Movatterモバイル変換

[0]ホーム
URL:

Skip to content

Navigation Menu

Sign in
Appearance settings

Search code, repositories, users, issues, pull requests...

Provide feedback

We read every piece of feedback, and take your input very seriously.

Saved searches

Use saved searches to filter your results more quickly

 Sign up
Appearance settings

Commit7022b29

Browse files
committed
Added back shape_changed
1 parentd6ed44e commit7022b29

File tree

2 files changed

+9
-5
lines changed

2 files changed

+9
-5
lines changed

‎core/runtime/execute_engine.cpp‎

Lines changed: 5 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -96,7 +96,8 @@ void setup_input_tensors(
9696
std::vector<at::Tensor> inputs,
9797
c10::intrusive_ptr<TRTEngine> compiled_engine,
9898
bool cudagraphs_enabled,
99-
bool need_cudagraphs_record) {
99+
bool need_cudagraphs_record,
100+
bool shape_changed) {
100101
// this is a buffer to store shape tensor input addresses throughout the runtime scope
101102
std::list<std::vector<int64_t>> inputShapeTensorValues;
102103
std::list<at::Tensor>formatted_inputs(compiled_engine->num_io.first);
@@ -145,7 +146,7 @@ void setup_input_tensors(
145146
// Create a new persistent input buffer
146147
compiled_engine->input_buffers[i] =std::move(formatted_inputs.back().clone());
147148
}
148-
if (need_cudagraphs_record) {
149+
if (shape_changed) {
149150
TORCHTRT_CHECK(
150151
compiled_engine->exec_ctx->setInputShape(name.c_str(), dims),"Error while setting the input shape");
151152
}
@@ -226,7 +227,7 @@ std::vector<at::Tensor> execute_engine(std::vector<at::Tensor> inputs, c10::intr
226227
input_profiler_guard =
227228
std::make_unique<torch::autograd::profiler::RecordProfile>(compiled_engine->input_profile_path);
228229
}
229-
setup_input_tensors(inputs, compiled_engine, cudagraphs_enabled, need_cudagraphs_record);
230+
setup_input_tensors(inputs, compiled_engine, cudagraphs_enabled, need_cudagraphs_record, shape_changed);
230231
// Check if input shapes can be inferred.
231232
int32_tconst io_size{compiled_engine->io_size};
232233
std::vector<charconst*>names(io_size);
@@ -361,7 +362,7 @@ std::vector<at::Tensor> execute_engine(std::vector<at::Tensor> inputs, c10::intr
361362
std::make_unique<torch::autograd::profiler::RecordProfile>(compiled_engine->input_profile_path);
362363
}
363364

364-
setup_input_tensors(inputs, compiled_engine,false,false);
365+
setup_input_tensors(inputs, compiled_engine,false,false,true);
365366
// Check if input shapes can be inferred.
366367
int32_tconst io_size{compiled_engine->cuda_engine->getNbIOTensors()};
367368
std::vector<charconst*>names(io_size);

‎py/torch_tensorrt/dynamo/runtime/_PythonTorchTensorRTModule.py‎

Lines changed: 4 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -384,6 +384,7 @@ def setup_input_tensors(
384384
contiguous_inputs:List[torch.Tensor],
385385
cudagraphs_enabled:bool,
386386
need_cudagraphs_record:bool,
387+
shape_changed:bool=True,
387388
)->None:
388389
fori,input_nameinenumerate(self.input_names):
389390
ifnotcontiguous_inputs[i].is_cuda:
@@ -417,7 +418,7 @@ def setup_input_tensors(
417418
inputs_cpu=contiguous_inputs[i].cpu().to(torch.int64).numpy().copy()
418419
self.context.set_tensor_address(input_name,inputs_cpu.ctypes.data)
419420
else:
420-
ifneed_cudagraphs_record:
421+
ifshape_changed:
421422
self.context.set_input_shape(
422423
input_name,tuple(contiguous_inputs[i].shape)
423424
)
@@ -491,6 +492,8 @@ def run_standard_execution() -> torch.Tensor | Tuple[torch.Tensor, ...]:
491492
contiguous_inputs,
492493
self.cudagraphs_enabled,
493494
need_cudagraphs_record,
495+
shape_changed
496+
orself.output_tensorsisNone,# First time execution
494497
)
495498

496499
ifshape_changed:

0 commit comments

Comments
 (0)
[8]ページ先頭
©2009-2025 Movatter.jp