import gradio as gr from gradio_leaderboard import Leaderboard, ColumnFilter, SelectColumns import pandas as pd from apscheduler.schedulers.background import BackgroundScheduler from huggingface_hub import snapshot_download import matplotlib.pyplot as plt import numpy as np import torch from transformers import AutoTokenizer, AutoModelForCausalLM from src.about import ( CITATION_BUTTON_LABEL, CITATION_BUTTON_TEXT, EVALUATION_QUEUE_TEXT, INTRODUCTION_TEXT, LLM_BENCHMARKS_TEXT, TITLE, REPORT_TEXT, ) from src.display.css_html_js import custom_css from src.display.utils import ( BENCHMARK_COLS, COLS, EVAL_COLS, EVAL_TYPES, AutoEvalColumn, ModelType, fields, WeightType, Precision ) from src.envs import API, EVAL_REQUESTS_PATH, EVAL_RESULTS_PATH, QUEUE_REPO, REPO_ID, RESULTS_REPO, TOKEN from src.populate import get_evaluation_queue_df, get_leaderboard_df from src.submission.submit import add_new_eval def restart_space(): API.restart_space(repo_id=REPO_ID) ### Space initialisation try: print(EVAL_REQUESTS_PATH) snapshot_download( repo_id=QUEUE_REPO, local_dir=EVAL_REQUESTS_PATH, repo_type="dataset", tqdm_class=None, etag_timeout=30, token=TOKEN ) except Exception: restart_space() try: print(EVAL_RESULTS_PATH) snapshot_download( repo_id=RESULTS_REPO, local_dir=EVAL_RESULTS_PATH, repo_type="dataset", tqdm_class=None, etag_timeout=30, token=TOKEN ) except Exception: restart_space() LEADERBOARD_DF = get_leaderboard_df(EVAL_RESULTS_PATH, EVAL_REQUESTS_PATH, COLS, BENCHMARK_COLS) ( finished_eval_queue_df, running_eval_queue_df, pending_eval_queue_df, ) = get_evaluation_queue_df(EVAL_REQUESTS_PATH, EVAL_COLS) def init_leaderboard(dataframe): if dataframe is None or dataframe.empty: raise ValueError("Leaderboard DataFrame is empty or None.") return Leaderboard( value=dataframe, datatype=[c.type for c in fields(AutoEvalColumn)], select_columns=SelectColumns( default_selection=[c.name for c in fields(AutoEvalColumn) if c.displayed_by_default], cant_deselect=[c.name for c in fields(AutoEvalColumn) if c.never_hidden], label="Select Columns to Display:", ), search_columns=[AutoEvalColumn.model.name, AutoEvalColumn.license.name], hide_columns=[c.name for c in fields(AutoEvalColumn) if c.hidden], filter_columns=[ ColumnFilter(AutoEvalColumn.model_type.name, type="checkboxgroup", label="Model types"), ColumnFilter(AutoEvalColumn.precision.name, type="checkboxgroup", label="Precision"), ColumnFilter( AutoEvalColumn.params.name, type="slider", min=0.01, max=150, label="Select the number of parameters (B)", ), ColumnFilter( AutoEvalColumn.still_on_hub.name, type="boolean", label="Deleted/incomplete", default=True ), ], bool_checkboxgroup_label="Hide models", interactive=False, ) def draw_grace_radar(): models = ["LLaMA-2-7b-chat", "Qwen-7B-Chat"] labels = ["Instruction Following", "Coding", "Math", "Reasoning", "Multilingual"] scores = [ [0.89, 0.87, 0.82, 0.92, 0.88], [0.85, 0.84, 0.80, 0.90, 0.91], ] angles = np.linspace(0, 2 * np.pi, len(labels), endpoint=False).tolist() angles += angles[:1] fig, ax = plt.subplots(figsize=(6, 6), subplot_kw=dict(polar=True)) for model, score in zip(models, scores): score += score[:1] ax.plot(angles, score, label=model) ax.fill(angles, score, alpha=0.25) ax.set_theta_offset(np.pi / 2) ax.set_theta_direction(-1) ax.set_thetagrids(np.degrees(angles[:-1]), labels) ax.set_ylim(0, 1) ax.legend(loc='upper right', bbox_to_anchor=(1.3, 1.1)) plt.title("GRACE Radar Evaluation") return fig # --- 本地加载两个大模型 --- device = "cuda" if torch.cuda.is_available() else "cpu" print(f"Using device: {device}") MODEL_A_PATH = "gpt2" MODEL_B_PATH = "distilgpt2" print("Loading Model A...") tokenizer_a = AutoTokenizer.from_pretrained(MODEL_A_PATH) model_a = AutoModelForCausalLM.from_pretrained(MODEL_A_PATH).to(device) print("Model A loaded.") print("Loading Model B...") tokenizer_b = AutoTokenizer.from_pretrained(MODEL_B_PATH) model_b = AutoModelForCausalLM.from_pretrained(MODEL_B_PATH).to(device) print("Model B loaded.") def model_a_infer(input_text: str) -> str: