Update app.py
Browse files
app.py
CHANGED
@@ -1,24 +1,7 @@
|
|
1 |
-
|
2 |
-
from
|
3 |
-
from typing import List, Tuple, Optional, Dict, Any, Union
|
4 |
-
from dataclasses import dataclass
|
5 |
-
from enum import Enum
|
6 |
-
import logging
|
7 |
-
|
8 |
-
from typing import List, Tuple, Optional, Dict, Any, Union, AsyncGenerator
|
9 |
-
from dataclasses import dataclass
|
10 |
from enum import Enum
|
11 |
-
import
|
12 |
-
import torch
|
13 |
-
from transformers import AutoModelForCausalLM, AutoTokenizer
|
14 |
-
from llama_cpp import Llama
|
15 |
-
|
16 |
-
from huggingface_hub import hf_hub_download
|
17 |
-
|
18 |
-
prm_model_path = hf_hub_download(
|
19 |
-
repo_id="tensorblock/Llama3.1-8B-PRM-Mistral-Data-GGUF",
|
20 |
-
filename="Llama3.1-8B-PRM-Mistral-Data-Q4_K_M.gguf"
|
21 |
-
)
|
22 |
|
23 |
class GenerationStrategy(str, Enum):
|
24 |
DEFAULT = "default"
|
@@ -26,189 +9,236 @@ class GenerationStrategy(str, Enum):
|
|
26 |
BEST_OF_N = "best_of_n"
|
27 |
BEAM_SEARCH = "beam_search"
|
28 |
DVTS = "dvts"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
29 |
|
30 |
@dataclass
|
31 |
class GenerationConfig:
|
32 |
num_samples: int = 5
|
33 |
depth: int = 3
|
34 |
breadth: int = 2
|
35 |
-
max_history_turns: int =
|
36 |
max_new_tokens: int = 50
|
37 |
temperature: float = 0.7
|
38 |
top_p: float = 0.9
|
|
|
|
|
|
|
|
|
39 |
strategy: GenerationStrategy = GenerationStrategy.DEFAULT
|
40 |
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
47 |
-
|
48 |
-
|
49 |
-
|
|
|
50 |
self.logger = logging.getLogger(__name__)
|
51 |
self.device = device or ("cuda" if torch.cuda.is_available() else "cpu")
|
52 |
-
self.
|
53 |
-
|
54 |
-
|
55 |
-
|
|
|
56 |
try:
|
57 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
58 |
except Exception as e:
|
59 |
-
self.logger.error(f"Failed to
|
60 |
raise
|
61 |
|
62 |
-
def
|
63 |
-
"""
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
|
69 |
-
)
|
70 |
-
if self.llama_tokenizer.pad_token is None:
|
71 |
-
self.llama_tokenizer.pad_token = self.llama_tokenizer.eos_token
|
72 |
-
|
73 |
-
self.llama_model = AutoModelForCausalLM.from_pretrained(
|
74 |
-
llama_model_name,
|
75 |
-
device_map="auto",
|
76 |
-
trust_remote_code=True
|
77 |
-
)
|
78 |
-
|
79 |
-
# Initialize PRM model
|
80 |
-
self.prm_model = self._load_quantized_model(prm_model_path)
|
81 |
-
|
82 |
-
# Enable token streaming
|
83 |
-
self.supports_streaming = hasattr(self.llama_model, "streamer")
|
84 |
-
|
85 |
-
async def generate_stream(
|
86 |
-
self,
|
87 |
-
prompt: str,
|
88 |
-
config: Optional[GenerationConfig] = None
|
89 |
-
) -> AsyncGenerator[str, None]:
|
90 |
-
"""Stream tokens as they're generated."""
|
91 |
-
if not self.supports_streaming:
|
92 |
-
raise NotImplementedError("This model doesn't support streaming")
|
93 |
-
|
94 |
-
config = config or self.default_config
|
95 |
-
input_ids = self.llama_tokenizer(prompt, return_tensors="pt").input_ids.to(self.device)
|
96 |
-
|
97 |
-
async for token in self.llama_model.streamer(input_ids, **self._get_generation_kwargs(config)):
|
98 |
-
yield self.llama_tokenizer.decode([token])
|
99 |
|
100 |
-
def
|
101 |
-
"""
|
102 |
-
return {
|
103 |
-
"max_new_tokens": config.max_new_tokens,
|
104 |
-
"temperature": config.temperature,
|
105 |
-
"top_p": config.top_p,
|
106 |
-
"do_sample": config.temperature > 0,
|
107 |
-
}
|
108 |
-
|
109 |
-
def _load_quantized_model(self, model_path: str) -> Llama:
|
110 |
-
"""Load a quantized GGUF model using llama-cpp-python.
|
111 |
-
|
112 |
-
Args:
|
113 |
-
model_path (str): Path to the GGUF model file
|
114 |
-
|
115 |
-
Returns:
|
116 |
-
Llama: Loaded model instance
|
117 |
-
"""
|
118 |
try:
|
119 |
-
# Configure GPU layers if CUDA is available
|
120 |
n_gpu_layers = -1 if torch.cuda.is_available() else 0
|
121 |
-
|
122 |
-
# Load the model
|
123 |
model = Llama(
|
124 |
model_path=model_path,
|
125 |
-
n_ctx=2048,
|
126 |
-
n_batch=512,
|
127 |
-
n_gpu_layers=n_gpu_layers,
|
128 |
-
verbose=False
|
129 |
)
|
130 |
-
|
131 |
-
self.logger.info(f"Successfully loaded GGUF model from {model_path}")
|
132 |
return model
|
133 |
-
|
134 |
except Exception as e:
|
135 |
self.logger.error(f"Failed to load GGUF model: {str(e)}")
|
136 |
raise
|
137 |
|
138 |
-
def _score_with_prm(self, text: str) -> float:
|
139 |
-
"""Score text using the PRM model.
