Feature/run jetson e2e tests (#4)

* Run tests

* Run tests

* Run tests

* Run tests

* Added rebuild

* Added files for e2e tests

* Added rebuild

* Added rebuild

* Added biuld TensorRT flag

* Changed to use NumPy 1.x for Jetson

* Make universal invocation

* Make Cython constans

* Changed to prepare onnx

* Changed smoke-test to wait AI conversion

* Added step for model conversion

* Changed to not run step in parallel

* Push model to docker registry

* Push model to docker registry

* Push model to docker registry

src/constants_inf.pxd

@@ -12,8 +12,8 @@ cdef str SPLIT_SUFFIX
 cdef double TILE_DUPLICATE_CONFIDENCE_THRESHOLD
 cdef int METERS_IN_TILE
 
-cdef log(str log_message)
-cdef logerror(str error)
+cpdef log(str log_message)
+cpdef logerror(str error)
 cdef format_time(long ms)
 
 cdef dict[int, AnnotationClass] annotations_dict

src/constants_inf.pyx

@@ -78,10 +78,10 @@ def get_annotation_name(int cls_id):
         return (<AnnotationClass>annotations_dict[cls_id]).name
     return ""
 
-cdef log(str log_message):
+cpdef log(str log_message):
     logger.info(log_message)
 
-cdef logerror(str error):
+cpdef logerror(str error):
     logger.error(error)
 
 cdef format_time(long ms):

src/engines/engine_factory.pyx

@@ -44,6 +44,10 @@ class EngineFactory:
     def build_and_cache(self, bytes source_bytes, LoaderHttpClient loader_client, str models_dir):
         cdef LoadResult res
         engine_bytes, engine_filename = self.build_from_source(source_bytes, loader_client, models_dir)
+        if engine_bytes is None:
+            raise RuntimeError("TensorRT conversion failed: no engine bytes produced")
+        if engine_filename is None:
+            raise RuntimeError("TensorRT conversion failed: engine filename could not be resolved")
         res = loader_client.upload_big_small_resource(engine_bytes, engine_filename, models_dir)
         if res.err is not None:
             constants_inf.log(f"Failed to upload converted model: {res.err}")
@@ -93,6 +97,22 @@ class JetsonTensorRTEngineFactory(TensorRTEngineFactory):
         from engines.jetson_tensorrt_engine import JetsonTensorRTEngine
         return JetsonTensorRTEngine(model_bytes)
 
+    def load_engine(self, LoaderHttpClient loader_client, str models_dir):
+        cdef str filename
+        cdef LoadResult res
+        from engines.tensorrt_engine import TensorRTEngine
+        for precision in ("int8", "fp16"):
+            filename = TensorRTEngine.get_engine_filename(precision)
+            if filename is None:
+                continue
+            try:
+                res = loader_client.load_big_small_resource(filename, models_dir)
+                if res.err is None:
+                    return self.create(res.data)
+            except Exception:
+                pass
+        return None
+
     def _get_ai_engine_filename(self):
         from engines.tensorrt_engine import TensorRTEngine
         return TensorRTEngine.get_engine_filename("int8")
@@ -100,5 +120,5 @@ class JetsonTensorRTEngineFactory(TensorRTEngineFactory):
     def build_from_source(self, onnx_bytes, LoaderHttpClient loader_client, str models_dir):
         from engines.jetson_tensorrt_engine import JetsonTensorRTEngine
         from engines.tensorrt_engine import TensorRTEngine
-        engine_bytes = JetsonTensorRTEngine.convert_from_source(onnx_bytes, loader_client, models_dir)
-        return engine_bytes, TensorRTEngine.get_engine_filename("int8")
+        engine_bytes, precision = JetsonTensorRTEngine.convert_from_source_with_precision(onnx_bytes, loader_client, models_dir)
+        return engine_bytes, TensorRTEngine.get_engine_filename(precision)

src/engines/jetson_tensorrt_engine.pyx

@@ -1,5 +1,6 @@
 import os
 import tempfile
+cimport constants_inf
 from engines.tensorrt_engine cimport TensorRTEngine
 from loader_http_client cimport LoaderHttpClient, LoadResult
 
