Refactor for streaming

app.py

@@ -3,25 +3,21 @@ import os
 import shutil
 
 import gradio as gr
-import requests
 from huggingface_hub import Repository
-from share_btn import community_icon_html, loading_icon_html, share_btn_css, share_js
+from text_generation import Client
 
+from dialogues import DialogueTemplate
+from share_btn import (community_icon_html, loading_icon_html, share_btn_css,
+                       share_js)
 
-HF_TOKEN = os.environ.get("H4_TOKEN", None)
+HF_TOKEN = os.environ.get("HF_TOKEN", None)
 API_TOKEN = os.environ.get("API_TOKEN", None)
-STAR_CHAT_API_URL = os.environ.get("STAR_CHAT_API_URL", None)
-STAR_CHAT_GPT_API_URL = os.environ.get("STAR_CHAT_GPT_API_URL", None)
+API_URL = os.environ.get("API_URL", None)
 
-API_TOKEN = "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"
-STAR_CHAT_API_URL = "https://i1qe9e7uv7jzsg8k.us-east-1.aws.endpoints.huggingface.cloud"
-STAR_CHAT_GPT_API_URL = "https://czpdnzuklyfoqjbs.us-east-1.aws.endpoints.huggingface.cloud"
-
-model_to_api = {
-    "StarChat": STAR_CHAT_API_URL,
-    "StarChatGPT": STAR_CHAT_GPT_API_URL,
-}
-PROMPT_TEMPLATE = "<|system|>\n{system}<|end|>\n<|user|>\n{prompt}<|end|>\n<|assistant|>"
+client = Client(
+    API_URL,
+    headers={"Authorization": f"Bearer {API_TOKEN}"},
+)
 
 theme = gr.themes.Monochrome(
     primary_hue="indigo",
@@ -31,57 +27,25 @@ theme = gr.themes.Monochrome(
     font=[gr.themes.GoogleFont("Open Sans"), "ui-sans-serif", "system-ui", "sans-serif"],
 )
 
-if HF_TOKEN:
-    try:
-        shutil.rmtree("./data/")
-    except:
-        pass
-
-    repo = Repository(
-        local_dir="./data/", clone_from="trl-lib/star-chat-prompts", use_auth_token=HF_TOKEN, repo_type="dataset"
-    )
-    repo.git_pull()
+# if HF_TOKEN:
+#     try:
+#         shutil.rmtree("./data/")
+#     except:
+#         pass
 
+#     repo = Repository(
+#         local_dir="./data/", clone_from="trl-lib/star-chat-prompts", use_auth_token=HF_TOKEN, repo_type="dataset"
+#     )
+#     repo.git_pull()
 
 
 def save_inputs_and_outputs(inputs, outputs, generate_kwargs):
     with open(os.path.join("data", "prompts.jsonl"), "a") as f:
-        json.dump({"inputs": inputs, "outputs": outputs,
-                  "generate_kwargs": generate_kwargs}, f, ensure_ascii=False)
+        json.dump({"inputs": inputs, "outputs": outputs, "generate_kwargs": generate_kwargs}, f, ensure_ascii=False)
         f.write("\n")
         repo.push_to_hub()
 
 
-def inference(
-    model, prompt, system_message, user_message, temperature, top_p, top_k, max_new_tokens, do_sample, eos_token_id
-):
-    headers = {"Authorization": f"Bearer {API_TOKEN}"}
-    api_url = model_to_api[model]
-    print(f"CUSTOM_LOG {model} - {api_url}")
-    response = requests.post(
-        api_url,
-        headers=headers,
-        json={
-            "inputs": prompt,
-            "parameters": {
-                "do_sample": do_sample,
-                "temperature": temperature,
-                "top_p": top_p,
-                "top_k": top_k,
-                "max_new_tokens": max_new_tokens,
-                "eos_token_id": eos_token_id,
-            },
-        },
-    )
-
-    if response.status_code != 200:
-        return None
-    completion = response.json()[0]["generated_text"]
-    if user_message in completion:
-        completion = completion.lstrip()[len(f"{system_message}\n{user_message}\n"):]
-    return completion
-
-
 def get_total_inputs(inputs, chatbot, preprompt, user_name, assistant_name, sep):
     past = []
     for data in chatbot:
@@ -107,7 +71,7 @@ def has_no_history(chatbot, history):
 
 
 def generate(
-    model,
+    # model,
     system_message,
     user_message,
     chatbot,
@@ -122,32 +86,74 @@ def generate(
     if not user_message:
         return chatbot, history, user_message, ""
 