|
140 |
-
|
141 |
-
Args:
|
142 |
-
text (str): Text to score
|
143 |
-
|
144 |
-
Returns:
|
145 |
-
float: Model score
|
146 |
-
"""
|
147 |
-
try:
|
148 |
-
# For GGUF models, we need to use the proper scoring interface
|
149 |
-
result = self.prm_model.eval(text)
|
150 |
-
return result['logprobs'] # Or another appropriate scoring metric
|
151 |
-
|
152 |
-
except Exception as e:
|
153 |
-
self.logger.error(f"Error scoring text with PRM: {str(e)}")
|
154 |
-
return float('-inf') # Return very low score on error
|
155 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
156 |
|
157 |
-
def
|
158 |
-
|
159 |
-
|
160 |
-
user_input: str,
|
161 |
-
chat_history: List[Tuple[str, str]],
|
162 |
-
max_history_turns: int = 3
|
163 |
-
) -> str:
|
164 |
-
"""Construct a formatted prompt from the input components."""
|
165 |
-
system_message = f"Please assist based on the following context: {context}"
|
166 |
-
prompt = f"<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\n{system_message}<|eot_id|>"
|
167 |
|
168 |
-
for user_msg, assistant_msg in chat_history[-max_history_turns:]:
|
169 |
-
prompt += f"<|start_header_id|>user<|end_header_id|>\n\n{user_msg}<|eot_id|>"
|
170 |
-
prompt += f"<|start_header_id|>assistant<|end_header_id|>\n\n{assistant_msg}<|eot_id|>"
|
171 |
|
172 |
-
|
173 |
-
|
174 |
-
|
175 |
|
176 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
177 |
self,
|
178 |
prompt: str,
|
179 |
-
|
180 |
-
|
181 |
-
|
182 |
-
depth: int = 3,
|
183 |
-
breadth: int = 2
|
184 |
-
) -> str:
|
185 |
-
"""Generate a response using the specified strategy.
|
186 |
|
187 |
-
|
188 |
-
|
189 |
-
|
190 |
-
|
191 |
-
|
192 |
-
|
193 |
-
|
194 |
-
|
195 |
-
|
196 |
-
|
197 |
-
|
198 |
-
|
199 |
-
|
200 |
-
|
201 |
-
|
202 |
-
|
203 |
-
|
204 |
-
|
205 |
-
|
206 |
-
|
207 |
-
|
208 |
-
|
209 |
-
|
210 |
-
|
211 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
212 |
scored_outputs = []
|
213 |
for _ in range(num_samples):
|
214 |
input_ids = self.llama_tokenizer(prompt, return_tensors="pt").input_ids.to(self.device)
|
@@ -217,8 +247,9 @@ class LlamaGenerator:
|
|
217 |
score = self.prm_model(**self.llama_tokenizer(response, return_tensors="pt").to(self.device)).logits.mean().item()
|
218 |
scored_outputs.append((response, score))
|
219 |
return max(scored_outputs, key=lambda x: x[1])[0]
|
220 |
-
|
221 |
-
|
|
|
222 |
input_ids = self.llama_tokenizer(prompt, return_tensors="pt").input_ids.to(self.device)
|
223 |
outputs = self.llama_model.generate(
|
224 |
input_ids,
|
@@ -227,8 +258,9 @@ class LlamaGenerator:
|
|
227 |
**model_kwargs
|
228 |
)
|
229 |
return [self.llama_tokenizer.decode(output, skip_special_tokens=True) for output in outputs]
|
230 |
-
|
231 |
-
|
|
|
232 |
results = []
|
233 |
for _ in range(breadth):
|
234 |
input_ids = self.llama_tokenizer(prompt, return_tensors="pt").input_ids.to(self.device)
|
@@ -247,97 +279,322 @@ class LlamaGenerator:
|
|
247 |
results.append((extended_response, score))
|
248 |
return max(results, key=lambda x: x[1])[0]
|
249 |
|
250 |
-
|
251 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
252 |
|
253 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
254 |
self,
|
255 |
-
|
256 |
-
|
257 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
258 |
model_kwargs: Dict[str, Any],
|
259 |
-
max_history_turns: int = 3,
|
260 |
strategy: str = "default",
|
261 |
-
|
262 |
-
depth: int = 3,
|
263 |
-
breadth: int = 2
|
264 |
) -> str:
|
265 |
-
|
266 |
-
|
267 |
-
Args:
|
268 |
-
context (str): Context for the conversation
|
269 |
-
user_input (str): Current user input
|
270 |
-
chat_history (List[Tuple[str, str]]): List of (user, assistant) message pairs
|
271 |
-
model_kwargs (dict): Additional arguments for model.generate()
|
272 |
-
max_history_turns (int): Maximum number of history turns to include
|
273 |
-
strategy (str): Generation strategy
|
274 |
-
num_samples (int): Number of samples for applicable strategies
|
275 |
-
depth (int): Depth for DVTS strategy
|
276 |
-
breadth (int): Breadth for DVTS strategy
|
277 |
|
278 |
-
|
279 |
-
|
280 |
-
"""
|
281 |
-
prompt = self._construct_prompt(
|
282 |
-
context,
|
283 |
-
user_input,
|
284 |
-
chat_history,
|
285 |
-
max_history_turns
|
286 |
-
)
|
287 |
-
return self.generate(
|
288 |
prompt,
|
289 |
model_kwargs,
|
290 |
-
|
291 |
-
num_samples,
|
292 |
-
depth,
|
293 |
-
breadth
|
294 |
)
|
295 |
-
|
296 |
-
|
297 |
-
|
298 |
-
|
299 |
-
|
|
|
|
|
|
|
|
|
|
|
300 |
from fastapi.middleware.cors import CORSMiddleware
|
301 |
-
from
|
302 |
-
from
|
|
|
303 |
import asyncio
|
304 |
import uuid
|
305 |
from datetime import datetime
|
306 |
import json
|
|
|
|
|
|
|
|
|
307 |
|
308 |
class ChatMessage(BaseModel):
|
309 |
-
|
|
|
|
|
|
|
|
|
|
|
310 |
content: str = Field(..., description="Content of the message")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
311 |
|
312 |
class GenerationRequest(BaseModel):
|
313 |
-
|
314 |
-
|
315 |
-
|
316 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
