Update (base update)

[ghstack-poisoned]

Author: JCNTH

backends/webgpu/runtime/ops/log_softmax/LogSoftmax.cpp

@@ -29,7 +29,9 @@ struct LogSoftmaxParams {
   uint32_t inner;
   uint32_t pad_;
 };
-static_assert(sizeof(LogSoftmaxParams) == 16, "LogSoftmaxParams must be 16 bytes");
+static_assert(
+    sizeof(LogSoftmaxParams) == 16,
+    "LogSoftmaxParams must be 16 bytes");
 
 // Decompose dims into [outer, R, inner] for a reduction along `dim`.
 void decompose(
@@ -85,7 +87,8 @@ void log_softmax_impl(WebGPUGraph& graph, const std::vector<int>& args) {
   for (int64_t d : in.dims) {
     numel *= static_cast<uint64_t>(d);
   }
-  if (in.nbytes != numel * sizeof(float) || out.nbytes != numel * sizeof(float)) {
+  if (in.nbytes != numel * sizeof(float) ||
+      out.nbytes != numel * sizeof(float)) {
     throw std::runtime_error(
         "WebGPU log_softmax: fp32-only (byte-size mismatch)");
   }
@@ -205,7 +208,9 @@ void log_softmax_impl(WebGPUGraph& graph, const std::vector<int>& args) {
         wgpuQueueWriteBuffer(g.queue(), params_buf, 0, &p, sizeof(p));
         g.dispatch_at(dispatch_idx).workgroup_count_x =
             utils::compute_1d_workgroup_count(
-                g.device(), static_cast<uint32_t>(live_lines), wg_size,
+                g.device(),
+                static_cast<uint32_t>(live_lines),
+                wg_size,
                 "log_softmax(resize)");
         g.set_cur_dims(out_id, std::vector<int64_t>(d.begin(), d.end()));
       });

backends/webgpu/runtime/ops/reduce/Reduce.cpp

@@ -227,7 +227,9 @@ void reduce_impl(
         wgpuQueueWriteBuffer(g.queue(), params_buf, 0, &p, sizeof(p));
         g.dispatch_at(dispatch_idx).workgroup_count_x =
             utils::compute_1d_workgroup_count(
-                g.device(), static_cast<uint32_t>(live_outputs), wg_size,
+                g.device(),
+                static_cast<uint32_t>(live_outputs),
+                wg_size,
                 "reduce(resize)");
         // Propagate reduced output dims for downstream resize hooks.
         int64_t nd = static_cast<int64_t>(d.size());

backends/webgpu/runtime/ops/softmax/Softmax.cpp

@@ -173,8 +173,7 @@ void softmax_impl(WebGPUGraph& graph, const std::vector<int>& args) {
 
   // Dynamic shapes: recompute the decomposition for the reduced dim + dispatch.
   WGPUBuffer params_buf = uniform_buffer;
-  const uint64_t build_numel =
-      static_cast<uint64_t>(outer) * r * inner;
+  const uint64_t build_numel = static_cast<uint64_t>(outer) * r * inner;
   graph.add_tensor_resize_hook(
       in_id,
       [in_id, out_id, dim, build_numel, wg_size, dispatch_idx, params_buf](
@@ -198,7 +197,9 @@ void softmax_impl(WebGPUGraph& graph, const std::vector<int>& args) {
         wgpuQueueWriteBuffer(g.queue(), params_buf, 0, &p, sizeof(p));
         g.dispatch_at(dispatch_idx).workgroup_count_x =
             utils::compute_1d_workgroup_count(
-                g.device(), static_cast<uint32_t>(live_lines), wg_size,
+                g.device(),
+                static_cast<uint32_t>(live_lines),
+                wg_size,
                 "softmax(resize)");
         g.set_cur_dims(out_id, d);
       });

backends/webgpu/test/ops/test_bmm.py

@@ -19,9 +19,7 @@
 
 import torch
 
-from executorch.backends.vulkan.partitioner.vulkan_partitioner import (
-    VulkanPartitioner,
-)
+from executorch.backends.vulkan.partitioner.vulkan_partitioner import VulkanPartitioner
 from executorch.exir import to_edge_transform_and_lower
 
 # name -> (shape_a [B, M, K], shape_b [B, K, N]). Output is [B, M, N].

backends/webgpu/test/ops/test_div.py

@@ -20,9 +20,7 @@
 
 import torch
 
-from executorch.backends.vulkan.partitioner.vulkan_partitioner import (
-    VulkanPartitioner,
-)
+from executorch.backends.vulkan.partitioner.vulkan_partitioner import VulkanPartitioner
 from executorch.exir import to_edge_transform_and_lower
 
 # name -> (shape_a, shape_b). Output shape is the broadcast of the two.

backends/webgpu/test/ops/test_linear.py

@@ -19,9 +19,7 @@
 
 import torch
 
-from executorch.backends.vulkan.partitioner.vulkan_partitioner import (
-    VulkanPartitioner,
-)
+from executorch.backends.vulkan.partitioner.vulkan_partitioner import VulkanPartitioner
 from executorch.exir import to_edge_transform_and_lower
 
 # name -> (M, K, N). weight is [N, K]; output is [M, N].

backends/webgpu/test/ops/test_log_softmax.py

@@ -20,9 +20,7 @@
 
 import torch
 
-from executorch.backends.vulkan.partitioner.vulkan_partitioner import (
-    VulkanPartitioner,
-)
+from executorch.backends.vulkan.partitioner.vulkan_partitioner import VulkanPartitioner
 from executorch.exir import to_edge_transform_and_lower
 
 

backends/webgpu/test/ops/test_reduce.py

@@ -19,9 +19,7 @@
 
 import torch
 
-from executorch.backends.vulkan.partitioner.vulkan_partitioner import (
-    VulkanPartitioner,
-)
+from executorch.backends.vulkan.partitioner.vulkan_partitioner import VulkanPartitioner
 from executorch.exir import to_edge_transform_and_lower
 
 

backends/webgpu/test/ops/test_softmax.py

@@ -21,9 +21,7 @@
 
 import torch
 
-from executorch.backends.vulkan.partitioner.vulkan_partitioner import (
-    VulkanPartitioner,
-)
+from executorch.backends.vulkan.partitioner.vulkan_partitioner import VulkanPartitioner
 from executorch.exir import to_edge_transform_and_lower
 
 # name -> (shape, dim). dim=-1 => inner=1 (attention); a middle dim => inner>1.

backends/webgpu/test/ops/test_sub.py

@@ -19,9 +19,7 @@
 
 import torch
 
-from executorch.backends.vulkan.partitioner.vulkan_partitioner import (
-    VulkanPartitioner,
-)
+from executorch.backends.vulkan.partitioner.vulkan_partitioner import VulkanPartitioner
 from executorch.exir import to_edge_transform_and_lower