tokenrace 1.0.0__py3-none-any.whl

tokenrace/__init__.py

@@ -0,0 +1,3 @@
+"""TokenRace: benchmark della velocità di server LLM locali compatibili OpenAI."""
+
+__version__ = "1.0.0"

tokenrace/_data/.streamlit/config.toml

@@ -0,0 +1,11 @@
+# Nessuna statistica d'uso inviata a Streamlit: l'app resta tutta locale.
+[browser]
+gatherUsageStats = false
+
+[client]
+# Nasconde il pulsante "Deploy" e lascia solo il menu essenziale.
+toolbarMode = "minimal"
+
+[theme]
+primaryColor = "#FF5A1F"
+baseRadius = "0.6rem"

tokenrace/_data/app.py

@@ -0,0 +1,414 @@
+"""TokenRace - Local LLM Speed Test.
+
+Interfaccia Streamlit. La logica di misura sta in tokenrace/bench.py,
+i valori modificabili in config/, i testi tradotti in locales/.
+
+Avvio: streamlit run app.py
+Licenza: MIT
+"""
+
+import base64
+import json
+import os
+import time
+import uuid
+from datetime import datetime
+
+import altair as alt  # incluso con Streamlit
+import pandas as pd
+import streamlit as st
+
+from tokenrace import bench
+from tokenrace.config import ROOT, load_defaults, load_prompts, load_servers
+from tokenrace.i18n import Translator, language_info
+
+# Metriche riassunte nella tabella di riepilogo (chiavi interne del CSV)
+SUMMARY_METRICS = ["ttft", "tok_s_gen", "tok_s_eff", "total", "prompt_tok_s"]
+# Colonne della tabella delle ripetizioni del test in corso
+RUN_COLUMNS = ["run", "cold"] + SUMMARY_METRICS + ["tok_out"]
+# Spiegazione (sezione [help]) di ciascuna colonna; le altre usano [colhelp]
+COLUMN_HELP = {"ttft": "help.ttft", "tok_s_gen": "help.gen", "tok_s_eff": "help.eff",
+               "total": "help.total", "prompt_tok_s": "help.prompt_speed",
+               "ram_before": "help.ram_before", "ram_peak": "help.ram_peak"}
+PRESET_IDS = ["short", "long", "cache"]
+
+
+# ---------------------------------------------------------------------------
+# Utilità
+# ---------------------------------------------------------------------------
+
+def fmt(v, dec=2):
+    return "—" if v is None else f"{v:.{dec}f}"
+
+
+def reasoning_md(t, text: str) -> str:
+    if not text:
+        return ""
+    return f"> 💭 *{t('label.reasoning')}:* " + text[-3000:].replace("\n", " ") + "\n\n"
+
+
+def status_badge(text: str, state: str) -> str:
+    """Etichetta dello stato della ripetizione (state: running, done)."""
+    return (f'<div class="tr-status-wrap"><span class="tr-status tr-status-{state}">'
+            f'<span class="tr-dot"></span>{text}</span></div>')
+
+
+def error_text(t, e: bench.ServerError) -> str:
+    return t(f"err.{e.code}", **e.params)
+
+
+def pick_language(languages: list[str], default: str) -> str:
+    """Lingua indicata nell'URL (?lang=), altrimenti quella predefinita."""
+    qp = st.query_params.get("lang")
+    env = os.environ.get("TOKENRACE_LANG")  # impostata dal comando `tokenrace --lang`
+    return qp if qp in languages else env if env in languages else default
+
+
+def show_table(t, df: pd.DataFrame):
+    """Mostra un DataFrame con colonne interne tradotte e spiegazioni nelle intestazioni."""
