groupuser/TinyLlama-1.1B-Chat-v1.0
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Automatically updated from 모델 배포(742:TinyLlama-1.1B(Triton, vLLM)) by 그룹사용자(groupuser)

Browse Source
This commit is contained in:
groupuser 2026-05-14 02:02:12 +00:00
parent 3d4383abf2
commit 2e65464067
1 changed files with 4 additions and 60 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

64
config.pbtxt
View File

@ -1,131 +1,75 @@
# Triton Backend for vLLM. # Triton Backend for Python.
backend: "python" backend: "python"
max_batch_size: 0 max_batch_size: 0
# Triton should expect as input a single string
# input of variable length named 'text_input'
input [ input [
{ {
name: "text_input" name: "text_input"
data_type: TYPE_STRING data_type: TYPE_STRING
dims: [ 1 ] dims: [ 1 ]
}, },
{ {
name: "max_length" name: "max_length"
data_type: TYPE_INT32 data_type: TYPE_INT32
dims: [ 1 ] dims: [ 1 ]
optional: true
}, },
{ {
name: "max_new_tokens" name: "max_new_tokens"
data_type: TYPE_INT32 data_type: TYPE_INT32
dims: [ 1 ] dims: [ 1 ]
optional: true
}, },
{ {
name: "do_sample" name: "do_sample"
data_type: TYPE_BOOL data_type: TYPE_BOOL
dims: [ 1 ] dims: [ 1 ]
optional: true
}, },
{ {
name: "top_k" name: "top_k"
data_type: TYPE_INT32 data_type: TYPE_INT32
dims: [ 1 ] dims: [ 1 ]
optional: true
}, },
{ {
name: "top_p" name: "top_p"
data_type: TYPE_FP32 data_type: TYPE_FP32
dims: [ 1 ] dims: [ 1 ]
optional: true
}, },
{ {
name: "temperature" name: "temperature"
data_type: TYPE_FP32 data_type: TYPE_FP32
dims: [ 1 ] dims: [ 1 ]
optional: true
}, },
{ {
name: "repetition_penalty" name: "repetition_penalty"
data_type: TYPE_FP32 data_type: TYPE_FP32
dims: [ 1 ] dims: [ 1 ]
optional: true
}, },
{ {
name: "stream" name: "stream"
data_type: TYPE_BOOL data_type: TYPE_BOOL
dims: [ 1 ] dims: [ 1 ]
optional: true
} }
] ]
# Triton should expect to respond with a single string
# output of variable length named 'text_output'
output [ output [
{ {
name: "text_output" name: "text_output"
data_type: TYPE_STRING data_type: TYPE_STRING
dims: [ 1 ] dims: [ 1 ]
} }
] ]
parameters: [ parameters: [
{ {
key: "base_model_path", key: "model_path",
value: {string_value: "/cheetah/input/model/groupuser/TinyLlama-1.1B-Chat-v1.0"} value: {string_value: "/cheetah/input/model/groupuser/TinyLlama-1.1B-Chat-v1.0"}
},
{
key: "is_adapter_model",
value: {string_value: "false"}
},
{
key: "adapter_model_path",
value: {string_value: ""}
},
{
key: "quantization",
value: {string_value: "none"}
} }
] ]
instance_group [ instance_group [
{ {
kind: KIND_AUTO kind: KIND_AUTO
count: 1 count: 1
} }
] ]