Spaces:
Running
Running
perf: improvement requests handling with async mode
Browse files- lightweight_embeddings/router.py +32 -21
- lightweight_embeddings/service.py +150 -82
- requirements.txt +1 -0
lightweight_embeddings/router.py
CHANGED
@@ -2,8 +2,8 @@ from __future__ import annotations
|
|
2 |
|
3 |
import logging
|
4 |
import os
|
5 |
-
from typing import Dict, List, Union
|
6 |
from datetime import datetime
|
|
|
7 |
|
8 |
from fastapi import APIRouter, BackgroundTasks, HTTPException
|
9 |
from pydantic import BaseModel, Field
|
@@ -27,45 +27,45 @@ router = APIRouter(
|
|
27 |
|
28 |
class EmbeddingRequest(BaseModel):
|
29 |
"""
|
30 |
-
|
31 |
"""
|
32 |
|
33 |
model: str = Field(
|
34 |
default=TextModelType.MULTILINGUAL_E5_SMALL.value,
|
35 |
description=(
|
36 |
"Which model ID to use? "
|
37 |
-
"Text: ['multilingual-e5-small', 'multilingual-e5-base', 'multilingual-e5-large',
|
38 |
-
"
|
|
|
|
|
39 |
),
|
40 |
)
|
41 |
input: Union[str, List[str]] = Field(
|
42 |
-
..., description="Text(s) or
|
43 |
)
|
44 |
|
45 |
|
46 |
class RankRequest(BaseModel):
|
47 |
"""
|
48 |
-
|
49 |
"""
|
50 |
|
51 |
model: str = Field(
|
52 |
default=TextModelType.MULTILINGUAL_E5_SMALL.value,
|
53 |
description=(
|
54 |
"Model ID for the queries. "
|
55 |
-
"
|
56 |
),
|
57 |
)
|
58 |
queries: Union[str, List[str]] = Field(
|
59 |
-
..., description="Query text or image(s) depending on the model type."
|
60 |
-
)
|
61 |
-
candidates: List[str] = Field(
|
62 |
-
..., description="Candidate texts to rank. Must be text."
|
63 |
)
|
|
|
64 |
|
65 |
|
66 |
class EmbeddingResponse(BaseModel):
|
67 |
"""
|
68 |
-
Response
|
69 |
"""
|
70 |
|
71 |
object: str
|
@@ -76,7 +76,7 @@ class EmbeddingResponse(BaseModel):
|
|
76 |
|
77 |
class RankResponse(BaseModel):
|
78 |
"""
|
79 |
-
Response
|
80 |
"""
|
81 |
|
82 |
probabilities: List[List[float]]
|
@@ -84,7 +84,9 @@ class RankResponse(BaseModel):
|
|
84 |
|
85 |
|
86 |
class StatsBucket(BaseModel):
|
87 |
-
"""
|
|
|
|
|
88 |
|
89 |
total: Dict[str, int]
|
90 |
daily: Dict[str, int]
|
@@ -94,12 +96,15 @@ class StatsBucket(BaseModel):
|
|
94 |
|
95 |
|
96 |
class StatsResponse(BaseModel):
|
97 |
-
"""
|
|
|
|
|
98 |
|
99 |
access: StatsBucket
|
100 |
tokens: StatsBucket
|
101 |
|
102 |
|
|
|
103 |
service_config = ModelConfig()
|
104 |
embeddings_service = EmbeddingsService(config=service_config)
|
105 |
|
@@ -115,16 +120,16 @@ async def create_embeddings(
|
|
115 |
request: EmbeddingRequest, background_tasks: BackgroundTasks
|
116 |
):
|
117 |
"""
|
118 |
-
|
119 |
"""
|
120 |
try:
|
121 |
modality = detect_model_kind(request.model)
|
122 |
embeddings = await embeddings_service.generate_embeddings(
|
123 |
-
inputs=request.input,
|
124 |
model=request.model,
|
|
|
125 |
)
|
126 |
|
127 |
-
# Estimate tokens
|
128 |
total_tokens = 0
|
129 |
if modality == ModelKind.TEXT:
|
130 |
total_tokens = embeddings_service.estimate_tokens(request.input)
|
@@ -148,6 +153,7 @@ async def create_embeddings(
|
|
148 |
}
|
149 |
)
|
150 |
|
|
|
151 |
background_tasks.add_task(
|
152 |
analytics.access, request.model, resp["usage"]["total_tokens"]
|
153 |
)
|
@@ -166,7 +172,7 @@ async def create_embeddings(
|
|
166 |
@router.post("/rank", response_model=RankResponse, tags=["rank"])
|
167 |
async def rank_candidates(request: RankRequest, background_tasks: BackgroundTasks):
|
168 |
"""
|
169 |
-
|
170 |
"""
|
171 |
try:
|
172 |
results = await embeddings_service.rank(
|
@@ -175,6 +181,7 @@ async def rank_candidates(request: RankRequest, background_tasks: BackgroundTask
|
|
175 |
candidates=request.candidates,
|
176 |
)
|
177 |
|
|
|
178 |
background_tasks.add_task(
|
179 |
analytics.access, request.model, results["usage"]["total_tokens"]
|
180 |
)
|
@@ -192,14 +199,18 @@ async def rank_candidates(request: RankRequest, background_tasks: BackgroundTask
|
|
192 |
|
193 |
@router.get("/stats", response_model=StatsResponse, tags=["stats"])
|
194 |
async def get_stats():
|
195 |
-
"""
|
|
|
|
|
196 |
try:
|
197 |
day_key = datetime.utcnow().strftime("%Y-%m-%d")
|
198 |
week_key = f"{datetime.utcnow().year}-W{datetime.utcnow().strftime('%U')}"
|
199 |
month_key = datetime.utcnow().strftime("%Y-%m")
|
200 |
year_key = datetime.utcnow().strftime("%Y")
|
201 |
|
202 |
-
stats_data =
|
|
|
|
|
203 |
|
204 |
return {
|
205 |
"access": {
|
|
|
2 |
|
3 |
import logging
|
4 |
import os
|
|
|
5 |
from datetime import datetime
|
6 |
+
from typing import Dict, List, Union
|
7 |
|
8 |
from fastapi import APIRouter, BackgroundTasks, HTTPException
|
9 |
from pydantic import BaseModel, Field
|
|
|
27 |
|
28 |
class EmbeddingRequest(BaseModel):
|
29 |
"""
|
30 |
+
Request model for generating embeddings.