+    names, config = {}, {}
+    for c in df.columns:
+        base, _, stat = c.partition("·")
+        names[c] = t(f"col.{base}") + (f" · {t('stat.' + stat)}" if stat else "")
+        help_key = COLUMN_HELP.get(base, f"colhelp.{base}")
+        if base in ("server", "model"):
+            config[names[c]] = st.column_config.TextColumn(width="medium")
+        elif t.has(help_key):
+            config[names[c]] = st.column_config.Column(help=t(help_key))
+    st.dataframe(df.rename(columns=names), hide_index=True, width="stretch", column_config=config)
+
+
+def make_row(ctx, rip: int, m: dict) -> dict:
+    p = ctx["params"]
+    return {
+        "ts": datetime.now().strftime("%Y-%m-%d %H:%M:%S"),
+        "server": ctx["server"].name,
+        "url": ctx["base_url"],
+        "model": ctx["model"],
+        "prompt_type": ctx["preset"],
+        "prompt_lang": ctx["lang"],
+        "run": rip,
+        "cold": rip == 1,
+        "temperature": p["temperature"],
+        "seed": p["seed"],
+        "max_tokens": p["max_tokens"],
+        "repetitions": p["repetitions"],
+        "no_reasoning": bool(p["no_reasoning"]),
+        "no_reasoning_fields": json.dumps(p["no_reasoning"]) if p["no_reasoning"] else "",
+        **{k: m[k] for k in ("ttft", "tok_s_gen", "tok_s_eff", "total", "tok_in", "tok_out",
+                             "tok_reason", "tok_answer", "prompt_tok_s", "ram_before",
+                             "ram_peak", "estimated", "split_estimated")},
+    }
+
+
+# ---------------------------------------------------------------------------
+# Sezioni della pagina
+# ---------------------------------------------------------------------------
+
+# Mascotte nell'intestazione, incorporata nella pagina come data URI
+LOGO = "data:image/svg+xml;base64," + base64.b64encode(
+    (ROOT / "assets" / "logo" / "tokenrace-icon.svg").read_bytes()).decode()
+
+
+def render_header(t, n_servers: int, languages: list[str], lang: str) -> str:
+    left, right = st.columns([6, 1], vertical_alignment="center")
+    left.markdown(
+        f"""<div class="tr-hero"><h1><img src="{LOGO}" alt="">Token<span>Race</span></h1>
+        <p>{t('app.tagline')}</p>
+        <span class="tr-pill">{t('app.servers_pill', n=n_servers)}</span></div>""",
+        unsafe_allow_html=True)
+    info = {code: language_info(code) for code in languages}
+    # Il valore iniziale arriva da st.session_state["lang_choice"], impostato in main()
+    choice = right.segmented_control(
+        t("label.language"), languages, key="lang_choice",
+        format_func=lambda c: f"{info[c]['flag']} {c.upper()}",
+        label_visibility="collapsed", help=t("help.language"))
+    return choice or lang
+
+
+def render_server_form(t, servers, measure):
+    ss = st.session_state
+    by_id = {s.id: s for s in servers}
+    with st.container(border=True):
+        st.markdown(f"##### {t('section.server')}")
+        c1, c2, c5, c3, c4 = st.columns([1.4, 1.9, 0.9, 2, 1.6], vertical_alignment="bottom")
+        server = by_id[c1.selectbox(t("label.server"), list(by_id), key="server",
+                                    format_func=lambda i: by_id[i].name, help=t("help.server"))]
+        base_url = c2.text_input(t("label.url"), value=server.url, key=f"url_{server.id}",
+                                 help=t("help.url"))
+        api_key = c5.text_input(t("label.api_key"), type="password", key=f"key_{server.id}",
+                                help=t("help.api_key"))
+        error = None
+        if c4.button(t("label.refresh"), help=t("help.refresh"), width="stretch"):
+            try:
+                ss.models[base_url] = bench.fetch_models(base_url, api_key,
+                                                         measure["models_timeout_s"])
+                if not ss.models[base_url]:
+                    error = ("warning", t("msg.no_models"))
+            except bench.ServerError as e:
+                ss.models.pop(base_url, None)
+                error = ("error", error_text(t, e))
+        models = ss.models.get(base_url, [])
+        model = c3.selectbox(t("label.model"), models, key=f"model_{base_url}",
+                             placeholder=t("label.model_placeholder"),
+                             index=0 if models else None, help=t("help.model"))
+        note = t(f"server_notes.{server.id}")
+        start = f"**{t('label.start_cmd')}:** `{server.start}` · " if server.start else ""
+        st.caption(start + note)
+        if error:
+            getattr(st, error[0])(error[1])
+    return server, base_url, api_key, model
+
+
+def render_prompt_form(t, server, presets, defaults, model):
+    ss = st.session_state
+    d = defaults["params"]
+    with st.container(border=True):
+        st.markdown(f"##### {t('section.prompt')}")
+        r = st.columns([1.3, 1, 1, 1, 1, 1.2, 1], vertical_alignment="bottom")
+        # Una chiave per lingua: le etichette cambiano con la lingua, il valore scelto
+        # (short/long/cache) è tenuto a parte in ss.preset.
