commitcopilot/infer-003

1{2  "model_id": "gemma-4-e2b",3  "model_path": "/models/model.gguf",4  "infer_api_key": "commit-copilot",5  "download": {6    "source": "hf_hub",7    "repo_id": "unsloth/gemma-4-E2B-it-GGUF",8    "filename": "gemma-4-E2B-it-UD-Q3_K_XL.gguf",9    "revision": "main",10    "output_filename": "model.gguf"11  },12  "llama": {13    "n_ctx": 16384,14    "n_threads": 2,15    "n_batch": 512,16    "n_parallel": 1,17    "extra_args": [18      "--jinja",19      "--flash-attn",20      "auto",21      "--swa-full",22      "--cache-reuse",23      "256",24      "--reasoning",25      "off",26      "--reasoning-budget",27      "0"28    ]29  },30  "generation": {31    "max_concurrent_requests": 1,32    "default_max_tokens": 384,33    "default_temperature": 0.2,34    "default_top_p": 0.95,35    "ready_wait_timeout_seconds": 180,36    "ready_poll_interval_seconds": 137  }38}39