| 103 | self.server_port = int(os.environ["PORT"]) |
| 104 | |
| 105 | def start(self, timeout_seconds: int | None = DEFAULT_HTTP_TIMEOUT) -> None: |
| 106 | if self.server_path is not None: |
| 107 | server_path = self.server_path |
| 108 | elif "LLAMA_SERVER_BIN_PATH" in os.environ: |
| 109 | server_path = os.environ["LLAMA_SERVER_BIN_PATH"] |
| 110 | elif os.name == "nt": |
| 111 | server_path = "../../../build/bin/Release/llama-server.exe" |
| 112 | else: |
| 113 | server_path = "../../../build/bin/llama-server" |
| 114 | server_args = [ |
| 115 | "--host", |
| 116 | self.server_host, |
| 117 | "--port", |
| 118 | self.server_port, |
| 119 | "--temp", |
| 120 | self.temperature, |
| 121 | "--seed", |
| 122 | self.seed, |
| 123 | ] |
| 124 | if self.model_file: |
| 125 | server_args.extend(["--model", self.model_file]) |
| 126 | if self.model_url: |
| 127 | server_args.extend(["--model-url", self.model_url]) |
| 128 | if self.model_draft: |
| 129 | server_args.extend(["--model-draft", self.model_draft]) |
| 130 | if self.model_hf_repo: |
| 131 | server_args.extend(["--hf-repo", self.model_hf_repo]) |
| 132 | if self.model_hf_file: |
| 133 | server_args.extend(["--hf-file", self.model_hf_file]) |
| 134 | if self.n_batch: |
| 135 | server_args.extend(["--batch-size", self.n_batch]) |
| 136 | if self.n_ubatch: |
| 137 | server_args.extend(["--ubatch-size", self.n_ubatch]) |
| 138 | if self.n_threads: |
| 139 | server_args.extend(["--threads", self.n_threads]) |
| 140 | if self.n_gpu_layer: |
| 141 | server_args.extend(["--n-gpu-layers", self.n_gpu_layer]) |
| 142 | if self.draft is not None: |
| 143 | server_args.extend(["--draft", self.draft]) |
| 144 | if self.server_continuous_batching: |
| 145 | server_args.append("--cont-batching") |
| 146 | if self.server_embeddings: |
| 147 | server_args.append("--embedding") |
| 148 | if self.server_reranking: |
| 149 | server_args.append("--reranking") |
| 150 | if self.server_metrics: |
| 151 | server_args.append("--metrics") |
| 152 | if self.server_slots: |
| 153 | server_args.append("--slots") |
| 154 | if self.pooling: |
| 155 | server_args.extend(["--pooling", self.pooling]) |
| 156 | if self.model_alias: |
| 157 | server_args.extend(["--alias", self.model_alias]) |
| 158 | if self.n_ctx: |
| 159 | server_args.extend(["--ctx-size", self.n_ctx]) |
| 160 | if self.n_slots: |
| 161 | server_args.extend(["--parallel", self.n_slots]) |
| 162 | if self.ctk: |