317 |
|
318 |
class GenerationResponse(BaseModel):
|
319 |
-
|
320 |
-
|
321 |
-
|
322 |
-
|
323 |
-
|
324 |
-
|
325 |
-
|
326 |
-
|
327 |
-
|
328 |
-
|
329 |
-
|
330 |
-
|
331 |
-
|
332 |
-
|
|
|
|
|
|
|
333 |
|
334 |
-
|
335 |
-
|
336 |
-
|
337 |
-
|
338 |
-
async def startup_event():
|
339 |
global generator
|
340 |
try:
|
|
|
341 |
generator = LlamaGenerator(
|
342 |
llama_model_name="meta-llama/Llama-3.2-1B-Instruct",
|
343 |
prm_model_path=prm_model_path,
|
@@ -346,30 +603,87 @@ async def startup_event():
|
|
346 |
temperature=0.7
|
347 |
)
|
348 |
)
|
349 |
-
|
350 |
-
|
351 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
352 |
|
353 |
-
|
354 |
-
|
355 |
if not generator:
|
356 |
-
raise HTTPException(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
357 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
358 |
try:
|
359 |
-
# Format chat history
|
360 |
chat_history = [(msg.role, msg.content) for msg in request.messages[:-1]]
|
361 |
user_input = request.messages[-1].content
|
362 |
|
363 |
-
|
364 |
-
config = GenerationConfig(**request.config) if request.config else None
|
365 |
|
366 |
-
# Generate response
|
367 |
response = await asyncio.to_thread(
|
368 |
generator.generate_with_context,
|
369 |
context=request.context or "",
|
370 |
user_input=user_input,
|
371 |
chat_history=chat_history,
|
372 |
-
model_kwargs={}, # Add any model-specific kwargs here
|
373 |
config=config
|
374 |
)
|
375 |
|
@@ -381,23 +695,30 @@ async def generate(request: GenerationRequest):
|
|
381 |
raise HTTPException(status_code=500, detail=str(e))
|
382 |
|
383 |
@app.websocket("/generate/stream")
|
384 |
-
async def generate_stream(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
385 |
await websocket.accept()
|
386 |
|
387 |
try:
|
388 |
while True:
|
389 |
-
# Receive and parse request
|
390 |
request_data = await websocket.receive_text()
|
391 |
request = GenerationRequest.parse_raw(request_data)
|
392 |
|
393 |
-
# Format chat history
|
394 |
chat_history = [(msg.role, msg.content) for msg in request.messages[:-1]]
|
395 |
user_input = request.messages[-1].content
|
396 |
|
397 |
-
|
398 |
-
config = GenerationConfig(**request.config) if request.config else None
|
399 |
|
400 |
-
# Stream response
|
401 |
async for token in generator.generate_stream(
|
402 |
prompt=generator._construct_prompt(
|
403 |
context=request.context or "",
|
@@ -410,8 +731,7 @@ async def generate_stream(websocket):
|
|
410 |
"token": token,
|
411 |
"finished": False
|
412 |
}))
|
413 |
-
|
414 |
-
# Send finished message
|
415 |
await websocket.send_text(json.dumps({
|
416 |
"token": "",
|
417 |
"finished": True
|
@@ -426,4 +746,5 @@ async def generate_stream(websocket):
|
|
426 |
|
427 |
if __name__ == "__main__":
|
428 |
import uvicorn
|
429 |
-
uvicorn.run(app, host="0.0.0.0", port=8000)
|
|
|
|
1 |
+
# config.py
|
2 |
+
from dataclasses import dataclass, field
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
3 |
from enum import Enum
|
4 |
+
from typing import Dict, Any, Optional
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
5 |
|
6 |
class GenerationStrategy(str, Enum):
|
7 |
DEFAULT = "default"
|
|
|
9 |
BEST_OF_N = "best_of_n"
|
10 |
BEAM_SEARCH = "beam_search"
|
11 |
DVTS = "dvts"
|
12 |
+
COT = "chain_of_thought"
|
13 |
+
REACT = "react"
|
14 |
+
|
15 |
+
@dataclass
|
16 |
+
class ModelConfig:
|
17 |
+
model_kwargs: Dict[str, Any] = field(default_factory=dict)
|
18 |
+
tokenizer_kwargs: Dict[str, Any] = field(default_factory=dict)
|
19 |
+
quantization_kwargs: Dict[str, Any] = field(default_factory=dict)
|
20 |
|
21 |
@dataclass
|
22 |
class GenerationConfig:
|
23 |
num_samples: int = 5
|
24 |
depth: int = 3
|
25 |
breadth: int = 2
|
26 |
+
max_history_turns: int = 1
|
27 |
max_new_tokens: int = 50
|
28 |
temperature: float = 0.7
|
29 |
top_p: float = 0.9
|
30 |
+
top_k: int = 50
|
31 |
+
repetition_penalty: float = 1.1
|
32 |
+
length_penalty: float = 1.0
|
33 |
+
do_sample: bool = True
|
34 |
strategy: GenerationStrategy = GenerationStrategy.DEFAULT
|
35 |
|
36 |
+
# model_manager.py
|
37 |
+
import torch
|
38 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
39 |
+
from llama_cpp import Llama
|
40 |
+
from typing import Optional, Dict
|
41 |
+
import logging
|
42 |
+
from functools import lru_cache
|
43 |
+
|
44 |
+
class ModelManager:
|
45 |
+
def __init__(self, device: Optional[str] = None):
|
46 |
self.logger = logging.getLogger(__name__)
|
47 |
self.device = device or ("cuda" if torch.cuda.is_available() else "cpu")
|
48 |
+
self.models: Dict[str, Any] = {}
|
49 |
+
self.tokenizers: Dict[str, Any] = {}
|
50 |
+
|
51 |
+
def load_model(self, model_id: str, model_path: str, model_type: str, config: ModelConfig) -> None:
|
52 |
+
"""Load a model with specified configuration."""
|
53 |
try:
|
54 |
+
##could be differnt models, so we can use a factory pattern to load the correct model - textgen, llama, gguf, text2video, text2image etc.