+        widget_key = f"preset_{t.lang}"
+
+        def on_preset():
+            ss.preset = ss[widget_key]
+            ss.prompt = presets[ss.preset]
+
+        r[0].selectbox(t("label.preset"), PRESET_IDS, index=PRESET_IDS.index(ss.preset),
+                       key=widget_key, format_func=lambda i: t(f"preset.{i}"),
+                       on_change=on_preset, help=t("help.preset"))
+        params = {
+            "temperature": r[1].number_input(t("label.temperature"), 0.0, 2.0, float(d["temperature"]),
+                                             0.1, key="temperature", help=t("help.temperature")),
+            "seed": int(r[2].number_input(t("label.seed"), value=int(d["seed"]), step=1, key="seed",
+                                          help=t("help.seed"))),
+            "max_tokens": int(r[3].number_input(t("label.max_tokens"), 1, 32768, int(d["max_tokens"]),
+                                                64, key="max_tokens", help=t("help.max_tokens"))),
+            "repetitions": int(r[4].number_input(t("label.repetitions"), 1, 50, int(d["repetitions"]),
+                                                 1, key="repetitions", help=t("help.repetitions"))),
+        }
+        start = r[5].button(t("label.start"), type="primary", disabled=not model,
+                            help=t("help.start"), width="stretch")
+        stop = r[6].button(t("label.stop"), help=t("help.stop"), width="stretch")
+
+        prompt = st.text_area(t("label.prompt"), key="prompt", height=90, help=t("help.prompt"))
+        n1, n2 = st.columns([3, 1.2], vertical_alignment="center")
+        n1.caption(f"**{t('preset.' + ss.preset)}**: {t('preset_descr.' + ss.preset)} · "
+                   + t("preset_descr.token_estimate",
+                       n=f"{bench.estimate_tokens(prompt):,}".replace(",", ".")))
+        fields = json.dumps(server.no_reasoning)
+        state = t("msg.nr_tested") if server.no_reasoning_tested else t("msg.nr_untested")
+        nr_on = n2.toggle(t("label.no_reasoning"), key="no_reasoning",
+                          help=t("help.no_reasoning") + "\n\n"
+                          + t("msg.nr_help_suffix", server=server.name, fields=fields) + f" ({state}).")
+        params["no_reasoning"] = server.no_reasoning if nr_on else None
+        if nr_on:
+            st.caption(t("msg.nr_sends", fields=fields, state=state))
+        if stop:
+            st.warning(t("msg.stopped"))
+    return prompt, params, start
+
+
+def run_test(t, ctx, prompt, run_id_label, measure, resp_box, status_box, live_box, reps_box):
+    """Esegue le ripetizioni aggiornando risposta e misure in tempo reale."""
+    ss = st.session_state
+    p = ctx["params"]
+    rows = []
+    for rip in range(1, p["repetitions"] + 1):
+        text = prompt
+        if ctx["preset"] in ("short", "long"):
+            text = f"[{run_id_label}: {uuid.uuid4()}]\n{prompt}"
+        last_ui = [0.0]
+        status = t("msg.run_status", i=rip, n=p["repetitions"]) + (t("msg.cold") if rip == 1 else "")
+        status_box.markdown(status_badge(f"{status} · {t('msg.in_progress')}", "running"),
+                            unsafe_allow_html=True)
+
+        def on_update(r_txt, c_txt, live, status=status):
+            now = time.perf_counter()
+            if now - last_ui[0] < measure["ui_refresh_s"]:
+                return
+            last_ui[0] = now
+            resp_box.markdown(reasoning_md(t, r_txt) + c_txt)
+            with live_box.container():
+                a, b = st.columns(2)
+                a.metric(t("metric.ttft"), f"{live['ttft']:.2f}", help=t("help.ttft"), border=True)
+                b.metric(t("metric.chunks"), live["chunks"], help=t("help.chunks"), border=True)
+                a.metric(t("metric.elapsed"), f"{live['elapsed']:.1f}", help=t("help.elapsed"),
+                         border=True)
+                b.metric(t("metric.ram_peak"), f"{live['ram_peak']:.1f}", help=t("help.ram_peak"),
+                         border=True)
+
+        try:
+            m = bench.run_once(ctx["base_url"], ctx["model"], text, p["temperature"], p["seed"],
