added custom hadnler
Browse files- export_to_onnx.py +6 -3
- handler.py +94 -0
- model.onnx +2 -2
- modeling.py +31 -17
- requirements.txt +4 -0
export_to_onnx.py
CHANGED
@@ -16,9 +16,12 @@ try:
|
|
16 |
model.eval()
|
17 |
print("✓ Model set to evaluation mode")
|
18 |
|
19 |
-
|
20 |
-
|
21 |
-
|
|
|
|
|
|
|
22 |
print("✓ Dummy input prepared")
|
23 |
|
24 |
print("Exporting model to ONNX format...")
|
|
|
16 |
model.eval()
|
17 |
print("✓ Model set to evaluation mode")
|
18 |
|
19 |
+
# Use the doc_maxlen from the *loaded model's* colbert_config
|
20 |
+
actual_doc_maxlen = model.colbert_config.doc_maxlen
|
21 |
+
print(f"DEBUG: model.colbert_config.doc_maxlen = {actual_doc_maxlen}")
|
22 |
+
print(f"Preparing dummy input for ONNX export with doc_maxlen={actual_doc_maxlen}...")
|
23 |
+
dummy_input_ids = torch.ones((1, actual_doc_maxlen), dtype=torch.long)
|
24 |
+
dummy_attention_mask = torch.ones((1, actual_doc_maxlen), dtype=torch.long)
|
25 |
print("✓ Dummy input prepared")
|
26 |
|
27 |
print("Exporting model to ONNX format...")
|
handler.py
ADDED
@@ -0,0 +1,94 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# handler.py
|
2 |
+
import os
|
3 |
+
import onnxruntime as ort
|
4 |
+
import numpy as np
|
5 |
+
from transformers import AutoTokenizer
|
6 |
+
from typing import Dict, List, Any
|
7 |
+
from colbert_configuration import ColBERTConfig # Import ColBERTConfig
|
8 |
+
|
9 |
+
# Assuming modeling.py and colbert_configuration.py are in the same directory
|
10 |
+
# We'll use local imports since this handler will run within the model's directory context
|
11 |
+
# For ConstBERT to be recognized, you need to ensure these are importable.
|
12 |
+
# If you run into issues, consider a custom Docker image or ensuring the model
|
13 |
+
# is loadable via AutoModel.from_pretrained if it has auto_map in config.json
|
14 |
+
# For simplicity, we're relying on ConstBERT.from_pretrained working with ONNXRuntime path.
|
15 |
+
|
16 |
+
# Note: The EndpointHandler class must be named exactly this.
|
17 |
+
class EndpointHandler:
|
18 |
+
def __init__(self, path=""): # path will be '/repository' on HF Endpoints
|
19 |
+
# `path` is the directory where your model files (model.onnx, tokenizer files) are located.
|
20 |
+
|
21 |
+
# Load the tokenizer
|
22 |
+
self.tokenizer = AutoTokenizer.from_pretrained(path)
|
23 |
+
print(f"Tokenizer loaded from: {path}")
|
24 |
+
|
25 |
+
# Load ColBERTConfig to get doc_maxlen for consistent padding
|
26 |
+
# IMPORTANT: Use load_from_checkpoint to get the *exact* config used for model export.
|
27 |
+
self.colbert_config = ColBERTConfig.load_from_checkpoint(path)
|
28 |
+
self.doc_max_length = self.colbert_config.doc_maxlen
|
29 |
+
print(f"ColBERTConfig doc_maxlen loaded as: {self.doc_max_length}")
|
30 |
+
|
31 |
+
# Load the ONNX model
|
32 |
+
onnx_model_path = os.path.join(path, "model.onnx")
|
33 |
+
self.session = ort.InferenceSession(onnx_model_path)
|
34 |
+
print(f"ONNX model loaded from: {onnx_model_path}")
|
35 |
+
|
36 |
+
# Get input names from the ONNX model
|
37 |
+
self.input_names = [input.name for input in self.session.get_inputs()]
|
38 |
+
print(f"ONNX input names: {self.input_names}")
|
39 |
+
|
40 |
+
|
41 |
+
def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
|
42 |
+
"""
|
43 |
+
Inference call for the endpoint.