|
55 |
+
if model_type == "llama":
|
56 |
+
self.tokenizers[model_id] = AutoTokenizer.from_pretrained(
|
57 |
+
model_path,
|
58 |
+
padding_side='left',
|
59 |
+
trust_remote_code=True,
|
60 |
+
**config.tokenizer_kwargs
|
61 |
+
)
|
62 |
+
if self.tokenizers[model_id].pad_token is None:
|
63 |
+
self.tokenizers[model_id].pad_token = self.tokenizers[model_id].eos_token
|
64 |
+
|
65 |
+
self.models[model_id] = AutoModelForCausalLM.from_pretrained(
|
66 |
+
model_path,
|
67 |
+
device_map="auto",
|
68 |
+
trust_remote_code=True,
|
69 |
+
**config.model_kwargs
|
70 |
+
)
|
71 |
+
elif model_type == "gguf":
|
72 |
+
#TODO load the model first from the cache, if not found load the model and save it in the cache
|
73 |
+
#from huggingface_hub import hf_hub_download
|
74 |
+
#prm_model_path = hf_hub_download(
|
75 |
+
# repo_id="tensorblock/Llama3.1-8B-PRM-Mistral-Data-GGUF",
|
76 |
+
# filename="Llama3.1-8B-PRM-Mistral-Data-Q4_K_M.gguf"
|
77 |
+
#)
|
78 |
+
|
79 |
+
|
80 |
+
self.models[model_id] = self._load_quantized_model(
|
81 |
+
model_path,
|
82 |
+
**config.quantization_kwargs
|
83 |
+
)
|
84 |
except Exception as e:
|
85 |
+
self.logger.error(f"Failed to load model {model_id}: {str(e)}")
|
86 |
raise
|
87 |
|
88 |
+
def unload_model(self, model_id: str) -> None:
|
89 |
+
"""Unload a model and free resources."""
|
90 |
+
if model_id in self.models:
|
91 |
+
del self.models[model_id]
|
92 |
+
if model_id in self.tokenizers:
|
93 |
+
del self.tokenizers[model_id]
|
94 |
+
torch.cuda.empty_cache()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
95 |
|
96 |
+
def _load_quantized_model(self, model_path: str, **kwargs) -> Llama:
|
97 |
+
"""Load a quantized GGUF model."""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
98 |
try:
|
|
|
99 |
n_gpu_layers = -1 if torch.cuda.is_available() else 0
|
|
|
|
|
100 |
model = Llama(
|
101 |
model_path=model_path,
|
102 |
+
n_ctx=kwargs.get('n_ctx', 2048),
|
103 |
+
n_batch=kwargs.get('n_batch', 512),
|
104 |
+
n_gpu_layers=kwargs.get('n_gpu_layers', n_gpu_layers),
|
105 |
+
verbose=kwargs.get('verbose', False)
|
106 |
)
|
|
|
|
|
107 |
return model
|
|
|
108 |
except Exception as e:
|
109 |
self.logger.error(f"Failed to load GGUF model: {str(e)}")
|
110 |
raise
|
111 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
112 |
|
113 |
+
# cache.py
|
114 |
+
from functools import lru_cache
|
115 |
+
from typing import Tuple, Any
|
116 |
+
|
117 |
+
# TODO explain howto use the cache
|
118 |
+
class ResponseCache:
|
119 |
+
def __init__(self, cache_size: int = 1000):
|
120 |
+
self.cache_size = cache_size
|
121 |
+
self._initialize_cache()
|
122 |
+
|
123 |
+
def _initialize_cache(self):
|
124 |
+
@lru_cache(maxsize=self.cache_size)
|
125 |
+
def cached_response(prompt: str, config_hash: str) -> Tuple[str, float]:
|
126 |
+
pass
|
127 |
+
self.get_cached_response = cached_response
|
128 |
+
|
129 |
+
def cache_response(self, prompt: str, config: GenerationConfig, response: str, score: float) -> None:
|
130 |
+
config_hash = hash(str(config.__dict__))
|
131 |
+
self.get_cached_response(prompt, str(config_hash))
|
132 |
|
133 |
+
def get_response(self, prompt: str, config: GenerationConfig) -> Optional[Tuple[str, float]]:
|
134 |
+
config_hash = hash(str(config.__dict__))
|
135 |
+
return self.get_cached_response(prompt, str(config_hash))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
136 |
|
|
|
|
|
|
|
137 |
|
138 |
+
# batch_processor.py
|
139 |
+
from typing import List, Dict
|
140 |
+
import asyncio
|
141 |
|
142 |
+
#TODO explain how to use the batch processor
|
143 |
+
class BatchProcessor:
|
144 |
+
def __init__(self, max_batch_size: int = 32, max_wait_time: float = 0.1):
|
145 |
+
self.max_batch_size = max_batch_size
|
146 |
+
self.max_wait_time = max_wait_time
|
147 |
+
self.pending_requests: List[Dict] = []
|
148 |
+
self.lock = asyncio.Lock()
|
149 |
+
|
150 |
+
async def add_request(self, request: Dict) -> Any:
|
151 |
+
async with self.lock:
|
152 |
+
self.pending_requests.append(request)
|
153 |
+
if len(self.pending_requests) >= self.max_batch_size:
|
154 |
+
return await self._process_batch()
|
155 |
+
else:
|
156 |
+
await asyncio.sleep(self.max_wait_time)
|
157 |
+
if self.pending_requests:
|
158 |
+
return await self._process_batch()
|
159 |
+
|
160 |
+
async def _process_batch(self) -> List[Any]:
|
161 |
+
batch = self.pending_requests[:self.max_batch_size]
|
162 |
+
self.pending_requests = self.pending_requests[self.max_batch_size:]
|
163 |
+
# TODO implement the batch processing logic
|
164 |
+
return batch
|
165 |
+
|
166 |
+
|
167 |
+
|
168 |
+
# base_generator.py
|
169 |
+
from abc import ABC, abstractmethod
|
170 |
+
from typing import AsyncGenerator, Dict, Any, Optional, List, Tuple
|
171 |
+
from dataclasses import dataclass
|
172 |
+
from logging import getLogger
|
173 |
+
|
174 |
+
from .config import GenerationConfig, ModelConfig
|
175 |
+
|
176 |
+
class BaseGenerator(ABC):
|
177 |
+
"""Base class for all generator implementations."""
