import os os.system("pip install gradio==4.31.5 pydantic==2.7.1 gradio_modal==0.0.3") import gradio as gr import pandas as pd import re from apscheduler.schedulers.background import BackgroundScheduler from huggingface_hub import snapshot_download from gradio_space_ci import enable_space_ci from src.display.about import ( CITATION_BUTTON_LABEL, CITATION_BUTTON_TEXT, EVALUATION_QUEUE_TEXT, INTRODUCTION_TEXT, LLM_BENCHMARKS_TEXT, FAQ_TEXT, TITLE, ) from src.display.css_html_js import custom_css from src.display.utils import ( BENCHMARK_COLS, COLS, EVAL_COLS, EVAL_TYPES, NUMERIC_INTERVALS, NUMERIC_MODELSIZE, TYPES, AutoEvalColumn, GroupDtype, ModelType, fields, WeightType, Precision, ComputeDtype, WeightDtype, QuantType ) from src.envs import API, EVAL_REQUESTS_PATH, DYNAMIC_INFO_REPO, DYNAMIC_INFO_FILE_PATH, DYNAMIC_INFO_PATH, EVAL_RESULTS_PATH, H4_TOKEN, IS_PUBLIC, QUEUE_REPO, REPO_ID, RESULTS_REPO, REPO, GIT_REQUESTS_PATH, GIT_STATUS_PATH, GIT_RESULTS_PATH from src.populate import get_evaluation_queue_df, get_leaderboard_df from src.submission.submit import add_new_eval from src.scripts.update_all_request_files import update_dynamic_files from src.tools.collections import update_collections from src.tools.plots import ( create_metric_plot_obj, create_plot_df, create_scores_df, ) from gradio_modal import Modal import plotly.graph_objects as go selected_indices = [] selected_values = {} # Start ephemeral Spaces on PRs (see config in README.md) #enable_space_ci() precision_to_dtype = { "2bit": ["int2"], "3bit": ["int3"], "4bit": ["int4", "nf4", "fp4"], "?": ["?"] } current_weightDtype = ["All", "int2", "int3", "int4", "nf4", "fp4", "?"] # Global variable to store the selected dtypes selected_dtypes = ["All"] init_select = False def quant_update_Weight_Dtype(selected_precisions): global current_weightDtype if '✖ None' in selected_precisions: if not any(dtype in ['float16', 'bfloat16', 'float32'] for dtype in current_weightDtype): current_weightDtype += ['float16', 'bfloat16', 'float32'] else: if any(dtype in ['float16', 'bfloat16', 'float32'] for dtype in current_weightDtype): current_weightDtype = [dtype for dtype in current_weightDtype if dtype not in ['float16', 'bfloat16', 'float32']] return gr.Dropdown(choices=current_weightDtype, value="All") def update_Weight_Dtype(selected_precisions): global selected_dtypes global current_weightDtype global init_select init_select = True if not selected_precisions: # If no precision is selected, return "All" selected_dtypes = ["All"] return gr.Dropdown(choices=["All"], value="All") selected_dtypes_set = set() for precision in selected_precisions: if precision in precision_to_dtype: selected_dtypes_set.update(precision_to_dtype[precision]) # Convert set to sorted list to maintain order selected_dtypes = sorted(selected_dtypes_set) if any(dtype in ['float16', 'bfloat16', 'float32'] for dtype in current_weightDtype) and not any(dtype in ['float16', 'bfloat16', 'float32'] for dtype in selected_dtypes): selected_dtypes += ['float16', 'bfloat16', 'float32'] # Add "All" to the beginning of the list for display purposes display_choices = ["All"] + selected_dtypes current_weightDtype = display_choices return gr.Dropdown(choices=display_choices, value="All") def restart_space(): API.restart_space(repo_id=REPO_ID, token=H4_TOKEN) def init_space(full_init: bool = True): if full_init: try: branch = REPO.active_branch.name REPO.remotes.origin.pull(branch) except Exception as e: print(str(e)) restart_space() try: print(DYNAMIC_INFO_PATH) snapshot_download( repo_id=DYNAMIC_INFO_REPO, local_dir=DYNAMIC_INFO_PATH, repo_type="dataset", tqdm_class=None, etag_timeout=30 ) except Exception: restart_space() raw_data, original_df = get_leaderboard_df( results_path=GIT_RESULTS_PATH, requests_path=GIT_STATUS_PATH, dynamic_path=DYNAMIC_INFO_FILE_PATH, cols=COLS, benchmark_cols=BENCHMARK_COLS ) # update_collections(original_df.copy()) leaderboard_df = original_df.copy() plot_df = create_plot_df(create_scores_df(raw_data)) ( finished_eval_queue_df, running_eval_queue_df, pending_eval_queue_df, ) = get_evaluation_queue_df(GIT_STATUS_PATH, EVAL_COLS) return leaderboard_df, original_df, plot_df, finished_eval_queue_df, running_eval_queue_df, pending_eval_queue_df leaderboard_df, original_df, plot_df, finished_eval_queue_df, running_eval_queue_df, pending_eval_queue_df = init_space() def str_to_bool(value): if str(value).lower() == "true": return True elif str(value).lower() == "false": return False else: return False # Searching and filtering def update_table( hidden_df: pd.DataFrame, columns: list, type_query: list, precision_query: str, size_query: list, params_query: list, hide_models: list, query: str, compute_dtype: str, weight_dtype: str, double_quant: str, group_dtype: str ): global init_select global current_weightDtype if selected_dtypes == ['All']: weight_dtype = current_weightDtype elif weight_dtype == ['All'] or weight_dtype == 'All' or init_select: weight_dtype = selected_dtypes init_select = False else: weight_dtype = [weight_dtype] if compute_dtype == 'All': compute_dtype = ['bfloat16', 'float16', 'int8', 'float32'] else: compute_dtype = [compute_dtype] if group_dtype == 'All': group_dtype = [-1, 1024, 256, 128, 64, 32] else: try: group_dtype = [int(group_dtype)] except ValueError: group_dtype = [-1] double_quant = [str_to_bool(double_quant)] filtered_df = filter_models(df=hidden_df, type_query=type_query, size_query=size_query, precision_query=precision_query, hide_models=hide_models, compute_dtype=compute_dtype, weight_dtype=weight_dtype, double_quant=double_quant, group_dtype=group_dtype, params_query=params_query) filtered_df = filter_queries(query, filtered_df) df = select_columns(filtered_df, columns) return df def load_query(request: gr.Request): # triggered only once at startup => read query parameter if it exists query = request.query_params.get("query") or "" return query, query # return one for the "search_bar", one for a hidden component that triggers a reload only if value has changed def search_table(df: pd.DataFrame, query: str) -> pd.DataFrame: return df[(df[AutoEvalColumn.dummy.name].str.contains(query, case=False))] def select_columns(df: pd.DataFrame, columns: list) -> pd.DataFrame: always_here_cols = [c.name for c in fields(AutoEvalColumn) if c.never_hidden] dummy_col = [AutoEvalColumn.dummy.name] # We use COLS to maintain sorting filtered_df = df[ always_here_cols + [c for c in COLS if c in df.columns and c in columns] + dummy_col ] return filtered_df def filter_queries(query: str, filtered_df: pd.DataFrame): """Added by Abishek""" final_df = [] if query != "": queries = [q.strip() for q in query.split(";")] for _q in queries: _q = _q.strip() if _q != "": temp_filtered_df = search_table(filtered_df, _q) if len(temp_filtered_df) > 0: final_df.append(temp_filtered_df) if len(final_df) > 0: filtered_df = pd.concat(final_df) filtered_df = filtered_df.drop_duplicates( subset=[AutoEvalColumn.model.name, AutoEvalColumn.precision.name, AutoEvalColumn.revision.name] ) return filtered_df def filter_models( df: pd.DataFrame, type_query: list, size_query: list, params_query:list, precision_query: list, hide_models: list, compute_dtype: list, weight_dtype: list, double_quant: list, group_dtype: list, ) -> pd.DataFrame: # Show all models if "Private or deleted" in hide_models: filtered_df = df[df[AutoEvalColumn.still_on_hub.name] == True] else: filtered_df = df if "Contains a merge/moerge" in hide_models: filtered_df = filtered_df[filtered_df[AutoEvalColumn.merged.name] == False] if "MoE" in hide_models: filtered_df = filtered_df[filtered_df[AutoEvalColumn.moe.name] == False] if "Flagged" in hide_models: filtered_df = filtered_df[filtered_df[AutoEvalColumn.flagged.name] == False] type_emoji = [t[0] for t in type_query] filtered_df = filtered_df.loc[df[AutoEvalColumn.model_type_symbol.name].isin(type_emoji)] filtered_df = filtered_df.loc[df[AutoEvalColumn.precision.name].isin(precision_query + ["None"])] filtered_df = filtered_df.loc[df[AutoEvalColumn.weight_dtype.name].isin(weight_dtype)] filtered_df = filtered_df.loc[df[AutoEvalColumn.compute_dtype.name].isin(compute_dtype)] filtered_df = filtered_df.loc[df[AutoEvalColumn.double_quant.name].isin(double_quant)] filtered_df = filtered_df.loc[df[AutoEvalColumn.group_size.name].isin(group_dtype)] numeric_interval = pd.IntervalIndex(sorted([NUMERIC_INTERVALS[s] for s in size_query])) params_column = pd.to_numeric(df[AutoEvalColumn.params.name], errors="coerce") mask = params_column.apply(lambda x: any(numeric_interval.contains(x))) filtered_df = filtered_df.loc[mask] numeric_interval_params = pd.IntervalIndex(sorted([NUMERIC_MODELSIZE[s] for s in params_query])) params_column_params = pd.to_numeric(df[AutoEvalColumn.model_size.name], errors="coerce") mask_params = params_column_params.apply(lambda x: any(numeric_interval_params.contains(x))) filtered_df = filtered_df.loc[mask_params] return filtered_df def select(df, data: gr.SelectData): global selected_indices global selected_values selected_index = data.index[0] if selected_index in selected_indices: selected_indices.remove(selected_index) value = df.iloc[selected_index].iloc[1] pattern = r']+>([^<]+)' match = re.search(pattern, value) if match: text_content = match.group(1) if text_content in selected_values: del selected_values[text_content] else: selected_indices.append(selected_index) value = df.iloc[selected_index].iloc[1] pattern = r']+>([^<]+)' match = re.search(pattern, value) if match: text_content = match.group(1) selected_values[text_content] = value print('selected_values', selected_values, selected_indices) return gr.CheckboxGroup(list(selected_values.keys()), value=list(selected_values.keys())) def init_comparison_data(): global selected_values return gr.CheckboxGroup(list(selected_values.keys()), value=list(selected_values.keys())) def generate_spider_chart(df, selected_keys): global selected_values print('generate_spider_chart', selected_values, selected_keys) current_selected_values = [selected_values[key] for key in selected_keys if key in selected_values] selected_rows = df[df.iloc[:, 1].isin(current_selected_values)] fig = go.Figure() for _, row in selected_rows.iterrows(): fig.add_trace(go.Scatterpolar( r=[row['Average ⬆️'], row['ARC-c'], row['ARC-e'], row['Boolq'], row['HellaSwag'], row['Lambada'], row['MMLU'], row['Openbookqa'], row['Piqa'], row['Truthfulqa'], row['Winogrande']], theta=['Average ⬆️', 'ARC-c', 'ARC-e', 'Boolq', 'HellaSwag', 'Lambada', 'MMLU', 'Openbookqa', 'Piqa', 'Truthfulqa', 'Winogrande'], fill='toself', name=str(row['Model']) )) fig.update_layout( polar=dict( radialaxis=dict( visible=True, )), showlegend=True ) return fig leaderboard_df = filter_models( df=leaderboard_df, type_query=[t.to_str(" : ") for t in QuantType], size_query=list(NUMERIC_INTERVALS.keys()), params_query=list(NUMERIC_MODELSIZE.keys()), precision_query=[i.value.name for i in Precision], hide_models=["Private or deleted", "Contains a merge/moerge", "Flagged"], # Deleted, merges, flagged, MoEs, compute_dtype=[i.value.name for i in ComputeDtype], weight_dtype=[i.value.name for i in WeightDtype], double_quant=[True, False], group_dtype=[-1, 1024, 256, 128, 64, 32] ) 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): with gr.Row(): with gr.Column(): with gr.Row(): search_bar = gr.Textbox( placeholder=" 🔍 Search for your model (separate multiple queries with `;`) and press ENTER...", show_label=False, elem_id="search-bar", ) with gr.Row(): shown_columns = gr.CheckboxGroup( choices=[ c.name for c in fields(AutoEvalColumn) if not c.hidden and not c.never_hidden