-    prompt = PROMPT_TEMPLATE.format(system=system_message, prompt=user_message)
-
     history.append(user_message)
 
+    past_messages = []
+    for data in chatbot:
+        user_data, model_data = data
+
+        past_messages.extend(
+            [{"role": "user", "content": user_data}, {"role": "assistant", "content": model_data.rstrip()}]
+        )
+
+    if len(past_messages) < 1:
+        dialogue_template = DialogueTemplate(
+            system=system_message, messages=[{"role": "user", "content": user_message}]
+        )
+        prompt = dialogue_template.get_inference_prompt()
+    else:
+        dialogue_template = DialogueTemplate(
+            system=system_message, messages=past_messages + [{"role": "user", "content": user_message}]
+        )
+        prompt = dialogue_template.get_inference_prompt()
+
     generate_kwargs = {
         "temperature": temperature,
         "top_p": top_p,
         "top_k": top_k,
         "max_new_tokens": max_new_tokens,
-        "do_sample": True,
-        "eos_token_id": [49155, 32003],
     }
 
-    response = inference(model, prompt, system_message, user_message, **generate_kwargs)
+    temperature = float(temperature)
+    if temperature < 1e-2:
+        temperature = 1e-2
+    top_p = float(top_p)
+
+    generate_kwargs = dict(
+        temperature=temperature,
+        max_new_tokens=max_new_tokens,
+        top_p=top_p,
+        do_sample=True,
+        truncate=999,
+        seed=42,
+        stop_sequences=["<|end|>"],
+    )
+
+    stream = client.generate_stream(
+        prompt,
+        **generate_kwargs,
+    )
+
+    output = ""
+    for idx, response in enumerate(stream):
+        if response.token.special:
+            continue
+        output += response.token.text
+        if idx == 0:
+            history.append(" " + output)
+        else:
+            history[-1] = output
 
-    history.append(response)
-    chat = [(history[i].strip(), history[i + 1].strip()) for i in range(0, len(history) - 1, 2)]
+        chat = [(history[i].strip(), history[i + 1].strip()) for i in range(0, len(history) - 1, 2)]
 
-    if HF_TOKEN and do_save:
-        try:
-            print("Pushing prompt and completion to the Hub")
-            save_inputs_and_outputs(prompt, output, generate_kwargs)
-        except Exception as e:
-            print(e)
+        # if HF_TOKEN and do_save:
+        #     try:
+        #         print("Pushing prompt and completion to the Hub")
+        #         save_inputs_and_outputs(prompt, output, generate_kwargs)
+        #     except Exception as e:
+        #         print(e)
 
-    return chat, history, user_message, ""
+        yield chat, history, user_message, ""
 
 
 examples = [
@@ -160,7 +166,6 @@ examples = [
 
 
 def regenerate(
-    model,
     system_message,
     user_message,
     chatbot,
@@ -183,33 +188,20 @@ def regenerate(
     chatbot = chatbot[:-1]
     history = history[:-2]
 
-    return generate(
-        model, system_message, user_message, chatbot, history, temperature, top_p, top_k, max_new_tokens, do_save
-    )
+    return generate(system_message, user_message, chatbot, history, temperature, top_p, top_k, max_new_tokens, do_save)
 
 
 def clear_chat():
     return [], []
 
 
-def radio_on_change():
-    return [], []
-
-# def radio_on_change(
-#     model, system_message, user_message, chatbot, history, temperature, top_p, top_k, max_new_tokens, do_save
-# ):
-#     return generate(
-#         model, system_message, user_message, chatbot, history, temperature, top_p, top_k, max_new_tokens, do_save
-#     )
-
-
 def process_example(args):
     for [x, y] in generate(args):
         pass
     return [x, y]
 
