2024-05-14 nightly release (c517fef)

Author: pytorchbot

.github/workflows/docgen.yml

@@ -11,6 +11,7 @@ on:
 jobs:
   build-docs:
     runs-on: linux.g5.4xlarge.nvidia.gpu
+    environment: pytorchbot-env
     container:
       image: docker.io/pytorch/manylinux-builder:cuda12.1
       options: --gpus all
@@ -22,6 +23,7 @@ jobs:
       - uses: actions/checkout@v3
         with:
           ref: ${{github.head_ref}}
+          token: ${{ secrets.GH_PYTORCHBOT_TOKEN }}
       - name: Select Python / CUDA
         run: |
           git config --global --add safe.directory /__w/TensorRT/TensorRT

.gitignore

@@ -69,3 +69,8 @@ bazel-tensorrt
 bazel-project
 build/
 wheelhouse/
+*_status.json
+tests/py/dynamo/models/*.ts
+tests/py/dynamo/models/*.ep
+*.deb
+*.tar.xz

docs/_cpp_api/classtorch__tensorrt_1_1DataType.html

@@ -43,7 +43,7 @@ def forward(self, x: torch.Tensor, y: torch.Tensor):
 # For the default settings, we can simply call torch.compile
 # with the backend "torch_tensorrt", and run the model on an
 # input to cause compilation, as so:
-optimized_model = torch.compile(model, backend="torch_tensorrt")
+optimized_model = torch.compile(model, backend="torch_tensorrt", dynamic=False)
 optimized_model(*sample_inputs)
 
 # %%
@@ -81,7 +81,10 @@ def forward(self, x: torch.Tensor, y: torch.Tensor):
 
 # Run the model on an input to cause compilation, as so:
 optimized_model_custom = torch.compile(
-    model_half, backend="torch_tensorrt", options=backend_kwargs
+    model_half,
+    backend="torch_tensorrt",
+    options=backend_kwargs,
+    dynamic=False,
 )
 optimized_model_custom(*sample_inputs_half)
 

docs/_cpp_api/classtorch__tensorrt_1_1Device_1_1DeviceType.html

@@ -18,9 +18,8 @@
 # ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 import torch
-from diffusers import DiffusionPipeline
-
 import torch_tensorrt
+from diffusers import DiffusionPipeline
 
 model_id = "CompVis/stable-diffusion-v1-4"
 device = "cuda:0"
@@ -39,7 +38,7 @@
     backend=backend,
     options={
         "truncate_long_and_double": True,
-        "precision": torch.float16,
+        "enabled_precisions": {torch.float32, torch.float16},
     },
     dynamic=False,
 )

docs/_cpp_api/classtorch__tensorrt_1_1TensorFormat.html

@@ -62,7 +62,7 @@
       },
       "outputs": [],
       "source": [
-        "# Next, we compile the model using torch.compile\n# For the default settings, we can simply call torch.compile\n# with the backend \"torch_tensorrt\", and run the model on an\n# input to cause compilation, as so:\noptimized_model = torch.compile(model, backend=\"torch_tensorrt\")\noptimized_model(*sample_inputs)"
+        "# Next, we compile the model using torch.compile\n# For the default settings, we can simply call torch.compile\n# with the backend \"torch_tensorrt\", and run the model on an\n# input to cause compilation, as so:\noptimized_model = torch.compile(model, backend=\"torch_tensorrt\", dynamic=False)\noptimized_model(*sample_inputs)"
       ]
     },
     {
@@ -91,7 +91,7 @@
       },
       "outputs": [],
       "source": [
-        "# If we want to customize certain options in the backend,\n# but still use the torch.compile call directly, we can provide\n# custom options to the backend via the \"options\" keyword\n# which takes in a dictionary mapping options to values.\n#\n# For accepted backend options, see the CompilationSettings dataclass:\n# py/torch_tensorrt/dynamo/_settings.py\nbackend_kwargs = {\n    \"enabled_precisions\": {torch.half},\n    \"debug\": True,\n    \"min_block_size\": 2,\n    \"torch_executed_ops\": {\"torch.ops.aten.sub.Tensor\"},\n    \"optimization_level\": 4,\n    \"use_python_runtime\": False,\n}\n\n# Run the model on an input to cause compilation, as so:\noptimized_model_custom = torch.compile(\n    model_half, backend=\"torch_tensorrt\", options=backend_kwargs\n)\noptimized_model_custom(*sample_inputs_half)"
+        "# If we want to customize certain options in the backend,\n# but still use the torch.compile call directly, we can provide\n# custom options to the backend via the \"options\" keyword\n# which takes in a dictionary mapping options to values.\n#\n# For accepted backend options, see the CompilationSettings dataclass:\n# py/torch_tensorrt/dynamo/_settings.py\nbackend_kwargs = {\n    \"enabled_precisions\": {torch.half},\n    \"debug\": True,\n    \"min_block_size\": 2,\n    \"torch_executed_ops\": {\"torch.ops.aten.sub.Tensor\"},\n    \"optimization_level\": 4,\n    \"use_python_runtime\": False,\n}\n\n# Run the model on an input to cause compilation, as so:\noptimized_model_custom = torch.compile(\n    model_half,\n    backend=\"torch_tensorrt\",\n    options=backend_kwargs,\n    dynamic=False,\n)\noptimized_model_custom(*sample_inputs_half)"
       ]
     },
     {

docs/_cpp_api/classtorch__tensorrt_1_1ptq_1_1Int8CacheCalibrator.html

@@ -22,7 +22,7 @@
       },
       "outputs": [],
       "source": [
-        "import torch\nfrom diffusers import DiffusionPipeline\n\nimport torch_tensorrt\n\nmodel_id = \"CompVis/stable-diffusion-v1-4\"\ndevice = \"cuda:0\"\n\n# Instantiate Stable Diffusion Pipeline with FP16 weights\npipe = DiffusionPipeline.from_pretrained(\n    model_id, revision=\"fp16\", torch_dtype=torch.float16\n)\npipe = pipe.to(device)\n\nbackend = \"torch_tensorrt\"\n\n# Optimize the UNet portion with Torch-TensorRT\npipe.unet = torch.compile(\n    pipe.unet,\n    backend=backend,\n    options={\n        \"truncate_long_and_double\": True,\n        \"precision\": torch.float16,\n    },\n    dynamic=False,\n)"
+        "import torch\nimport torch_tensorrt\nfrom diffusers import DiffusionPipeline\n\nmodel_id = \"CompVis/stable-diffusion-v1-4\"\ndevice = \"cuda:0\"\n\n# Instantiate Stable Diffusion Pipeline with FP16 weights\npipe = DiffusionPipeline.from_pretrained(\n    model_id, revision=\"fp16\", torch_dtype=torch.float16\n)\npipe = pipe.to(device)\n\nbackend = \"torch_tensorrt\"\n\n# Optimize the UNet portion with Torch-TensorRT\npipe.unet = torch.compile(\n    pipe.unet,\n    backend=backend,\n    options={\n        \"truncate_long_and_double\": True,\n        \"enabled_precisions\": {torch.float32, torch.float16},\n    },\n    dynamic=False,\n)"
       ]
     },
     {