|
31 |
"""
|
32 |
|
33 |
model: str = Field(
|
34 |
default=TextModelType.MULTILINGUAL_E5_SMALL.value,
|
35 |
description=(
|
36 |
"Which model ID to use? "
|
37 |
+
"Text options: ['multilingual-e5-small', 'multilingual-e5-base', 'multilingual-e5-large', "
|
38 |
+
"'snowflake-arctic-embed-l-v2.0', 'paraphrase-multilingual-MiniLM-L12-v2', "
|
39 |
+
"'paraphrase-multilingual-mpnet-base-v2', 'bge-m3']. "
|
40 |
+
"Image option: ['siglip-base-patch16-256-multilingual']."
|
41 |
),
|
42 |
)
|
43 |
input: Union[str, List[str]] = Field(
|
44 |
+
..., description="Text(s) or image URL(s)/path(s)."
|
45 |
)
|
46 |
|
47 |
|
48 |
class RankRequest(BaseModel):
|
49 |
"""
|
50 |
+
Request model for ranking candidates.
|
51 |
"""
|
52 |
|
53 |
model: str = Field(
|
54 |
default=TextModelType.MULTILINGUAL_E5_SMALL.value,
|
55 |
description=(
|
56 |
"Model ID for the queries. "
|
57 |
+
"Can be a text or image model (e.g. 'siglip-base-patch16-256-multilingual' for images)."
|
58 |
),
|
59 |
)
|
60 |
queries: Union[str, List[str]] = Field(
|
61 |
+
..., description="Query text(s) or image(s) depending on the model type."
|
|
|
|
|
|
|
62 |
)
|
63 |
+
candidates: List[str] = Field(..., description="Candidate texts to rank.")
|
64 |
|
65 |
|
66 |
class EmbeddingResponse(BaseModel):
|
67 |
"""
|
68 |
+
Response model for embeddings.
|
69 |
"""
|
70 |
|
71 |
object: str
|
|
|
76 |
|
77 |
class RankResponse(BaseModel):
|
78 |
"""
|
79 |
+
Response model for ranking results.
|
80 |
"""
|
81 |
|
82 |
probabilities: List[List[float]]
|
|
|
84 |
|
85 |
|
86 |
class StatsBucket(BaseModel):
|
87 |
+
"""
|
88 |
+
Model for daily/weekly/monthly/yearly stats.
|
89 |
+
"""
|
90 |
|
91 |
total: Dict[str, int]
|
92 |
daily: Dict[str, int]
|
|
|
96 |
|
97 |
|
98 |
class StatsResponse(BaseModel):
|
99 |
+
"""
|
100 |
+
Analytics stats response model, including both access and token counts.
|
101 |
+
"""
|
102 |
|
103 |
access: StatsBucket
|
104 |
tokens: StatsBucket
|
105 |
|
106 |
|
107 |
+
# Initialize the embeddings service and analytics.
|
108 |
service_config = ModelConfig()
|
109 |
embeddings_service = EmbeddingsService(config=service_config)
|
110 |
|
|
|
120 |
request: EmbeddingRequest, background_tasks: BackgroundTasks
|
121 |
):
|
122 |
"""
|
123 |
+
Generate embeddings for the given text or image inputs.
|
124 |
"""
|
125 |
try:
|
126 |
modality = detect_model_kind(request.model)
|
127 |
embeddings = await embeddings_service.generate_embeddings(
|
|
|
128 |
model=request.model,
|
129 |
+
inputs=request.input,
|
130 |
)
|
131 |
|
132 |
+
# Estimate tokens if using a text model.
|
133 |
total_tokens = 0
|
134 |
if modality == ModelKind.TEXT:
|
135 |
total_tokens = embeddings_service.estimate_tokens(request.input)
|
|
|
153 |
}
|
154 |
)
|
155 |
|
156 |
+
# Record analytics in the background.
|
157 |
background_tasks.add_task(
|
158 |
analytics.access, request.model, resp["usage"]["total_tokens"]
|
159 |
)
|
|
|
172 |
@router.post("/rank", response_model=RankResponse, tags=["rank"])
|
173 |
async def rank_candidates(request: RankRequest, background_tasks: BackgroundTasks):
|
174 |
"""
|
175 |
+
Rank candidate texts against the given queries.