and not c.dummy ], value=[ c.name for c in fields(AutoEvalColumn) if c.displayed_by_default and not c.hidden and not c.never_hidden ], label="Select columns to show", elem_id="column-select", interactive=True, ) with gr.Row(): filter_columns_parameters = gr.CheckboxGroup( label="Model parameters (in billions of parameters)", choices=list(NUMERIC_INTERVALS.keys()), value=list(NUMERIC_INTERVALS.keys()), interactive=True, elem_id="filter-columns-size", ) with gr.Row(): filter_columns_size = gr.CheckboxGroup( label="Model sizes (GB, int4)", choices=list(NUMERIC_MODELSIZE.keys()), value=list(NUMERIC_MODELSIZE.keys()), interactive=True, elem_id="filter-columns-size", ) with gr.Column(min_width=320): #with gr.Box(elem_id="box-filter"): filter_columns_type = gr.CheckboxGroup( label="Quantization types", choices=[t.to_str() for t in QuantType], value=[t.to_str() for t in QuantType if t != QuantType.QuantType_None], interactive=True, elem_id="filter-columns-type", ) filter_columns_precision = gr.CheckboxGroup( label="Weight precision", choices=[i.value.name for i in Precision], value=[i.value.name for i in Precision], interactive=True, elem_id="filter-columns-precision", ) with gr.Group() as config: # gr.HTML("""

Quantization config

""") gr.HTML("""

Quantization config

""") with gr.Row(): filter_columns_computeDtype = gr.Dropdown(choices=[i.value.name for i in ComputeDtype], label="Compute Dtype", multiselect=False, value="All", interactive=True,) filter_columns_weightDtype = gr.Dropdown(choices=[i.value.name for i in WeightDtype], label="Weight Dtype", multiselect=False, value="All", interactive=True,) filter_columns_doubleQuant = gr.Dropdown(choices=["True", "False"], label="Double Quant", multiselect=False, value="False", interactive=True) filter_columns_groupDtype = gr.Dropdown(choices=[i.value.name for i in GroupDtype], label="Group Size", multiselect=False, value="All", interactive=True,) with gr.Row(): with gr.Column(): model_comparison = gr.CheckboxGroup(label="Accuracy Comparison on Selected Models", choices=list(selected_values.keys()), value=list(selected_values.keys()), interactive=True, elem_id="model_comparison") with gr.Column(): spider_btn = gr.Button("Compare") leaderboard_table = gr.components.Dataframe( value=leaderboard_df[ [c.name for c in fields(AutoEvalColumn) if c.never_hidden] + shown_columns.value + [AutoEvalColumn.dummy.name] ], headers=[c.name for c in fields(AutoEvalColumn) if c.never_hidden] + shown_columns.value, datatype=TYPES, elem_id="leaderboard-table", interactive=False, visible=True, #column_widths=["2%", "33%"] ) with Modal(visible=False) as modal: map = gr.Plot() leaderboard_table.select(select, leaderboard_table, model_comparison) spider_btn.click(generate_spider_chart, [leaderboard_table, model_comparison], map) spider_btn.click(lambda: Modal(visible=True), None, modal) demo.load(init_comparison_data, None, model_comparison) # Dummy leaderboard for handling the case when the user uses backspace key hidden_leaderboard_table_for_search = gr.components.Dataframe( value=original_df[COLS], headers=COLS, datatype=TYPES, visible=False, ) hide_models = gr.Textbox( placeholder="", show_label=False, elem_id="search-bar", value="", visible=False, ) search_bar.submit( update_table, [ hidden_leaderboard_table_for_search, shown_columns, filter_columns_type, filter_columns_precision, filter_columns_parameters, filter_columns_size, hide_models, search_bar, filter_columns_computeDtype, filter_columns_weightDtype, filter_columns_doubleQuant, filter_columns_groupDtype ], leaderboard_table, ) """ # Define a hidden component that will trigger a reload only if a query parameter has been set hidden_search_bar = gr.Textbox(value="", visible=False) hidden_search_bar.change( update_table, [ hidden_leaderboard_table_for_search, shown_columns, filter_columns_type, filter_columns_precision, filter_columns_size, hide_models, search_bar, ], leaderboard_table, ) # Check