feat: lowering replace aten.full_like with aten.full

[chohk88]

Description

1. Lowering: Replaced aten.full_like with aten.full.
2. Decomposition: For aten.scatter_add, replaced torch.zeros_like with torch.zeros.
3. Removed expected_ops_param for the aten.scatter_add test case.
4. Linting for py/torch_tensorrt/dynamo/lowering/_decompositions.py file.

Type of change

Please delete options that are not relevant and/or add your own.

• New feature (non-breaking change which adds functionality)

Checklist:

• My code follows the style guidelines of this project (You can use the linters)
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas and hacks
• I have made corresponding changes to the documentation
• I have added tests to verify my fix or my feature
• New and existing unit tests pass locally with my changes
• I have added the relevant labels to my PR in so that relevant reviewers are notified

[chohk88]

One concern is that the test case takes a significant amount of runtime, as shown below:

DEBUG: [Torch-TensorRT - Debug Build] - Torch-TensorRT TensorRT Engine:
  Name: _run_on_acc_0_engine
  Inputs: [
  ]
  Outputs: [
    id: 0
      name: output0
      shape: [3, 3]
      dtype: Float
  ]
  Device: Device(ID: 0, Name: NVIDIA GeForce RTX 3080 Ti, SM Capability: 8.6, Type: GPU)
  Hardware Compatibility: Disabled

DEBUG: [Torch-TensorRT - Debug Build] - Attempting to run engine (ID: _run_on_acc_0_engine); Hardware Compatible: 0
INFO: [Torch-TensorRT - Debug Build] - Execution profiling is enabled, find results here:
  Device selection profile: /tmp/_run_on_acc_0_engine_device_config_profile.trace
  Input packing profile: /tmp/_run_on_acc_0_engine_input_profile.trace
  Output packing profile: /tmp/_run_on_acc_0_engine_output_profile.trace
  TRT enqueue profile: /tmp/_run_on_acc_0_engine_enqueue_profile.trace
  Engine execution profile: /tmp/_run_on_acc_0_engine_engine_exectuion_profile.trace

DEBUG: [Torch-TensorRT - Debug Build] - Output Name: output0 Shape: [3, 3]
INFO: [Torch-TensorRT - Debug Build] - 
========== _run_on_acc_0_engine profile ==========
                                                   TensorRT layer name    Runtime, %  Invocations  Runtime, ms
                                                                output          0.0%            1         0.00
                   Reformatting CopyNode for Output Tensor 0 to output        100.0%            1         2.28
========== _run_on_acc_0_engine total runtime = 2.27635 ms ==========

INFO: [Torch-TensorRT - Debug Build] - The profiling verbosity was set to ProfilingVerbosity::kLAYER_NAMES_ONLY when the engine was built, so only the layer names will be returned. Rebuild the engine with ProfilingVerbosity::kDETAILED to get more verbose layer information.
..INFO:torch_tensorrt.dynamo.utils:Using Default Torch-TRT Runtime (as requested by user)
INFO:torch_tensorrt.dynamo.utils:Device not specified, using Torch default current device - cuda:0. If this is incorrect, please specify an input device, via the device keyword.
INFO:torch_tensorrt.dynamo.utils:Compilation Settings: CompilationSettings(enabled_precisions={<dtype.f32: 7>}, debug=False, workspace_size=0, min_block_size=1, torch_executed_ops=set(), pass_through_build_failures=True, max_aux_streams=None, version_compatible=False, optimization_level=None, use_python_runtime=False, truncate_double=False, use_fast_partitioner=True, enable_experimental_decompositions=False, device=Device(type=DeviceType.GPU, gpu_id=0), require_full_compilation=False, disable_tf32=False, assume_dynamic_shape_support=False, sparse_weights=False, make_refitable=False, engine_capability=<EngineCapability.STANDARD: 1>, num_avg_timing_iters=1, dla_sram_size=1048576, dla_local_dram_size=1073741824, dla_global_dram_size=536870912, dryrun=False, hardware_compatible=False, timing_cache_path='/tmp/timing_cache.bin', lazy_engine_init=False)