|
178 |
+
|
179 |
+
def __init__(
|
180 |
+
self,
|
181 |
+
model_name: str,
|
182 |
+
device: Optional[str] = None,
|
183 |
+
default_generation_config: Optional[GenerationConfig] = None,
|
184 |
+
model_config: Optional[ModelConfig] = None,
|
185 |
+
cache_size: int = 1000,
|
186 |
+
max_batch_size: int = 32
|
187 |
+
):
|
188 |
+
self.logger = getLogger(__name__)
|
189 |
+
self.model_manager = ModelManager(device)
|
190 |
+
self.cache = ResponseCache(cache_size)
|
191 |
+
self.batch_processor = BatchProcessor(max_batch_size)
|
192 |
+
self.health_check = HealthCheck()
|
193 |
+
|
194 |
+
self.default_config = default_generation_config or GenerationConfig()
|
195 |
+
self.model_config = model_config or ModelConfig()
|
196 |
+
|
197 |
+
@abstractmethod
|
198 |
+
async def generate_stream(
|
199 |
self,
|
200 |
prompt: str,
|
201 |
+
config: Optional[GenerationConfig] = None
|
202 |
+
) -> AsyncGenerator[str, None]:
|
203 |
+
pass
|
|
|
|
|
|
|
|
|
204 |
|
205 |
+
@abstractmethod
|
206 |
+
def _get_generation_kwargs(self, config: GenerationConfig) -> Dict[str, Any]:
|
207 |
+
pass
|
208 |
+
|
209 |
+
@abstractmethod
|
210 |
+
def generate(self, prompt: str, model_kwargs: Dict[str, Any], **kwargs) -> str:
|
211 |
+
pass
|
212 |
+
|
213 |
+
# strategy.py
|
214 |
+
#TODO UPDATE Paths
|
215 |
+
from abc import ABC, abstractmethod
|
216 |
+
from typing import List, Tuple
|
217 |
+
|
218 |
+
class GenerationStrategy(ABC):
|
219 |
+
"""Base class for generation strategies."""
|
220 |
+
|
221 |
+
@abstractmethod
|
222 |
+
def generate(self, generator: 'BaseGenerator', prompt: str, model_kwargs: Dict[str, Any], **kwargs) -> str:
|
223 |
+
pass
|
224 |
+
|
225 |
+
class DefaultStrategy(GenerationStrategy):
|
226 |
+
def generate(self, generator: 'BaseGenerator', prompt: str, model_kwargs: Dict[str, Any], **kwargs) -> str:
|
227 |
+
input_ids = generator.tokenizer(prompt, return_tensors="pt").input_ids.to(generator.device)
|
228 |
+
output = generator.model.generate(input_ids, **model_kwargs)
|
229 |
+
return generator.tokenizer.decode(output[0], skip_special_tokens=True)
|
230 |
+
|
231 |
+
class MajorityVotingStrategy(GenerationStrategy):
|
232 |
+
def generate(self, generator: 'BaseGenerator', prompt: str, model_kwargs: Dict[str, Any], num_samples: int = 5, **kwargs) -> str:
|
233 |
+
outputs = []
|
234 |
+
for _ in range(num_samples):
|
235 |
+
input_ids = generator.tokenizer(prompt, return_tensors="pt").input_ids.to(generator.device)
|
236 |
+
output = generator.model.generate(input_ids, **model_kwargs)
|
237 |
+
outputs.append(generator.tokenizer.decode(output[0], skip_special_tokens=True))
|
238 |
+
return max(set(outputs), key=outputs.count)
|
239 |
+
|
240 |
+
class BestOfN(GenerationStrategy):
|
241 |
+
def generate(self, generator: 'BaseGenerator', prompt: str, model_kwargs: Dict[str, Any], num_samples: int = 5, **kwargs) -> str:
|
242 |
scored_outputs = []
|
243 |
for _ in range(num_samples):
|
244 |
input_ids = self.llama_tokenizer(prompt, return_tensors="pt").input_ids.to(self.device)
|
|
|
247 |
score = self.prm_model(**self.llama_tokenizer(response, return_tensors="pt").to(self.device)).logits.mean().item()
|
248 |
scored_outputs.append((response, score))
|
249 |
return max(scored_outputs, key=lambda x: x[1])[0]
|
250 |
+
|
251 |
+
class BeamSearch(GenerationStrategy):
|
252 |
+
def generate(self, generator: 'BaseGenerator', prompt: str, model_kwargs: Dict[str, Any], num_samples: int = 5, **kwargs) -> str:
|
253 |
input_ids = self.llama_tokenizer(prompt, return_tensors="pt").input_ids.to(self.device)
|
254 |
outputs = self.llama_model.generate(
|
255 |
input_ids,
|
|
|
258 |
**model_kwargs
|
259 |
)
|
260 |
return [self.llama_tokenizer.decode(output, skip_special_tokens=True) for output in outputs]
|
261 |
+
|
262 |
+
class DVT(GenerationStrategy):
|
263 |
+
def generate(self, generator: 'BaseGenerator', prompt: str, model_kwargs: Dict[str, Any], num_samples: int = 5, **kwargs) -> str:
|
264 |
results = []
|
265 |
for _ in range(breadth):
|
266 |
input_ids = self.llama_tokenizer(prompt, return_tensors="pt").input_ids.to(self.device)
|
|
|
279 |
results.append((extended_response, score))
|
280 |
return max(results, key=lambda x: x[1])[0]
|
281 |
|
282 |
+
class COT(GenerationStrategy):
|
283 |
+
def generate(self, generator: 'BaseGenerator', prompt: str, model_kwargs: Dict[str, Any], num_samples: int = 5, **kwargs) -> str:
|
284 |
+
#TODO implement the chain of thought strategy
|
285 |
+
|
286 |
+
return "Not implemented yet"
|
287 |
+
|
288 |
+
class ReAct(GenerationStrategy):
|
289 |
+
def generate(self, generator: 'BaseGenerator', prompt: str, model_kwargs: Dict[str, Any], num_samples: int = 5, **kwargs) -> str:
|
290 |
+
#TODO implement the ReAct framework
|
291 |
+
return "Not implemented yet"
|
292 |
+
# Add other strategy implementations...
|
293 |
+
|
294 |
+
# prompt_builder.py
|
295 |
+
from typing import Protocol, List, Tuple
|
296 |
+
from transformers import AutoTokenizer
|
297 |
+
|
298 |
+
class PromptTemplate(Protocol):
|
299 |
+
"""Protocol for prompt templates."""
