From 8f32adc90a7fee83583c9ba89dbdfabb317e0452 Mon Sep 17 00:00:00 2001
From: Nikita Shulga <nikita.shulga@gmail.com>
Date: Sun, 28 Sep 2025 19:00:49 -0700
Subject: [PATCH] [MPSHooks] Release pending command encoder (#164093)

Before returning a comand buffer, as subsequent calle are very likely to allocate their own encoder, which results in the following runtime error
```
 tryCoalescingPreviousComputeCommandEncoderWithConfig:nextEncoderClass:]:1090: failed assertion `A command encoder is already encoding to this command buffer'
```

Added regression test to `test_mps_extension`

Please note, that `torch::mps::get_command_buffer()` should be called with dispatch_queue held, both before and after this change, but many implementations skip that

Fixes https://github.com/pytorch/pytorch/issues/163721
Pull Request resolved: https://github.com/pytorch/pytorch/pull/164093
Approved by: https://github.com/atalman, https://github.com/Skylion007
---
 aten/src/ATen/mps/MPSHooks.mm        |  5 ++++-
 test/cpp_extensions/mps_extension.mm | 29 ++++++++++++++++++++++++++++
 test/test_cpp_extensions_jit.py      |  6 ++++++
 3 files changed, 39 insertions(+), 1 deletion(-)

diff --git a/aten/src/ATen/mps/MPSHooks.mm b/aten/src/ATen/mps/MPSHooks.mm
index a2ec221c1bfe..34fbd31af91d 100644
--- a/aten/src/ATen/mps/MPSHooks.mm
+++ b/aten/src/ATen/mps/MPSHooks.mm
@@ -70,7 +70,10 @@ void MPSHooks::commitStream() const {
 }
 
 void* MPSHooks::getCommandBuffer() const {
-  return at::mps::getDefaultMPSStream()->commandBuffer();
+  auto stream = at::mps::getDefaultMPSStream();
+  // Release pending computeCommandEncoder, as extensions is likely to allocate new one
+  stream->endKernelCoalescing();
+  return stream->commandBuffer();
 }
 
 void* MPSHooks::getDispatchQueue() const {
diff --git a/test/cpp_extensions/mps_extension.mm b/test/cpp_extensions/mps_extension.mm
index 882e5c5603e2..30b70a76563d 100644
--- a/test/cpp_extensions/mps_extension.mm
+++ b/test/cpp_extensions/mps_extension.mm
@@ -13,6 +13,11 @@ kernel void add_arrays(device const float* inA,
 {
     result[index] = inA[index] + inB[index];
 }
+
+kernel void add_one(device float* data,
+                    uint index [[thread_position_in_grid]]) {
+  data[index] += 1.0;
+}
 )MPS_ADD_ARRAYS");
 
 at::Tensor get_cpu_add_output(at::Tensor & cpu_input1, at::Tensor & cpu_input2) {
@@ -50,7 +55,31 @@ at::Tensor get_mps_add_output(at::Tensor & mps_input1, at::Tensor & mps_input2)
   return mps_output;
 }
 
+void mps_add_one_new_encoder(const at::Tensor& input) {
+  using namespace at::native::mps;
+  TORCH_CHECK(input.is_mps());
+  TORCH_CHECK(input.numel() > 0);
+
+  @autoreleasepool {
+  auto kernelPSO = lib.getPipelineStateForFunc("add_one");
+  auto serialQueue = torch::mps::get_dispatch_queue();
+
+  dispatch_sync(serialQueue, ^(){
+    auto commandBuffer = torch::mps::get_command_buffer();
+    // Start a compute pass.
+    auto computeEncoder = [commandBuffer computeCommandEncoder];
+    TORCH_CHECK(computeEncoder, "Failed to create compute command encoder");
+    [computeEncoder setComputePipelineState: kernelPSO];
+    mtl_setArgs(computeEncoder, input);
+    mtl_dispatch1DJob(computeEncoder, kernelPSO, input.numel());
+    [computeEncoder endEncoding];
+     torch::mps::commit();
+  });
+  }
+}
+
 PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
   m.def("get_cpu_add_output", &get_cpu_add_output);
   m.def("get_mps_add_output", &get_mps_add_output);
+  m.def("mps_add_one_new_context", &mps_add_one_new_encoder);
 }
diff --git a/test/test_cpp_extensions_jit.py b/test/test_cpp_extensions_jit.py
index fd80c7fa565a..e93167296a00 100644
--- a/test/test_cpp_extensions_jit.py
+++ b/test/test_cpp_extensions_jit.py
@@ -220,6 +220,12 @@ class TestCppExtensionJIT(common.TestCase):
 
         self.assertEqual(cpu_output, mps_output.to("cpu"))
 
+        # Regression test for https://github.com/pytorch/pytorch/issues/163721
+        lib = torch.mps.compile_shader("void kernel noop(device float *x) {}")
+        lib.noop(mps_output)
+        module.mps_add_one_new_context(mps_output)
+        self.assertEqual(cpu_output + 1.0, mps_output.to("cpu"))
+
     def _run_jit_cuda_archflags(self, flags, expected):
         # Compile an extension with given `flags`
         def _check_cuobjdump_output(expected_values, is_ptx=False):