(self, model_config: ModelConfig)
| 53 | |
| 54 | class FFNWrapper(paddle.nn.Layer): |
| 55 | def __init__(self, model_config: ModelConfig): |
| 56 | super().__init__() |
| 57 | self.model_config = model_config |
| 58 | |
| 59 | self.intermediate_size = self.model_config.intermediate_size |
| 60 | self.hidden_size = self.model_config.hidden_size |
| 61 | self.prefix = "hahahha" |
| 62 | self.fd_config = FDConfig( |
| 63 | model_config=self.model_config, |
| 64 | parallel_config=ParallelConfig( |
| 65 | { |
| 66 | "tensor_parallel_size": 1, |
| 67 | "expert_parallel_size": 1, |
| 68 | "expert_parallel_rank": 0, |
| 69 | "data_parallel_size": 1, |
| 70 | } |
| 71 | ), |
| 72 | quant_config=BlockWiseFP8Config(weight_block_size=[128, 128]), |
| 73 | # quant_config = WINT8Config({}), |
| 74 | scheduler_config=SchedulerConfig({}), |
| 75 | cache_config=CacheConfig({}), |
| 76 | graph_opt_config=GraphOptimizationConfig({}), |
| 77 | load_config=LoadConfig({}), |
| 78 | ips="0.0.0.0", |
| 79 | ) |
| 80 | self.fd_config.parallel_config.tp_group = None |
| 81 | self.fd_config.parallel_config.tensor_parallel_rank = 0 |
| 82 | self.fd_config.parallel_config.tensor_parallel_size = 1 |
| 83 | |
| 84 | self.ffn = Ernie4_5_MLP( |
| 85 | fd_config=self.fd_config, |
| 86 | intermediate_size=self.intermediate_size, |
| 87 | prefix=self.prefix, |
| 88 | ) |
| 89 | |
| 90 | up_gate_proj_weight_shape = [self.hidden_size, self.intermediate_size * 2] |
| 91 | down_proj_weight_shape = [self.intermediate_size, self.hidden_size] |
| 92 | |
| 93 | up_gate_proj_weight = paddle.randn(up_gate_proj_weight_shape, paddle.bfloat16) |
| 94 | down_proj_weight = paddle.randn(down_proj_weight_shape, paddle.bfloat16) |
| 95 | |
| 96 | state_dict = { |
| 97 | f"{self.prefix}.up_gate_proj.weight": up_gate_proj_weight, |
| 98 | f"{self.prefix}.down_proj.weight": down_proj_weight, |
| 99 | } |
| 100 | |
| 101 | self.ffn.load_state_dict(state_dict) |
| 102 | |
| 103 | |
| 104 | class TestFusedMoE(unittest.TestCase): |
nothing calls this directly
no test coverage detected