@atomic-chat/core 0.18.3 → 0.18.4

package/src/settings/schema/vllm.json

@@ -1,273 +0,0 @@
-[
-  {
-    "key": "gpu_id",
-    "title": "GPU",
-    "description": "UUID of the NVIDIA GPU models run on (GPU-… or MIG-…). Leave empty to use the card with the most free memory at load time; if the saved card is gone, the load uses that card and says so. Applies from the next model load.",
-    "controllerType": "input",
-    "controllerProps": {
-      "value": "",
-      "placeholder": "Card with the most free memory",
-      "type": "text"
-    }
-  },
-  {
-    "key": "context_length",
-    "title": "Context Length",
-    "description": "Maximum prompt plus reply length in tokens, fixed when the model starts. A longer request is refused with context_length_exceeded instead of restarting the model. Applies from the next model load.",
-    "controllerType": "input",
-    "controllerProps": {
-      "value": 8192,
-      "placeholder": "8192",
-      "type": "number",
-      "min": 512,
-      "max": 1048576
-    }
-  },
-  {
-    "key": "max_output_tokens",
-    "title": "Max Output Tokens",
-    "description": "Longest reply per request. Atomic Chat caps each request's max_tokens to this value on its way to the model; it is not an engine setting. Must be less than the context length. Applies from the next model load.",
-    "controllerType": "input",
-    "controllerProps": {
-      "value": 4096,
-      "placeholder": "4096",
-      "type": "number",
-      "min": 1,
-      "max": 1048576
-    }
-  },
-  {
-    "key": "max_num_seqs",
-    "title": "Parallel Requests",
-    "description": "How many requests the engine serves at once (chat, agent and API together); more wait in a queue. Atomic Chat reserves KV cache memory for that many full contexts unless a KV cache token limit is set, so 1 fits the most on a desktop card. Applies from the next model load.",
-    "controllerType": "input",
-    "controllerProps": {
-      "value": 1,
-      "placeholder": "1",
-      "type": "number",
-      "min": 1,
-      "max": 256
-    }
-  },
-  {
-    "key": "kv_cache_memory_gib",
-    "title": "KV Cache Size (GB)",
-    "description": "Memory for the KV cache, in GB. 0 lets vLLM size it itself from what its share of the card leaves after the weights — the default, right for every model including hybrid ones such as Qwen3.5. Set a number only to cap the KV cache by hand; too little refuses the context length at load. Applies from the next model load.",
-    "controllerType": "input",
-    "controllerProps": {
-      "value": 0,
-      "placeholder": "0",
-      "type": "number",
-      "min": 0,
-      "max": 1024
-    }
-  },
-  {
-    "key": "gpu_memory_utilization",
-    "title": "GPU Memory Share",
-    "description": "The share of the card's memory vLLM may take, from 0 to 1. 0 is auto: Atomic Chat gives vLLM what is free on the card right before the model starts, less 1.5 GB for vLLM's own start-up and other programs. vLLM refuses to start when the card has less free memory than this share. Applies from the next model load.",
-    "controllerType": "input",
-    "controllerProps": {
-      "value": 0,
-      "placeholder": "0 (auto)",
-      "type": "number",
-      "min": 0,
-      "max": 1,
-      "step": 0.01
-    }
-  },
-  {
-    "key": "cuda_graphs",
-    "title": "CUDA Graphs",
-    "description": "Speeds up token generation, but takes extra GPU memory and time at the first load (compiled graphs are cached for the next one). Auto turns them on only on cards with 12 GB of memory or more, so a model still fits on a small card. Applies from the next model load.",
-    "controllerType": "dropdown",
-    "controllerProps": {
-      "value": "auto",
-      "options": [
-        {
-          "value": "auto",
-          "name": "Auto"
-        },
-        {
-          "value": "on",
-          "name": "On"
-        },
-        {
-          "value": "off",
-          "name": "Off"
-        }
-      ]
-    }
-  },
-  {
-    "key": "kv_cache_dtype",
-    "title": "KV Cache Precision",
-    "description": "FP8 halves the KV cache memory, so a longer context fits, at a small quality cost on very long contexts. Needs an Ada (RTX 40) or newer card; on older cards it is ignored and the model's own precision is used. Applies from the next model load.",
-    "controllerType": "dropdown",
-    "controllerProps": {
-      "value": "auto",
-      "options": [
-        {
-          "value": "auto",
-          "name": "Model precision"
-        },
-        {
-          "value": "fp8",
-          "name": "FP8"
-        }
-      ]
-    }
-  },
-  {
-    "key": "load_timeout_seconds",
-    "title": "Load Timeout (seconds)",
-    "description": "How long a load may take before it fails. 0 estimates it from the model's size, with room for the first graph compilation. Applies from the next model load.",
-    "controllerType": "input",
-    "controllerProps": {
-      "value": 0,
-      "placeholder": "0",
-      "type": "number",
-      "min": 0,
-      "max": 3600
-    }
-  },
-  {
-    "key": "max_num_batched_tokens",
-    "title": "Max Batched Tokens",