I think we could optimize here

TensorRT/py/torch_tensorrt/dynamo/conversion/impl/full.py

Line 28 in 6321710

return np.full(shape, fill_value)

[peri044]

Can you use "val" key instead of "tensor_meta" ? If "val" isn't available, then use tensor_meta

[peri044]

One concern is that the test case takes a significant amount of runtime, as shown below:

DEBUG: [Torch-TensorRT - Debug Build] - Torch-TensorRT TensorRT Engine:
  Name: _run_on_acc_0_engine
  Inputs: [
  ]
  Outputs: [
    id: 0
      name: output0
      shape: [3, 3]
      dtype: Float
  ]
  Device: Device(ID: 0, Name: NVIDIA GeForce RTX 3080 Ti, SM Capability: 8.6, Type: GPU)
  Hardware Compatibility: Disabled

DEBUG: [Torch-TensorRT - Debug Build] - Attempting to run engine (ID: _run_on_acc_0_engine); Hardware Compatible: 0
INFO: [Torch-TensorRT - Debug Build] - Execution profiling is enabled, find results here:
  Device selection profile: /tmp/_run_on_acc_0_engine_device_config_profile.trace
  Input packing profile: /tmp/_run_on_acc_0_engine_input_profile.trace
  Output packing profile: /tmp/_run_on_acc_0_engine_output_profile.trace
  TRT enqueue profile: /tmp/_run_on_acc_0_engine_enqueue_profile.trace
  Engine execution profile: /tmp/_run_on_acc_0_engine_engine_exectuion_profile.trace

DEBUG: [Torch-TensorRT - Debug Build] - Output Name: output0 Shape: [3, 3]
INFO: [Torch-TensorRT - Debug Build] - 
========== _run_on_acc_0_engine profile ==========
                                                   TensorRT layer name    Runtime, %  Invocations  Runtime, ms
                                                                output          0.0%            1         0.00
                   Reformatting CopyNode for Output Tensor 0 to output        100.0%            1         2.28
========== _run_on_acc_0_engine total runtime = 2.27635 ms ==========

INFO: [Torch-TensorRT - Debug Build] - The profiling verbosity was set to ProfilingVerbosity::kLAYER_NAMES_ONLY when the engine was built, so only the layer names will be returned. Rebuild the engine with ProfilingVerbosity::kDETAILED to get more verbose layer information.
..INFO:torch_tensorrt.dynamo.utils:Using Default Torch-TRT Runtime (as requested by user)
INFO:torch_tensorrt.dynamo.utils:Device not specified, using Torch default current device - cuda:0. If this is incorrect, please specify an input device, via the device keyword.
INFO:torch_tensorrt.dynamo.utils:Compilation Settings: CompilationSettings(enabled_precisions={<dtype.f32: 7>}, debug=False, workspace_size=0, min_block_size=1, torch_executed_ops=set(), pass_through_build_failures=True, max_aux_streams=None, version_compatible=False, optimization_level=None, use_python_runtime=False, truncate_double=False, use_fast_partitioner=True, enable_experimental_decompositions=False, device=Device(type=DeviceType.GPU, gpu_id=0), require_full_compilation=False, disable_tf32=False, assume_dynamic_shape_support=False, sparse_weights=False, make_refitable=False, engine_capability=<EngineCapability.STANDARD: 1>, num_avg_timing_iters=1, dla_sram_size=1048576, dla_local_dram_size=1073741824, dla_global_dram_size=536870912, dryrun=False, hardware_compatible=False, timing_cache_path='/tmp/timing_cache.bin', lazy_engine_init=False)

I think we could optimize here

TensorRT/py/torch_tensorrt/dynamo/conversion/impl/full.py

Line 28 in 6321710

return np.full(shape, fill_value)

Do you mean the static case which uses np.full takes more time or the dynamic one ? For the static, I'm not sure what control we have for optimization. Also, there would be variation b/w debug and release build. So, you might want to try the release build for perf measurement.

[peri044]

@chohk88 Opened a new issue for the perf improvement you were mentioning #3107