Spaces:
Runtime error
Runtime error
Commit
•
dd7ed3d
1
Parent(s):
2a9b1c8
Upload 4 files
Browse files
api-examples/api-example-chat-stream.py
ADDED
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import asyncio
|
2 |
+
import json
|
3 |
+
import sys
|
4 |
+
|
5 |
+
try:
|
6 |
+
import websockets
|
7 |
+
except ImportError:
|
8 |
+
print("Websockets package not found. Make sure it's installed.")
|
9 |
+
|
10 |
+
# For local streaming, the websockets are hosted without ssl - ws://
|
11 |
+
HOST = 'localhost:5005'
|
12 |
+
URI = f'ws://{HOST}/api/v1/chat-stream'
|
13 |
+
|
14 |
+
# For reverse-proxied streaming, the remote will likely host with ssl - wss://
|
15 |
+
# URI = 'wss://your-uri-here.trycloudflare.com/api/v1/stream'
|
16 |
+
|
17 |
+
|
18 |
+
async def run(user_input, history):
|
19 |
+
# Note: the selected defaults change from time to time.
|
20 |
+
request = {
|
21 |
+
'user_input': user_input,
|
22 |
+
'max_new_tokens': 250,
|
23 |
+
'history': history,
|
24 |
+
'mode': 'instruct', # Valid options: 'chat', 'chat-instruct', 'instruct'
|
25 |
+
'character': 'Example',
|
26 |
+
'instruction_template': 'Vicuna-v1.1', # Will get autodetected if unset
|
27 |
+
# 'context_instruct': '', # Optional
|
28 |
+
'your_name': 'You',
|
29 |
+
|
30 |
+
'regenerate': False,
|
31 |
+
'_continue': False,
|
32 |
+
'stop_at_newline': False,
|
33 |
+
'chat_generation_attempts': 1,
|
34 |
+
'chat-instruct_command': 'Continue the chat dialogue below. Write a single reply for the character "<|character|>".\n\n<|prompt|>',
|
35 |
+
|
36 |
+
# Generation params. If 'preset' is set to different than 'None', the values
|
37 |
+
# in presets/preset-name.yaml are used instead of the individual numbers.
|
38 |
+
'preset': 'None',
|
39 |
+
'do_sample': True,
|
40 |
+
'temperature': 0.7,
|
41 |
+
'top_p': 0.1,
|
42 |
+
'typical_p': 1,
|
43 |
+
'epsilon_cutoff': 0, # In units of 1e-4
|
44 |
+
'eta_cutoff': 0, # In units of 1e-4
|
45 |
+
'tfs': 1,
|
46 |
+
'top_a': 0,
|
47 |
+
'repetition_penalty': 1.18,
|
48 |
+
'repetition_penalty_range': 0,
|
49 |
+
'top_k': 40,
|
50 |
+
'min_length': 0,
|
51 |
+
'no_repeat_ngram_size': 0,
|
52 |
+
'num_beams': 1,
|
53 |
+
'penalty_alpha': 0,
|
54 |
+
'length_penalty': 1,
|
55 |
+
'early_stopping': False,
|
56 |
+
'mirostat_mode': 0,
|
57 |
+
'mirostat_tau': 5,
|
58 |
+
'mirostat_eta': 0.1,
|
59 |
+
|
60 |
+
'seed': -1,
|
61 |
+
'add_bos_token': True,
|
62 |
+
'truncation_length': 2048,
|
63 |
+
'ban_eos_token': False,
|
64 |
+
'skip_special_tokens': True,
|
65 |
+
'stopping_strings': []
|
66 |
+
}
|
67 |
+
|
68 |
+
async with websockets.connect(URI, ping_interval=None) as websocket:
|
69 |
+
await websocket.send(json.dumps(request))
|
70 |
+
|
71 |
+
while True:
|
72 |
+
incoming_data = await websocket.recv()
|
73 |
+
incoming_data = json.loads(incoming_data)
|
74 |
+
|
75 |
+
match incoming_data['event']:
|
76 |
+
case 'text_stream':
|
77 |
+
yield incoming_data['history']
|
78 |
+
case 'stream_end':
|
79 |
+
return
|
80 |
+
|
81 |
+
|
82 |
+
async def print_response_stream(user_input, history):
|
83 |
+
cur_len = 0
|
84 |
+
async for new_history in run(user_input, history):
|
85 |
+
cur_message = new_history['visible'][-1][1][cur_len:]
|
86 |
+
cur_len += len(cur_message)
|
87 |
+
print(cur_message, end='')
|
88 |
+
sys.stdout.flush() # If we don't flush, we won't see tokens in realtime.