|
300 |
+
def format(self, context: str, user_input: str, chat_history: List[Tuple[str, str]], **kwargs) -> str:
|
301 |
+
pass
|
302 |
+
|
303 |
+
class LlamaPromptTemplate:
|
304 |
+
def format(self, context: str, user_input: str, chat_history: List[Tuple[str, str]], max_history_turns: int = 1) -> str:
|
305 |
+
system_message = f"Please assist based on the following context: {context}"
|
306 |
+
prompt = f"<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\n{system_message}<|eot_id|>"
|
307 |
+
|
308 |
+
for user_msg, assistant_msg in chat_history[-max_history_turns:]:
|
309 |
+
prompt += f"<|start_header_id|>user<|end_header_id|>\n\n{user_msg}<|eot_id|>"
|
310 |
+
prompt += f"<|start_header_id|>assistant<|end_header_id|>\n\n{assistant_msg}<|eot_id|>"
|
311 |
+
|
312 |
+
prompt += f"<|start_header_id|>user<|end_header_id|>\n\n{user_input}<|eot_id|>"
|
313 |
+
prompt += "<|start_header_id|>assistant<|end_header_id|>\n\n"
|
314 |
+
return prompt
|
315 |
|
316 |
+
class TransformersPromptTemplate:
|
317 |
+
def __init__(self, model_path: str):
|
318 |
+
self.tokenizer = AutoTokenizer.from_pretrained(model_path)
|
319 |
+
|
320 |
+
def format(self, context: str, user_input: str, chat_history: List[Tuple[str, str]], **kwargs) -> str:
|
321 |
+
messages = [
|
322 |
+
{
|
323 |
+
"role": "system",
|
324 |
+
"content": f"Please assist based on the following context: {context}",
|
325 |
+
}
|
326 |
+
]
|
327 |
+
|
328 |
+
for user_msg, assistant_msg in chat_history:
|
329 |
+
messages.extend([
|
330 |
+
{"role": "user", "content": user_msg},
|
331 |
+
{"role": "assistant", "content": assistant_msg}
|
332 |
+
])
|
333 |
+
|
334 |
+
messages.append({"role": "user", "content": user_input})
|
335 |
+
|
336 |
+
tokenized_chat = self.tokenizer.apply_chat_template(
|
337 |
+
messages,
|
338 |
+
tokenize=False,
|
339 |
+
add_generation_prompt=True
|
340 |
+
)
|
341 |
+
return tokenized_chat
|
342 |
+
|
343 |
+
# health_check.py
|
344 |
+
import psutil
|
345 |
+
from dataclasses import dataclass
|
346 |
+
from typing import Dict, Any
|
347 |
+
|
348 |
+
@dataclass
|
349 |
+
class HealthStatus:
|
350 |
+
status: str
|
351 |
+
gpu_memory: Dict[str, float]
|
352 |
+
cpu_usage: float
|
353 |
+
ram_usage: float
|
354 |
+
model_status: Dict[str, str]
|
355 |
+
|
356 |
+
class HealthCheck:
|
357 |
+
@staticmethod
|
358 |
+
def check_gpu_memory() -> Dict[str, float]:
|
359 |
+
if torch.cuda.is_available():
|
360 |
+
return {
|
361 |
+
f"gpu_{i}": torch.cuda.memory_allocated(i) / 1024**3
|
362 |
+
for i in range(torch.cuda.device_count())
|
363 |
+
}
|
364 |
+
return {}
|
365 |
+
|
366 |
+
@staticmethod
|
367 |
+
def check_system_resources() -> HealthStatus:
|
368 |
+
return HealthStatus(
|
369 |
+
status="healthy",
|
370 |
+
gpu_memory=HealthCheck.check_gpu_memory(),
|
371 |
+
cpu_usage=psutil.cpu_percent(),
|
372 |
+
ram_usage=psutil.virtual_memory().percent,
|
373 |
+
#TODO add more system resources like disk, network, etc.
|
374 |
+
model_status={} # To be filled by the model manager
|
375 |
+
)
|
376 |
+
|
377 |
+
|
378 |
+
# llama_generator.py
|
379 |
+
class LlamaGenerator(BaseGenerator):
|
380 |
+
def __init__(
|
381 |
self,
|
382 |
+
llama_model_name: str,
|
383 |
+
prm_model_path: str,
|
384 |
+
device: Optional[str] = None,
|
385 |
+
default_generation_config: Optional[GenerationConfig] = None,
|
386 |
+
model_config: Optional[ModelConfig] = None,
|
387 |
+
cache_size: int = 1000,
|
388 |
+
max_batch_size: int = 32
|
389 |
+
):
|
390 |
+
super().__init__(
|
391 |
+
llama_model_name,
|
392 |
+
device,
|
393 |
+
default_generation_config,
|
394 |
+
model_config,
|
395 |
+
cache_size,
|
396 |
+
max_batch_size
|
397 |
+
)
|
398 |
+
|
399 |
+
# Initialize models
|
400 |
+
self.model_manager.load_model(
|
401 |
+
"llama",
|
402 |
+
llama_model_name,
|
403 |
+
"llama",
|
404 |
+
self.model_config
|
405 |
+
)
|
406 |
+
self.model_manager.load_model(
|
407 |
+
"prm",
|
408 |
+
prm_model_path,
|
409 |
+
"gguf",
|
410 |
+
self.model_config
|
411 |
+
)
|
412 |
+
|
413 |
+
self.prompt_builder = LlamaPromptTemplate()
|
414 |
+
self._init_strategies()
|
415 |
+
|
416 |
+
def _init_strategies(self):
|
417 |
+
self.strategies = {
|
418 |
+
"default": DefaultStrategy(),
|
419 |
+
"majority_voting": MajorityVotingStrategy(),
|
420 |
+
"best_of_n": BestOfN(),
|
421 |
+
"beam_search": BeamSearch(),
|
422 |
+
"dvts": DVT(),
|
423 |
+
}
|
424 |
+
|
425 |
+
def _get_generation_kwargs(self, config: GenerationConfig) -> Dict[str, Any]:
|
426 |
+
"""Get generation kwargs based on config."""