 
-title = """<h1 align="center">⭐ StarChat Demo 💬</h1>"""
+title = """<h1 align="center">⭐ Chat with StarCoder Demo 💬</h1>"""
 custom_css = """
 #banner-image {
   display: block;
@@ -253,42 +245,34 @@ with gr.Blocks(theme=theme, analytics_enabled=False, css=custom_css) as demo:
     gr.HTML(title)
     gr.Image("StarCoderBanner.png", elem_id="banner-image", show_label=False)
     gr.Markdown(
-    """
-            StarChat is an instruction fine-tuned model based on [StarCoder](https://huggingface.co/bigcode/starcoder), a 16B parameter model trained on one trillion tokens sourced from 80+ programming languages, GitHub issues, Git commits, and Jupyter notebooks (all permissively licensed). With an enterprise-friendly license, 8,192 token context length, and fast large-batch inference via [multi-query attention](https://arxiv.org/abs/1911.02150), StarCoder is currently the best open-source choice for code-based applications. For more details, check out our [blog post]().
+        """
+            This demo showcases an instruction fine-tuned model based on [StarCoder](https://huggingface.co/bigcode/starcoder), a 16B parameter model trained on one trillion tokens sourced from 80+ programming languages, GitHub issues, Git commits, and Jupyter notebooks (all permissively licensed). With an enterprise-friendly license, 8,192 token context length, and fast large-batch inference via [multi-query attention](https://arxiv.org/abs/1911.02150), StarCoder is currently the best open-source choice for code-based applications. For more details, check out our [blog post]().
+
+            ⚠️ **Intended Use**: this app and its [supporting model](https://huggingface.co/HuggingFaceH4/starcoderbase-finetuned-oasst1) are provided as educational tools to explain instruction fine-tuning; not to serve as replacement for human expertise. For more details on the model's limitations in terms of factuality and biases, see the [model card](https://huggingface.co/HuggingFaceH4/starcoderbase-finetuned-oasst1#bias-risks-and-limitations).
 
-            ⚠️ **Intended Use**: this app and its supporting models ([StarChat](https://huggingface.co/HuggingFaceH4/starchat) and [StarChatGPT](https://huggingface.co/HuggingFaceH4/starchatgpt)) are provided as educational tools to explain instruction fine-tuning; not to serve as replacement for human expertise. For more details on the model's limitations in terms of factuality and biases, see the model cards: [StarChat](https://huggingface.co/HuggingFaceH4/starchat#bias-risks-and-limitations) and [StarChatGPT](https://huggingface.co/HuggingFaceH4/starchatgpt#bias-risks-and-limitations).
-            
-            ⚠️ **Data Collection**: by default, we are collecting the prompts entered in this app to further improve and evaluate the model. Do not share any personal or sensitive information while using the app! You can opt out of this data collection by removing the checkbox below.
+            ⚠️ **Data Collection**: by default, we are collecting the prompts entered in this app to further improve and evaluate the model. Do NOT share any personal or sensitive information while using the app! You can opt out of this data collection by removing the checkbox below.
     """
     )
 
+    with gr.Row():
+        do_save = gr.Checkbox(
+            value=True,
+            label="Store data",
+            info="You agree to the storage of your prompt and generated text for research and development purposes:",
+        )
+
     with gr.Row():
         with gr.Column(scale=1):
             system_message = gr.Textbox(elem_id="system-message", label="System prompt")
 
         with gr.Column(scale=2):
             with gr.Box():
-                model = gr.Radio(
-                    value="StarChat",
-                    choices=[
-                        "StarChat",
-                        "StarChatGPT",
-                    ],
-                    label="Model",
-                    interactive=True,
-                )
                 output = gr.Markdown()
                 chatbot = gr.Chatbot(elem_id="chat-message", label="Chat")
 