@@ -7,10 +8,19 @@ from loader_http_client cimport LoaderHttpClient, LoadResult
 cdef class JetsonTensorRTEngine(TensorRTEngine):
     @staticmethod
     def convert_from_source(bytes onnx_model, LoaderHttpClient loader_client, str models_dir):
+        engine_bytes, precision = JetsonTensorRTEngine.convert_from_source_with_precision(
+            onnx_model, loader_client, models_dir
+        )
+        return engine_bytes
+
+    @staticmethod
+    def convert_from_source_with_precision(bytes onnx_model, LoaderHttpClient loader_client, str models_dir):
         cdef str calib_cache_path
         calib_cache_path = JetsonTensorRTEngine._download_calib_cache(loader_client, models_dir)
         try:
-            return TensorRTEngine.convert_from_source(onnx_model, calib_cache_path)
+            engine_bytes = TensorRTEngine.convert_from_source(onnx_model, calib_cache_path, True)
+            precision = "int8" if calib_cache_path is not None else "fp16"
+            return engine_bytes, precision
         finally:
             if calib_cache_path is not None:
                 try:
@@ -21,7 +31,6 @@ cdef class JetsonTensorRTEngine(TensorRTEngine):
     @staticmethod
     def _download_calib_cache(LoaderHttpClient loader_client, str models_dir):
         cdef LoadResult res
-        import constants_inf
         try:
             res = loader_client.load_big_small_resource(
                 constants_inf.INT8_CALIB_CACHE_FILE, models_dir

src/engines/onnx_tensorrt_compat.py

@@ -0,0 +1,111 @@
+import ast
+import io
+
+import onnx
+from onnx import helper, numpy_helper
+
+
+_REDUCE_OPS_WITH_AXES_INPUT = {
+    "ReduceL1",
+    "ReduceL2",
+    "ReduceLogSum",
+    "ReduceLogSumExp",
+    "ReduceMax",
+    "ReduceMean",
+    "ReduceMin",
+    "ReduceProd",
+    "ReduceSum",
+    "ReduceSumSquare",
+}
+
+
+def _metadata(model):
+    return {p.key: p.value for p in model.metadata_props}
+
+
+def _input_size(model):
+    try:
+        imgsz = _metadata(model).get("imgsz")
+        parsed = ast.literal_eval(imgsz)
+        if isinstance(parsed, (list, tuple)) and len(parsed) == 2:
+            h, w = int(parsed[0]), int(parsed[1])
+            if h > 0 and w > 0:
+                return h, w
+    except Exception:
+        pass
+    return 1280, 1280
+
+
+def _constant_values(graph):
+    values = {init.name: numpy_helper.to_array(init) for init in graph.initializer}
+    for node in graph.node:
+        if node.op_type != "Constant" or not node.output:
+            continue
+        for attr in node.attribute:
+            if attr.name == "value":
+                values[node.output[0]] = numpy_helper.to_array(attr.t)
+                break
+    return values
+
+
+def _as_int_list(value):
+    if value is None:
+        return None
+    if getattr(value, "shape", ()) == ():
+        return [int(value)]
+    return [int(v) for v in value.reshape(-1).tolist()]
+
+
+def _set_static_input_shape(model, batch=1):
+    h, w = _input_size(model)
+    for graph_input in model.graph.input:
+        tensor_type = graph_input.type.tensor_type
+        if tensor_type.elem_type != onnx.TensorProto.FLOAT:
+            continue
+        dims = tensor_type.shape.dim
+        if len(dims) != 4:
+            continue
+        for dim, value in zip(dims, (batch, 3, h, w)):
+            dim.dim_value = value
+        return True
+    return False
+
+
+def _rewrite_reduce_axes_inputs(model):
+    constants = _constant_values(model.graph)
+    changed = False
+    for node in model.graph.node:
+        if node.op_type not in _REDUCE_OPS_WITH_AXES_INPUT or len(node.input) < 2:
+            continue
+        axes = _as_int_list(constants.get(node.input[1]))
+        if axes is None:
+            continue
+        kept_attrs = [attr for attr in node.attribute if attr.name != "axes"]
+        del node.attribute[:]
+        node.attribute.extend(kept_attrs)
+        node.attribute.extend([helper.make_attribute("axes", axes)])
+        del node.input[1:]
+        changed = True
+    return changed
+
+
+def _cap_default_opset(model, max_opset=17):
+    for opset in model.opset_import:
+        if opset.domain in ("", "ai.onnx") and opset.version > max_opset:
+            opset.version = max_opset
+            return True
+    return False
+
+
+def prepare_for_tensorrt(model_bytes):
+    model = onnx.load_model_from_string(model_bytes)
+    changed = False
+    changed = _set_static_input_shape(model) or changed
+    changed = _rewrite_reduce_axes_inputs(model) or changed
+    changed = _cap_default_opset(model) or changed
+    if not changed:
+        return model_bytes
+
+    buffer = io.BytesIO()
+    onnx.save_model(model, buffer)
+    return buffer.getvalue()

