kernel-memory-dump/HuggingFaceAgentsCourse_SmolAgent1

1#!/usr/bin/env python2# coding=utf-83# Copyright 2024 The HuggingFace Inc. team. All rights reserved.4#5# Licensed under the Apache License, Version 2.0 (the "License");6# you may not use this file except in compliance with the License.7# You may obtain a copy of the License at8#9#     http://www.apache.org/licenses/LICENSE-2.010#11# Unless required by applicable law or agreed to in writing, software12# distributed under the License is distributed on an "AS IS" BASIS,13# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.14# See the License for the specific language governing permissions and15# limitations under the License.16import mimetypes17import os18import re19import shutil20from typing import Optional21 22from smolagents.agent_types import (23    AgentAudio,24    AgentImage,25    AgentText,26    handle_agent_output_types,27)28from smolagents.agents import ActionStep, MultiStepAgent29from smolagents.memory import MemoryStep30from smolagents.utils import _is_package_available31 32 33def pull_messages_from_step(34    step_log: MemoryStep,35):36    """Extract ChatMessage objects from agent steps with proper nesting"""37    import gradio as gr38 39    if isinstance(step_log, ActionStep):40        # Output the step number41        step_number = (42            f"Step {step_log.step_number}" if step_log.step_number is not None else ""43        )44        yield gr.ChatMessage(role="assistant", content=f"**{step_number}**")45 46        # First yield the thought/reasoning from the LLM47        if hasattr(step_log, "model_output") and step_log.model_output is not None:48            # Clean up the LLM output49            model_output = step_log.model_output.strip()50            # Remove any trailing <end_code> and extra backticks, handling multiple possible formats51            model_output = re.sub(52                r"```\s*<end_code>", "```", model_output53            )  # handles ```<end_code>54            model_output = re.sub(55                r"<end_code>\s*```", "```", model_output56            )  # handles <end_code>```57            model_output = re.sub(58                r"```\s*\n\s*<end_code>", "```", model_output59            )  # handles ```\n<end_code>60            model_output = model_output.strip()61            yield gr.ChatMessage(role="assistant", content=model_output)62 63        # For tool calls, create a parent message64        if hasattr(step_log, "tool_calls") and step_log.tool_calls is not None:65            first_tool_call = step_log.tool_calls[0]66            used_code = first_tool_call.name == "python_interpreter"67            parent_id = f"call_{len(step_log.tool_calls)}"68 69            # Tool call becomes the parent message with timing info70            # First we will handle arguments based on type71            args = first_tool_call.arguments72            if isinstance(args, dict):73                content = str(args.get("answer", str(args)))74            else:75                content = str(args).strip()76 77            if used_code:78                # Clean up the content by removing any end code tags79                content = re.sub(80                    r"```.*?\n", "", content81                )  # Remove existing code blocks82                content = re.sub(83                    r"\s*<end_code>\s*", "", content84                )  # Remove end_code tags85                content = content.strip()86                if not content.startswith("```python"):87                    content = f"```python\n{content}\n```"88 89            parent_message_tool = gr.ChatMessage(90                role="assistant",91                content=content,92                metadata={93                    "title": f"🛠️ Used tool {first_tool_call.name}",94                    "id": parent_id,95                    "status": "pending",96                },97            )98            yield parent_message_tool99 100            # Nesting execution logs under the tool call if they exist101            if hasattr(step_log, "observations") and (102                step_log.observations is not None and step_log.observations.strip()103            ):  # Only yield execution logs if there's actual content104                log_content = step_log.observations.strip()105                if log_content:106                    log_content = re.sub(r"^Execution logs:\s*", "", log_content)107                    yield gr.ChatMessage(108                        role="assistant",109                        content=f"{log_content}",110                        metadata={111                            "title": "📝 Execution Logs",112                            "parent_id": parent_id,113                            "status": "done",114                        },115                    )116 117            # Nesting any errors under the tool call118            if hasattr(step_log, "error") and step_log.error is not None:119                yield gr.ChatMessage(120                    role="assistant",121                    content=str(step_log.error),122                    metadata={123                        "title": "💥 Error",124                        "parent_id": parent_id,125                        "status": "done",126                    },127                )128 129            # Update parent message metadata to done status without yielding a new message130            parent_message_tool.metadata["status"] = "done"131 132        # Handle standalone errors but not from tool calls133        elif hasattr(step_log, "error") and step_log.error is not None:134            yield gr.ChatMessage(135                role="assistant",136                content=str(step_log.error),137                metadata={"title": "💥 Error"},138            )139 140        # Calculate duration and token information141        step_footnote = f"{step_number}"142        if hasattr(step_log, "input_token_count") and hasattr(143            step_log, "output_token_count"144        ):145            token_str = f" | Input-tokens:{step_log.input_token_count:,} | Output-tokens:{step_log.output_token_count:,}"146            step_footnote += token_str147        if hasattr(step_log, "duration"):148            step_duration = (149                f" | Duration: {round(float(step_log.duration), 2)}"150                if step_log.duration151                else None152            )153            step_footnote += step_duration154        step_footnote = f"""<span style="color: #bbbbc2; font-size: 12px;">{step_footnote}</span> """155        yield gr.ChatMessage(role="assistant", content=f"{step_footnote}")156        yield gr.ChatMessage(role="assistant", content="-----")157 158 159def stream_to_gradio(160    agent,161    task: str,162    reset_agent_memory: bool = False,163    additional_args: Optional[dict] = None,164):165    """Runs an agent with the given task and streams the messages from the agent as gradio ChatMessages."""166    if not _is_package_available("gradio"):167        raise ModuleNotFoundError(168            "Please install 'gradio' extra to use the GradioUI: `pip install 'smolagents[gradio]'`"169        )170    import gradio as gr171 172    total_input_tokens = 0173    total_output_tokens = 0174 175    for step_log in agent.run(176        task, stream=True, reset=reset_agent_memory, additional_args=additional_args177    ):178        # Track tokens if model provides them179        if hasattr(agent.model, "last_input_token_count"):180            total_input_tokens += agent.model.last_input_token_count181            total_output_tokens += agent.model.last_output_token_count182            if isinstance(step_log, ActionStep):183                step_log.input_token_count = agent.model.last_input_token_count184                step_log.output_token_count = agent.model.last_output_token_count185 186        for message in pull_messages_from_step(187            step_log,188        ):189            yield message190 191    final_answer = step_log  # Last log is the run's final_answer192    final_answer = handle_agent_output_types(final_answer)193 194    if isinstance(final_answer, AgentText):195        yield gr.ChatMessage(196            role="assistant",197            content=f"**Final answer:**\n{final_answer.to_string()}\n",198        )199    elif isinstance(final_answer, AgentImage):200        yield gr.ChatMessage(201            role="assistant",202            content={"path": final_answer.to_string(), "mime_type": "image/png"},203        )204    elif isinstance(final_answer, AgentAudio):205        yield gr.ChatMessage(206            role="assistant",207            content={"path": final_answer.to_string(), "mime_type": "audio/wav"},208        )209    else:210        yield gr.ChatMessage(211            role="assistant", content=f"**Final answer:** {str(final_answer)}"212        )213 214 215class GradioUI:216    """A one-line interface to launch your agent in Gradio"""217 218    def __init__(219        self,220        agent: MultiStepAgent,221        file_upload_folder: str | None = None,222        initial_message=None,223    ):224        if not _is_package_available("gradio"):225            raise ModuleNotFoundError(226                "Please install 'gradio' extra to use the GradioUI: `pip install 'smolagents[gradio]'`"227            )228        self.agent = agent229        self.file_upload_folder = file_upload_folder230        self.initial_message = initial_message231        if self.file_upload_folder is not None:232            if not