|
89 |
+
|
90 |
+
|
91 |
+
if __name__ == '__main__':
|
92 |
+
user_input = "Please give me a step-by-step guide on how to plant a tree in my backyard."
|
93 |
+
|
94 |
+
# Basic example
|
95 |
+
history = {'internal': [], 'visible': []}
|
96 |
+
|
97 |
+
# "Continue" example. Make sure to set '_continue' to True above
|
98 |
+
# arr = [user_input, 'Surely, here is']
|
99 |
+
# history = {'internal': [arr], 'visible': [arr]}
|
100 |
+
|
101 |
+
asyncio.run(print_response_stream(user_input, history))
|
api-examples/api-example-model.py
ADDED
@@ -0,0 +1,176 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python3
|
2 |
+
|
3 |
+
import requests
|
4 |
+
|
5 |
+
HOST = '0.0.0.0:5000'
|
6 |
+
|
7 |
+
|
8 |
+
def generate(prompt, tokens=200):
|
9 |
+
request = {'prompt': prompt, 'max_new_tokens': tokens}
|
10 |
+
response = requests.post(f'http://{HOST}/api/v1/generate', json=request)
|
11 |
+
|
12 |
+
if response.status_code == 200:
|
13 |
+
return response.json()['results'][0]['text']
|
14 |
+
|
15 |
+
|
16 |
+
def model_api(request):
|
17 |
+
response = requests.post(f'http://{HOST}/api/v1/model', json=request)
|
18 |
+
return response.json()
|
19 |
+
|
20 |
+
|
21 |
+
# print some common settings
|
22 |
+
def print_basic_model_info(response):
|
23 |
+
basic_settings = ['truncation_length', 'instruction_template']
|
24 |
+
print("Model: ", response['result']['model_name'])
|
25 |
+
print("Lora(s): ", response['result']['lora_names'])
|
26 |
+
for setting in basic_settings:
|
27 |
+
print(setting, "=", response['result']['shared.settings'][setting])
|
28 |
+
|
29 |
+
|
30 |
+
# model info
|
31 |
+
def model_info():
|
32 |
+
response = model_api({'action': 'info'})
|
33 |
+
print_basic_model_info(response)
|
34 |
+
|
35 |
+
|
36 |
+
# simple loader
|
37 |
+
def model_load(model_name):
|
38 |
+
return model_api({'action': 'load', 'model_name': model_name})
|
39 |
+
|
40 |
+
|
41 |
+
# complex loader
|
42 |
+
def complex_model_load(model):
|
43 |
+
|
44 |
+
def guess_groupsize(model_name):
|
45 |
+
if '1024g' in model_name:
|
46 |
+
return 1024
|
47 |
+
elif '128g' in model_name:
|
48 |
+
return 128
|
49 |
+
elif '32g' in model_name:
|
50 |
+
return 32
|
51 |
+
else:
|
52 |
+
return -1
|
53 |
+
|
54 |
+
req = {
|
55 |
+
'action': 'load',
|
56 |
+
'model_name': model,
|
57 |
+
'args': {
|
58 |
+
'loader': 'AutoGPTQ',
|
59 |
+
|
60 |
+
'bf16': False,
|
61 |
+
'load_in_8bit': False,
|
62 |
+
'groupsize': 0,
|
63 |
+
'wbits': 0,
|
64 |
+
|
65 |
+
# llama.cpp
|
66 |
+
'threads': 0,
|
67 |
+
'n_batch': 512,
|
68 |
+
'no_mmap': False,
|
69 |
+
'mlock': False,
|
70 |