src/engines/tensorrt_engine.pxd

@@ -4,6 +4,8 @@ from engines.inference_engine cimport InferenceEngine
 cdef class TensorRTEngine(InferenceEngine):
 
     cdef public object context
+    cdef object cuda_context
+    cdef object cuda_lock
 
     cdef public object d_input
     cdef public object d_output

src/engines/tensorrt_engine.pyx

@@ -1,10 +1,10 @@
 from engines.inference_engine cimport InferenceEngine
 import tensorrt as trt  # pyright: ignore[reportMissingImports]
 import pycuda.driver as cuda  # pyright: ignore[reportMissingImports]
-import pycuda.autoinit  # pyright: ignore[reportMissingImports]
 import pynvml
 import numpy as np
 import os
+import threading
 cimport constants_inf
 
 GPU_MEMORY_FRACTION = 0.8
@@ -32,48 +32,64 @@ class _CacheCalibrator(trt.IInt8EntropyCalibrator2):
 cdef class TensorRTEngine(InferenceEngine):
     def __init__(self, model_bytes: bytes, max_batch_size: int = 8, **kwargs):
         InferenceEngine.__init__(self, model_bytes, max_batch_size, engine_name="tensorrt")
+        self.cuda_context = TensorRTEngine.create_cuda_context()
+        self.cuda_lock = threading.Lock()
         try:
-            logger = trt.Logger(trt.Logger.WARNING)
-            runtime = trt.Runtime(logger)
-            engine = runtime.deserialize_cuda_engine(model_bytes)
-            if engine is None:
-                raise RuntimeError("Failed to load TensorRT engine from bytes")
+            with self.cuda_lock:
+                self.cuda_context.push()
+                try:
+                    logger = trt.Logger(trt.Logger.WARNING)
+                    runtime = trt.Runtime(logger)
+                    engine = runtime.deserialize_cuda_engine(model_bytes)
+                    if engine is None:
+                        raise RuntimeError("Failed to load TensorRT engine from bytes")
 
-            self.context = engine.create_execution_context()
+                    self.context = engine.create_execution_context()
 
-            self.input_name = engine.get_tensor_name(0)
-            engine_input_shape = engine.get_tensor_shape(self.input_name)
+                    self.input_name = engine.get_tensor_name(0)
+                    engine_input_shape = engine.get_tensor_shape(self.input_name)
 
