Update handler.py
Browse files- handler.py +90 -2
handler.py
CHANGED
@@ -22,10 +22,22 @@ class EndpointHandler:
|
|
22 |
self.classifier.bias.data = torch.tensor(head["scorer_bias"]).to(self.device)
|
23 |
|
24 |
self.model.eval()
|
|
|
|
|
|
|
|
|
25 |
|
26 |
def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
|
27 |
payload = data.get("inputs", data)
|
28 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
29 |
query = payload["query"]
|
30 |
candidates = payload["candidates"]
|
31 |
results = []
|
@@ -38,7 +50,7 @@ class EndpointHandler:
|
|
38 |
return_tensors="pt",
|
39 |
padding="max_length",
|
40 |
truncation=True,
|
41 |
-
max_length=
|
42 |
).to(self.device)
|
43 |
|
44 |
out = self.model(**tokens)
|
@@ -51,3 +63,79 @@ class EndpointHandler:
|
|
51 |
})
|
52 |
|
53 |
return sorted(results, key=lambda x: x["score"], reverse=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
22 |
self.classifier.bias.data = torch.tensor(head["scorer_bias"]).to(self.device)
|
23 |
|
24 |
self.model.eval()
|
25 |
+
|
26 |
+
# Batch processing configuration
|
27 |
+
self.max_batch_size = 128 # Adjust based on GPU memory
|
28 |
+
self.max_length = 64
|
29 |
|
30 |
def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
|
31 |
payload = data.get("inputs", data)
|
32 |
+
|
33 |
+
# Check if this is batch processing (multiple queries) or single query
|
34 |
+
if "queries" in payload:
|
35 |
+
return self._process_batch(payload)
|
36 |
+
else:
|
37 |
+
return self._process_single(payload)
|
38 |
+
|
39 |
+
def _process_single(self, payload: Dict[str, Any]) -> List[Dict[str, Any]]:
|
40 |
+
"""Original single query processing for backward compatibility"""
|
41 |
query = payload["query"]
|
42 |
candidates = payload["candidates"]
|
43 |
results = []
|
|
|
50 |
return_tensors="pt",
|
51 |
padding="max_length",
|
52 |
truncation=True,
|
53 |
+
max_length=self.max_length
|
54 |
).to(self.device)
|
55 |
|
56 |
out = self.model(**tokens)
|
|
|
63 |
})
|
64 |
|
65 |
return sorted(results, key=lambda x: x["score"], reverse=True)
|
66 |
+
|
67 |
+
def _process_batch(self, payload: Dict[str, Any]) -> List[List[Dict[str, Any]]]:
|
68 |
+
"""True batch processing for multiple queries"""
|
69 |
+
queries = payload["queries"]
|
70 |
+
candidates = payload["candidates"]
|
71 |
+
|
72 |
+
# Create all query-candidate combinations
|
73 |
+
all_texts = []
|
74 |
+
query_indices = []
|
75 |
+
candidate_indices = []
|
76 |
+
|
77 |
+
for q_idx, query in enumerate(queries):
|
78 |
+
for c_idx, candidate in enumerate(candidates):
|
79 |
+
text = f"[QUERY] {query} [LABEL_NAME] {candidate['label']} [LABEL_DESCRIPTION] {candidate['description']}"
|
80 |
+
all_texts.append(text)
|
81 |
+
query_indices.append(q_idx)
|
82 |
+
candidate_indices.append(c_idx)
|
83 |
+
|
84 |
+
# Process in batches to avoid memory issues
|
85 |
+
all_scores = []
|
86 |
+
total_combinations = len(all_texts)
|
87 |
+
|
88 |
+
with torch.no_grad():
|
89 |
+
for i in range(0, total_combinations, self.max_batch_size):
|
90 |
+
batch_texts = all_texts[i:i + self.max_batch_size]
|
91 |
+
|
92 |
+
# Tokenize batch
|
93 |
+
tokens = self.tokenizer(
|
94 |
+
batch_texts,
|
95 |
+
return_tensors="pt",
|
96 |
+
padding="max_length",
|
97 |
+
truncation=True,
|
98 |
+
max_length=self.max_length
|
99 |
+
).to(self.device)
|
100 |
+
|
101 |
+
# Single forward pass for entire batch
|
102 |
+
out = self.model(**tokens)
|
103 |
+
cls = out.last_hidden_state[:, 0, :]
|
104 |
+
scores = torch.sigmoid(self.classifier(cls)).squeeze()
|
105 |
+
|
106 |
+
# Handle single item case
|
107 |
+
if scores.dim() == 0:
|
108 |
+
scores = scores.unsqueeze(0)
|
109 |
+
|
110 |
+
all_scores.extend(scores.cpu().tolist())
|
111 |
+
|
112 |
+
# Reshape results back to query structure
|
113 |
+
results = []
|
114 |
+
for q_idx in range(len(queries)):
|
115 |
+
query_results = []
|
116 |
+
for c_idx, candidate in enumerate(candidates):
|
117 |
+
# Find the score for this query-candidate combination
|
118 |
+
combination_idx = q_idx * len(candidates) + c_idx
|
119 |
+
score = all_scores[combination_idx]
|
120 |
+
|
121 |
+
query_results.append({
|
122 |
+
"label": candidate["label"],
|
123 |
+
"description": candidate["description"],
|
124 |
+
"score": round(score, 4)
|
125 |
+
})
|
126 |
+
|
127 |
+
# Sort by score for this query
|
128 |
+
query_results.sort(key=lambda x: x["score"], reverse=True)
|
129 |
+
results.append(query_results)
|
130 |
+
|
131 |
+
return results
|
132 |
+
|
133 |
+
def get_batch_stats(self) -> Dict[str, Any]:
|
134 |
+
"""Return batch processing statistics"""
|
135 |
+
return {
|
136 |
+
"max_batch_size": self.max_batch_size,
|
137 |
+
"max_length": self.max_length,
|
138 |
+
"device": str(self.device),
|
139 |
+
"model_name": self.model.config.name_or_path if hasattr(self.model.config, 'name_or_path') else "unknown"
|
140 |
+
}
|
141 |
+
|