root/gpustack-ui
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

chore: update vllm parameters, do not change backend when file type unknown

Browse Source
This commit is contained in:
jialin
2025-04-16 12:34:18 +08:00
parent ced45bec65
commit c831c19848
14 changed files with 298 additions and 211 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/pages/llmodels/config/vllm-config.ts
+121 -96
View File
@@ -4,6 +4,11 @@ const options = [
value: '--uvicorn-log-level',
options: ['debug', 'info', 'warning', 'error', 'critical', 'trace']
},
{
label: '--disable-uvicorn-access-log',
value: '--disable-uvicorn-access-log',
options: []
},
{
label: '--allow-credentials',
value: '--allow-credentials',
@@ -59,6 +64,11 @@ const options = [
value: '--ssl-certfile',
options: []
},
{
label: '--enable-ssl-refresh',
value: '--enable-ssl-refresh',
options: []
},
{
label: '--ssl-ca-certs',
value: '--ssl-ca-certs',
@@ -89,6 +99,11 @@ const options = [
value: '--disable-frontend-multiprocessing',
options: []
},
{
label: '--enable-request-id-headers',
value: '--enable-request-id-headers',
options: []
},
{
label: '--enable-auto-tool-choice',
value: '--enable-auto-tool-choice',
@@ -104,6 +119,7 @@ const options = [
'jamba',
'llama3_json',
'granite-20b-fc',
'phi4_mini_json',
'granite',
'pythonic'
]
@@ -118,7 +134,8 @@ const options = [
'embed',
'classify',
'score',
'reward'
'reward',
'transcription'
]
},
{
@@ -141,6 +158,11 @@ const options = [
value: '--tokenizer',
options: []
},
{
label: '--hf-config-path',
value: '--hf-config-path',
options: []
},
{
label: '--skip-tokenizer-init',
value: '--skip-tokenizer-init',
@@ -164,7 +186,7 @@ const options = [
{
label: '--tokenizer-mode',
value: '--tokenizer-mode',
options: ['auto', 'slow', 'mistral']
options: ['auto', 'slow', 'mistral', 'custom']
},
{
label: '--trust-remote-code',
@@ -190,7 +212,8 @@ const options = [
'gguf',
'bitsandbytes',
'mistral',
'runai_streamer'
'runai_streamer',
'fastsafetensors'
]
},
{
@@ -213,6 +236,11 @@ const options = [
value: '--config-format',
options: ['auto', 'hf', 'mistral']
},
{
label: '--disable-cascade-attn',
value: '--disable-cascade-attn',
options: []
},
{
label: '--dtype',
value: '--dtype',
@@ -223,6 +251,16 @@ const options = [
value: '--kv-cache-dtype',
options: ['auto', 'fp8', 'fp8_e5m2', 'fp8_e4m3']
},
{
label: '--disable-chunked-mm-input',
value: '--disable-chunked-mm-input',
options: []
},
{
label: 'DISABLE_CHUNKED_MM_INPUT',
value: 'DISABLE_CHUNKED_MM_INPUT',
options: []
},
{
label: '--quantization-param-path',
value: '--quantization-param-path',
@@ -259,6 +297,16 @@ const options = [
value: '--worker-use-ray',
options: []
},
{
label: '--data-parallel-size',
value: '--data-parallel-size',
options: []
},
{
label: '--enable-expert-parallel',
value: '--enable-expert-parallel',
options: []
},
{
label: '--pipeline-parallel-size',
value: '--pipeline-parallel-size',
@@ -289,6 +337,11 @@ const options = [
value: '--enable-prefix-caching',
options: []
},
{
label: '--prefix-caching-hash-algo',
value: '--prefix-caching-hash-algo',
options: ['builtin', 'sha256']
},
{
label: '--disable-sliding-window',
value: '--disable-sliding-window',
@@ -334,6 +387,23 @@ const options = [
value: '--max-num-batched-tokens',
options: []
},
{
label: '--max-num-partial-prefills',
value: '--max-num-partial-prefills',
options: []
},
{
label: '--max-long-partial-prefills',
value: '--max-long-partial-prefills',
options: []
},
{
label: '--long-prefill-token-threshold',
value: '--long-prefill-token-threshold',
options: []
},
{
label: '--max-num-seqs',
value: '--max-num-seqs',
@@ -354,6 +424,17 @@ const options = [