     with gr.Row():
         with gr.Column(scale=3):
-            do_save = gr.Checkbox(
-                value=True,
-                label="Store data",
-                info="You agree to the storage of your prompt and generated text for research and development purposes:",
-            )
-            user_message = gr.Textbox(placeholder="Enter your message here",
-                                      show_label=False, elem_id="q-input")
+            user_message = gr.Textbox(placeholder="Enter your message here", show_label=False, elem_id="q-input")
             with gr.Row():
                 send_button = gr.Button("Send", elem_id="send-btn", visible=True)
                 regenerate_button = gr.Button("Regenerate", elem_id="send-btn", visible=True)
@@ -298,7 +282,7 @@ with gr.Blocks(theme=theme, analytics_enabled=False, css=custom_css) as demo:
             # with gr.Group(elem_id="share-btn-container"):
             #     community_icon = gr.HTML(community_icon_html, visible=True)
             #     loading_icon = gr.HTML(loading_icon_html, visible=True)
-                # share_button = gr.Button("Share to community", elem_id="share-btn", visible=True)
+            # share_button = gr.Button("Share to community", elem_id="share-btn", visible=True)
             with gr.Row():
                 gr.Examples(
                     examples=examples,
@@ -311,9 +295,9 @@ with gr.Blocks(theme=theme, analytics_enabled=False, css=custom_css) as demo:
         with gr.Column(scale=1):
             temperature = gr.Slider(
                 label="Temperature",
-                value=0.8,
+                value=0.2,
                 minimum=0.0,
-                maximum=2.0,
+                maximum=1.0,
                 step=0.1,
                 interactive=True,
                 info="Higher values produce more diverse outputs",
@@ -329,7 +313,7 @@ with gr.Blocks(theme=theme, analytics_enabled=False, css=custom_css) as demo:
             )
             top_p = gr.Slider(
                 label="Top-p (nucleus sampling)",
-                value=0.25,
+                value=0.95,
                 minimum=0.0,
                 maximum=1,
                 step=0.05,
@@ -338,7 +322,7 @@ with gr.Blocks(theme=theme, analytics_enabled=False, css=custom_css) as demo:
             )
             max_new_tokens = gr.Slider(
                 label="Max new tokens",
-                value=512,
+                value=384,
                 minimum=0,
                 maximum=2048,
                 step=4,
@@ -353,7 +337,6 @@ with gr.Blocks(theme=theme, analytics_enabled=False, css=custom_css) as demo:
     user_message.submit(
         generate,
         inputs=[
-            model,
             system_message,
             user_message,
             chatbot,
@@ -370,7 +353,6 @@ with gr.Blocks(theme=theme, analytics_enabled=False, css=custom_css) as demo:
     send_button.click(
         generate,
         inputs=[
-            model,
             system_message,
             user_message,
             chatbot,
@@ -387,7 +369,6 @@ with gr.Blocks(theme=theme, analytics_enabled=False, css=custom_css) as demo:
     regenerate_button.click(
         regenerate,
         inputs=[
-            model,
             system_message,
             last_user_message,
             chatbot,
@@ -402,8 +383,6 @@ with gr.Blocks(theme=theme, analytics_enabled=False, css=custom_css) as demo:
     )
 
     clear_chat_button.click(clear_chat, outputs=[chatbot, history])
-
-    model.change(radio_on_change, outputs=[chatbot, history])
     # share_button.click(None, [], [], _js=share_js)
 
 demo.queue(concurrency_count=16).launch(debug=True)

