56 lines
1.5 KiB
JSON
56 lines
1.5 KiB
JSON
{
|
|
"default_model": "Qwen3.5-35B-A3B-GPTQ-Int4",
|
|
"Qwen3-Next-80B-A3B-Instruct-AWQ-4bit": {
|
|
"ctx": "24576",
|
|
"trust_remote": true,
|
|
"valid_tp": [2],
|
|
"max_num_seqs": "32",
|
|
"max_tokens": "16384",
|
|
"gpu_util": "0.98",
|
|
"enforce_eager": false,
|
|
"env": {
|
|
"VLLM_USE_TRITON_AWQ": "1"
|
|
},
|
|
"tool_call_parser": "qwen3_xml",
|
|
"enable_auto_tool_choice": true,
|
|
"served_model_name": "Qwen3-Next-80B-A3B-Instruct-AWQ-4bit",
|
|
"hf_model_id": "cpatonn/Qwen3-Next-80B-A3B-Instruct-AWQ-4bit"
|
|
},
|
|
"GLM-4.7-Flash-AWQ": {
|
|
"ctx": "32768",
|
|
"trust_remote": true,
|
|
"valid_tp": [1, 2],
|
|
"max_num_seqs": "64",
|
|
"max_tokens": "32768",
|
|
"gpu_util": "0.98",
|
|
"tool_call_parser": "qwen3_xml",
|
|
"enable_auto_tool_choice": true,
|
|
"served_model_name": "GLM-4.7-Flash-AWQ",
|
|
"hf_model_id": "THUDM/GLM-4.7-Flash-AWQ"
|
|
},
|
|
"Qwen3.5-27B-FP8": {
|
|
"ctx": "32768",
|
|
"trust_remote": true,
|
|
"valid_tp": [1, 2],
|
|
"max_num_seqs": "64",
|
|
"max_tokens": "32768",
|
|
"gpu_util": "0.98",
|
|
"tool_call_parser": "qwen3_xml",
|
|
"enable_auto_tool_choice": true,
|
|
"served_model_name": "Qwen3.5-27B-FP8",
|
|
"hf_model_id": "RedHatAI/Qwen3.5-27B-FP8-dynamic"
|
|
},
|
|
"Qwen3.5-35B-A3B-GPTQ-Int4": {
|
|
"ctx": "32768",
|
|
"trust_remote": true,
|
|
"valid_tp": [1, 2],
|
|
"max_num_seqs": "64",
|
|
"max_tokens": "32768",
|
|
"gpu_util": "0.98",
|
|
"tool_call_parser": "qwen3_xml",
|
|
"enable_auto_tool_choice": true,
|
|
"served_model_name": "Qwen3.5-35B-A3B-GPTQ-Int4",
|
|
"hf_model_id": "Qwen/Qwen3.5-35B-A3B-GPTQ-Int4"
|
|
}
|
|
}
|