|
44 |
+
|
45 |
+
Args:
|
46 |
+
data (Dict[str, Any]): The request payload.
|
47 |
+
Expected to contain "inputs" (str or list of str).
|
48 |
+
|
49 |
+
Returns:
|
50 |
+
List[Dict[str, Any]]: A list of dictionaries, where each dict
|
51 |
+
contains the raw multi-vector output for an input.
|
52 |
+
Example: [{"embedding": [[...], [...], ...]}, ...]
|
53 |
+
"""
|
54 |
+
inputs = data.pop("inputs", None)
|
55 |
+
if inputs is None:
|
56 |
+
raise ValueError("No 'inputs' found in the request payload.")
|
57 |
+
|
58 |
+
# Ensure inputs is a list
|
59 |
+
if isinstance(inputs, str):
|
60 |
+
inputs = [inputs]
|
61 |
+
|
62 |
+
# Tokenize the inputs, ensuring consistent padding/truncation to doc_max_length
|
63 |
+
tokenized_inputs = self.tokenizer(
|
64 |
+
inputs,
|
65 |
+
padding="max_length", # Use max_length padding
|
66 |
+
truncation=True,
|
67 |
+
max_length=self.doc_max_length, # Use the loaded doc_max_length
|
68 |
+
return_tensors="np"
|
69 |
+
)
|
70 |
+
|
71 |
+
input_ids = tokenized_inputs["input_ids"]
|
72 |
+
attention_mask = tokenized_inputs["attention_mask"]
|
73 |
+
|
74 |
+
# Prepare ONNX input dictionary
|
75 |
+
onnx_inputs = {
|
76 |
+
"input_ids": input_ids,
|
77 |
+
"attention_mask": attention_mask
|
78 |
+
}
|
79 |
+
|
80 |
+
# Run ONNX inference
|
81 |
+
outputs = self.session.run(None, onnx_inputs)
|
82 |
+
|
83 |
+
# The first output is your multi-vector embedding
|
84 |
+
multi_vector_embeddings = outputs[0]
|
85 |
+
|
86 |
+
# Convert to list of lists (JSON serializable)
|
87 |
+
# Assuming batch_size will be 1 for typical endpoint requests, but handling potential batching from client for robustness.
|
88 |
+
result_list = []
|
89 |
+
for i in range(multi_vector_embeddings.shape[0]):
|
90 |
+
# Each element in the result_list will be a dictionary for one input,
|
91 |
+
# containing its multi-vector embedding (fixed 32 x 128)
|
92 |
+
result_list.append({"embedding": multi_vector_embeddings[i].tolist()})
|
93 |
+
|
94 |
+
return result_list
|
model.onnx
CHANGED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
-
size
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:d515b85a59a302d13d04b3a45c6211b3e1893a2718c13598231acc18825f0f02
|
3 |
+
size 436300888
|
modeling.py
CHANGED
@@ -60,6 +60,7 @@ class ConstBERT(BertPreTrainedModel):
|
|
60 |
super().__init__(config)
|
61 |
|
62 |
self.config = config
|
|
|
63 |
self.dim = colbert_config.dim
|
64 |
self.linear = nn.Linear(config.hidden_size, colbert_config.dim, bias=False)
|
65 |
self.doc_project = nn.Linear(colbert_config.doc_maxlen, 32, bias=False)
|
@@ -132,33 +133,46 @@ class ConstBERT(BertPreTrainedModel):
|
|
132 |
def forward(self, input_ids, attention_mask):
|
133 |
"""
|
134 |
Forward method for ONNX export and PyTorch compatibility.
|
135 |
-
This
|
136 |
"""
|
137 |
-
return self.
