@atomic-chat/core 0.18.3 → 0.18.4

package/src/settings/schema/sglang.json

@@ -1,251 +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_running_requests",
-    "title": "Parallel Requests",
-    "description": "How many requests the engine serves at once (chat, agent and API together); more wait in a queue. The memory check reserves KV cache for that many full contexts, and CUDA graphs are captured up to that many, 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": "mem_fraction_static",
-    "title": "GPU Memory Share",
-    "description": "The share of the card's free memory SGLang keeps for the model weights and the KV cache, from 0 to 1; the rest is its room for activations and CUDA graphs. 0 is auto: Atomic Chat measures the card right before the model starts and leaves SGLang about 2 GB of that room. Applies from the next model load.",
-    "controllerType": "input",
-    "controllerProps": {
-      "value": 0,
-      "placeholder": "0 (auto)",
-      "type": "number",
-      "min": 0,
-      "max": 0.99,
-      "step": 0.01
-    }
-  },
-  {
-    "key": "max_total_tokens",
-    "title": "KV Cache Size (tokens)",
-    "description": "Upper bound of the KV cache, in tokens. 0 lets SGLang 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. Applies from the next model load.",
-    "controllerType": "input",
-    "controllerProps": {
-      "value": 0,
-      "placeholder": "0",
-      "type": "number",
-      "min": 0,
-      "max": 16777216
-    }
-  },
-  {
-    "key": "cuda_graphs",
-    "title": "CUDA Graphs",
-    "description": "Speeds up token generation, but takes extra GPU memory and time at the first load. Auto captures graphs only for as many requests as Parallel Requests allows, which costs little memory; On uses SGLang's own range; Off turns them off. 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": "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": "chunked_prefill_size",
-    "title": "Prefill Chunk Size",
-    "description": "Most prompt tokens the engine reads in one step; a longer prompt is split into chunks of this size. Lower takes less activation memory, higher reads a long prompt faster. 0 leaves SGLang's own value. Applies from the next model load.",
-    "controllerType": "input",
-    "controllerProps": {
-      "value": 0,
-      "placeholder": "0",
-      "type": "number",
-      "min": 0,
-      "max": 1048576
-    }
-  },
-  {
-    "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 SGLang's own. Applies from the next model load.",
-    "controllerType": "input",
-    "controllerProps": {
-      "value": 0,
-      "placeholder": "0",
-      "type": "number",
-      "min": 0,
-      "max": 2147483647
-    }
-  },
-  {
-    "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 SGLang'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"
-    }
-  }
-]

package/src/settings/schema/tensorrt-llm.json

@@ -1,253 +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": "kv_cache_free_gpu_memory_fraction",
-    "title": "KV Cache Memory Fraction",
-    "description": "Share of the GPU memory still free after the weights load that the engine reserves for its KV cache. Applies from the next model load.",
-    "controllerType": "input",
-    "controllerProps": {
-      "value": 0.8,
-      "placeholder": "0.8",
-      "type": "number",
-      "min": 0.1,
-      "max": 0.95,
-      "step": 0.05
-    }
-  },
-  {
-    "key": "max_batch_size",
-    "title": "Parallel Requests",
-    "description": "How many requests the engine serves at once (chat, agent and API together); more wait in a queue. Hybrid models such as Qwen3.5 reserve memory for each one up front, so a higher value needs more GPU memory. Applies from the next model load.",
-    "controllerType": "input",
-    "controllerProps": {
-      "value": 1,
-      "placeholder": "1",
-      "type": "number",
-      "min": 1,
-      "max": 256
-    }
-  },
-  {
-    "key": "kv_cache_max_tokens",
-    "title": "KV Cache Token Limit",
-    "description": "Upper bound of the KV cache in tokens. 0 sizes it for every parallel request at the full context length. The engine uses the smaller of this and the KV cache memory fraction; hybrid models such as Qwen3.5 need it to start. Applies from the next model load.",
-    "controllerType": "input",
-    "controllerProps": {
-      "value": 0,
-      "placeholder": "0",
-      "type": "number",
-      "min": 0,
-      "max": 16777216
-    }
-  },
-  {
-    "key": "cuda_graphs",
-    "title": "CUDA Graphs",
-    "description": "Speeds up token generation, but takes extra GPU memory (about 2 GB for a small model) and a few seconds at load. 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. Applies from the next model load.",
-    "controllerType": "input",
-    "controllerProps": {
-      "value": 0,
-      "placeholder": "0",
-      "type": "number",
-      "min": 0,
-      "max": 3600
-    }
-  },
-  {
-    "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": "overlap_scheduler",
-    "title": "Overlap Scheduler",
-    "description": "Plans the engine's next step while the card runs the current one, which speeds up generation. Turn it off only if a model misbehaves with it. Applies from the next model load.",
-    "controllerType": "checkbox",
-    "controllerProps": {
-      "value": true
-    }
-  },
-  {
-    "key": "capacity_scheduler_policy",
-    "title": "Scheduler Policy",
-    "description": "How the engine admits parallel requests. Guaranteed no evict starts a request only when the KV cache can hold it to the end; max utilization starts more at once and pauses some when the cache runs out. Matters only with more than one parallel request. Applies from the next model load.",
-    "controllerType": "dropdown",
-    "controllerProps": {
-      "value": "guaranteed_no_evict",
-      "options": [
-        {
-          "value": "guaranteed_no_evict",
-          "name": "Guaranteed no evict"
-        },
-        {
-          "value": "max_utilization",
-          "name": "Max utilization"
-        }
-      ]
-    }
-  },
-  {
-    "key": "dtype",
-    "title": "Weight Precision",
-    "description": "Precision of the model's unquantized weights and activations. Auto takes the model's own. 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": "default_temperature",
-    "title": "Default Temperature",
-    "description": "Temperature for a request that does not set one, from 0 to 2. Empty leaves it to the engine. A request's own value always wins. Applies from the next model load or reload.",
-    "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 leaves it to the engine. A request's own value always wins. Applies from the next model load or reload.",
-    "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 leaves it to the engine. A request's own value always wins. Applies from the next model load or reload.",
-    "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 leaves it to the engine. A request's own value always wins. Applies from the next model load or reload.",
-    "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 leaves it to the engine. A request's own value always wins. Applies from the next model load or reload.",
-    "controllerType": "input",
-    "controllerProps": {
-      "value": "",
-      "placeholder": "Model default",
-      "type": "text"
-    }
-  }
-]