-            C = engine_input_shape[1]
-            H = 1280 if engine_input_shape[2] == -1 else engine_input_shape[2]
-            W = 1280 if engine_input_shape[3] == -1 else engine_input_shape[3]
+                    C = engine_input_shape[1]
+                    H = 1280 if engine_input_shape[2] == -1 else engine_input_shape[2]
+                    W = 1280 if engine_input_shape[3] == -1 else engine_input_shape[3]
 
-            if engine_input_shape[0] == -1:
-                gpu_mem = TensorRTEngine.get_gpu_memory_bytes(0)
-                self.max_batch_size = TensorRTEngine.calculate_max_batch_size(gpu_mem, H, W)
-            else:
-                self.max_batch_size = engine_input_shape[0]
+                    if engine_input_shape[0] == -1:
+                        gpu_mem = TensorRTEngine.get_gpu_memory_bytes(0)
+                        self.max_batch_size = TensorRTEngine.calculate_max_batch_size(gpu_mem, H, W)
+                    else:
+                        self.max_batch_size = engine_input_shape[0]
 
-            self.input_shape = [self.max_batch_size, C, H, W]
-            self.context.set_input_shape(self.input_name, self.input_shape)
-            input_size = trt.volume(self.input_shape) * np.dtype(np.float32).itemsize
-            self.d_input = cuda.mem_alloc(input_size)
+                    self.input_shape = [self.max_batch_size, C, H, W]
+                    self.context.set_input_shape(self.input_name, self.input_shape)
+                    input_size = trt.volume(self.input_shape) * np.dtype(np.float32).itemsize
+                    self.d_input = cuda.mem_alloc(input_size)
 
-            self.output_name = engine.get_tensor_name(1)
-            engine_output_shape = tuple(engine.get_tensor_shape(self.output_name))
-            self.output_shape = [
-                self.max_batch_size,
-                300 if engine_output_shape[1] == -1 else engine_output_shape[1],
-                6 if engine_output_shape[2] == -1 else engine_output_shape[2],
-            ]
-            self.h_output = cuda.pagelocked_empty(tuple(self.output_shape), dtype=np.float32)
-            self.d_output = cuda.mem_alloc(self.h_output.nbytes)
-
-            self.stream = cuda.Stream()
+                    self.output_name = engine.get_tensor_name(1)
+                    engine_output_shape = tuple(engine.get_tensor_shape(self.output_name))
+                    self.output_shape = [
+                        self.max_batch_size,
+                        300 if engine_output_shape[1] == -1 else engine_output_shape[1],
+                        6 if engine_output_shape[2] == -1 else engine_output_shape[2],
+                    ]
+                    self.h_output = cuda.pagelocked_empty(tuple(self.output_shape), dtype=np.float32)
+                    self.d_output = cuda.mem_alloc(self.h_output.nbytes)
 
+                    self.stream = cuda.Stream()
+                finally:
+                    try:
+                        self.cuda_context.pop()
+                    except Exception:
+                        pass
         except Exception as e:
             raise RuntimeError(f"Failed to initialize TensorRT engine: {str(e)}")
 
+    def __dealloc__(self):
+        try:
+            if self.cuda_context is not None:
+                self.cuda_context.detach()
+        except Exception:
+            pass
+
     @staticmethod
     def calculate_max_batch_size(gpu_memory_bytes, int input_h, int input_w):
         frame_input_bytes = 3 * input_h * input_w * 4
@@ -99,9 +115,18 @@ cdef class TensorRTEngine(InferenceEngine):
                 pass
         return 2 * 1024 * 1024 * 1024 if total_memory is None else total_memory
 
+    @staticmethod
+    def create_cuda_context():
+        cuda.init()
+        from engines import tensor_gpu_index
+        ctx = cuda.Device(max(tensor_gpu_index, 0)).make_context()
+        ctx.pop()
+        return ctx
+
     @staticmethod
     def get_engine_filename(str precision="fp16"):
         try:
+            cuda.init()
             from engines import tensor_gpu_index
             device = cuda.Device(max(tensor_gpu_index, 0))
             sm_count = device.multiprocessor_count
@@ -114,82 +139,114 @@ cdef class TensorRTEngine(InferenceEngine):
             return None
 
