| 37 | |
| 38 | |
| 39 | class _DummyCfg: |
| 40 | def __init__( |
| 41 | self, |
| 42 | speculative_method=None, |
| 43 | enable_logprob=False, |
| 44 | max_num_seqs=2, |
| 45 | enable_prefix_caching=False, |
| 46 | enable_output_caching=False, |
| 47 | ): |
| 48 | self.parallel_config = types.SimpleNamespace( |
| 49 | local_data_parallel_id=0, |
| 50 | enable_expert_parallel=False, |
| 51 | data_parallel_size=1, |
| 52 | ) |
| 53 | self.speculative_config = types.SimpleNamespace( |
| 54 | method=speculative_method, |
| 55 | num_speculative_tokens=2, |
| 56 | enable_draft_logprob=True, |
| 57 | ) |
| 58 | self.model_config = types.SimpleNamespace(enable_logprob=enable_logprob) |
| 59 | self.scheduler_config = types.SimpleNamespace(name="default", splitwise_role="decode") |
| 60 | self.cache_config = types.SimpleNamespace( |
| 61 | enable_prefix_caching=enable_prefix_caching, |
| 62 | enable_output_caching=enable_output_caching, |
| 63 | block_size=64, |
| 64 | ) |
| 65 | self.max_num_seqs = max_num_seqs |
| 66 | self.splitwise_version = "v1" |
| 67 | |
| 68 | |
| 69 | class _DummyResourceManager: |
no outgoing calls