make skip_tokens an input for VocabDecode (parametrize detokenization/decoding)

python/openvino_tokenizers/hf_parser.py

@@ -375,8 +375,8 @@ def decoding(self) -> None:
         if self.tokenizer_json["decoder"] is None or self.tokenizer_json["model"]["type"] == "WordPiece":
             return
 
-        skip_tokens = parse_special_tokens(self.original_tokenizer) if self.skip_special_tokens else {}
-        self.pipeline.add_steps(VocabDecoderStep(skip_tokens=list(skip_tokens)))
+        skip_tokens = parse_special_tokens(self.original_tokenizer)
+        self.pipeline.add_steps(VocabDecoderStep(skip_tokens=list(skip_tokens), do_skip_tokens=self.skip_special_tokens))
 
         if self.tokenizer_json["decoder"]["type"] == "Sequence":
             for decoder_dict in self.tokenizer_json["decoder"]["decoders"]:
@@ -1013,9 +1013,7 @@ def convert_tiktoken_model_tokenizer(
     split_pattern = encoding._pat_str
 
     pipeline = TokenizerPipeline()
-    skip_tokens = []
-    if params.skip_special_tokens:
-        skip_tokens = list(parse_special_tokens(hf_tokenizer))
+    skip_tokens = list(parse_special_tokens(hf_tokenizer))
 
     add_prefix_steps = []
     if hasattr(hf_tokenizer, "get_prefix_tokens") and params.add_special_tokens:
@@ -1048,7 +1046,7 @@ def convert_tiktoken_model_tokenizer(
 
     pipeline.add_steps(
         [
-            VocabDecoderStep(vocab=decoder_vocab, skip_tokens=skip_tokens),
+            VocabDecoderStep(vocab=decoder_vocab, skip_tokens=skip_tokens, do_skip_tokens=params.skip_special_tokens),
             FuseStep(),
         ]
     )

python/openvino_tokenizers/tokenizer_pipeline.py

@@ -14,7 +14,7 @@
 from typing import Any, Dict, Iterable, List, Optional, Tuple, Union
 
 import numpy as np
-from openvino.runtime import Model, Output, PartialShape, Type, op
+from openvino.runtime import Model, Output, PartialShape, Type, op, Shape
 from openvino.runtime import opset12 as opset
 from openvino.runtime.exceptions import OVTypeError, UserInputError
 from openvino.runtime.utils.types import as_node, make_constant_node
@@ -1008,6 +1008,7 @@ class DecodingStep(BasePipelineStep):
 class VocabDecoderStep(DecodingStep):
     vocab: Optional[List[str]] = None
     skip_tokens: Optional[List[int]] = None
+    do_skip_tokens: Optional[bool] = True
 
     def finalize(self) -> None:
         pipeline = self.get_pipeline()
@@ -1025,7 +1026,18 @@ def get_ov_subgraph(self, input_nodes: List[Output]) -> List[Output]:
         else:
             vocab_outputs = self.create_string_constant_node(self.vocab).outputs()
         input_nodes.extend(vocab_outputs)
-        return _get_factory().create("VocabDecoder", input_nodes, {"skip_tokens": self.skip_tokens}).outputs()
+
+        # Put constant with skip tokens even if do_skip_tokens=False, so that it can be switched on/off at runtime.
+        # Slice through all skip tokens if flag is true, else slice to get an empty tensor.
+        stop_const = op.Constant(Type.i32, Shape([1]), [np.iinfo(np.int32).max if self.do_skip_tokens else 0])
+
+        zero_const = op.Constant(Type.i32, Shape([1]), [0])
+        one_const = op.Constant(Type.i32, Shape([1]), [1])
+        skip_tokens_const = op.Constant(Type.i32, Shape([len(self.skip_tokens)]), self.skip_tokens)
+        sliced_skips = opset.slice(skip_tokens_const, zero_const, stop_const, one_const).outputs()
+        input_nodes.extend(sliced_skips)
+
+        return _get_factory().create("VocabDecoder", input_nodes).outputs()
 
 
 @dataclass

src/vocab_decoder.cpp

@@ -25,12 +25,20 @@ bool VocabDecoder::evaluate(ov::TensorVector& outputs, const ov::TensorVector& i
     auto seq_len    = inputs[0].get_shape()[1];
     auto input_data = inputs[0].data<const int32_t>();
 
+    auto vocab_size   = inputs[1].get_size();
     auto vocab_begins = inputs[1].data<const int32_t>();
     auto vocab_ends   = inputs[2].data<const int32_t>();
     auto vocab_chars  = inputs[3].data<const uint8_t>();
-    auto vocab_size   = inputs[1].get_size();
 
-    OPENVINO_ASSERT(inputs.size() == 4, "Too few inputs passed to VocabDecoder, it means it is not converted properly or it is not used in the supported pattern");
+    OPENVINO_ASSERT(inputs.size() == 4 || inputs.size() == 5, "Too few inputs passed to VocabDecoder, it means it is not converted properly or it is not used in the supported pattern");
+
+    // Use skip tokens from input if specified, otherwise use the attribute.
+    std::vector<int> skip_tokens;
+    if (inputs.size() == 5) {
+        skip_tokens = std::vector<int>(inputs[4].data<int32_t>(), inputs[4].data<int32_t>() + inputs[4].get_shape()[0]);
+    } else {
+        skip_tokens = m_skip_tokens;
+    }
 
     // Set output shapes
     outputs[0].set_shape({batch_size});
@@ -61,7 +69,7 @@ bool VocabDecoder::evaluate(ov::TensorVector& outputs, const ov::TensorVector& i
             new_begins[seq] = buffer.size();
             if (
                 token_id < vocab_size
-                && std::find(m_skip_tokens.begin(), m_skip_tokens.end(), token_id) == m_skip_tokens.end()
+                && std::find(skip_tokens.begin(), skip_tokens.end(), token_id) == skip_tokens.end()
             ) {
                 buffer.insert(
                     buffer.end(),

src/vocab_decoder.hpp

@@ -6,11 +6,24 @@
 #include <vector>
 #include <openvino/op/op.hpp>
 
+/**
+ * @class VocabDecoder
+ * @brief The VocabDecoder class is an OpenVINO operation that decodes vocabulary tokens.
+ *
+ * This class inherits from the ov::op::Op base class and provides functionality to decode
+ * vocabulary tokens while skipping specified tokens.
+ * @param input_data
+ * @param vocab_begins
+ * @param vocab_ends
+ * @param vocab_chars
+ * @param skip_tokens input has priority over the attribute.
+ * @param skip_tokens attribute is used only when skip_tokens input is not provided.
+ */
 class VocabDecoder : public ov::op::Op {
 public:
     OPENVINO_OP("VocabDecoder");
 
-    VocabDecoder () = default;
+    VocabDecoder () = default;    
     VocabDecoder(
         const ov::OutputVector& arguments,
         std::vector<int> skip_tokens