TensorRT-LLM 0.5.0 源码之十七

sessoin.py

@contextlib.contextmanagerdef _scoped_stream():    '''Create a scoped cuda stream, and synchronize it when the context is destroyed    '''    #TODO: delete torch, use cuda native python bindings    import torch    stream = torch.cuda.current_stream()    try:        # return a handle, trt and other lib does not recognize torch.cuda.Stream        yield stream.cuda_stream    finally:        stream.synchronize()

@dataclassclass TensorInfo:    name: str    dtype: trt.DataType    shape: tuple    # add more info like strides, formats if needed

class Session(object):    ''' Session is a managed TensorRT runtime.  '''    def __init__(self, **kwargs):        # use Session.from_serialized_engine to create a session        pass    def _init(self, engine_buffer=None):        '''        @brief: Setup TensorRT engines and context from a serialized engine file        @param engine_buffer: a buffer holds the serialized TRT engine        '''        self._runtime = trt.Runtime(logger.trt_logger)        if engine_buffer is not None:            self._engine = self.runtime.deserialize_cuda_engine(engine_buffer)        self._context = self.engine.create_execution_context()        with _scoped_stream() as stream:            self._context.set_optimization_profile_async(0, stream)        return self    @staticmethod    def from_serialized_engine(engine) -> Session:        '''        @brief: Create a session from a serialized engine        @param engine: a serialized engine        @return: a Session object        '''        session = Session()        return session._init(engine)    @staticmethod    def from_engine(engine) -> Session:        '''        @brief: Create a session from an existing ICudaEngine engine        @param engine: an ICudaEngine        @return: a Session object        '''        session = Session()        session.engine = engine        return session._init()    @property    def runtime(self) -> trt.Runtime:        return self._runtime    @property    def engine(self) -> trt.ICudaEngine:        return self._engine    @engine.setter    def engine(self, engine: trt.ICudaEngine):        self._engine = engine    @property    def context(self) -> trt.IExecutionContext:        '''        @brief: Get the default TensorRT execution context,            use self.engine.create_execution_context() to create a new context if needed        @return: one TensorRT execution context object        '''        return self._context    def _print_engine_info(self):        '''print engine info for debug purpose, internal use only.        '''        refitable = self.engine.refittable        num_layers = self.engine.num_layers        device_memory_size = self.engine.device_memory_size        name = self.engine.name        nb_profiles = self.engine.num_optimization_profiles        logger.info(            f"Engine:{name=:}, {refitable=:}, {num_layers=:}, {device_memory_size=:}, {nb_profiles=:}"        )        self._print_io_info()    def _print_io_info(self):        '''print engine i/o info for debug purpose, internal use only.        '''        for i in range(self.engine.num_bindings):            name = self.engine.get_binding_name(i)            dtype = self.engine.get_binding_dtype(i)            shape = self.engine.get_binding_shape(i)            is_input = self.engine.binding_is_input(i)            logger.info(                f"Binding:{i=:}, {name=:}, {dtype=:}, {shape=:}, {is_input=:}")        for i in range(self.engine.num_io_tensors):            name = self.engine.get_tensor_name(i)            mode = self.engine.get_tensor_mode(name)            shape = self.engine.get_tensor_shape(name)            dtype = self.engine.get_tensor_dtype(name)            tformat = ";".join([                self.engine.get_tensor_format_desc(name, p)                for p in range(self.engine.num_optimization_profiles)            ])            logger.info(                f"Tensor:{name=:}, {mode=:}, {shape=:}, {dtype=:}, {tformat=:}")    def infer_shapes(self,                     inputs: List[TensorInfo],                     context=None) -> List[TensorInfo]:        '''        @brief: Set input shapes to given context, and infer the output shapes from the given input shapes.               This function should be called every time when the input shapes are changed before calling run().               Or call the context.set_input_shape on all dynamic shaped input tensors manually.        @param inputs: list of TensorInfo object, each item represents an input tensor        @param context: TensorRT execution context, if None, use the default context        @return: list of TensorInfo object, each item represents an output tensor, returns None if failed        '''        # set shape to the default context if context is not specified        if context is None:            context = self.context        for i in inputs:            if self.engine.get_tensor_mode(i.name) != trt.TensorIOMode.INPUT:                logger.error(f"Tensor:{i.name} is not an input tensor")                return None            if self.engine.get_tensor_dtype(i.name) != i.dtype:                logger.error(f"Tensor:{i.name} has wrong dtype")                return None            context.set_input_shape(i.name, i.shape)        outputs = []        for i in range(self.engine.num_io_tensors):            name = self.engine.get_tensor_name(i)            if self.engine.get_tensor_mode(name) == trt.TensorIOMode.OUTPUT:                shape = context.get_tensor_shape(name)                dtype = self.engine.get_tensor_dtype(name)                outputs.append(TensorInfo(name, dtype, shape))        return outputs    def run(self,            inputs: Dict[str, Any],            outputs: Dict[str, Any],            stream,            context=None) -> bool:        '''        @brief: Run the TensorRT engine with the given inputs and outputs        @param inputs: dict of input tensors, key is tensor name, value is tensor pointer or torch tensor        @param outputs: dict of output tensors, key is tensor name, value is tensor pointer or torch tensor        @param stream: cuda stream to enqueue the TensorRT engine on        @param context: TensorRT execution context, if None, use the default context        @return: True if enqueue succeeded, note the enqueue is an async call,            returning True does not mean the execution is finished        '''        # enqueue to the default context if context is not specified        if context is None:            context = self.context        import torch        for tensor_name in inputs:            tensor = inputs[tensor_name]            ptr = tensor.data_ptr() if isinstance(tensor,                                                  torch.Tensor) else tensor            context.set_tensor_address(tensor_name, ptr)        for tensor_name in outputs:            tensor = outputs[tensor_name]            ptr = tensor.data_ptr() if isinstance(tensor,                                                  torch.Tensor) else tensor            context.set_tensor_address(tensor_name, ptr)        ok = context.execute_async_v3(stream)        return ok    def _debug_run(self,                   inputs: Dict[str, "torch.Tensor"],                   context=None) -> Dict[str, "torch.Tensor"]:        '''Run the engine enqueue with allocated output tensors, for debug purpose, since it is a sync call and slower than run        '''        import torch        torch_dtype_to_trt = {            torch.float16: trt.float16,            torch.float32: trt.float32,            torch.int32: trt.int32        }        inputs_info = [            TensorInfo(name, torch_dtype_to_trt[tensor.dtype], tensor.shape)            for name, tensor in inputs.items()        ]        outputs_info = self.infer_shapes(inputs_info)        outputs = {            t.name: torch.empty(tuple(t.shape),                                dtype=trt_dtype_to_torch(t.dtype),                                device='cuda')            for t in outputs_info        }        with _scoped_stream() as stream:            self.run(inputs=inputs,                     outputs=outputs,                     stream=stream,                     context=context)        return outputs