+
'cache_capacity': None,
|
71 |
+
'n_gpu_layers': 0,
|
72 |
+
'n_ctx': 2048,
|
73 |
+
|
74 |
+
# RWKV
|
75 |
+
'rwkv_strategy': None,
|
76 |
+
'rwkv_cuda_on': False,
|
77 |
+
|
78 |
+
# b&b 4-bit
|
79 |
+
# 'load_in_4bit': False,
|
80 |
+
# 'compute_dtype': 'float16',
|
81 |
+
# 'quant_type': 'nf4',
|
82 |
+
# 'use_double_quant': False,
|
83 |
+
|
84 |
+
# "cpu": false,
|
85 |
+
# "auto_devices": false,
|
86 |
+
# "gpu_memory": null,
|
87 |
+
# "cpu_memory": null,
|
88 |
+
# "disk": false,
|
89 |
+
# "disk_cache_dir": "cache",
|
90 |
+
},
|
91 |
+
}
|
92 |
+
|
93 |
+
model = model.lower()
|
94 |
+
|
95 |
+
if '4bit' in model or 'gptq' in model or 'int4' in model:
|
96 |
+
req['args']['wbits'] = 4
|
97 |
+
req['args']['groupsize'] = guess_groupsize(model)
|
98 |
+
elif '3bit' in model:
|
99 |
+
req['args']['wbits'] = 3
|
100 |
+
req['args']['groupsize'] = guess_groupsize(model)
|
101 |
+
else:
|
102 |
+
req['args']['gptq_for_llama'] = False
|
103 |
+
|
104 |
+
if '8bit' in model:
|
105 |
+
req['args']['load_in_8bit'] = True
|
106 |
+
elif '-hf' in model or 'fp16' in model:
|
107 |
+
if '7b' in model:
|
108 |
+
req['args']['bf16'] = True # for 24GB
|
109 |
+
elif '13b' in model:
|
110 |
+
req['args']['load_in_8bit'] = True # for 24GB
|
111 |
+
elif 'ggml' in model:
|
112 |
+
# req['args']['threads'] = 16
|
113 |
+
if '7b' in model:
|
114 |
+
req['args']['n_gpu_layers'] = 100
|
115 |
+
elif '13b' in model:
|
116 |
+
req['args']['n_gpu_layers'] = 100
|
117 |
+
elif '30b' in model or '33b' in model:
|
118 |
+
req['args']['n_gpu_layers'] = 59 # 24GB
|
119 |
+
elif '65b' in model:
|
120 |
+
req['args']['n_gpu_layers'] = 42 # 24GB
|
121 |
+
elif 'rwkv' in model:
|
122 |
+
req['args']['rwkv_cuda_on'] = True
|
123 |
+
if '14b' in model:
|
124 |
+
req['args']['rwkv_strategy'] = 'cuda f16i8' # 24GB
|
125 |
+
else:
|
126 |
+
req['args']['rwkv_strategy'] = 'cuda f16' # 24GB
|
127 |
+
|
128 |
+
return model_api(req)
|
129 |
+
|
130 |
+
|
131 |
+
if __name__ == '__main__':
|
132 |
+
for model in model_api({'action': 'list'})['result']:
|
133 |
+
try:
|
134 |
+
resp = complex_model_load(model)
|
135 |
+
|
136 |
+
if 'error' in resp:
|
137 |
+
print(f"❌ {model} FAIL Error: {resp['error']['message']}")
|
138 |
+
continue
|
139 |
+
else:
|
140 |
+
print_basic_model_info(resp)
|
141 |
+
|
142 |
+
ans = generate("0,1,1,2,3,5,8,13,", tokens=2)
|
143 |
+
|
144 |
+
if '21' in ans:
|
145 |
+
print(f"✅ {model} PASS ({ans})")
|
146 |
+
else:
|
147 |
+
print(f"❌ {model} FAIL ({ans})")
|
148 |
+
|
149 |
+
except Exception as e:
|
150 |
+
print(f"❌ {model} FAIL Exception: {repr(e)}")
