🐛 [Bug] [weight-stripped engine] doesn't work for torch.compile

[zewenli98]

Bug Description

There are three interfaces to compile a model: 1) torch export + torch_trt.dynamo.compile 2) torch_trt.compile(ir="dynamo") 3) torch.compile(backend="tensorrt").
The PR #3167 is supporting weight-stripped engines, which works for 1) and 2) but not 3). Not sure the reason.

I observed the issue in the test:

TensorRT/tests/py/dynamo/models/test_weight_stripped_engine.py

Lines 83 to 115 in 76bdf5e

	# 1. Compile with torch_trt.dynamo.compile
	gm1 = torch_trt.dynamo.compile(
	exp_program,
	example_inputs,
	**settings,
	)
	gm1_output = gm1(*example_inputs)
	# 2. Compile with torch_trt.compile using dynamo backend
	gm2 = torch_trt.compile(
	pyt_model, ir="dynamo", inputs=example_inputs, **settings
	)
	gm2_output = gm2(*example_inputs)
	# 3. Compile with torch.compile using tensorrt backend
	gm3 = torch.compile(
	pyt_model,
	backend="tensorrt",
	options=settings,
	)
	gm3_output = gm3(*example_inputs)
	assertions.assertEqual(
	gm1_output.sum(), 0, msg="gm1_output should be all zeros"
	)
	assertions.assertEqual(
	gm2_output.sum(), 0, msg="gm2_output should be all zeros"
	)
	assertions.assertEqual(
	gm3_output.sum(), 0, msg="gm3_output should be all zeros"
	)

The CI test reports the error:

FAILED models/test_weight_stripped_engine.py::TestWeightStrippedEngine::test_three_ways_to_compile_weight_stripped_engine - AssertionError: tensor(0.5406, device='cuda:0') != 0 : gm3_output should be all zeros

Only torch.compile failed so I output the result of gm3_output, which is the same as the output of the weight-included engine. So I suspect the engine was not getting its weights stripped. I also tried torch_trt.compile(ir="torch_compile") which is just another way to call torch.compile so it failed as well. Not sure the reason.

[narendasan]

Have you pulled the engines out to see what they look like?

[zewenli98]

@narendasan It seems like the main difference between workflows of torch.compile() and torch_trt.compile() is that torch.compile() calls py/torch_tensorrt/dynamo/backend/backends.py::pretraced_backend() while torch_trt.compile() calls py/torch_tensorrt/_compile.py::compile() before both calling py/torch_tensorrt/dynamo/_compiler.py::compile_module().

I pulled the engine out and outputs some attributes.

• When building the engine with strip_engine_weights=True:

  • torch.compile():
    • engine.refittable: True
      engine.num_layers: 186
      engine size: 2038852
  • torch_trt.compile():
    engine.refittable: True
    engine.num_layers: 276
    engine size: 2802668

• When building the engine with strip_engine_weights=False:

  • torch.compile():
    • engine.refittable: True
      engine.num_layers: 186
      engine size: 2038332
  • torch_trt.compile():
    • engine.refittable: True
      engine.num_layers: 276
      engine size: 77122580

We can see that torch.compile() has less num of layers and its engine sizes are almost same, but we expect the engine size to be smaller when strip_engine_weights=True. Is this because of the pretraced_backend()?

[narendasan]

torch.compile and torch.export graphs arent going to be identical that is expected. Is the argument being correct parsed from torch.compile(options=...)?

[zewenli98]

I'm calling three compiles like this:

settings = {
    "use_python_runtime": True,
    "enabled_precisions": {torch.float},
    "debug": False,
    "min_block_size": 1,
    "strip_engine_weights": True,
    "refit_identical_engine_weights": False,
}

# 1. Compile with torch_trt.dynamo.compile
gm1 = torch_trt.dynamo.compile(
    exp_program,
    example_inputs,
    **settings,
)
gm1_output = gm1(*example_inputs)

# 2. Compile with torch_trt.compile using dynamo backend
gm2 = torch_trt.compile(
    pyt_model, ir="dynamo", inputs=example_inputs, **settings
)
gm2_output = gm2(*example_inputs)

# 3. Compile with torch.compile using tensorrt backend
gm3 = torch.compile(
    pyt_model,
    backend="tensorrt",
    options=settings,
)
gm3_output = gm3(*example_inputs)

I also output strip_engine_weights after building the TRT engine, which is as expected.