dialogues.py

@@ -0,0 +1,241 @@
+# coding=utf-8
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import json
+import os
+from dataclasses import asdict, dataclass
+from pathlib import Path
+from typing import Any, Dict, List, Optional, Type, TypeVar, Union
+
+from huggingface_hub import ModelHubMixin, hf_hub_download
+
+# Generic variable that is either ModelHubMixin or a subclass thereof
+T = TypeVar("T", bound="ModelHubMixin")
+
+TEMPLATE_FILENAME = "dialogue_template.json"
+IGNORE_INDEX = -100
+
+
+@dataclass
+class DialogueTemplate(ModelHubMixin):
+    """Converts all turns of a dialogue between a user and assistant to a standardized format.
+
+    Adapted from OpenAI's ChatML (https://github.com/openai/openai-python/blob/main/chatml.md) and Vicuna (https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py)
+    """
+
+    system: str
+    messages: List[Dict[str, str]] = None
+    system_token: str = "<|system|>"
+    user_token: str = "<|user|>"
+    assistant_token: str = "<|assistant|>"
+    end_token: str = "<|end|>"
+
+    def get_training_prompt(self) -> str:
+        prompt = self.system_token + "\n" + self.system + self.end_token + "\n"
+        if self.messages is None:
+            raise ValueError("Dialogue template must have at least one message.")
+        for message in self.messages:
+            if message["role"] == "user":
+                prompt += self.user_token + "\n" + message["content"] + self.end_token + "\n"
+            else:
+                prompt += self.assistant_token + "\n" + message["content"] + self.end_token + "\n"
+        return prompt
+
+    def get_inference_prompt(self) -> str:
+        prompt = self.system_token + "\n" + self.system + self.end_token + "\n"
+        if self.messages is None:
+            raise ValueError("Dialogue template must have at least one message.")
+        for message in self.messages:
+            if message["role"] == "user":
+                prompt += self.user_token + "\n" + message["content"] + self.end_token + "\n"
+            else:
+                prompt += self.assistant_token + "\n" + message["content"] + self.end_token + "\n"
+        prompt += self.assistant_token
+        return prompt
+
+    def get_dialogue(self):
+        """Helper function to format the messages as an easy-to-read dialogue."""
+        prompt = ""
+        if self.messages is None:
+            raise ValueError("Dialogue template must have at least one message.")
+        for message in self.messages:
+            if message["role"] == "user":
+                prompt += "\n\nHuman: " + message["content"]
+            else:
+                prompt += "\n\nAssistant: " + message["content"]
+        return prompt
+
+    def get_special_tokens(self) -> List[str]:
+        return [self.system_token, self.user_token, self.assistant_token, self.end_token]
+
+    def copy(self):
+        return DialogueTemplate(
+            system=self.system,
+            messages=self.messages,
+            system_token=self.system_token,
+            user_token=self.user_token,
+            assistant_token=self.assistant_token,
+            end_token=self.end_token,
+        )
+
+    def to_dict(self) -> Dict[str, Any]:
+        return {k: v for k, v in asdict(self).items()}
+
+    @classmethod
+    def from_dict(cls, data):
+        return DialogueTemplate(
+            system=data["system"] if "system" in data else "",
+            messages=data["messages"] if "messages" in data else None,
+            system_token=data["system_token"] if "system_token" in data else "<|system|>",
+            user_token=data["user_token"] if "user_token" in data else "<|user|>",
+            assistant_token=data["assistant_token"] if "assistant_token" in data else "<|assistant|>",
+            end_token=data["end_token"] if "end_token" in data else "<|end|>",
+        )
+
+    def _save_pretrained(self, save_directory: Union[str, Path]) -> None:
+        save_directory = Path(save_directory)
+        save_directory.mkdir(exist_ok=True)
+        with open(save_directory / "dialogue_template.json", "w") as f:
+            json.dump(self.to_dict(), f, indent=2)
+
+    @classmethod
+    def _from_pretrained(
+        cls: Type[T],
+        *,
+        model_id: str,
+        revision: Optional[str],
+        cache_dir: Optional[Union[str, Path]],
+        force_download: bool,
+        proxies: Optional[Dict],
+        resume_download: bool,
+        local_files_only: bool,
+        token: Optional[Union[str, bool]],
+        **model_kwargs,
+    ) -> T:
+        """Loads the dialogue template from a local directory or the Huggingface Hub.
+
+        Args:
+            model_id (`str`):
+                ID of the model to load from the Huggingface Hub (e.g. `bigscience/bloom`).
+            revision (`str`, *optional*):
+                Revision of the model on the Hub. Can be a branch name, a git tag or any commit id. Defaults to the
+                latest commit on `main` branch.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether to force (re-)downloading the model weights and configuration files from the Hub, overriding
