group_norm.default` as a 2-pass, 3-output group normalization, the second Phase B (vision) op. Pass 1 reduces per group to `mean`/`rstd`; pass 2 applies the per-channel affine. Key changes: - `runtime/ops/native_group_norm/{GroupNorm.cpp, group_norm_reduce.wgsl, group_norm.wgsl}` (+ generated `_wgsl.h`) — reduce: one thread per `(n, group)` serially scans that group's contiguous NCHW block (`base = n*C*HxW + g*(C/G)*HxW`, `group_size = (C/G)*HxW` elements), writing `mean = s/count`, `rstd = inverseSqrt(ss/count - mean^2 + eps)`; normalize: one thread per element, `g = c / (C/G)`, `out = (x - mean[n*G+g]) * rstd[n*G+g] * weight[c] + bias[c]`. A file-local `add_gn_dispatch` helper builds each of the two dispatches from an explicit binding list sharing one params UBO. Registered `aten.native_group_norm.default`. - `CMakeLists.txt` — `runtime/ops/native_group_norm/GroupNorm.cpp` in `WEBGPU_SRCS`. Mirrors Vulkan `impl/GroupNorm.cpp` (2 nodes: reduce -> mean/rstd, then normalize) + `glsl/group_norm_reduce_texture.glsl` (`mean=sum/count`, `variance=sumsq/count - mean^2`, `rstd=1/sqrt(var+eps)`). Buffer NCHW re-derivation (Vulkan is texture/channels-packed), so a serial per-group reduce (a group's channels x HxW are contiguous in NCHW) replaces the cooperative texture reduction. All 3 outputs are real. The inter-pass RAW ordering (normalize reads the reduce's mean/rstd) is guaranteed because `execute()` runs one compute pass per dispatch (`WebGPUGraph.cpp`; proven by `test_dispatch_order.cpp` to 1M elems). A dynamic-shape resize hook recomputes params + both dispatch counts + propagates cur_dims for out/mean/rstd. Fail-loud guards: 4D, `C % group == 0`, fp32, weight/bias len `== C`, mean/rstd size `== N*group`. ghstack-source-id: 406360859 @exported-using-ghexport Differential Revision: [D112257596](https://our.internmc.facebook.com/intern/diff/D112257596/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21173 Ports `aten.native_group_norm.default` as a 2-pass, 3-output group normalization, the second Phase B (vision) op. Pass 1 reduces per group to `mean`/`rstd`; pass 2 applies the per-channel affine. Key changes: - `runtime/ops/native_group_norm/{GroupNorm.cpp, group_norm_reduce.wgsl, group_norm.wgsl}` (+ generated `_wgsl.h`) — reduce: one thread per `(n, group)` serially scans that group's contiguous NCHW block (`base = n*C*HxW + g*(C/G)*HxW`, `group_size = (C/G)*HxW` elements), writing `mean = s/count`, `rstd = inverseSqrt(ss/count - mean^2 + eps)`; normalize: one thread per element, `g = c / (C/G)`, `out = (x - mean[n*G+g]) * rstd[n*G+g] * weight[c] + bias[c]`. A file-local `add_gn_dispatch` helper builds each of the two dispatches from an explicit binding list sharing one params UBO. Registered `aten.native_group_norm.default`. - `CMakeLists.txt` — `runtime/ops/native_group_norm/GroupNorm.cpp` in `WEBGPU_SRCS`. Mirrors Vulkan `impl/GroupNorm.cpp` (2 nodes: reduce -> mean/rstd, then normalize) + `glsl/group_norm_reduce_texture.glsl` (`mean=sum/count`, `variance=sumsq/count - mean^2`, `rstd=1/sqrt(var+eps)`). Buffer NCHW re-derivation (Vulkan is texture/channels-packed), so a serial per-group reduce (a group's channels x HxW are contiguous in NCHW) replaces the cooperative texture reduction. All 3 outputs are real. The inter-pass RAW ordering (normalize reads the reduce's mean/rstd) is guaranteed because `execute()` runs one compute pass per dispatch (`WebGPUGraph.cpp`; proven by `test_dispatch_order.cpp` to 1M elems). A dynamic-shape resize hook recomputes params + both dispatch counts + propagates cur_dims for out/mean/rstd. Fail-loud guards: 4D, `C % group == 0`, fp32, weight/bias len `== C`, mean/rstd size `== N*group`. ghstack-source-id: 406360859 @exported-using-ghexport Differential Revision: [D112257596](https://our.internmc.facebook.com/intern/diff/D112257596/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21173 Ports `aten.native_group_norm.default` as a 2-pass, 3-output group normalization, the second Phase B (vision) op. Pass 1 reduces per group to `mean`/`rstd`; pass 2 applies the per-channel affine. Key changes: - `runtime/ops/native_group_norm/{GroupNorm.cpp, group_norm_reduce.wgsl, group_norm.wgsl}` (+ generated `_wgsl.h`) — reduce: one thread per `(n, group)` serially scans that group's contiguous NCHW block (`base = n*C*HxW + g*(C/G)*HxW`, `group_size = (C/G)*HxW` elements), writing `mean = s/count`, `rstd = inverseSqrt(ss/count - mean^2 + eps)`; normalize: one thread per element, `g = c / (C/G)`, `out = (x - mean[n*G+g]) * rstd[n*G+g] * weight[c] + bias[c]`. A file-local `add_gn_dispatch` helper builds each of the two dispatches from an explicit binding list sharing one params UBO. Registered `aten.native_group_norm.default`. - `CMakeLists.txt` — `runtime/ops/native_group_norm/GroupNorm.cpp` in `WEBGPU_SRCS`. Mirrors Vulkan `impl/GroupNorm.cpp` (2 nodes: reduce -> mean/rstd, then normalize) + `glsl/group_norm_reduce_texture.glsl` (`mean=sum/count`, `variance=sumsq/count - mean^2`, `rstd=1/sqrt(var+eps)`). Buffer NCHW re-derivation (Vulkan is texture/channels-packed), so a serial per-group reduce (a group's channels x HxW are contiguous in NCHW) replaces the cooperative texture reduction. All 3 outputs are real. The inter-pass RAW ordering (normalize reads the reduce's mean/rstd) is guaranteed because `execute()` runs one compute pass per dispatch (`WebGPUGraph.cpp`; proven by `test_dispatch_order.cpp` to 1M elems). A dynamic-shape resize hook recomputes params + both dispatch counts + propagates cur_dims for out/mean/rstd. Fail-loud guards: 4D, `C % group == 0`, fp32, weight/bias len `== C`, mean/rstd size `== N*group`. ghstack-source-id: 406360859 @exported-using-ghexport Differential Revision: [D112257596](https://our.internmc.facebook.com/intern/diff/D112257596/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21173 Ports `aten.native_group_norm.default` as a 2-pass, 3-output group normalization, the second Phase B (vision) op. Pass 1 reduces per group to `mean`/`rstd`; pass 2 applies the per-channel affine. Key changes: - `runtime/ops/native_group_norm/{GroupNorm.cpp, group_norm_reduce.wgsl, group_norm.wgsl}` (+ generated `_wgsl.h`) — reduce: one thread per `(n, group)` serially scans that group's contiguous NCHW block (`base = n*C*HxW + g*(C/G)*HxW`, `group_size = (C/G)*HxW` elements), writing `mean = s/count`, `rstd = inverseSqrt(ss/count - mean^2 + eps)`; normalize: one thread per element, `g = c / (C/G)`, `out = (x - mean[n*G+g]) * rstd[n*G+g] * weight[c] + bias[c]`. A file-local `add_gn_dispatch` helper builds each of the two dispatches from an explicit binding list sharing one params UBO. Registered `aten.native_group_norm.default`. - `CMakeLists.txt` — `runtime/ops/native_group_norm/GroupNorm.cpp` in `WEBGPU_SRCS`. Mirrors Vulkan `impl/GroupNorm.cpp` (2 nodes: reduce -> mean/rstd, then normalize) + `glsl/group_norm_reduce_texture.glsl` (`mean=sum/count`, `variance=sumsq/count - mean^2`, `rstd=1/sqrt(var+eps)`). Buffer NCHW re-derivation (Vulkan is texture/channels-packed), so a serial per-group reduce (a group's channels x HxW are contiguous in NCHW) replaces the cooperative texture reduction. All 3 outputs are real. The inter-pass RAW ordering (normalize reads the reduce's mean/rstd) is guaranteed because `execute()` runs one compute pass per dispatch (`WebGPUGraph.cpp`; proven by `test_dispatch_order.cpp` to 1M elems). A dynamic-shape resize hook recomputes params + both dispatch counts + propagates cur_dims for out/mean/rstd. Fail-loud guards: 4D, `C % group == 0`, fp32, weight/bias len `== C`, mean/rstd size `== N*group`. ghstack-source-id: 406360859 @exported-using-ghexport Differential Revision: [D112257596](https://our.internmc.facebook.com/intern/diff/D112257596/)
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stack from ghstack (oldest at bottom):
Ports
aten.native_group_norm.defaultas a 2-pass, 3-output group normalization, the second Phase B (vision) op. Pass 1 reduces per group tomean/rstd; pass 2 applies the per-channel affine.Key changes:
runtime/ops/native_group_norm/{GroupNorm.cpp, group_norm_reduce.wgsl, group_norm.wgsl}(+ generated_wgsl.h) — reduce: one thread per(n, group)serially scans that group's contiguous NCHW block (base = n*C*HxW + g*(C/G)*HxW,group_size = (C/G)*HxWelements), writingmean = s/count,rstd = inverseSqrt(ss/count - mean^2 + eps); normalize: one thread per element,g = c / (C/G),out = (x - mean[n*G+g]) * rstd[n*G+g] * weight[c] + bias[c]. A file-localadd_gn_dispatchhelper builds each of the two dispatches from an explicit binding list sharing one params UBO. Registeredaten.native_group_norm.default.CMakeLists.txt—runtime/ops/native_group_norm/GroupNorm.cppinWEBGPU_SRCS.Mirrors Vulkan
impl/GroupNorm.cpp(2 nodes: reduce -> mean/rstd, then normalize) +glsl/group_norm_reduce_texture.glsl(mean=sum/count,variance=sumsq/count - mean^2,rstd=1/sqrt(var+eps)). Buffer NCHW re-derivation (Vulkan is texture/channels-packed), so a serial per-group reduce (a group's channels x HxW are contiguous in NCHW) replaces the cooperative texture reduction. All 3 outputs are real. The inter-pass RAW ordering (normalize reads the reduce's mean/rstd) is guaranteed becauseexecute()runs one compute pass per dispatch (WebGPUGraph.cpp; proven bytest_dispatch_order.cppto 1M elems). A dynamic-shape resize hook recomputes params + both dispatch counts + propagates cur_dims for out/mean/rstd. Fail-loud guards: 4D,C % group == 0, fp32, weight/bias len== C, mean/rstd size== N*group.@exported-using-ghexport
Differential Revision: D112257596
Differential Revision: D112257596