|
151 |
+
|
152 |
+
|
153 |
+
# 0,1,1,2,3,5,8,13, is the fibonacci sequence, the next number is 21.
|
154 |
+
# Some results below.
|
155 |
+
""" $ ./model-api-example.py
|
156 |
+
Model: 4bit_gpt4-x-alpaca-13b-native-4bit-128g-cuda
|
157 |
+
Lora(s): []
|
158 |
+
truncation_length = 2048
|
159 |
+
instruction_template = Alpaca
|
160 |
+
✅ 4bit_gpt4-x-alpaca-13b-native-4bit-128g-cuda PASS (21)
|
161 |
+
Model: 4bit_WizardLM-13B-Uncensored-4bit-128g
|
162 |
+
Lora(s): []
|
163 |
+
truncation_length = 2048
|
164 |
+
instruction_template = WizardLM
|
165 |
+
✅ 4bit_WizardLM-13B-Uncensored-4bit-128g PASS (21)
|
166 |
+
Model: Aeala_VicUnlocked-alpaca-30b-4bit
|
167 |
+
Lora(s): []
|
168 |
+
truncation_length = 2048
|
169 |
+
instruction_template = Alpaca
|
170 |
+
✅ Aeala_VicUnlocked-alpaca-30b-4bit PASS (21)
|
171 |
+
Model: alpaca-30b-4bit
|
172 |
+
Lora(s): []
|
173 |
+
truncation_length = 2048
|
174 |
+
instruction_template = Alpaca
|
175 |
+
✅ alpaca-30b-4bit PASS (21)
|
176 |
+
"""
|
api-examples/api-example-stream.py
ADDED
@@ -0,0 +1,80 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import asyncio
|
2 |
+
import json
|
3 |
+
import sys
|
4 |
+
|
5 |
+
try:
|
6 |
+
import websockets
|
7 |
+
except ImportError:
|
8 |
+
print("Websockets package not found. Make sure it's installed.")
|
9 |
+
|
10 |
+
# For local streaming, the websockets are hosted without ssl - ws://
|
11 |
+
HOST = 'localhost:5005'
|
12 |
+
URI = f'ws://{HOST}/api/v1/stream'
|
13 |
+
|
14 |
+
# For reverse-proxied streaming, the remote will likely host with ssl - wss://
|
15 |
+
# URI = 'wss://your-uri-here.trycloudflare.com/api/v1/stream'
|
16 |
+
|
17 |
+
|
18 |
+
async def run(context):
|
19 |
+
# Note: the selected defaults change from time to time.
|
20 |
+
request = {
|
21 |
+
'prompt': context,
|
22 |
+
'max_new_tokens': 250,
|
23 |
+
|
24 |
+
# Generation params. If 'preset' is set to different than 'None', the values
|
25 |
+
# in presets/preset-name.yaml are used instead of the individual numbers.
|
26 |
+
'preset': 'None',
|
27 |
+
'do_sample': True,
|
28 |
+
'temperature': 0.7,
|
29 |
+
'top_p': 0.1,
|
30 |
+
'typical_p': 1,
|
31 |
+
'epsilon_cutoff': 0, # In units of 1e-4
|
32 |
+
'eta_cutoff': 0, # In units of 1e-4
|
33 |
+
'tfs': 1,
|
34 |
+
'top_a': 0,
|
35 |
+
'repetition_penalty': 1.18,
|
36 |
+
'repetition_penalty_range': 0,
|
37 |
+
'top_k': 40,
|
38 |
+
'min_length': 0,
|
39 |
+
'no_repeat_ngram_size': 0,
|
40 |
+
'num_beams': 1,
|
41 |
+
'penalty_alpha': 0,
|
42 |
+
'length_penalty': 1,
|
43 |
+
'early_stopping': False,
|
44 |
+
'mirostat_mode': 0,
|
45 |
+
'mirostat_tau': 5,
|
46 |
+
'mirostat_eta': 0.1,
|
47 |
+
|
48 |
+
'seed': -1,
|
49 |
+
'add_bos_token': True,
|
50 |
+
'truncation_length': 2048,
|
51 |
+
'ban_eos_token': False,
|
52 |
+
'skip_special_tokens': True,
|
53 |
+
'stopping_strings': []
|
54 |
+
}
|
55 |
+
|
56 |
+
async with websockets.connect(URI, ping_interval=None) as websocket:
|
57 |
+
await websocket.send(json.dumps(request))
|
58 |
+
|
59 |
+
yield context # Remove this if you just want to see the reply
|
60 |
+
|
61 |
+
while True:
|
62 |
+
incoming_data = await websocket.recv()
|
63 |
+
incoming_data = json.loads(incoming_data)
|
64 |
+
|
65 |
+
match incoming_data['event']:
|
66 |
+
case 'text_stream':
|
67 |
+
yield incoming_data['text']
|
68 |
+
case 'stream_end':
|
69 |
+
return
|
70 |
+
|
71 |
+
|
72 |
+
async def print_response_stream(prompt):
|
73 |
+
async for response in run(prompt):
|
74 |
+
print(response, end='')
|
75 |
+
sys.stdout.flush() # If we don't flush, we won't see tokens in realtime.