     @staticmethod
-    def convert_from_source(bytes onnx_model, str calib_cache_path=None):
+    def convert_from_source(bytes onnx_model, str calib_cache_path=None, bint force_static_input=False):
+        cuda_context = TensorRTEngine.create_cuda_context()
+        cuda_context.push()
         gpu_mem = TensorRTEngine.get_gpu_memory_bytes(0)
         workspace_bytes = int(gpu_mem * 0.9)
 
         explicit_batch_flag = 1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)
         trt_logger = trt.Logger(trt.Logger.WARNING)
 
-        with trt.Builder(trt_logger) as builder, \
-                builder.create_network(explicit_batch_flag) as network, \
-                trt.OnnxParser(network, trt_logger) as parser, \
-                builder.create_builder_config() as config:
+        if force_static_input:
+            try:
+                from engines.onnx_tensorrt_compat import prepare_for_tensorrt
+                onnx_model = prepare_for_tensorrt(onnx_model)
+                constants_inf.log(<str>'Prepared ONNX model for TensorRT static Jetson build')
+            except Exception as e:
+                constants_inf.logerror(<str>f'ONNX TensorRT compatibility preparation failed: {str(e)}')
 
-            config.set_memory_pool_limit(trt.MemoryPoolType.WORKSPACE, workspace_bytes)
+        try:
+            with trt.Builder(trt_logger) as builder, \
+                    builder.create_network(explicit_batch_flag) as network, \
+                    trt.OnnxParser(network, trt_logger) as parser, \
+                    builder.create_builder_config() as config:
 
-            if not parser.parse(onnx_model):
-                return None
+                config.set_memory_pool_limit(trt.MemoryPoolType.WORKSPACE, workspace_bytes)
 
-            input_tensor = network.get_input(0)
-            shape = input_tensor.shape
-            C = shape[1]
-            H = max(shape[2], 1280) if shape[2] != -1 else 1280
-            W = max(shape[3], 1280) if shape[3] != -1 else 1280
+                if not parser.parse(onnx_model):
+                    for i in range(parser.num_errors):
+                        constants_inf.logerror(<str>f'TensorRT ONNX parser error: {parser.get_error(i)}')
+                    return None
 
-            if shape[0] == -1:
-                max_batch = TensorRTEngine.calculate_max_batch_size(gpu_mem, H, W)
-                profile = builder.create_optimization_profile()
-                profile.set_shape(
-                    input_tensor.name,
-                    (1, C, H, W),
-                    (max_batch, C, H, W),
-                    (max_batch, C, H, W),
-                )
-                config.add_optimization_profile(profile)
+                input_tensor = network.get_input(0)
+                shape = input_tensor.shape
+                C = shape[1]
+                H = max(shape[2], 1280) if shape[2] != -1 else 1280
+                W = max(shape[3], 1280) if shape[3] != -1 else 1280
 
