[ExecuTorch][WebGPU] Add q8ta_add op (int8 elementwise add)#21191
Merged
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/21191
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit ff9b3d7 with merge base 4a26c64 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This PR needs a
|
This was referenced Jul 22, 2026
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
JCNTH
added a commit
that referenced
this pull request
Jul 24, 2026
Pull Request resolved: #21191 Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established. Solution: Port `et_vk.q8ta_add` (two int8 tensors + per-tensor qparams -> int8): dequantize both, compute `a + alpha*b`, requantize. Reuses the C0 int8 buffer path (`array<u32>` pack/unpack, `is_int8` guard, multiply-by-`inv_scale`). Implementation: `Q8taAdd.cpp` registers `et_vk.q8ta_add.default` (`out = args.back()`), guards all three tensors int8 + equal `numel` + `numel % 4 == 0`, fail-loud; `q8ta_add.wgsl` unpacks 4 int8/word, dequantizes `scale*(q - zero_point)`, forms `a + alpha*b`, and requantizes `clamp(round(v * inv_output_scale) + zero_point, -128, 127)`. Mirrors Vulkan `q8ta_binary.glsl` + `common.glslh`, with one deliberate divergence: the Vulkan glsl buffer kernel drops `alpha` (computes `a + b`), so this port implements the full `a + alpha*b` to match the CPU reference for any `alpha` (surfaced in the handler comment). Constraints: same-shape only (no broadcast — the equal-`numel` guard rejects it, matching the `add` op baseline); `supports_resize=False` (no resize hook), mirroring the partitioner. ghstack-source-id: 406366820 @exported-using-ghexport Differential Revision: [D112257597](https://our.internmc.facebook.com/intern/diff/D112257597/)
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):
Problem: The WebGPU delegate has no quantized elementwise add — the first C1 q8ta op, on the int8 activation path C0 established.
Solution: Port
et_vk.q8ta_add(two int8 tensors + per-tensor qparams -> int8): dequantize both, computea + alpha*b, requantize. Reuses the C0 int8 buffer path (array<u32>pack/unpack,is_int8guard, multiply-by-inv_scale).Implementation:
Q8taAdd.cppregisterset_vk.q8ta_add.default(out = args.back()), guards all three tensors int8 + equalnumel+numel % 4 == 0, fail-loud;q8ta_add.wgslunpacks 4 int8/word, dequantizesscale*(q - zero_point), formsa + alpha*b, and requantizesclamp(round(v * inv_output_scale) + zero_point, -128, 127). Mirrors Vulkanq8ta_binary.glsl+common.glslh, with one deliberate divergence: the Vulkan glsl buffer kernel dropsalpha(computesa + b), so this port implements the fulla + alpha*bto match the CPU reference for anyalpha(surfaced in the handler comment).Constraints: same-shape only (no broadcast — the equal-
numelguard rejects it, matching theaddop baseline);supports_resize=False(no resize hook), mirroring the partitioner.@exported-using-ghexport
Differential Revision: D112257597
Differential Revision: D112257597