inputs = tokenizer_a(input_text, return_tensors="pt").to(device) with torch.no_grad(): outputs = model_a.generate( **inputs, max_new_tokens=64, do_sample=True, temperature=0.8, pad_token_id=tokenizer_a.eos_token_id, ) input_len = len(inputs['input_ids'][0]) generated_only = tokenizer_a.decode(outputs[0][input_len:], skip_special_tokens=True) return generated_only.strip() def model_b_infer(input_text: str) -> str: inputs = tokenizer_b(input_text, return_tensors="pt").to(device) with torch.no_grad(): outputs = model_b.generate( **inputs, max_new_tokens=64, do_sample=True, temperature=0.8, pad_token_id=tokenizer_b.eos_token_id, ) input_len = len(inputs['input_ids'][0]) generated_only = tokenizer_b.decode(outputs[0][input_len:], skip_special_tokens=True) return generated_only.strip() demo = gr.Blocks(css=custom_css) with demo: gr.HTML(TITLE) gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text") with gr.Tabs(elem_classes="tab-buttons") as tabs: with gr.TabItem("🏅 LLM Benchmark", elem_id="llm-benchmark-tab-table", id=0): leaderboard = init_leaderboard(LEADERBOARD_DF) radar_plot = gr.Plot(value=draw_grace_radar(), label="GRACE Radar Evaluation") gr.Markdown("本图展示了两个模型在 GRACE 五大任务维度下的性能对比。") with gr.TabItem("🧪 Arena", elem_id="arena-tab-table", id=4): gr.Markdown("## 🔁 Arena: 模型同台竞技") arena_input = gr.Textbox(label="输入文本 (适用于所有模型)", lines=3) arena_output_a = gr.Textbox(label="模型 A 输出 (GPT2)", lines=6) arena_output_b = gr.Textbox(label="模型 B 输出 (DistilGPT2)", lines=6) arena_button = gr.Button("运行 Arena 对比") def run_arena(text): if not text.strip(): return "请输入内容", "请输入内容" return model_a_infer(text), model_b_infer(text) arena_button.click(run_arena, inputs=arena_input, outputs=[arena_output_a, arena_output_b]) with gr.TabItem("📝 About", elem_id="about-tab-table", id=2): gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text") gr.Markdown(REPORT_TEXT, elem_classes="markdown-text") with gr.TabItem("🚀 Submit here!", elem_id="submit-tab-table", id=3): with gr.Column(): gr.Markdown(EVALUATION_QUEUE_TEXT, elem_classes="markdown-text") with gr.Accordion( f"✅ Finished Evaluations ({len(finished_eval_queue_df)})", open=False, ): gr.Dataframe( value=finished_eval_queue_df, headers=EVAL_COLS, datatype=EVAL_TYPES, row_count=5, ) with gr.Accordion( f"🔄 Running Evaluation Queue ({len(running_eval_queue_df)})", open=False, ): gr.Dataframe( value=running_eval_queue_df, headers=EVAL_COLS, datatype=EVAL_TYPES, row_count=5, ) with gr.Accordion( f"⏳ Pending Evaluation Queue ({len(pending_eval_queue_df)})", open=False, ): gr.Dataframe( value=pending_eval_queue_df, headers=EVAL_COLS, datatype=EVAL_TYPES, row_count=5, ) with gr.Row(): with gr.Column(): model_name_textbox = gr.Textbox(label="Model name") revision_name_textbox = gr.Textbox(label="Revision commit", placeholder="main") model_type = gr.Dropdown( choices=[t.to_str(" : ") for t in ModelType if t != ModelType.Unknown], label="Model type", multiselect=False, value=None, interactive=True, ) with gr.Column(): precision = gr.Dropdown( choices=[i.value.name for i in Precision if i != Precision.Unknown], label="Precision", multiselect=False, value="float16", interactive=True, ) weight_type = gr.Dropdown( choices=[i.value.name for i in WeightType], label="Weights type", multiselect=False, value="Original", interactive=True, ) base_model_name_textbox = gr.Textbox(label="Base model (for delta or adapter weights)") submit_button = gr.Button("Submit Eval") submission_result = gr.Markdown() submit_button.click( add_new_eval, [ model_name_textbox, base_model_name_textbox, revision_name_textbox, precision, weight_type, model_type, ], submission_result, ) with gr.Row(): with gr.Accordion("📙 Citation", open=False): citation_button = gr.Textbox( value=CITATION_BUTTON_TEXT, label=CITATION_BUTTON_LABEL, lines=20, elem_id="citation-button", show_copy_button=True, ) scheduler = BackgroundScheduler() scheduler.add_job(restart_space, "interval", seconds=1800) scheduler.start() demo.queue(default_concurrency_limit=40).launch()