os.path.exists(file_upload_folder):233                os.mkdir(file_upload_folder)234 235    def interact_with_agent(self, prompt, messages):236        import gradio as gr237 238        messages.append(gr.ChatMessage(role="user", content=prompt))239        yield messages240        for msg in stream_to_gradio(self.agent, task=prompt, reset_agent_memory=False):241            messages.append(msg)242            yield messages243        yield messages244 245    def upload_file(246        self,247        file,248        file_uploads_log,249        allowed_file_types=[250            "application/pdf",251            "application/vnd.openxmlformats-officedocument.wordprocessingml.document",252            "text/plain",253        ],254    ):255        """256        Handle file uploads, default allowed types are .pdf, .docx, and .txt257        """258        import gradio as gr259 260        if file is None:261            return gr.Textbox("No file uploaded", visible=True), file_uploads_log262 263        try:264            mime_type, _ = mimetypes.guess_type(file.name)265        except Exception as e:266            return gr.Textbox(f"Error: {e}", visible=True), file_uploads_log267 268        if mime_type not in allowed_file_types:269            return gr.Textbox("File type disallowed", visible=True), file_uploads_log270 271        # Sanitize file name272        original_name = os.path.basename(file.name)273        sanitized_name = re.sub(274            r"[^\w\-.]", "_", original_name275        )  # Replace any non-alphanumeric, non-dash, or non-dot characters with underscores276 277        type_to_ext = {}278        for ext, t in mimetypes.types_map.items():279            if t not in type_to_ext:280                type_to_ext[t] = ext281 282        # Ensure the extension correlates to the mime type283        sanitized_name = sanitized_name.split(".")[:-1]284        sanitized_name.append("" + type_to_ext[mime_type])285        sanitized_name = "".join(sanitized_name)286 287        # Save the uploaded file to the specified folder288        file_path = os.path.join(289            self.file_upload_folder, os.path.basename(sanitized_name)290        )291        shutil.copy(file.name, file_path)292 293        return gr.Textbox(294            f"File uploaded: {file_path}", visible=True295        ), file_uploads_log + [file_path]296 297    def log_user_message(self, text_input, file_uploads_log):298        return (299            text_input300            + (301                f"\nYou have been provided with these files, which might be helpful or not: {file_uploads_log}"302                if len(file_uploads_log) > 0303                else ""304            ),305            "",306        )307 308    def launch(self, **kwargs):309        import gradio as gr310 311        with gr.Blocks(fill_height=True) as demo:312            gr.State(self.initial_message)313            stored_messages = gr.State([])314            file_uploads_log = gr.State([])315            chatbot = gr.Chatbot(316                label="Agent",317                type="messages",318                value=(319                    [{"role": "assistant", "content": self.initial_message}]320                    if self.initial_message321                    else []322                ),323                avatar_images=(324                    None,325                    "https://huggingface.co/datasets/agents-course/course-images/resolve/main/en/communication/Alfred.png",326                ),327                resizeable=True,328                scale=1,329            )330            # If an upload folder is provided, enable the upload feature331            if self.file_upload_folder is not None:332                upload_file = gr.File(label="Upload a file")333                upload_status = gr.Textbox(334                    label="Upload Status", interactive=False, visible=False335                )336                upload_file.change(337                    self.upload_file,338                    [upload_file, file_uploads_log],339                    [upload_status, file_uploads_log],340                )341            text_input = gr.Textbox(342                submit_btn="Start the ritual of code analysis",343                lines=10,344                label="Sacred Code wrapped in ```python ``` click on button to start the ritual of code analysis, shift+enter also works as submit",345            )346            text_input.submit(347                self.log_user_message,348                [text_input, file_uploads_log],349                [stored_messages, text_input],350            ).then(self.interact_with_agent, [stored_messages, chatbot], [chatbot])351 352        demo.launch(debug=True, share=True, **kwargs)353 354 355__all__ = ["stream_to_gradio", "GradioUI"]356