| 96 | self.enable_pd_reorder = False |
| 97 | |
| 98 | def init_share_inputs(self): |
| 99 | max_num_seqs = self.scheduler_config.max_num_seqs |
| 100 | |
| 101 | self.pre_ids = paddle.full( |
| 102 | [max_num_seqs, self.model_config.max_model_len], |
| 103 | -1, |
| 104 | dtype="int64", |
| 105 | ) |
| 106 | self.input_ids = paddle.full( |
| 107 | [max_num_seqs, self.model_config.max_model_len], |
| 108 | self.model_config.pad_token_id, |
| 109 | dtype="int64", |
| 110 | ) |
| 111 | self.prompt_ids = paddle.full( |
| 112 | [max_num_seqs, self.model_config.max_model_len], |
| 113 | self.model_config.pad_token_id, |
| 114 | dtype="int64", |
| 115 | ) |
| 116 | self.eos_token_id = paddle.full([self.model_config.eos_tokens_lens, 1], 0, dtype="int64") |
| 117 | self.top_p = paddle.full([max_num_seqs, 1], self.model_config.top_p, dtype="float32") |
| 118 | self.top_k = paddle.full([max_num_seqs, 1], 0, dtype="int64") |
| 119 | self.top_k_list = [0] * max_num_seqs |
| 120 | self.min_p = paddle.full([max_num_seqs, 1], 0.0, dtype="float32") |
| 121 | self.min_p_list = [0.0] * max_num_seqs |
| 122 | self.temperature = paddle.full([max_num_seqs, 1], self.model_config.temperature, dtype="float32") |
| 123 | self.penalty_score = paddle.full([max_num_seqs, 1], self.model_config.penalty_score, dtype="float32") |
| 124 | self.frequency_score = paddle.full( |
| 125 | [max_num_seqs, 1], |
| 126 | self.model_config.frequency_score, |
| 127 | dtype="float32", |
| 128 | ) |
| 129 | self.presence_score = paddle.full([max_num_seqs, 1], self.model_config.presence_score, dtype="float32") |
| 130 | self.temp_scaled_logprobs = paddle.full([max_num_seqs, 1], False, dtype="bool") |
| 131 | self.top_p_normalized_logprobs = paddle.full([max_num_seqs, 1], False, dtype="bool") |
| 132 | |
| 133 | self.min_dec_len = paddle.full([max_num_seqs, 1], self.model_config.min_length, dtype="int64") |
| 134 | self.max_dec_len = paddle.full([max_num_seqs, 1], self.model_config.max_model_len, dtype="int64") |
| 135 | self.seq_lens_this_time_buffer = paddle.full([max_num_seqs, 1], 0, dtype="int32") |
| 136 | if self.enable_expert_parallel: |
| 137 | self.seq_lens_this_time = paddle.full([max_num_seqs, 1], 0, dtype="int32") |
| 138 | self.seq_lens_encoder = paddle.full([max_num_seqs, 1], 0, dtype="int32") |
| 139 | self.seq_lens_decoder = paddle.full([max_num_seqs, 1], 0, dtype="int32") |
| 140 | self.step_seq_lens_encoder = paddle.full([max_num_seqs, 1], 0, dtype="int32") |
| 141 | self.step_seq_lens_decoder = paddle.full([max_num_seqs, 1], 0, dtype="int32") |
| 142 | self.prompt_lens = paddle.full([max_num_seqs, 1], 0, dtype="int64") |
| 143 | self.step_idx = paddle.full([max_num_seqs, 1], 0, dtype="int64") |
| 144 | if current_platform.is_maca(): |
| 145 | self.not_need_stop = paddle.full([1], False, dtype="bool").cpu() |
| 146 | self.sampled_token_ids = paddle.full([max_num_seqs, 1], -1, dtype="int64").cpu() |
| 147 | self.seq_lens_this_time_cpu = paddle.full([max_num_seqs, 1], 0, dtype="int32").cpu() |
| 148 | self.is_block_step_cpu = paddle.full([max_num_seqs], False, dtype="bool").cpu() |
| 149 | else: |
| 150 | self.not_need_stop = paddle.full([1], False, dtype="bool").pin_memory() |
| 151 | self.sampled_token_ids = paddle.full([max_num_seqs, 1], -1, dtype="int64").pin_memory() |
| 152 | self.seq_lens_this_time_cpu = paddle.full([max_num_seqs, 1], 0, dtype="int32").pin_memory() |
| 153 | self.is_block_step_cpu = paddle.full([max_num_seqs], False, dtype="bool").pin_memory() |
| 154 | self.not_need_stop_device = paddle.full([1], False, dtype="bool") |
| 155 | self.stop_flags = paddle.full([max_num_seqs, 1], True, dtype="bool") |