|
427 |
+
return {
|
428 |
+
key: getattr(config, key)
|
429 |
+
for key in [
|
430 |
+
"max_new_tokens",
|
431 |
+
"temperature",
|
432 |
+
"top_p",
|
433 |
+
"top_k",
|
434 |
+
"repetition_penalty",
|
435 |
+
"length_penalty",
|
436 |
+
"do_sample"
|
437 |
+
]
|
438 |
+
if hasattr(config, key)
|
439 |
+
}
|
440 |
+
|
441 |
+
def generate(
|
442 |
+
self,
|
443 |
+
prompt: str,
|
444 |
model_kwargs: Dict[str, Any],
|
|
|
445 |
strategy: str = "default",
|
446 |
+
**kwargs
|
|
|
|
|
447 |
) -> str:
|
448 |
+
if strategy not in self.strategies:
|
449 |
+
raise ValueError(f"Unknown strategy: {strategy}")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
450 |
|
451 |
+
return self.strategies[strategy].generate(
|
452 |
+
self,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
453 |
prompt,
|
454 |
model_kwargs,
|
455 |
+
**kwargs
|
|
|
|
|
|
|
456 |
)
|
457 |
+
|
458 |
+
def check_health(self) -> HealthStatus:
|
459 |
+
"""Check the health status of the generator."""
|
460 |
+
return self.health_check.check_system_resources() # TODO add model status
|
461 |
+
|
462 |
+
|
463 |
+
###################
|
464 |
+
#################
|
465 |
+
|
466 |
+
from fastapi import FastAPI, HTTPException, BackgroundTasks, WebSocket, Depends
|
467 |
from fastapi.middleware.cors import CORSMiddleware
|
468 |
+
from fastapi.responses import StreamingResponse
|
469 |
+
from pydantic import BaseModel, Field, ConfigDict
|
470 |
+
from typing import List, Optional, Dict, Any, AsyncGenerator
|
471 |
import asyncio
|
472 |
import uuid
|
473 |
from datetime import datetime
|
474 |
import json
|
475 |
+
from huggingface_hub import hf_hub_download
|
476 |
+
from contextlib import asynccontextmanager
|
477 |
+
|
478 |
+
|
479 |
|
480 |
class ChatMessage(BaseModel):
|
481 |
+
"""A single message in the chat history."""
|
482 |
+
role: str = Field(
|
483 |
+
...,
|
484 |
+
description="Role of the message sender",
|
485 |
+
examples=["user", "assistant"]
|
486 |
+
)
|
487 |
content: str = Field(..., description="Content of the message")
|
488 |
+
|
489 |
+
model_config = ConfigDict(
|
490 |
+
json_schema_extra={
|
491 |
+
"example": {
|
492 |
+
"role": "user",
|
493 |
+
"content": "What is the capital of France?"
|
494 |
+
}
|
495 |
+
}
|
496 |
+
)
|
497 |
+
|
498 |
+
class GenerationConfig(BaseModel):
|
499 |
+
"""Configuration for text generation."""
|
500 |
+
temperature: float = Field(
|
501 |
+
0.7,
|
502 |
+
ge=0.0,
|
503 |
+
le=2.0,
|
504 |
+
description="Controls randomness in the output. Higher values (e.g., 0.8) make the output more random, lower values (e.g., 0.2) make it more focused and deterministic."
|
505 |
+
)
|
506 |
+
max_new_tokens: int = Field(
|
507 |
+
100,
|
508 |
+
ge=1,
|
509 |
+
le=2048,
|
510 |
+
description="Maximum number of tokens to generate"
|
511 |
+
)
|
512 |
+
top_p: float = Field(
|
513 |
+
0.9,
|
514 |
+
ge=0.0,
|
515 |
+
le=1.0,
|
516 |
+
description="Nucleus sampling parameter. Only tokens with cumulative probability < top_p are considered."
|
517 |
+
)
|
518 |
+
top_k: int = Field(
|
519 |
+
50,
|
520 |
+
ge=0,
|
521 |
+
description="Only consider the top k tokens for text generation"
|
522 |
+
)
|
523 |
+
strategy: str = Field(
|
524 |
+
"default",
|
525 |
+
description="Generation strategy to use",
|
526 |
+
examples=["default", "majority_voting", "best_of_n", "beam_search", "dvts"]
|
527 |
+
)
|
528 |
+
num_samples: int = Field(
|
529 |
+
5,
|
530 |
+
ge=1,
|
531 |
+
le=10,
|
532 |
+
description="Number of samples to generate (used in majority_voting and best_of_n strategies)"
|
533 |
+
)
|
534 |
|
535 |
class GenerationRequest(BaseModel):
|
536 |
+
"""Request model for text generation."""
|
537 |
+
context: Optional[str] = Field(
|
538 |
+
None,
|
539 |
+
description="Additional context to guide the generation",
|
540 |
+
examples=["You are a helpful assistant skilled in Python programming"]
|
541 |
+
)
|
542 |
+
messages: List[ChatMessage] = Field(
|
543 |
+
...,
|
544 |
+
description="Chat history including the current message",
|
545 |
+
min_items=1
|
546 |
+
)
|
547 |
+
config: Optional[GenerationConfig] = Field(
|
548 |
+
None,
|
549 |
+
description="Generation configuration parameters"
|
550 |
+
)
|
551 |
+
stream: bool = Field(
|
552 |
+
False,
|
553 |
+
description="Whether to stream the response token by token"
|
554 |
+
)
|
555 |
+
|
556 |
+
model_config = ConfigDict(
|
557 |
+
json_schema_extra={
|
558 |
+
"example": {
|
559 |
+
"context": "You are a helpful assistant",
|
560 |
+
"messages": [
|
561 |
+
{"role": "user", "content": "What is the capital of France?"}
|
562 |
+
],
|
563 |
+
"config": {
|
564 |
+
"temperature": 0.7,
|
565 |
+
"max_new_tokens": 100
|
566 |
+
},
|
567 |
+
"stream": False
|
568 |
+
}
|
569 |
+
}
|
570 |
+
)
|
571 |
|
572 |
class GenerationResponse(BaseModel):
|
573 |
+
"""Response model for text generation."""
|
574 |
+
id: str = Field(..., description="Unique generation ID")
|
575 |
+
content: str = Field(..., description="Generated text content")
|
576 |
+
created_at: datetime = Field(
|
577 |
+
default_factory=datetime.now,
|
578 |
+
description="Timestamp of generation"
|
579 |
+
)
|
580 |
+
|
581 |
+
|
582 |
+
# Model and cache management
|
583 |
+
async def get_prm_model_path():
|
584 |
+
"""Download and cache the PRM model."""
|
585 |
+
return await asyncio.to_thread(
|
586 |
+
hf_hub_download,
|
587 |
+
repo_id="tensorblock/Llama3.1-8B-PRM-Mistral-Data-GGUF",
|
588 |
+
filename="Llama3.1-8B-PRM-Mistral-Data-Q4_K_M.gguf"
|
589 |
+
)
|
590 |
|
591 |
+
@asynccontextmanager
|
592 |
+
async def lifespan(app: FastAPI):
|
593 |
+
"""Lifecycle management for the FastAPI application."""
