Revert "Deallocate HostIr Op and Test"

[wujingyue]

Reverts #4286 because of http://nv/eF0

[wujingyue]

!test

[github-actions]

Description

• Reverts removal of HostIr Deallocate Op and Test

• Restores Deallocate handling in HostIrEvaluator

• Removes unused preallocated_outputs check

• Updates dispatch.h and host_ir.h to include Deallocate

---

Changes walkthrough 📝

Relevant files

Enhancement

executor.cpp Restore Deallocate Handling                                                            csrc/host_ir/executor.cpp Restored Deallocate handling in HostIrEvaluator Removed preallocated_outputs check and related code +12/-36  host_ir.cpp Remove Deallocate Class                                                                    csrc/host_ir/host_ir.cpp Removed Deallocate class definition and related methods +0/-23    test_host_ir_integration.cpp Remove Deallocate Test                                                                      tests/cpp/test_host_ir_integration.cpp Removed Deallocate test case +0/-26    dispatch.h Restore Deallocate in Dispatch                                                      csrc/dispatch.h Restored Deallocate in dispatch macro +1/-2      executor.h Restore Deallocate Method                                                                csrc/host_ir/executor.h Restored Deallocate method declaration +0/-1      host_ir.h Remove Deallocate Class                                                                    csrc/host_ir/host_ir.h Removed Deallocate class definition +0/-21

---

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🧪 No relevant tests

⚡ Recommended focus areas for review Output Handling The new code directly binds outputs without checking if they are preallocated, which could lead to issues if outputs are expected to be preallocated. KernelArgumentHolder outputs = container_->getKernelExecutor(launch_kernel->getIndex()) ->run( args, {}, launch_kernel->launch_params(), launch_kernel->compile_params()); // Store the outputs in the context for (auto output_idx : arange(outputs.size())) { bind(launch_kernel->outputs().at(output_idx), outputs[output_idx]); } Removed Functionality The Deallocate class and its related methods have been removed, which might affect memory management and could lead to memory leaks if not handled elsewhere. std::string LaunchKernel::toInlineString(int indent_size) const { NVF_CHECK(false, "Can not be printed inline"); } Stream::Stream(IrBuilderPasskey passkey, Val* index) : Val(passkey, ValType::Stream), index_(index) {} Stream::Stream(const Stream* src, IrCloner* ir_cloner) : Val(src, ir_cloner), index_(src->index()) {} NVFUSER_DEFINE_CLONE(Stream) std::string Stream::toString(int indent_size) const { std::stringstream ss; indent(ss, indent_size) << "Stream "; if (index() == nullptr) { ss << name(); } else { ss << index()->toInlineString(); } return ss.str(); } std::string Stream::toInlineString(int indent_size) const { return toString(indent_size); } bool Stream::sameAs(const Statement* other) const { if (other == this) { return true; } Removed Test The Deallocate test has been removed, which could lead to a gap in testing memory deallocation functionality. {in_tensor}, __LINE__, __FILE__, ""); } } // namespace hir } // namespace nvfuser

[wujingyue]

@nsarka I suspect you need something like

Fuser/tests/cpp/test_runtime.cpp

Lines 27 to 29 in ce3d607

	at::cuda::clearCublasWorkspaces();
	releaseZeroedMemory();
	ASSERT_EQ(memoryAllocated(0), 0) << "Previous tests leaked memory.";

[jjsjann123]

🙇