-            use_int8 = calib_cache_path is not None and os.path.isfile(calib_cache_path)
-            if use_int8:
-                constants_inf.log(<str>'Converting to INT8 with calibration cache')
-                calibrator = _CacheCalibrator(calib_cache_path)
-                config.set_flag(trt.BuilderFlag.INT8)
-                if builder.platform_has_fast_fp16:
+                if force_static_input:
+                    input_tensor.shape = (1, C, H, W)
+                elif shape[0] == -1 or shape[2] == -1 or shape[3] == -1:
+                    max_batch = TensorRTEngine.calculate_max_batch_size(gpu_mem, H, W)
+                    profile = builder.create_optimization_profile()
+                    profile.set_shape(
+                        input_tensor.name,
+                        (1, C, H, W),
+                        (max_batch, C, H, W),
+                        (max_batch, C, H, W),
+                    )
+                    config.add_optimization_profile(profile)
+
+                use_int8 = calib_cache_path is not None and os.path.isfile(calib_cache_path)
+                if use_int8:
+                    constants_inf.log(<str>'Converting to INT8 with calibration cache')
+                    calibrator = _CacheCalibrator(calib_cache_path)
+                    config.set_flag(trt.BuilderFlag.INT8)
+                    if builder.platform_has_fast_fp16:
+                        config.set_flag(trt.BuilderFlag.FP16)
+                    config.int8_calibrator = calibrator
+                elif builder.platform_has_fast_fp16:
+                    constants_inf.log(<str>'Converting to supported fp16')
                     config.set_flag(trt.BuilderFlag.FP16)
-                config.int8_calibrator = calibrator
-            elif builder.platform_has_fast_fp16:
-                constants_inf.log(<str>'Converting to supported fp16')
-                config.set_flag(trt.BuilderFlag.FP16)
-            else:
-                constants_inf.log(<str>'Converting to supported fp32. (fp16 is not supported)')
+                else:
+                    constants_inf.log(<str>'Converting to supported fp32. (fp16 is not supported)')
 
-            plan = builder.build_serialized_network(network, config)
-            if plan is None:
-                constants_inf.logerror(<str>'Conversion failed.')
-                return None
-            constants_inf.log('conversion done!')
-            return bytes(plan)
+                plan = builder.build_serialized_network(network, config)
+                if plan is None:
+                    constants_inf.logerror(<str>'Conversion failed.')
+                    return None
+                constants_inf.log('conversion done!')
+                return bytes(plan)
+        finally:
+            try:
+                cuda_context.pop()
+            except Exception:
+                pass
+            try:
+                cuda_context.detach()
+            except Exception:
+                pass
 
     cdef tuple get_input_shape(self):
         return <tuple>(self.input_shape[2], self.input_shape[3])
 
     cdef run(self, input_data):
         try:
-            actual_batch = input_data.shape[0]
-            if actual_batch != self.input_shape[0]:
-                actual_shape = [actual_batch, self.input_shape[1], self.input_shape[2], self.input_shape[3]]
-                self.context.set_input_shape(self.input_name, actual_shape)
+            with self.cuda_lock:
+                self.cuda_context.push()
+                try:
+                    actual_batch = input_data.shape[0]
+                    if actual_batch != self.input_shape[0]:
+                        actual_shape = [actual_batch, self.input_shape[1], self.input_shape[2], self.input_shape[3]]
+                        self.context.set_input_shape(self.input_name, actual_shape)
 
-            cuda.memcpy_htod_async(self.d_input, input_data, self.stream)
-            self.context.set_tensor_address(self.input_name, int(self.d_input))
-            self.context.set_tensor_address(self.output_name, int(self.d_output))
+                    cuda.memcpy_htod_async(self.d_input, input_data, self.stream)
+                    self.context.set_tensor_address(self.input_name, int(self.d_input))
+                    self.context.set_tensor_address(self.output_name, int(self.d_output))
 
-            self.context.execute_async_v3(stream_handle=self.stream.handle)
-            self.stream.synchronize()
+                    self.context.execute_async_v3(stream_handle=self.stream.handle)
+                    self.stream.synchronize()
 
-            cuda.memcpy_dtoh(self.h_output, self.d_output)
-            output_shape = [actual_batch, self.output_shape[1], self.output_shape[2]]
-            output = self.h_output[:actual_batch].reshape(output_shape)
-            return [output]
+                    cuda.memcpy_dtoh(self.h_output, self.d_output)
+                    output_shape = [actual_batch, self.output_shape[1], self.output_shape[2]]
+                    output = self.h_output[:actual_batch].reshape(output_shape)
+                    return [output]
+                finally:
+                    try:
+                        self.cuda_context.pop()
+                    except Exception:
+                        pass
 
         except Exception as e:
             raise RuntimeError(f"Failed to run TensorRT inference: {str(e)}")