|
594 |
+
# Startup: Initialize generator
|
|
|
595 |
global generator
|
596 |
try:
|
597 |
+
prm_model_path = await get_prm_model_path()
|
598 |
generator = LlamaGenerator(
|
599 |
llama_model_name="meta-llama/Llama-3.2-1B-Instruct",
|
600 |
prm_model_path=prm_model_path,
|
|
|
603 |
temperature=0.7
|
604 |
)
|
605 |
)
|
606 |
+
yield
|
607 |
+
finally:
|
608 |
+
# Shutdown: Clean up resources
|
609 |
+
if generator:
|
610 |
+
await asyncio.to_thread(generator.cleanup)
|
611 |
+
|
612 |
+
# FastAPI application
|
613 |
+
app = FastAPI(
|
614 |
+
title="Inference Deluxe Service",
|
615 |
+
description="""
|
616 |
+
A service for generating text using LLaMA models with various generation strategies.
|
617 |
+
|
618 |
+
Generation Strategies:
|
619 |
+
- default: Standard autoregressive generation
|
620 |
+
- majority_voting: Generates multiple responses and selects the most common one
|
621 |
+
- best_of_n: Generates multiple responses and selects the best based on a scoring metric
|
622 |
+
- beam_search: Uses beam search for more coherent text generation
|
623 |
+
- dvts: Dynamic vocabulary tree search for efficient generation
|
624 |
+
""",
|
625 |
+
version="1.0.0",
|
626 |
+
lifespan=lifespan
|
627 |
+
)
|
628 |
+
|
629 |
+
# CORS middleware
|
630 |
+
app.add_middleware(
|
631 |
+
CORSMiddleware,
|
632 |
+
allow_origins=["*"],
|
633 |
+
allow_credentials=True,
|
634 |
+
allow_methods=["*"],
|
635 |
+
allow_headers=["*"],
|
636 |
+
)
|
637 |
|
638 |
+
async def get_generator():
|
639 |
+
"""Dependency to get the generator instance."""
|
640 |
if not generator:
|
641 |
+
raise HTTPException(
|
642 |
+
status_code=503,
|
643 |
+
detail="Generator not initialized"
|
644 |
+
)
|
645 |
+
return generator
|
646 |
+
|
647 |
+
@app.post(
|
648 |
+
"/generate",
|
649 |
+
response_model=GenerationResponse,
|
650 |
+
tags=["generation"],
|
651 |
+
summary="Generate text response",
|
652 |
+
response_description="Generated text with unique identifier"
|
653 |
+
)
|
654 |
+
async def generate(
|
655 |
+
request: GenerationRequest,
|
656 |
+
generator: Any = Depends(get_generator)
|
657 |
+
):
|
658 |
+
"""
|
659 |
+
Generate a text response based on the provided context and chat history.
|
660 |
+
|
661 |
+
The generation process can be customized using various parameters in the config:
|
662 |
+
- temperature: Controls randomness (0.0 to 2.0)
|
663 |
+
- max_new_tokens: Maximum length of generated text
|
664 |
+
- top_p: Nucleus sampling parameter
|
665 |
+
- top_k: Top-k sampling parameter
|
666 |
+
- strategy: Generation strategy to use
|
667 |
+
- num_samples: Number of samples for applicable strategies
|
668 |
|
669 |
+
Generation Strategies:
|
670 |
+
- default: Standard generation
|
671 |
+
- majority_voting: Generates multiple responses and uses the most common one
|
672 |
+
- best_of_n: Generates multiple responses and picks the best
|
673 |
+
- beam_search: Uses beam search for coherent generation
|
674 |
+
- dvts: Dynamic vocabulary tree search
|
675 |
+
"""
|
676 |
try:
|
|
|
677 |
chat_history = [(msg.role, msg.content) for msg in request.messages[:-1]]
|
678 |
user_input = request.messages[-1].content
|
679 |
|
680 |
+
config = request.config or GenerationConfig()
|
|
|
681 |
|
|
|
682 |
response = await asyncio.to_thread(
|
683 |
generator.generate_with_context,
|
684 |
context=request.context or "",
|
685 |
user_input=user_input,
|
686 |
chat_history=chat_history,
|
|
|
687 |
config=config
|
688 |
)
|
689 |
|
|
|
695 |
raise HTTPException(status_code=500, detail=str(e))
|
696 |
|
697 |
@app.websocket("/generate/stream")
|
698 |
+
async def generate_stream(
|
699 |
+
websocket: WebSocket,
|
700 |
+
generator: Any = Depends(get_generator)
|
701 |
+
):
|
702 |
+
"""
|
703 |
+
Stream generated text tokens over a WebSocket connection.
|
704 |
+
|
705 |
+
The stream sends JSON messages with the following format:
|
706 |
+
- During generation: {"token": "generated_token", "finished": false}
|
707 |
+
- End of generation: {"token": "", "finished": true}
|
708 |
+
- Error: {"error": "error_message"}
|
709 |
+
"""
|
710 |
await websocket.accept()
|
711 |
|
712 |
try:
|
713 |
while True:
|
|
|
714 |
request_data = await websocket.receive_text()
|
715 |
request = GenerationRequest.parse_raw(request_data)
|
716 |
|
|
|
717 |
chat_history = [(msg.role, msg.content) for msg in request.messages[:-1]]
|
718 |
user_input = request.messages[-1].content
|
719 |
|
720 |
+
config = request.config or GenerationConfig()
|
|
|
721 |
|
|
|
722 |
async for token in generator.generate_stream(
|
723 |
prompt=generator._construct_prompt(
|
724 |
context=request.context or "",
|
|
|
731 |
"token": token,
|
732 |
"finished": False
|
733 |
}))
|
734 |
+
|
|
|
735 |
await websocket.send_text(json.dumps({
|
736 |
"token": "",
|
737 |
"finished": True
|
|
|
746 |
|
747 |
if __name__ == "__main__":
|
748 |
import uvicorn
|
749 |
+
uvicorn.run(app, host="0.0.0.0", port=8000)
|
750 |
+
|