+                the existing cache.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether to delete incompletely received files. Will attempt to resume the download if such a file exists.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint (e.g., `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`).
+            token (`str` or `bool`, *optional*):
+                The token to use as HTTP bearer authorization for remote files. By default, it will use the token
+                cached when running `huggingface-cli login`.
+            cache_dir (`str`, `Path`, *optional*):
+                Path to the folder where cached files are stored.
+            local_files_only (`bool`, *optional*, defaults to `False`):
+                If `True`, avoid downloading the file and return the path to the local cached file if it exists.
+            model_kwargs:
+                Additional keyword arguments passed along to the [`~ModelHubMixin._from_pretrained`] method.
+        """
+        if os.path.isdir(model_id):  # Can either be a local directory
+            print("Loading dialogue template from local directory")
+            template_file = os.path.join(model_id, TEMPLATE_FILENAME)
+        else:  # Or a template on the Hub
+            template_file = hf_hub_download(  # Download from the hub, passing same input args
+                repo_id=model_id,
+                filename=TEMPLATE_FILENAME,
+                revision=revision,
+                cache_dir=cache_dir,
+                force_download=force_download,
+                proxies=proxies,
+                resume_download=resume_download,
+                token=token,
+                local_files_only=local_files_only,
+            )
+
+        # Load template
+        with open(template_file, "r") as f:
+            data = json.load(f)
+        return cls.from_dict(data=data)
+
+
+# A shortened version of the system message in Anthropic's HHH prompt: https://gist.github.com/jareddk/2509330f8ef3d787fc5aaac67aab5f11#file-hhh_prompt-txt
+default_template = DialogueTemplate(
+    system="Below is a dialogue between a human user and an AI assistant. The assistant is happy to help with almost anything, and will do its best to understand exactly what is needed.",
+)
+
+# OpenAI and OpenAssistant train on few to no system messages.
+# TODO: consider defining this as the `default` template
+no_system_template = DialogueTemplate(
+    system="",
+)
+
+alpaca_template = DialogueTemplate(
+    system="Below is an instruction that describes a task. Write a response that appropriately completes the request.",
+    user_token="### Instruction:",
+    assistant_token="### Response:",
+)
+
+SUPPORTED_DIALOGUE_TEMPLATES = {
+    "default": default_template,
+    "no_system": no_system_template,
+    "alpaca": alpaca_template,
+}
+
+
+def get_dialogue_template(template: str) -> DialogueTemplate:
+    if template not in SUPPORTED_DIALOGUE_TEMPLATES.keys():
+        raise ValueError(f"Template {template} is not supported!")
+    return SUPPORTED_DIALOGUE_TEMPLATES[template].copy()
+
+
+def prepare_dialogue(example, dialogue_template, is_train=True):
+    """Format example to single- or multi-turn dialogue."""
+    # TODO: make this simpler by just ensuring every dataset has a messages column
+    if "messages" in example.keys() and example["messages"] is not None:
+        dialogue_template.messages = example["messages"]
+    elif all(k in example.keys() for k in ("prompt", "completion")):
+        # Construct single-turn dialogue from prompt and completion
+        dialogue_template.messages = [
+            {"role": "user", "content": example["prompt"]},
+            {"role": "assistant", "content": example["completion"]},
+        ]
+    elif "prompt" in example.keys():
+        # Construct single-turn dialogue from prompt (inference only)
+        dialogue_template.messages = [
+            {"role": "user", "content": example["prompt"]},
+        ]
+    else:
+        raise ValueError(
+            f"Could not format example as dialogue! Require either `messages` or `[prompt, completion]` or `[prompt]` keys but found {list(example.keys())}"
+        )
+    if is_train:
+        example["text"] = dialogue_template.get_training_prompt()
+    else:
+        example["text"] = dialogue_template.get_inference_prompt()
+    return example
+
+
+def mask_user_labels(tokenizer, dialogue_template, labels):
+    """Masks the user turns of a dialogue from the loss"""
+    user_token_id = tokenizer.convert_tokens_to_ids(dialogue_template.user_token)
+    assistant_token_id = tokenizer.convert_tokens_to_ids(dialogue_template.assistant_token)
+    for idx, label_id in enumerate(labels):
+        if label_id == user_token_id:
+            current_idx = idx
+            while labels[current_idx] != assistant_token_id and current_idx < len(labels):
+                labels[current_idx] = IGNORE_INDEX
+                current_idx += 1