|
176 |
"""
|
177 |
try:
|
178 |
results = await embeddings_service.rank(
|
|
|
181 |
candidates=request.candidates,
|
182 |
)
|
183 |
|
184 |
+
# Record analytics in the background.
|
185 |
background_tasks.add_task(
|
186 |
analytics.access, request.model, results["usage"]["total_tokens"]
|
187 |
)
|
|
|
199 |
|
200 |
@router.get("/stats", response_model=StatsResponse, tags=["stats"])
|
201 |
async def get_stats():
|
202 |
+
"""
|
203 |
+
Retrieve usage statistics for all models, including access counts and token usage.
|
204 |
+
"""
|
205 |
try:
|
206 |
day_key = datetime.utcnow().strftime("%Y-%m-%d")
|
207 |
week_key = f"{datetime.utcnow().year}-W{datetime.utcnow().strftime('%U')}"
|
208 |
month_key = datetime.utcnow().strftime("%Y-%m")
|
209 |
year_key = datetime.utcnow().strftime("%Y")
|
210 |
|
211 |
+
stats_data = (
|
212 |
+
await analytics.stats()
|
213 |
+
) # Expected to return a dict with 'access' and 'tokens' keys
|
214 |
|
215 |
return {
|
216 |
"access": {
|
lightweight_embeddings/service.py
CHANGED
@@ -1,5 +1,6 @@
|
|
1 |
from __future__ import annotations
|
2 |
|
|
|
3 |
import logging
|
4 |
from enum import Enum
|
5 |
from typing import List, Union, Dict, Optional, NamedTuple, Any
|
@@ -9,7 +10,7 @@ from io import BytesIO
|
|
9 |
from hashlib import md5
|
10 |
from cachetools import LRUCache
|
11 |
|
12 |
-
import
|
13 |
import numpy as np
|
14 |
import torch
|
15 |
from PIL import Image
|
@@ -45,9 +46,7 @@ class ImageModelType(str, Enum):
|
|
45 |
|
46 |
class ModelInfo(NamedTuple):
|
47 |
"""
|
48 |
-
|
49 |
-
- model_id: Hugging Face model ID (or local path)
|
50 |
-
- onnx_file: Path to ONNX file (if available)
|
51 |
"""
|
52 |
|
53 |
model_id: str
|
@@ -69,7 +68,7 @@ class ModelConfig:
|
|
69 |
@property
|
70 |
def text_model_info(self) -> ModelInfo:
|
71 |
"""
|
72 |
-
|
73 |
"""
|
74 |
text_configs = {
|
75 |
TextModelType.MULTILINGUAL_E5_SMALL: ModelInfo(
|
@@ -110,7 +109,7 @@ class ModelConfig:
|
|
110 |
@property
|
111 |
def image_model_info(self) -> ModelInfo:
|
112 |
"""
|
113 |
-
|
114 |
"""
|
115 |
image_configs = {
|
116 |
ImageModelType.SIGLIP_BASE_PATCH16_256_MULTILINGUAL: ModelInfo(
|
@@ -121,14 +120,20 @@ class ModelConfig:
|
|
121 |
|
122 |
|
123 |
class ModelKind(str, Enum):
|
|
|
|
|
|
|
|
|
124 |
TEXT = "text"
|
125 |
IMAGE = "image"
|
126 |
|
127 |
|
128 |
def detect_model_kind(model_id: str) -> ModelKind:
|
129 |
"""
|
130 |
-
Detect whether
|
131 |
-
|
|
|
|
|
132 |
"""
|
133 |
if model_id in [m.value for m in TextModelType]:
|
134 |
return ModelKind.TEXT
|
@@ -145,32 +150,38 @@ def detect_model_kind(model_id: str) -> ModelKind:
|
|
145 |
class EmbeddingsService:
|
146 |
"""
|
147 |
Service for generating text/image embeddings and performing similarity ranking.
