| 120 | self.external_server = "DEBUG_EXTERNAL" in os.environ |
| 121 | |
| 122 | def start(self, timeout_seconds: int = DEFAULT_HTTP_TIMEOUT) -> None: |
| 123 | if self.external_server: |
| 124 | print(f"[external_server]: Assuming external server running on {self.server_host}:{self.server_port}") |
| 125 | return |
| 126 | if self.server_path is not None: |
| 127 | server_path = self.server_path |
| 128 | elif "LLAMA_SERVER_BIN_PATH" in os.environ: |
| 129 | server_path = os.environ["LLAMA_SERVER_BIN_PATH"] |
| 130 | elif os.name == "nt": |
| 131 | server_path = "../../../build/bin/Release/llama-server.exe" |
| 132 | else: |
| 133 | server_path = "../../../build/bin/llama-server" |
| 134 | server_args = [ |
| 135 | "--host", |
| 136 | self.server_host, |
| 137 | "--port", |
| 138 | self.server_port, |
| 139 | "--temp", |
| 140 | self.temperature, |
| 141 | "--seed", |
| 142 | self.seed, |
| 143 | ] |
| 144 | if self.offline: |
| 145 | server_args.append("--offline") |
| 146 | if self.model_file: |
| 147 | server_args.extend(["--model", self.model_file]) |
| 148 | if self.model_url: |
| 149 | server_args.extend(["--model-url", self.model_url]) |
| 150 | if self.model_draft: |
| 151 | server_args.extend(["--model-draft", self.model_draft]) |
| 152 | if self.model_hf_repo: |
| 153 | server_args.extend(["--hf-repo", self.model_hf_repo]) |
| 154 | if self.model_hf_file: |
| 155 | server_args.extend(["--hf-file", self.model_hf_file]) |
| 156 | if self.models_dir: |
| 157 | server_args.extend(["--models-dir", self.models_dir]) |
| 158 | if self.models_max is not None: |
| 159 | server_args.extend(["--models-max", self.models_max]) |
| 160 | if self.n_batch: |
| 161 | server_args.extend(["--batch-size", self.n_batch]) |
| 162 | if self.n_ubatch: |
| 163 | server_args.extend(["--ubatch-size", self.n_ubatch]) |
| 164 | if self.n_threads: |
| 165 | server_args.extend(["--threads", self.n_threads]) |
| 166 | if self.n_gpu_layer: |
| 167 | server_args.extend(["--n-gpu-layers", self.n_gpu_layer]) |
| 168 | if self.draft is not None: |
| 169 | server_args.extend(["--draft", self.draft]) |
| 170 | if self.server_continuous_batching: |
| 171 | server_args.append("--cont-batching") |
| 172 | if self.server_embeddings: |
| 173 | server_args.append("--embedding") |
| 174 | if self.server_reranking: |
| 175 | server_args.append("--reranking") |
| 176 | if self.server_metrics: |
| 177 | server_args.append("--metrics") |
| 178 | if self.kv_unified: |
| 179 | server_args.append("--kv-unified") |