-    "description": "Most tokens the engine processes in one step across all requests; long prompts are split into chunks of this size. Lower takes less activation memory and keeps replies to other requests flowing, higher reads a long prompt faster. 0 leaves vLLM's own value. Applies from the next model load.",
-    "controllerType": "input",
-    "controllerProps": {
-      "value": 0,
-      "placeholder": "0",
-      "type": "number",
-      "min": 0,
-      "max": 1048576
-    }
-  },
-  {
-    "key": "enable_prefix_caching",
-    "title": "Prefix Caching",
-    "description": "Reuses the KV cache of a prompt start the engine has already seen (the system prompt, earlier turns of a chat), so a follow-up reads only what is new. Turn it off only to rule it out when chasing a problem. Applies from the next model load.",
-    "controllerType": "checkbox",
-    "controllerProps": {
-      "value": true
-    }
-  },
-  {
-    "key": "cpu_offload_gb",
-    "title": "CPU Offload (GB)",
-    "description": "Keeps this many GB of the model's weights in system memory instead of on the card, so a model slightly larger than the card fits, at a large speed cost: they cross PCIe on every step. Atomic Chat counts only the rest of the weights against the card. 0 keeps every weight on the card. Applies from the next model load.",
-    "controllerType": "input",
-    "controllerProps": {
-      "value": 0,
-      "placeholder": "0",
-      "type": "number",
-      "min": 0,
-      "max": 1024
-    }
-  },
-  {
-    "key": "dtype",
-    "title": "Weight Precision",
-    "description": "Precision of the model's unquantized weights and activations. Auto takes the model's own (bfloat16 falls back to float16 on cards before Ampere). A quantized model keeps its quantized weights either way. Applies from the next model load.",
-    "controllerType": "dropdown",
-    "controllerProps": {
-      "value": "auto",
-      "options": [
-        {
-          "value": "auto",
-          "name": "Model precision"
-        },
-        {
-          "value": "float16",
-          "name": "float16"
-        },
-        {
-          "value": "bfloat16",
-          "name": "bfloat16"
-        },
-        {
-          "value": "float32",
-          "name": "float32"
-        }
-      ]
-    }
-  },
-  {
-    "key": "seed",
-    "title": "Seed",
-    "description": "Seed of the engine's random sampling, for replies that repeat across restarts. 0 leaves vLLM's own. Applies from the next model load.",
-    "controllerType": "input",
-    "controllerProps": {
-      "value": 0,
-      "placeholder": "0",
-      "type": "number",
-      "min": 0,
-      "max": 2147483647
-    }
-  },
-  {
-    "key": "async_scheduling",
-    "title": "Async Scheduling",
-    "description": "Plans the engine's next step while the card runs the current one, which speeds up generation a little. Experimental in vLLM: some models and features do not support it, and a load that fails with it on may succeed with it off. Applies from the next model load.",
-    "controllerType": "checkbox",
-    "controllerProps": {
-      "value": false
-    }
-  },
-  {
-    "key": "default_temperature",
-    "title": "Default Temperature",
-    "description": "Temperature for a request that does not set one, from 0 to 2. Empty takes the model's own generation_config.json, else vLLM's 1.0. A request's own value always wins. Applies from the next model load.",
-    "controllerType": "input",
-    "controllerProps": {
-      "value": "",
-      "placeholder": "Model default",
-      "type": "text"
-    }
-  },
-  {
-    "key": "default_top_p",
-    "title": "Default Top P",
-    "description": "Top P for a request that does not set one, above 0 and up to 1. Empty takes the model's own default. A request's own value always wins. Applies from the next model load.",
-    "controllerType": "input",
-    "controllerProps": {
-      "value": "",
-      "placeholder": "Model default",
-      "type": "text"
-    }
-  },
-  {
-    "key": "default_top_k",
-    "title": "Default Top K",
-    "description": "Top K for a request that does not set one; -1 or 0 turns it off. Empty takes the model's own default. A request's own value always wins. Applies from the next model load.",
-    "controllerType": "input",
-    "controllerProps": {
-      "value": "",
-      "placeholder": "Model default",
-      "type": "text"
-    }
-  },
-  {
-    "key": "default_min_p",
-    "title": "Default Min P",
-    "description": "Min P for a request that does not set one, from 0 to 1. Empty takes the model's own default. A request's own value always wins. Applies from the next model load.",
-    "controllerType": "input",
-    "controllerProps": {
-      "value": "",
-      "placeholder": "Model default",
-      "type": "text"
-    }
-  },
-  {
-    "key": "default_repetition_penalty",
-    "title": "Default Repetition Penalty",
-    "description": "Repetition penalty for a request that does not set one, above 0 (1 is none). Empty takes the model's own default. A request's own value always wins. Applies from the next model load.",
-    "controllerType": "input",
-    "controllerProps": {
-      "value": "",
-      "placeholder": "Model default",
-      "type": "text"
-    }
-  }
-]