|
148 |
-
|
149 |
"""
|
150 |
|
151 |
def __init__(self, config: Optional[ModelConfig] = None):
|
|
|
|
|
|
|
|
|
152 |
self.lru_cache = LRUCache(maxsize=10_000)
|
153 |
self.device = "cuda" if torch.cuda.is_available() else "cpu"
|
154 |
self.config = config or ModelConfig()
|
155 |
|
156 |
-
# Dictionaries to hold preloaded models
|
157 |
self.text_models: Dict[TextModelType, SentenceTransformer] = {}
|
158 |
self.image_models: Dict[ImageModelType, AutoModel] = {}
|
159 |
self.image_processors: Dict[ImageModelType, AutoProcessor] = {}
|
160 |
|
161 |
-
#
|
|
|
|
|
|
|
162 |
self._load_all_models()
|
163 |
|
164 |
def _load_all_models(self) -> None:
|
165 |
"""
|
166 |
-
Pre-load all
|
167 |
"""
|
168 |
try:
|
169 |
-
# Preload text models
|
170 |
for t_model_type in TextModelType:
|
171 |
info = ModelConfig(text_model_type=t_model_type).text_model_info
|
172 |
logger.info("Loading text model: %s", info.model_id)
|
173 |
-
|
174 |
if info.onnx_file:
|
175 |
logger.info("Using ONNX file: %s", info.onnx_file)
|
176 |
self.text_models[t_model_type] = SentenceTransformer(
|
@@ -190,16 +201,15 @@ class EmbeddingsService:
|
|
190 |
trust_remote_code=True,
|
191 |
)
|
192 |
|
193 |
-
# Preload image models
|
194 |
for i_model_type in ImageModelType:
|
195 |
model_id = ModelConfig(
|
196 |
image_model_type=i_model_type
|
197 |
).image_model_info.model_id
|
198 |
logger.info("Loading image model: %s", model_id)
|
199 |
-
|
200 |
model = AutoModel.from_pretrained(model_id).to(self.device)
|
|
|
201 |
processor = AutoProcessor.from_pretrained(model_id)
|
202 |
-
|
203 |
self.image_models[i_model_type] = model
|
204 |
self.image_processors[i_model_type] = processor
|
205 |
|
@@ -212,8 +222,10 @@ class EmbeddingsService:
|
|
212 |
@staticmethod
|
213 |
def _validate_text_list(input_text: Union[str, List[str]]) -> List[str]:
|
214 |
"""
|
215 |
-
|
216 |
-
|
|
|
|
|
217 |
"""
|
218 |
if isinstance(input_text, str):
|
219 |
if not input_text.strip():
|
@@ -233,8 +245,10 @@ class EmbeddingsService:
|
|
233 |
@staticmethod
|
234 |
def _validate_image_list(input_images: Union[str, List[str]]) -> List[str]:
|
235 |
"""
|
236 |
-
|
237 |
-
|
|
|
|
|
238 |
"""
|
239 |
if isinstance(input_images, str):
|
240 |
if not input_images.strip():
|
@@ -251,24 +265,51 @@ class EmbeddingsService:
|
|
251 |
|
252 |
return input_images
|
253 |
|
254 |
-
def
|
255 |
"""
|
256 |
-
|
257 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
258 |
"""
|
259 |
try:
|
260 |
if path_or_url.startswith("http"):
|
261 |
-
|
262 |
-
|
263 |
-
|
|
|
|
|
264 |
else:
|
265 |
-
|
266 |
-
|
267 |
-
|
268 |
-
processed_data = processor(images=img, return_tensors="pt").to(self.device)
|
269 |
-
return processed_data
|
270 |
except Exception as e:
|
271 |
-
raise ValueError(f"Error
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
272 |
|
273 |
def _generate_text_embeddings(
|
274 |
self,
|
@@ -276,8 +317,14 @@ class EmbeddingsService:
|
|
276 |
texts: List[str],
|
277 |
) -> np.ndarray:
|
278 |
"""
|
279 |
-
|
280 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
281 |
"""
|
282 |
try:
|
283 |
if len(texts) == 1:
|
@@ -285,48 +332,54 @@ class EmbeddingsService:
|
|
285 |
key = md5(f"{model_id}:{single_text}".encode("utf-8")).hexdigest()[:8]
|
286 |
if key in self.lru_cache:
|
287 |
return self.lru_cache[key]
|
288 |
-
|
289 |
model = self.text_models[model_id]
|
290 |
emb = model.encode([single_text])
|
291 |
self.lru_cache[key] = emb
|
292 |
return emb
|
293 |
|
294 |
-
# For multiple texts, no LRU cache is used
|
295 |
model = self.text_models[model_id]
|
296 |
return model.encode(texts)
|
297 |
-
|
298 |
except Exception as e:
|
299 |
raise RuntimeError(
|
300 |
f"Error generating text embeddings with model '{model_id}': {e}"
|
301 |
) from e
|
302 |
|
303 |
-
def
|
304 |
self,
|
305 |
model_id: ImageModelType,
|
306 |
images: List[str],
|
307 |
) -> np.ndarray:
|
308 |
"""
|
309 |
-
|
310 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
311 |
"""
|
312 |
try:
|
313 |
-
|
314 |
-
|
315 |
-
|
316 |
-
|
317 |
-
|
318 |
-
|
319 |
-
# Keys should be the same for all processed outputs
|
320 |
keys = processed_tensors[0].keys()
|
321 |
-
# Concatenate along the batch dimension
|
322 |
combined = {
|
323 |
k: torch.cat([pt[k] for pt in processed_tensors], dim=0) for k in keys
|
324 |
}
|
325 |
|
326 |
-
|
327 |
-
|
328 |
-
|
|
|
|
|
|
|
329 |
|
|
|
330 |
except Exception as e:
|
331 |
raise RuntimeError(
|
332 |
f"Error generating image embeddings with model '{model_id}': {e}"
|
@@ -338,19 +391,28 @@ class EmbeddingsService:
|
|
338 |
inputs: Union[str, List[str]],
|
339 |
) -> np.ndarray:
|
340 |
"""
|
341 |
-
Asynchronously
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
342 |
"""
|
343 |
modality = detect_model_kind(model)
|
344 |
-
|
345 |
if modality == ModelKind.TEXT:
|
346 |
text_model_id = TextModelType(model)
|
347 |
text_list = self._validate_text_list(inputs)
|
348 |
-
return
|
349 |
-
|
|
|
350 |
elif modality == ModelKind.IMAGE:
|
351 |
image_model_id = ImageModelType(model)
|
352 |
image_list = self._validate_image_list(inputs)
|
353 |
-
return self.
