701ae503
 tangwang
docs
|
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
|
"""
Qwen3-Reranker-0.6B backend using vLLM.
Reference: https://huggingface.co/Qwen/Qwen3-Reranker-0.6B
Requires: vllm>=0.8.5, transformers; GPU recommended.
"""
from __future__ import annotations
import logging
import math
import time
from typing import Any, Dict, List, Optional, Tuple
logger = logging.getLogger("reranker.backends.qwen3_vllm")
try:
import torch
from transformers import AutoTokenizer
from vllm import LLM, SamplingParams
from vllm.inputs import TokensPrompt
except ImportError as e:
raise ImportError(
"Qwen3-vLLM reranker backend requires vllm>=0.8.5 and transformers. "
"Install with: pip install vllm transformers"
) from e
def _format_instruction(instruction: str, query: str, doc: str) -> List[Dict[str, str]]:
"""Build chat messages for one (query, doc) pair."""
return [
{
"role": "system",
"content": "Judge whether the Document meets the requirements based on the Query and the Instruct provided. Note that the answer can only be \"yes\" or \"no\".",
},
{
"role": "user",
"content": f"<Instruct>: {instruction}\n\n<Query>: {query}\n\n<Document>: {doc}",
},
]
class Qwen3VLLMRerankerBackend:
"""
Qwen3-Reranker-0.6B with vLLM inference.
Config from services.rerank.backends.qwen3_vllm.
"""
def __init__(self, config: Dict[str, Any]) -> None:
self._config = config or {}
model_name = str(self._config.get("model_name") or "Qwen/Qwen3-Reranker-0.6B")
|
|
701ae503
tangwang
docs
|
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
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
213
214
215
216
217
218
219
220
221
222
223
|
)
self._tokenizer = AutoTokenizer.from_pretrained(model_name)
self._tokenizer.padding_side = "left"
self._tokenizer.pad_token = self._tokenizer.eos_token
# Suffix for generation prompt (assistant answer)
self._suffix = "<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n"
self._suffix_tokens = self._tokenizer.encode(
self._suffix, add_special_tokens=False
)
self._max_prompt_len = max_model_len - len(self._suffix_tokens)
self._true_token = self._tokenizer("yes", add_special_tokens=False).input_ids[0]
self._false_token = self._tokenizer("no", add_special_tokens=False).input_ids[0]
self._sampling_params = SamplingParams(
temperature=0,
max_tokens=1,
logprobs=20,
allowed_token_ids=[self._true_token, self._false_token],
)
self._model_name = model_name
logger.info("[Qwen3_VLLM] Model ready | model=%s", model_name)
def _process_inputs(
self,
pairs: List[Tuple[str, str]],
) -> List[TokensPrompt]:
"""Build tokenized prompts for vLLM from (query, doc) pairs."""
prompts = []
for q, d in pairs:
messages = _format_instruction(self._instruction, q, d)
# One conversation per call (apply_chat_template expects single conversation)
token_ids = self._tokenizer.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=False,
enable_thinking=False,
)
if isinstance(token_ids, list) and token_ids and isinstance(token_ids[0], list):
token_ids = token_ids[0]
ids = token_ids[: self._max_prompt_len] + self._suffix_tokens
prompts.append(TokensPrompt(prompt_token_ids=ids))
return prompts
def _compute_scores(
self,
prompts: List[TokensPrompt],
) -> List[float]:
"""Run vLLM generate and compute yes/no probability per prompt."""
if not prompts:
return []
outputs = self._llm.generate(prompts, self._sampling_params, use_tqdm=False)
scores = []
for i in range(len(outputs)):
out = outputs[i]
if not out.outputs:
scores.append(0.0)
continue
logprobs = out.outputs[0].logprobs
if not logprobs:
scores.append(0.0)
continue
last = logprobs[-1]
true_logp = last.get(self._true_token)
false_logp = last.get(self._false_token)
true_p = math.exp(true_logp.logprob) if true_logp else 1e-10
false_p = math.exp(false_logp.logprob) if false_logp else 1e-10
score = true_p / (true_p + false_p)
scores.append(float(score))
return scores
def score_with_meta(
self,
query: str,
docs: List[str],
normalize: bool = True,
) -> Tuple[List[float], Dict[str, Any]]:
start_ts = time.time()
total_docs = len(docs) if docs else 0
output_scores: List[float] = [0.0] * total_docs
query = "" if query is None else str(query).strip()
indexed: List[Tuple[int, str]] = []
for i, doc in enumerate(docs or []):
if doc is None:
continue
text = str(doc).strip()
if not text:
continue
indexed.append((i, text))
if not query or not indexed:
elapsed_ms = (time.time() - start_ts) * 1000.0
return output_scores, {
"input_docs": total_docs,
"usable_docs": len(indexed),
"unique_docs": 0,
"dedup_ratio": 0.0,
"elapsed_ms": round(elapsed_ms, 3),
"model": self._model_name,
"backend": "qwen3_vllm",
"normalize": normalize,
}
# Deduplicate by text, keep mapping to original indices
unique_texts: List[str] = []
position_to_unique: List[int] = []
prev: Optional[str] = None
for _idx, text in indexed:
if text != prev:
unique_texts.append(text)
prev = text
position_to_unique.append(len(unique_texts) - 1)
pairs = [(query, t) for t in unique_texts]
prompts = self._process_inputs(pairs)
unique_scores = self._compute_scores(prompts)
for (orig_idx, _), unique_idx in zip(indexed, position_to_unique):
# Score is already P(yes) in [0,1] from yes/(yes+no)
output_scores[orig_idx] = float(unique_scores[unique_idx])
elapsed_ms = (time.time() - start_ts) * 1000.0
dedup_ratio = 0.0
if indexed:
dedup_ratio = 1.0 - (len(unique_texts) / float(len(indexed)))
meta = {
"input_docs": total_docs,
"usable_docs": len(indexed),
"unique_docs": len(unique_texts),
"dedup_ratio": round(dedup_ratio, 4),
"elapsed_ms": round(elapsed_ms, 3),
"model": self._model_name,
"backend": "qwen3_vllm",
"normalize": normalize,
}
return output_scores, meta
|