value: '--hf-overrides',
options: []
},
{
label: '--hf-token',
value: '--hf-token',
options: []
},
{
label: 'HF_TOKEN',
value: 'HF_TOKEN',
options: []
},
{
label: '--disable-mm-preprocessor-cache',
value: '--disable-mm-preprocessor-cache',
@@ -368,23 +449,26 @@ const options = [
'deepspeedfp',
'tpu_int8',
'fp8',
'ptpc_fp8',
'fbgemm_fp8',
'modelopt',
'nvfp4',
'marlin',
'gguf',
'hqq',
'gptq_marlin_24',
'gptq_marlin',
'awq_marlin',
'gptq',
'quark',
'moe_wna16',
'compressed-tensors',
'bitsandbytes',
'qqq',
'hqq',
'experts_int8',
'neuron_quant',
'ipex',
'quark',
'moe_wna16',
'torchao',
'None'
]
},
@@ -506,7 +590,12 @@ const options = [
{
label: '--device',
value: '--device',
options: ['auto', 'cuda', 'neuron', 'cpu', 'openvino', 'tpu', 'xpu', 'hpu']
options: ['auto', 'cuda', 'neuron', 'cpu', 'tpu', 'xpu', 'hpu']
},
{
label: '--speculative-config',
value: '--speculative-config',
options: []
},
{
label: '--num-scheduler-steps',
@@ -528,105 +617,31 @@ const options = [
value: '--enable-chunked-prefill',
options: []
},
{
label: '--speculative-model',
value: '--speculative-model',
options: []
},
{
label: '--speculative-model-quantization',
value: '--speculative-model-quantization',
options: [
'aqlm',
'awq',
'deepspeedfp',
'tpu_int8',
'fp8',
'fbgemm_fp8',
'modelopt',
'marlin',
'gguf',
'gptq_marlin_24',
'gptq_marlin',
'awq_marlin',
'gptq',
'compressed-tensors',
'bitsandbytes',
'qqq',
'hqq',
'experts_int8',
'neuron_quant',
'ipex',
'quark',
'moe_wna16',
'None'
]
},
{
label: '--num-speculative-tokens',
value: '--num-speculative-tokens',
options: []
},
{
label: '--speculative-disable-mqa-scorer',
value: '--speculative-disable-mqa-scorer',
options: []
},
{
label: '--speculative-draft-tensor-parallel-size',
value: '--speculative-draft-tensor-parallel-size',
options: []
},
{
label: '--speculative-max-model-len',
value: '--speculative-max-model-len',
options: []
},
{
label: '--speculative-disable-by-batch-size',
value: '--speculative-disable-by-batch-size',
options: []
},
{
label: '--ngram-prompt-lookup-max',
value: '--ngram-prompt-lookup-max',
options: []
},
{
label: '--ngram-prompt-lookup-min',
value: '--ngram-prompt-lookup-min',
options: []
},
{
label: '--spec-decoding-acceptance-method',
value: '--spec-decoding-acceptance-method',
options: ['rejection_sampler', 'typical_acceptance_sampler']
},
{
label: '--typical-acceptance-sampler-posterior-threshold',
value: '--typical-acceptance-sampler-posterior-threshold',
options: []
},
{
label: '--typical-acceptance-sampler-posterior-alpha',
value: '--typical-acceptance-sampler-posterior-alpha',
options: []
},
{
label: '--disable-logprobs-during-spec-decoding',
value: '--disable-logprobs-during-spec-decoding',
options: []
},
{
label: '--model-loader-extra-config',
value: '--model-loader-extra-config',
options: []
},
{
label: '--scheduler-cls',
value: '--scheduler-cls',
options: []
},
{
label: '--use-tqdm-on-load',
value: '--use-tqdm-on-load',
options: []
},
{
label: '--ignore-patterns',
value: '--ignore-patterns',
options: []
},
{
label: '--show-hidden-metrics-for-version',
value: '--show-hidden-metrics-for-version',
options: []
},
{
label: '--preemption-mode',
value: '--preemption-mode',
@@ -713,6 +728,11 @@ const options = [
value: '--disable-log-requests',
options: []
},
{
label: '--additional-config',
value: '--additional-config',
options: []
},
{
label: '--max-log-len',
value: '--max-log-len',
@@ -727,6 +747,11 @@ const options = [
label: '--enable-prompt-tokens-details',
value: '--enable-prompt-tokens-details',
options: []
},
{
label: '--enable-server-load-tracking',
value: '--enable-server-load-tracking',
options: []
}
];