|
|
|
|
|
354 |
|
355 |
async def rank(
|
356 |
self,
|
@@ -359,35 +421,32 @@ class EmbeddingsService:
|
|
359 |
candidates: Union[str, List[str]],
|
360 |
) -> Dict[str, Any]:
|
361 |
"""
|
362 |
-
|
363 |
-
|
364 |
|
365 |
-
|
366 |
-
|
367 |
"""
|
368 |
modality = detect_model_kind(model)
|
369 |
-
|
370 |
-
# Convert the string model to the appropriate enum
|
371 |
if modality == ModelKind.TEXT:
|
372 |
model_enum = TextModelType(model)
|
373 |
else:
|
374 |
model_enum = ImageModelType(model)
|
375 |
|
376 |
-
#
|
377 |
-
|
378 |
-
|
379 |
-
|
380 |
-
|
381 |
-
candidate_embeds = await
|
|
|
|
|
382 |
|
383 |
-
#
|
384 |
sim_matrix = self.cosine_similarity(query_embeds, candidate_embeds)
|
385 |
-
|
386 |
-
# 4) Apply logit scale + softmax to obtain probabilities
|
387 |
scaled = np.exp(self.config.logit_scale) * sim_matrix
|
388 |
probs = self.softmax(scaled)
|
389 |
|
390 |
-
# 5) Estimate token usage if we're dealing with text
|
391 |
if modality == ModelKind.TEXT:
|
392 |
query_tokens = self.estimate_tokens(queries)
|
393 |
candidate_tokens = self.estimate_tokens(candidates)
|
@@ -408,32 +467,41 @@ class EmbeddingsService:
|
|
408 |
|
409 |
def estimate_tokens(self, input_data: Union[str, List[str]]) -> int:
|
410 |
"""
|
411 |
-
|
412 |
-
|
|
|
|
|
413 |
"""
|
414 |
texts = self._validate_text_list(input_data)
|
415 |
model = self.text_models[self.config.text_model_type]
|
416 |
tokenized = model.tokenize(texts)
|
417 |
-
# Summing over the lengths of input_ids for each example
|
418 |
return sum(len(ids) for ids in tokenized["input_ids"])
|
419 |
|
420 |
@staticmethod
|
421 |
def softmax(scores: np.ndarray) -> np.ndarray:
|
422 |
"""
|
423 |
-
|
|
|
|
|
|
|
424 |
"""
|
425 |
-
# Stabilize scores by subtracting max
|
426 |
exps = np.exp(scores - np.max(scores, axis=-1, keepdims=True))
|
427 |
return exps / np.sum(exps, axis=-1, keepdims=True)
|
428 |
|
429 |
@staticmethod
|
430 |
def cosine_similarity(a: np.ndarray, b: np.ndarray) -> np.ndarray:
|
431 |
"""
|
432 |
-
|
433 |
-
|
434 |
-
|
435 |
-
|
436 |
"""
|
437 |
a_norm = a / (np.linalg.norm(a, axis=1, keepdims=True) + 1e-9)
|
438 |
b_norm = b / (np.linalg.norm(b, axis=1, keepdims=True) + 1e-9)
|
439 |
return np.dot(a_norm, b_norm.T)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
from __future__ import annotations
|
2 |
|
3 |
+
import asyncio
|
4 |
import logging
|
5 |
from enum import Enum
|
6 |
from typing import List, Union, Dict, Optional, NamedTuple, Any
|
|
|
10 |
from hashlib import md5
|
11 |
from cachetools import LRUCache
|
12 |
|
13 |
+
import httpx
|
14 |
import numpy as np
|
15 |
import torch
|
16 |
from PIL import Image
|
|
|
46 |
|
47 |
class ModelInfo(NamedTuple):
|
48 |
"""
|
49 |
+
Container mapping a model type to its model identifier and optional ONNX file.
|
|
|
|
|
50 |
"""
|
51 |
|
52 |
model_id: str
|
|
|
68 |
@property
|
69 |
def text_model_info(self) -> ModelInfo:
|
70 |
"""
|
71 |
+
Return model information for the configured text model.
|
72 |
"""
|
73 |
text_configs = {
|
74 |
TextModelType.MULTILINGUAL_E5_SMALL: ModelInfo(
|
|
|
109 |
@property
|
110 |
def image_model_info(self) -> ModelInfo:
|
111 |
"""
|
112 |
+
Return model information for the configured image model.