|
138 |
|
139 |
def _doc(self, input_ids, attention_mask, keep_dims=True):
|
140 |
assert keep_dims in [True, False, 'return_mask']
|
141 |
|
142 |
input_ids, attention_mask = input_ids.to(self.device), attention_mask.to(self.device)
|
143 |
-
D = self.bert(input_ids, attention_mask=attention_mask)[0]
|
144 |
-
|
145 |
-
|
146 |
-
D =
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
|
151 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
152 |
D = torch.nn.functional.normalize(D, p=2, dim=2)
|
153 |
if self.use_gpu:
|
154 |
D = D.half()
|
155 |
|
156 |
-
|
157 |
-
D, mask = D.cpu(), mask.bool().cpu().squeeze(-1)
|
158 |
-
D = [d[mask[idx]] for idx, d in enumerate(D)]
|
159 |
-
|
160 |
-
elif keep_dims == 'return_mask':
|
161 |
-
return D, mask.bool()
|
162 |
|
163 |
return D
|
164 |
|
|
|
60 |
super().__init__(config)
|
61 |
|
62 |
self.config = config
|
63 |
+
self.colbert_config = colbert_config
|
64 |
self.dim = colbert_config.dim
|
65 |
self.linear = nn.Linear(config.hidden_size, colbert_config.dim, bias=False)
|
66 |
self.doc_project = nn.Linear(colbert_config.doc_maxlen, 32, bias=False)
|
|
|
133 |
def forward(self, input_ids, attention_mask):
|
134 |
"""
|
135 |
Forward method for ONNX export and PyTorch compatibility.
|
136 |
+
This will now call _doc to produce a fixed number of vectors.
|
137 |
"""
|
138 |
+
return self._doc(input_ids, attention_mask)
|
139 |
|
140 |
def _doc(self, input_ids, attention_mask, keep_dims=True):
|
141 |
assert keep_dims in [True, False, 'return_mask']
|
142 |
|
143 |
input_ids, attention_mask = input_ids.to(self.device), attention_mask.to(self.device)
|
144 |
+
D = self.bert(input_ids, attention_mask=attention_mask)[0] # Shape: (batch_size, seq_len, hidden_size)
|
145 |
+
|
146 |
+
# First, apply linear layer to project hidden_size to colbert_config.dim (128)
|
147 |
+
D = self.linear(D) # Shape: (batch_size, seq_len, dim)
|
148 |
+
|
149 |
+
# Now, permute to put seq_len in the feature dimension for doc_project
|
150 |
+
D = D.permute(0, 2, 1) # Shape: (batch_size, dim, seq_len)
|
151 |
+
|
152 |
+
# Apply doc_project to reduce seq_len (e.g., 250) to fixed length (32)
|
153 |
+
# The nn.Linear(in_features, out_features) operates on the last dimension.
|
154 |
+
# So it expects the last dimension to be seq_len (doc_maxlen).
|
155 |
+
# It will transform it to (batch_size, dim, 32)
|
156 |
+
D = self.doc_project(D) # Shape: (batch_size, dim, 32)
|
157 |
+
|
158 |
+
# Permute back to (batch_size, 32, dim)
|
159 |
+
D = D.permute(0, 2, 1) # Shape: (batch_size, 32, dim)
|
160 |
+
|
161 |
+
# Apply mask (assuming it's still needed in this part of the flow)
|
162 |
+
# The mask now needs to be applied correctly to the (batch_size, 32, dim) shape
|
163 |
+
# For now, let's simplify mask application or ensure it's handled correctly if it remains a static shape.
|
164 |
+
# Given the fixed output, the original masking might be less critical here, or needs to be re-evaluated.
|
165 |
+
|
166 |
+
# Temporarily removing original mask logic in _doc to avoid immediate conflict.
|
167 |
+
# If a learned mask is needed on the 32 vectors, it needs separate logic.
|
168 |
+
# mask = torch.ones(D.shape[0], D.shape[1], device=self.device).unsqueeze(2).float()
|
169 |
+
# D = D * mask
|
170 |
+
|
171 |
D = torch.nn.functional.normalize(D, p=2, dim=2)
|
172 |
if self.use_gpu:
|
173 |
D = D.half()
|
174 |
|
175 |
+
# Removed keep_dims conditional branches as _doc now consistently returns fixed 32 vectors.
|
|
|
|
|
|
|
|
|
|
|
176 |
|
177 |
return D
|
178 |
|
requirements.txt
ADDED
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
1 |
+
onnxruntime
|
2 |
+
transformers
|
3 |
+
numpy
|
4 |
+
torch # Required by your modeling.py for ConstBERT logic
|