|
76 |
+
|
77 |
+
|
78 |
+
if __name__ == '__main__':
|
79 |
+
prompt = "In order to make homemade bread, follow these steps:\n1)"
|
80 |
+
asyncio.run(print_response_stream(prompt))
|
api-examples/api-example.py
ADDED
@@ -0,0 +1,57 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import requests
|
2 |
+
|
3 |
+
# For local streaming, the websockets are hosted without ssl - http://
|
4 |
+
HOST = 'localhost:5000'
|
5 |
+
URI = f'http://{HOST}/api/v1/generate'
|
6 |
+
|
7 |
+
# For reverse-proxied streaming, the remote will likely host with ssl - https://
|
8 |
+
# URI = 'https://your-uri-here.trycloudflare.com/api/v1/generate'
|
9 |
+
|
10 |
+
|
11 |
+
def run(prompt):
|
12 |
+
request = {
|
13 |
+
'prompt': prompt,
|
14 |
+
'max_new_tokens': 250,
|
15 |
+
|
16 |
+
# Generation params. If 'preset' is set to different than 'None', the values
|
17 |
+
# in presets/preset-name.yaml are used instead of the individual numbers.
|
18 |
+
'preset': 'None',
|
19 |
+
'do_sample': True,
|
20 |
+
'temperature': 0.7,
|
21 |
+
'top_p': 0.1,
|
22 |
+
'typical_p': 1,
|
23 |
+
'epsilon_cutoff': 0, # In units of 1e-4
|
24 |
+
'eta_cutoff': 0, # In units of 1e-4
|
25 |
+
'tfs': 1,
|
26 |
+
'top_a': 0,
|
27 |
+
'repetition_penalty': 1.18,
|
28 |
+
'repetition_penalty_range': 0,
|
29 |
+
'top_k': 40,
|
30 |
+
'min_length': 0,
|
31 |
+
'no_repeat_ngram_size': 0,
|
32 |
+
'num_beams': 1,
|
33 |
+
'penalty_alpha': 0,
|
34 |
+
'length_penalty': 1,
|
35 |
+
'early_stopping': False,
|
36 |
+
'mirostat_mode': 0,
|
37 |
+
'mirostat_tau': 5,
|
38 |
+
'mirostat_eta': 0.1,
|
39 |
+
|
40 |
+
'seed': -1,
|
41 |
+
'add_bos_token': True,
|
42 |
+
'truncation_length': 2048,
|
43 |
+
'ban_eos_token': False,
|
44 |
+
'skip_special_tokens': True,
|
45 |
+
'stopping_strings': []
|
46 |
+
}
|
47 |
+
|
48 |
+
response = requests.post(URI, json=request)
|
49 |
+
|
50 |
+
if response.status_code == 200:
|
51 |
+
result = response.json()['results'][0]['text']
|
52 |
+
print(prompt + result)
|
53 |
+
|
54 |
+
|
55 |
+
if __name__ == '__main__':
|
56 |
+
prompt = "In order to make homemade bread, follow these steps:\n1)"
|
57 |
+
run(prompt)
|