|
113 |
"""
|
114 |
image_configs = {
|
115 |
ImageModelType.SIGLIP_BASE_PATCH16_256_MULTILINGUAL: ModelInfo(
|
|
|
120 |
|
121 |
|
122 |
class ModelKind(str, Enum):
|
123 |
+
"""
|
124 |
+
Indicates the type of model: text or image.
|
125 |
+
"""
|
126 |
+
|
127 |
TEXT = "text"
|
128 |
IMAGE = "image"
|
129 |
|
130 |
|
131 |
def detect_model_kind(model_id: str) -> ModelKind:
|
132 |
"""
|
133 |
+
Detect whether the model identifier corresponds to a text or image model.
|
134 |
+
|
135 |
+
Raises:
|
136 |
+
ValueError: If the model identifier is unrecognized.
|
137 |
"""
|
138 |
if model_id in [m.value for m in TextModelType]:
|
139 |
return ModelKind.TEXT
|
|
|
150 |
class EmbeddingsService:
|
151 |
"""
|
152 |
Service for generating text/image embeddings and performing similarity ranking.
|
153 |
+
Asynchronous methods are used to maximize throughput and avoid blocking the event loop.
|
154 |
"""
|
155 |
|
156 |
def __init__(self, config: Optional[ModelConfig] = None):
|
157 |
+
"""
|
158 |
+
Initialize the service by setting up model caches, device configuration,
|
159 |
+
and asynchronous HTTP client.
|
160 |
+
"""
|
161 |
self.lru_cache = LRUCache(maxsize=10_000)
|
162 |
self.device = "cuda" if torch.cuda.is_available() else "cpu"
|
163 |
self.config = config or ModelConfig()
|
164 |
|
165 |
+
# Dictionaries to hold preloaded models.
|
166 |
self.text_models: Dict[TextModelType, SentenceTransformer] = {}
|
167 |
self.image_models: Dict[ImageModelType, AutoModel] = {}
|
168 |
self.image_processors: Dict[ImageModelType, AutoProcessor] = {}
|
169 |
|
170 |
+
# Create a persistent asynchronous HTTP client.
|
171 |
+
self.async_http_client = httpx.AsyncClient(timeout=10)
|
172 |
+
|
173 |
+
# Preload all models.
|
174 |
self._load_all_models()
|
175 |
|
176 |
def _load_all_models(self) -> None:
|
177 |
"""
|
178 |
+
Pre-load all text and image models to minimize latency at request time.
|
179 |
"""
|
180 |
try:
|
181 |
+
# Preload text models.
|
182 |
for t_model_type in TextModelType:
|
183 |
info = ModelConfig(text_model_type=t_model_type).text_model_info
|
184 |
logger.info("Loading text model: %s", info.model_id)
|
|
|
185 |
if info.onnx_file:
|
186 |
logger.info("Using ONNX file: %s", info.onnx_file)
|
187 |
self.text_models[t_model_type] = SentenceTransformer(
|
|
|
201 |
trust_remote_code=True,
|
202 |
)
|
203 |
|
204 |
+
# Preload image models.
|
205 |
for i_model_type in ImageModelType:
|
206 |
model_id = ModelConfig(
|
207 |
image_model_type=i_model_type
|
208 |
).image_model_info.model_id
|
209 |
logger.info("Loading image model: %s", model_id)
|
|
|
210 |
model = AutoModel.from_pretrained(model_id).to(self.device)
|
211 |
+
model.eval() # Set the model to evaluation mode.
|
212 |
processor = AutoProcessor.from_pretrained(model_id)
|
|
|
213 |
self.image_models[i_model_type] = model
|
214 |
self.image_processors[i_model_type] = processor
|
215 |
|
|
|
222 |
@staticmethod
|
223 |
def _validate_text_list(input_text: Union[str, List[str]]) -> List[str]:
|
224 |
"""
|
225 |
+
Validate and convert text input into a non-empty list of strings.
|
226 |
+
|
227 |
+
Raises:
|
228 |
+
ValueError: If the input is invalid.
|
229 |
"""
|
230 |
if isinstance(input_text, str):
|
231 |
if not input_text.strip():
|
|
|
245 |
@staticmethod
|
246 |
def _validate_image_list(input_images: Union[str, List[str]]) -> List[str]:
|
247 |
"""
|
248 |
+
Validate and convert image input into a non-empty list of image paths/URLs.
|
249 |
+
|
250 |
+
Raises:
|
251 |
+
ValueError: If the input is invalid.
|
252 |
"""
|
253 |
if isinstance(input_images, str):
|
254 |
if not input_images.strip():
|
|
|
265 |
|
266 |
return input_images
|
267 |
|
268 |
+
async def _fetch_image(self, path_or_url: str) -> Image.Image:
|
269 |
"""
|
270 |
+
Asynchronously fetch an image from a URL or load from a local path.
|
271 |
+
|
272 |
+
Args:
|
273 |
+
path_or_url: The URL or file path of the image.
|
274 |
+
|
275 |
+
Returns:
|
276 |
+
A PIL Image in RGB mode.
|
277 |
+
|
278 |
+
Raises:
|
279 |
+
ValueError: If image fetching or processing fails.
