Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,133 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# app.py (FastAPI server to host the Jina Embedding model)
|
2 |
+
# Must be set before importing Hugging Face libs
|
3 |
+
import os
|
4 |
+
os.environ["HF_HOME"] = "/tmp/huggingface"
|
5 |
+
os.environ["HF_HUB_CACHE"] = "/tmp/huggingface/hub"
|
6 |
+
os.environ["TRANSFORMERS_CACHE"] = "/tmp/huggingface/transformers"
|
7 |
+
from fastapi import FastAPI
|
8 |
+
from pydantic import BaseModel
|
9 |
+
from typing import List, Optional
|
10 |
+
import torch
|
11 |
+
from transformers import AutoModel, AutoTokenizer
|
12 |
+
|
13 |
+
app = FastAPI()
|
14 |
+
|
15 |
+
# -----------------------------
|
16 |
+
# Load model once on startup
|
17 |
+
# -----------------------------
|
18 |
+
MODEL_NAME = "jinaai/jina-embeddings-v4"
|
19 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
20 |
+
|
21 |
+
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True)
|
22 |
+
model = AutoModel.from_pretrained(
|
23 |
+
MODEL_NAME, trust_remote_code=True, torch_dtype=torch.float16
|
24 |
+
).to(device)
|
25 |
+
model.eval()
|
26 |
+
|
27 |
+
|
28 |
+
# -----------------------------
|
29 |
+
# Request / Response Models
|
30 |
+
# -----------------------------
|
31 |
+
class EmbedRequest(BaseModel):
|
32 |
+
text: str
|
33 |
+
task: str = "retrieval" # "retrieval", "text-matching", "code", etc.
|
34 |
+
prompt_name: Optional[str] = None
|
35 |
+
return_token_embeddings: bool = True # False → for queries (pooled embedding)
|
36 |
+
|
37 |
+
|
38 |
+
class EmbedResponse(BaseModel):
|
39 |
+
embeddings: List[List[float]] # (num_tokens, hidden_dim) if token-level
|
40 |
+
# (1, hidden_dim) if pooled query
|
41 |
+
|
42 |
+
|
43 |
+
class TokenizeRequest(BaseModel):
|
44 |
+
text: str
|
45 |
+
|
46 |
+
|
47 |
+
class TokenizeResponse(BaseModel):
|
48 |
+
input_ids: List[int]
|
49 |
+
|
50 |
+
|
51 |
+
class DecodeRequest(BaseModel):
|
52 |
+
input_ids: List[int]
|
53 |
+
|
54 |
+
|
55 |
+
class DecodeResponse(BaseModel):
|
56 |
+
text: str
|
57 |
+
|
58 |
+
|
59 |
+
# -----------------------------
|
60 |
+
# Embedding Endpoint
|
61 |
+
# -----------------------------
|
62 |
+
@app.post("/embed", response_model=EmbedResponse)
|
63 |
+
def embed(req: EmbedRequest):
|
64 |
+
text = req.text
|
65 |
+
|
66 |
+
# -----------------------------
|
67 |
+
# Case 1: Query → directly pooled embedding
|
68 |
+
# -----------------------------
|
69 |
+
if not req.return_token_embeddings:
|
70 |
+
with torch.no_grad():
|
71 |
+
emb = model.encode_text(
|
72 |
+
texts=[text],
|
73 |
+
task=req.task,
|
74 |
+
prompt_name=req.prompt_name or "query",
|
75 |
+
return_multivector=False
|
76 |
+
)
|
77 |
+
return {"embeddings": emb.tolist()} # shape: (1, hidden_dim)
|
78 |
+
|
79 |
+
# -----------------------------
|
80 |
+
# Case 2: Long passages → sliding window token embeddings
|
81 |
+
# -----------------------------
|
82 |
+
enc = tokenizer(text, add_special_tokens=False, return_tensors="pt")
|
83 |
+
input_ids = enc["input_ids"].squeeze(0).to(device) # (total_tokens,)
|
84 |
+
total_tokens = input_ids.size(0)
|
85 |
+
|
86 |
+
max_len = model.config.max_position_embeddings # e.g., 32k for v4
|
87 |
+
stride = 50 # overlap for sliding window
|
88 |
+
embeddings = []
|
89 |
+
position = 0
|
90 |
+
|
91 |
+
while position < total_tokens:
|
92 |
+
end = min(position + max_len, total_tokens)
|
93 |
+
window_ids = input_ids[position:end].unsqueeze(0).to(device)
|
94 |
+
|
95 |
+
with torch.no_grad():
|
96 |
+
outputs = model.encode_text(
|
97 |
+
texts=[tokenizer.decode(window_ids[0])],
|
98 |
+
task=req.task,
|
99 |
+
prompt_name=req.prompt_name or "passage",
|
100 |
+
return_multivector=True,
|
101 |
+
)
|
102 |
+
|
103 |
+
window_embeds = outputs.squeeze(0).cpu() # (window_len, hidden_dim)
|
104 |
+
|
105 |
+
# Drop overlapping tokens except in first window
|
106 |
+
if position > 0:
|
107 |
+
window_embeds = window_embeds[stride:]
|
108 |
+
|
109 |
+
embeddings.append(window_embeds)
|
110 |
+
|
111 |
+
# Advance window
|
112 |
+
position += max_len - stride
|
113 |
+
|
114 |
+
full_embeddings = torch.cat(embeddings, dim=0) # (total_tokens, hidden_dim)
|
115 |
+
return {"embeddings": full_embeddings.tolist()}
|
116 |
+
|
117 |
+
|
118 |
+
# -----------------------------
|
119 |
+
# Tokenize Endpoint
|
120 |
+
# -----------------------------
|
121 |
+
@app.post("/tokenize", response_model=TokenizeResponse)
|
122 |
+
def tokenize(req: TokenizeRequest):
|
123 |
+
enc = tokenizer(req.text, add_special_tokens=False)
|
124 |
+
return {"input_ids": enc["input_ids"]}
|
125 |
+
|
126 |
+
|
127 |
+
# -----------------------------
|
128 |
+
# Decode Endpoint
|
129 |
+
# -----------------------------
|
130 |
+
@app.post("/decode", response_model=DecodeResponse)
|
131 |
+
def decode(req: DecodeRequest):
|
132 |
+
decoded = tokenizer.decode(req.input_ids)
|
133 |
+
return {"text": decoded}
|