query parameter once at startup and update search bar + hidden component demo.load(load_query, inputs=[], outputs=[search_bar, hidden_search_bar]) """ filter_columns_precision.change( update_Weight_Dtype, [filter_columns_precision], [filter_columns_weightDtype] ) filter_columns_type.change( quant_update_Weight_Dtype, [filter_columns_type], [filter_columns_weightDtype] ) for selector in [shown_columns, filter_columns_type, filter_columns_precision, filter_columns_size, filter_columns_parameters, hide_models, filter_columns_computeDtype, filter_columns_weightDtype, filter_columns_doubleQuant, filter_columns_groupDtype]: selector.change( update_table, [ hidden_leaderboard_table_for_search, shown_columns, filter_columns_type, filter_columns_precision, filter_columns_parameters, filter_columns_size, hide_models, search_bar, filter_columns_computeDtype, filter_columns_weightDtype, filter_columns_doubleQuant, filter_columns_groupDtype ], leaderboard_table, queue=True, ) with gr.TabItem("📈 Metrics through time", elem_id="llm-benchmark-tab-table", id=2): with gr.Row(): with gr.Column(): chart = create_metric_plot_obj( plot_df, [AutoEvalColumn.average.name], title="Average of Top Scores and Human Baseline Over Time (from last update)", ) gr.Plot(value=chart, min_width=500) with gr.Column(): chart = create_metric_plot_obj( plot_df, BENCHMARK_COLS, title="Top Scores and Human Baseline Over Time (from last update)", ) gr.Plot(value=chart, min_width=500) with gr.TabItem("📝 About", elem_id="llm-benchmark-tab-table", id=3): gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text") with gr.TabItem("❗FAQ", elem_id="llm-benchmark-tab-table", id=4): gr.Markdown(FAQ_TEXT, elem_classes="markdown-text") with gr.TabItem("🚀 Submit ", elem_id="llm-benchmark-tab-table", id=5): with gr.Column(): with gr.Row(): gr.Markdown(EVALUATION_QUEUE_TEXT, elem_classes="markdown-text") with gr.Row(): gr.Markdown("# ✉️✨ Submit your model here!", elem_classes="markdown-text") with gr.Row(): with gr.Column(): model_name_textbox = gr.Textbox(label="Model name") revision_name_textbox = gr.Textbox(label="Revision commit", placeholder="main") private = gr.Checkbox(False, label="Private", visible=not IS_PUBLIC) with gr.Column(): """ precision = gr.Dropdown( choices=[i.value.name for i in Precision if i != Precision.Unknown], label="Precision", multiselect=False, value="4bit", interactive=True, ) weight_type = gr.Dropdown( choices=[i.value.name for i in WeightDtype], label="Weights dtype", multiselect=False, value="int4", interactive=True, ) """ base_model_name_textbox = gr.Textbox(label="Base model (for delta or adapter weights)", visible=not IS_PUBLIC) compute_type = gr.Dropdown( choices=[i.value.name for i in ComputeDtype if i.value.name != "All"], label="Compute dtype", multiselect=False, value="float16", interactive=True, ) submit_button = gr.Button("Submit Eval") submission_result = gr.Markdown() submit_button.click( add_new_eval, [ model_name_textbox, revision_name_textbox, private, compute_type, ], submission_result, ) with gr.Column(): with gr.Accordion( f"✅ Finished Evaluations ({len(finished_eval_queue_df)})", open=False, ): with gr.Row(): finished_eval_table = gr.components.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, ): with gr.Row(): running_eval_table = gr.components.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, ): with gr.Row(): pending_eval_table = gr.components.Dataframe( value=pending_eval_queue_df, headers=EVAL_COLS, datatype=EVAL_TYPES, row_count=5, ) 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", hours=3) # restarted every 3h scheduler.add_job(update_dynamic_files, "interval", hours=12) # launched every 2 hour scheduler.start() demo.queue(default_concurrency_limit=40).launch() # demo.queue(concurrency_count=40).launch()