Merge pull request #27297 from vespa-engine/bjorncs/bert-embedder-services-xml

Bjorncs/bert embedder services xml

2 files changed, 34 insertions, 0 deletions

diff --git a/configdefinitions/src/vespa/embedding.bert-base-embedder.def b/configdefinitions/src/vespa/embedding.bert-base-embedder.def
new file mode 100644
index 00000000000..2d8e840377b
--- /dev/null
+++ b/configdefinitions/src/vespa/embedding.bert-base-embedder.def
@@ -0,0 +1,32 @@
+
+namespace=embedding
+
+# Wordpiece tokenizer
+tokenizerVocab model
+
+transformerModel  model
+
+# Max length of token sequence model can handle
+transformerMaxTokens int default=384
+
+# Pooling strategy
+poolingStrategy enum { cls, mean } default=mean
+
+# Input names
+transformerInputIds      string default=input_ids
+transformerAttentionMask string default=attention_mask
+transformerTokenTypeIds  string default=token_type_ids
+
+# special token ids
+transformerStartSequenceToken int default=101
+transformerEndSequenceToken   int default=102
+
+# Output name
+transformerOutput string default=output_0
+
+# Settings for ONNX model evaluation
+onnxExecutionMode enum { parallel, sequential } default=sequential
+onnxInterOpThreads int default=1
+onnxIntraOpThreads int default=-4  # n=number of threads -> n<0: CPUs/(-n), n==0: CPUs, n>0: n
+# GPU device id, -1 for CPU
+onnxGpuDevice      int default=0
diff --git a/configdefinitions/src/vespa/hugging-face-embedder.def b/configdefinitions/src/vespa/hugging-face-embedder.def
index 36957004e02..7ea4227b3cd 100644
--- a/configdefinitions/src/vespa/hugging-face-embedder.def
+++ b/configdefinitions/src/vespa/hugging-face-embedder.def
@@ -21,6 +21,8 @@ transformerOutput string default=last_hidden_state
 # Normalize tensors from tokenizer
 normalize bool default=false
 
+poolingStrategy enum { cls, mean } default=mean
+
 # Settings for ONNX model evaluation
 transformerExecutionMode enum { parallel, sequential } default=sequential
 transformerInterOpThreads int default=1