|
280 |
"""
|
281 |
try:
|
282 |
if path_or_url.startswith("http"):
|
283 |
+
# Asynchronously fetch the image bytes.
|
284 |
+
response = await self.async_http_client.get(path_or_url)
|
285 |
+
response.raise_for_status()
|
286 |
+
# Offload the blocking I/O (PIL image opening) to a thread.
|
287 |
+
img = await asyncio.to_thread(Image.open, BytesIO(response.content))
|
288 |
else:
|
289 |
+
# Offload file I/O to a thread.
|
290 |
+
img = await asyncio.to_thread(Image.open, Path(path_or_url))
|
291 |
+
return img.convert("RGB")
|
|
|
|
|
292 |
except Exception as e:
|
293 |
+
raise ValueError(f"Error fetching image '{path_or_url}': {str(e)}") from e
|
294 |
+
|
295 |
+
async def _process_image(self, path_or_url: str) -> Dict[str, torch.Tensor]:
|
296 |
+
"""
|
297 |
+
Asynchronously load and process a single image.
|
298 |
+
|
299 |
+
Args:
|
300 |
+
path_or_url: The image URL or local path.
|
301 |
+
|
302 |
+
Returns:
|
303 |
+
A dictionary of processed tensors ready for model input.
|
304 |
+
|
305 |
+
Raises:
|
306 |
+
ValueError: If image processing fails.
|
307 |
+
"""
|
308 |
+
img = await self._fetch_image(path_or_url)
|
309 |
+
processor = self.image_processors[self.config.image_model_type]
|
310 |
+
# Note: Processor may perform CPU-intensive work; if needed, offload to thread.
|
311 |
+
processed_data = processor(images=img, return_tensors="pt").to(self.device)
|
312 |
+
return processed_data
|
313 |
|
314 |
def _generate_text_embeddings(
|
315 |
self,
|
|
|
317 |
texts: List[str],
|
318 |
) -> np.ndarray:
|
319 |
"""
|
320 |
+
Generate text embeddings using the SentenceTransformer model.
|
321 |
+
Single-text requests are cached using an LRU cache.
|
322 |
+
|
323 |
+
Returns:
|
324 |
+
A NumPy array of text embeddings.
|
325 |
+
|
326 |
+
Raises:
|
327 |
+
RuntimeError: If text embedding generation fails.
|
328 |
"""
|
329 |
try:
|
330 |
if len(texts) == 1:
|
|
|
332 |
key = md5(f"{model_id}:{single_text}".encode("utf-8")).hexdigest()[:8]
|
333 |
if key in self.lru_cache:
|
334 |
return self.lru_cache[key]
|
|
|
335 |
model = self.text_models[model_id]
|
336 |
emb = model.encode([single_text])
|
337 |
self.lru_cache[key] = emb
|
338 |
return emb
|
339 |
|
|
|
340 |
model = self.text_models[model_id]
|
341 |
return model.encode(texts)
|
|
|
342 |
except Exception as e:
|
343 |
raise RuntimeError(
|
344 |
f"Error generating text embeddings with model '{model_id}': {e}"
|
345 |
) from e
|
346 |
|
347 |
+
async def _async_generate_image_embeddings(
|
348 |
self,
|
349 |
model_id: ImageModelType,
|
350 |
images: List[str],
|
351 |
) -> np.ndarray:
|
352 |
"""
|
353 |
+
Asynchronously generate image embeddings.
|
354 |
+
|
355 |
+
This method concurrently processes multiple images and offloads
|
356 |
+
the blocking model inference to a separate thread.
|
357 |
+
|
358 |
+
Returns:
|
359 |
+
A NumPy array of image embeddings.
|
360 |
+
|
361 |
+
Raises:
|
362 |
+
RuntimeError: If image embedding generation fails.
|
363 |
"""
|
364 |
try:
|
365 |
+
# Concurrently process all images.
|
366 |
+
processed_tensors = await asyncio.gather(
|
367 |
+
*[self._process_image(img_path) for img_path in images]
|
368 |
+
)
|
369 |
+
# Assume all processed outputs have the same keys.
|
|
|
|
|
370 |
keys = processed_tensors[0].keys()
|
|
|
371 |
combined = {
|
372 |
k: torch.cat([pt[k] for pt in processed_tensors], dim=0) for k in keys
|
373 |
}
|
374 |
|
375 |
+
def infer():
|
376 |
+
with torch.no_grad():
|
377 |
+
embeddings = self.image_models[model_id].get_image_features(
|
378 |
+
**combined
|
379 |
+
)
|
380 |
+
return embeddings.cpu().numpy()
|
381 |
|
382 |
+
return await asyncio.to_thread(infer)
|
383 |
except Exception as e:
|
384 |
raise RuntimeError(
|
385 |
f"Error generating image embeddings with model '{model_id}': {e}"
|
|
|
391 |
inputs: Union[str, List[str]],
|
392 |
) -> np.ndarray:
|
393 |
"""
|
394 |
+
Asynchronously generate embeddings for text or image inputs based on model type.
|
395 |
+
|
396 |
+
Args:
|
397 |
+
model: The model identifier.
|
398 |
+
inputs: The text or image input(s).
|
399 |
+
|
400 |
+
Returns:
|
401 |
+
A NumPy array of embeddings.
