| 122 | |
| 123 | |
| 124 | class Network(object): |
| 125 | |
| 126 | def __init__(self, **kwargs): |
| 127 | # intentionally use **kwargs, user should never call this ctor directly |
| 128 | # use Builder.create_network() instead |
| 129 | |
| 130 | # Holds the removed layers and disable them in graph rewriting and other phases. |
| 131 | # This is a hacky way since INetwork python API doesn't provide a way to remove a layer. |
| 132 | # TODO: remove this when TensorRT provides a better way to remove a layer |
| 133 | self._removed_layers: Set[str] = set() |
| 134 | |
| 135 | self.is_graph_altered = False |
| 136 | |
| 137 | from .graph_rewriting import FLayerInfoMemo |
| 138 | self.flayer_memo = FLayerInfoMemo() # holds the functional metadata |
| 139 | self._parameter_tensors = {} # holds the parameter tensors |
| 140 | |
| 141 | def _init(self, trt_network): |
| 142 | self._trt_network = trt_network |
| 143 | self._inputs = {} |
| 144 | self._named_parameters = None |
| 145 | # layer precision of a given scope, this is used together with precision(dtype) context manager |
| 146 | self._dtype = None |
| 147 | self._name_generator = _UniqueNameGenerator() |
| 148 | self._plugin_config = PluginConfig() |
| 149 | self._module_call_stack = _TrtLlmModuleCallStack() |
| 150 | self._registered_ndarrays = [] |
| 151 | self._strongly_typed = trt.INetworkDefinition.get_flag( |
| 152 | self._trt_network, trt.NetworkDefinitionCreationFlag.STRONGLY_TYPED) |
| 153 | self._unfilled_weights: Dict[str, Tuple[np.array, np.array]] = {} |
| 154 | |
| 155 | return self |
| 156 | |
| 157 | def _register_unfilled_weights(self, layer_name: str, weights: np.array, |
| 158 | values: np.array): |
| 159 | self._unfilled_weights[layer_name] = (weights, values) |
| 160 | |
| 161 | def _fill_weights(self): |
| 162 | from tensorrt_llm.parameter import Parameter |
| 163 | |
| 164 | for layer_name in list(self._unfilled_weights.keys()): |
| 165 | weights, values = self._unfilled_weights.pop(layer_name) |
| 166 | self.register_ndarray(weights) |
| 167 | if values is not None: |
| 168 | np.copyto(weights, values, casting='no') |
| 169 | else: |
| 170 | Parameter.xavier_init(weights) |
| 171 | |
| 172 | @property |
| 173 | def parameter_tensors(self): |
| 174 | return self._parameter_tensors |
| 175 | |
| 176 | def get_parameter_tensor(self, param): |
| 177 | return self.parameter_tensors.get(param, None) |
| 178 | |
| 179 | def set_parameter_tensor(self, param, tensor): |
| 180 | assert param not in self.parameter_tensors |
| 181 | self.parameter_tensors[param] = tensor |