Skip to content

Enable inference mode by default #105

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
May 15, 2023
Merged

Enable inference mode by default #105

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -123,7 +123,7 @@ key: "DISABLE_OPTIMIZED_EXECUTION"
```

* `INFERENCE_MODE`: Boolean flag to enable the Inference Mode execution
of TorchScript models. By default, the inference mode is disabled.
of TorchScript models. By default, the inference mode is enabled.

[InferenceMode](https://pytorch.org/cppdocs/notes/inference_mode.html) is a new
RAII guard analogous to NoGradMode to be used when you are certain your operations
Expand Down
28 changes: 15 additions & 13 deletions src/libtorch.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -176,7 +176,7 @@ ModelState::Create(TRITONBACKEND_Model* triton_model, ModelState** state)

ModelState::ModelState(TRITONBACKEND_Model* triton_model)
: BackendModel(triton_model), enable_optimized_execution_(true),
enable_inference_mode_(false), enable_cache_cleaning_(false),
enable_inference_mode_(true), enable_cache_cleaning_(false),
enable_weight_sharing_(false), enable_tensor_fuser_pair_({false, true}),
enable_jit_profiling_pair_({false, true}),
enable_jit_executor_pair_({false, true}),
Expand Down Expand Up @@ -1312,12 +1312,12 @@ ModelInstanceState::Execute(
torch::jit::overrideCanFuseOnCPU(false);
torch::jit::overrideCanFuseOnGPU(false);
torch::jit::setTensorExprFuserEnabled(false);
torch::jit::fuser::cuda::setEnabled(true);
torch::jit::fuser::cuda::setEnabled(true);
} else {
torch::jit::overrideCanFuseOnCPU(true);
torch::jit::overrideCanFuseOnGPU(true);
torch::jit::setTensorExprFuserEnabled(true);
torch::jit::fuser::cuda::setEnabled(false);
torch::jit::fuser::cuda::setEnabled(false);
}
}

Expand Down Expand Up @@ -1761,9 +1761,9 @@ ModelInstanceState::SetInputTensors(

batchn_shape[0] += GetElementCount(input_shape, input_dims_count);
}
}
else {
batchn_shape = std::vector<int64_t>(input_shape, input_shape + input_dims_count);
} else {
batchn_shape =
std::vector<int64_t>(input_shape, input_shape + input_dims_count);
if (supports_batching_) {
batchn_shape[0] = total_batch_size;
}
Expand Down Expand Up @@ -1887,9 +1887,11 @@ ModelInstanceState::ReadOutputTensors(

// Output tensors may not reside on the same device as model
torch::Device tensor_device = output_flat.device();
const auto memory_type = (tensor_device.type() == torch::kCPU) ? TRITONSERVER_MEMORY_CPU
: TRITONSERVER_MEMORY_GPU;
const auto memory_id = (tensor_device.type() == torch::kCPU) ? 0 : tensor_device.index();
const auto memory_type = (tensor_device.type() == torch::kCPU)
? TRITONSERVER_MEMORY_CPU
: TRITONSERVER_MEMORY_GPU;
const auto memory_id =
(tensor_device.type() == torch::kCPU) ? 0 : tensor_device.index();

// Batch output doesn't support string data type yet, as it is not trivial
// to parse string output
Expand All @@ -1906,16 +1908,16 @@ ModelInstanceState::ReadOutputTensors(
return TRITONSERVER_ErrorNew(
TRITONSERVER_ERROR_INVALID_ARG,
(std::string("output '") + name +
"' is a scalar which is not supported.")
"' is a scalar which is not supported.")
.c_str());
}

responder.ProcessTensor(
name, output_dtype, batchn_shape, output_buffer,
memory_type, memory_id);
name, output_dtype, batchn_shape, output_buffer, memory_type,
memory_id);
} else {
responder.ProcessBatchOutput(
name, *batch_output, output_buffer, memory_type, memory_id);
name, *batch_output, output_buffer, memory_type, memory_id);
}
} else if (output_tensors[op_index].isList()) {
// Custom handling for string/bytes tensor...
Expand Down