|
402 |
"""
|
403 |
modality = detect_model_kind(model)
|
|
|
404 |
if modality == ModelKind.TEXT:
|
405 |
text_model_id = TextModelType(model)
|
406 |
text_list = self._validate_text_list(inputs)
|
407 |
+
return await asyncio.to_thread(
|
408 |
+
self._generate_text_embeddings, text_model_id, text_list
|
409 |
+
)
|
410 |
elif modality == ModelKind.IMAGE:
|
411 |
image_model_id = ImageModelType(model)
|
412 |
image_list = self._validate_image_list(inputs)
|
413 |
+
return await self._async_generate_image_embeddings(
|
414 |
+
image_model_id, image_list
|
415 |
+
)
|
416 |
|
417 |
async def rank(
|
418 |
self,
|
|
|
421 |
candidates: Union[str, List[str]],
|
422 |
) -> Dict[str, Any]:
|
423 |
"""
|
424 |
+
Asynchronously rank candidate texts/images against the provided queries.
|
425 |
+
Embeddings for queries and candidates are generated concurrently.
|
426 |
|
427 |
+
Returns:
|
428 |
+
A dictionary containing probabilities, cosine similarities, and usage statistics.
|
429 |
"""
|
430 |
modality = detect_model_kind(model)
|
|
|
|
|
431 |
if modality == ModelKind.TEXT:
|
432 |
model_enum = TextModelType(model)
|
433 |
else:
|
434 |
model_enum = ImageModelType(model)
|
435 |
|
436 |
+
# Concurrently generate embeddings.
|
437 |
+
query_task = asyncio.create_task(self.generate_embeddings(model, queries))
|
438 |
+
candidate_task = asyncio.create_task(
|
439 |
+
self.generate_embeddings(model, candidates)
|
440 |
+
)
|
441 |
+
query_embeds, candidate_embeds = await asyncio.gather(
|
442 |
+
query_task, candidate_task
|
443 |
+
)
|
444 |
|
445 |
+
# Compute cosine similarity.
|
446 |
sim_matrix = self.cosine_similarity(query_embeds, candidate_embeds)
|
|
|
|
|
447 |
scaled = np.exp(self.config.logit_scale) * sim_matrix
|
448 |
probs = self.softmax(scaled)
|
449 |
|
|
|
450 |
if modality == ModelKind.TEXT:
|
451 |
query_tokens = self.estimate_tokens(queries)
|
452 |
candidate_tokens = self.estimate_tokens(candidates)
|
|
|
467 |
|
468 |
def estimate_tokens(self, input_data: Union[str, List[str]]) -> int:
|
469 |
"""
|
470 |
+
Estimate the token count for the given text input using the SentenceTransformer tokenizer.
|
471 |
+
|
472 |
+
Returns:
|
473 |
+
The total number of tokens.
|
474 |
"""
|
475 |
texts = self._validate_text_list(input_data)
|
476 |
model = self.text_models[self.config.text_model_type]
|
477 |
tokenized = model.tokenize(texts)
|
|
|
478 |
return sum(len(ids) for ids in tokenized["input_ids"])
|
479 |
|
480 |
@staticmethod
|
481 |
def softmax(scores: np.ndarray) -> np.ndarray:
|
482 |
"""
|
483 |
+
Compute the softmax over the last dimension of the input array.
|
484 |
+
|
485 |
+
Returns:
|
486 |
+
The softmax probabilities.
|
487 |
"""
|
|
|
488 |
exps = np.exp(scores - np.max(scores, axis=-1, keepdims=True))
|
489 |
return exps / np.sum(exps, axis=-1, keepdims=True)
|
490 |
|
491 |
@staticmethod
|
492 |
def cosine_similarity(a: np.ndarray, b: np.ndarray) -> np.ndarray:
|
493 |
"""
|
494 |
+
Compute the pairwise cosine similarity between all rows of arrays a and b.
|
495 |
+
|
496 |
+
Returns:
|
497 |
+
A (N x M) matrix of cosine similarities.
|
498 |
"""
|
499 |
a_norm = a / (np.linalg.norm(a, axis=1, keepdims=True) + 1e-9)
|
500 |
b_norm = b / (np.linalg.norm(b, axis=1, keepdims=True) + 1e-9)
|
501 |
return np.dot(a_norm, b_norm.T)
|
502 |
+
|
503 |
+
async def close(self) -> None:
|
504 |
+
"""
|
505 |
+
Close the asynchronous HTTP client.
|
506 |
+
"""
|
507 |
+
await self.async_http_client.aclose()
|
requirements.txt
CHANGED
@@ -5,6 +5,7 @@ requests
|
|
5 |
pydantic
|
6 |
cachetools
|
7 |
pandas
|
|
|
8 |
sentence-transformers[onnx]==3.3.1
|
9 |
sentencepiece==0.2.0
|
10 |
torch==2.4.0
|
|
|
5 |
pydantic
|
6 |
cachetools
|
7 |
pandas
|
8 |
+
httpx
|
9 |
sentence-transformers[onnx]==3.3.1
|
10 |
sentencepiece==0.2.0
|
11 |
torch==2.4.0
|