update

build/torch28-metal-aarch64-darwin/gptoss_kernels/{_gptoss_kernels_c35553c_dirty.abi3.so → _gptoss_kernels_c15b761_dirty.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4dade83d26cdd3717057e51d61a30d4fb194c64f51221e060dc07e4a3c258a65
+oid sha256:0d781d989726f3295f1400d4ebe88c30cd223a070a5bc01b197fe531ee626373
 size 425784

build/torch28-metal-aarch64-darwin/gptoss_kernels/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _gptoss_kernels_c35553c_dirty
-ops = torch.ops._gptoss_kernels_c35553c_dirty
+from . import _gptoss_kernels_c15b761_dirty
+ops = torch.ops._gptoss_kernels_c15b761_dirty
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_gptoss_kernels_c35553c_dirty::{op_name}"
+    return f"_gptoss_kernels_c15b761_dirty::{op_name}"

build/torch29-metal-aarch64-darwin/gptoss_kernels/{_gptoss_kernels_c35553c_dirty.abi3.so → _gptoss_kernels_c15b761_dirty.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:bebf67ec9862c99e944560c8c4e7fc351b51a82399ef28d213f0322cbb5ac5b4
+oid sha256:b373b3470df558ef09f3f8e3ecb830a2d21a39b62904f3d0fe2cfcd581e071d3
 size 410328

build/torch29-metal-aarch64-darwin/gptoss_kernels/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _gptoss_kernels_c35553c_dirty
-ops = torch.ops._gptoss_kernels_c35553c_dirty
+from . import _gptoss_kernels_c15b761_dirty
+ops = torch.ops._gptoss_kernels_c15b761_dirty
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_gptoss_kernels_c35553c_dirty::{op_name}"
+    return f"_gptoss_kernels_c15b761_dirty::{op_name}"

gptoss_kernels/source/tensor_wrappers.cpp

@@ -138,7 +138,7 @@ void run_metal_kernel(const char* kernel_symbol, EncodeFn&& encode_fn) {
             "function_create failed");
         stream = at::mps::getCurrentMPSStream();
         if (stream != nullptr) {
-            stream->synchronize(at::mps::SyncType::COMMIT_AND_WAIT);
+            // stream->synchronize(at::mps::SyncType::COMMIT_AND_WAIT);
             stream->endKernelCoalescing();
             void* command_queue_obj = stream->commandQueue();
             if (command_queue_obj != nullptr) {