diff --git a/third_party/xla/xla/stream_executor/gpu/gpu_command_buffer.cc b/third_party/xla/xla/stream_executor/gpu/gpu_command_buffer.cc index bcfc6a3790507f..00d463498c131c 100644 --- a/third_party/xla/xla/stream_executor/gpu/gpu_command_buffer.cc +++ b/third_party/xla/xla/stream_executor/gpu/gpu_command_buffer.cc @@ -91,10 +91,6 @@ static int64_t NotifyExecDestroyed() { return alive_execs.fetch_sub(1, std::memory_order_relaxed) - 1; } -/*static*/ int64_t GpuCommandBuffer::AllocatedExecs() { - return allocated_execs.load(std::memory_order_relaxed); -} - /*static*/ int64_t GpuCommandBuffer::AliveExecs() { return alive_execs.load(std::memory_order_relaxed); } diff --git a/third_party/xla/xla/stream_executor/gpu/gpu_command_buffer.h b/third_party/xla/xla/stream_executor/gpu/gpu_command_buffer.h index 86b1fdea5bb210..2808fe6364c047 100644 --- a/third_party/xla/xla/stream_executor/gpu/gpu_command_buffer.h +++ b/third_party/xla/xla/stream_executor/gpu/gpu_command_buffer.h @@ -158,7 +158,6 @@ class GpuCommandBuffer : public CommandBuffer { // allocates resources on a GPU devices (rule of thumb is ~8kb per node), so // we have to be careful not to keep too many of them alive for too long, or // we have a higher risk of OOM errors. - static int64_t AllocatedExecs(); static int64_t AliveExecs(); private: @@ -357,10 +356,6 @@ class GpuCommandBuffer : public CommandBuffer { // Implementation details device kernels required by GpuCommandBuffer. //===----------------------------------------------------------------------===// -// A no-op kernel required for creating barriers inside command buffers because -// empty nodes are not supported within conditional CUDA graphs (in CUDA 12.3). -void* GetNoOpKernel(); - // See `cuda_conditional_kernels.cc` for CUDA implementation. These are // various kernels that update Gpu conditionals based on the device memory // values, and allow implementing on-device control flow via conditional command