+                               p["max_tokens"], on_update, p["no_reasoning"], ctx["api_key"],
+                               ram_interval=measure["ram_interval_s"],
+                               timeout=(measure["connect_timeout_s"], measure["read_timeout_s"]))
+        except bench.ServerError as e:
+            status_box.empty()
+            st.error(error_text(t, e))
+            break
+
+        resp_box.markdown(reasoning_md(t, m["reasoning"]) + m["content"])
+        status_box.markdown(status_badge(f"{status} · {t('msg.completed')}", "done"),
+                            unsafe_allow_html=True)
+        row = make_row(ctx, rip, m)
+        ss.history.append(row)
+        rows.append(row)
+
+        with live_box.container():
+            a, b, c = st.columns(3)
+            a.metric(t("metric.ttft"), fmt(m["ttft"]), help=t("help.ttft"), border=True)
+            b.metric(t("metric.gen"), fmt(m["tok_s_gen"], 1), help=t("help.gen"), border=True)
+            c.metric(t("metric.eff"), fmt(m["tok_s_eff"], 1), help=t("help.eff"), border=True)
+            a.metric(t("metric.total"), fmt(m["total"]), help=t("help.total"), border=True)
+            b.metric(t("metric.tokens_io"), f"{m['tok_in']}/{m['tok_out']}", help=t("help.tokens_io"),
+                     border=True)
+            c.metric(t("metric.prompt_speed"), fmt(m["prompt_tok_s"], 0), help=t("help.prompt_speed"),
+                     border=True)
+            a.metric(t("metric.split"), f"{m['tok_reason']}/{m['tok_answer']}", help=t("help.split"),
+                     border=True)
+            b.metric(t("metric.ram_before"), f"{m['ram_before']:.1f}", help=t("help.ram_before"),
+                     border=True)
+            c.metric(t("metric.ram_peak"), f"{m['ram_peak']:.1f}", help=t("help.ram_peak"), border=True)
+            st.caption(t("msg.ram_note"))
+            if m["estimated"]:
+                st.warning(t("msg.estimated"))
+            if p["no_reasoning"] and m["tok_reason"] > 0:
+                st.warning(t("msg.reasoning_still"))
+            if m["split_estimated"]:
+                st.caption(t("msg.split_estimated"))
+
+        with reps_box.container():
+            show_table(t, pd.DataFrame(rows)[RUN_COLUMNS].round(2))
+
+
+def render_results(t, primary_color: str):
+    ss = st.session_state
+    with st.container(border=True):
+        st.markdown(f"##### {t('section.results')}")
+        if not ss.history:
+            st.info(t("msg.no_results"))
+            return
+        df = pd.DataFrame(ss.history)
+        df["prompt_type"] = df["prompt_type"].map(lambda i: t(f"preset.{i}"))
+        keys = ["server", "model", "prompt_type"]
+        warm = df[~df["cold"]]
+        medians, mins, maxs = [], [], []
+        for m in SUMMARY_METRICS:
+            medians.append(warm.groupby(keys)[m].median().rename(f"{m}·median"))
+            mins.append(df.groupby(keys)[m].min().rename(f"{m}·min"))
+            maxs.append(df.groupby(keys)[m].max().rename(f"{m}·max"))
+        summary = pd.concat(medians + mins + maxs, axis=1).reset_index()
+
+        g1, g2 = st.columns([3, 2])
+        with g1:
+            st.markdown(t("label.summary"))
+            show_table(t, summary.round(2))
+        with g2:
+            st.markdown(t("label.chart"))
+            if warm.empty:
+                st.caption(t("msg.need_two"))
+            else:
+                # Una barra per server e modello; il tipo di prompt si aggiunge
+                # all'etichetta solo se nello storico ce n'è più di uno.
+                groups = keys if warm["prompt_type"].nunique() > 1 else keys[:2]
+                bars = warm.groupby(groups)["tok_s_eff"].median().rename("tok/s").reset_index()
+                axis = t("label.chart_axis")
+                bars[axis] = bars[groups].astype(str).agg(" · ".join, axis=1)
+                base = alt.Chart(bars).encode(
+                    x=alt.X("tok/s:Q", title="tok/s"),
+                    y=alt.Y(f"{axis}:N", sort="-x", title=None, axis=alt.Axis(labelLimit=0)),
+                    tooltip=[axis, alt.Tooltip("tok/s:Q", format=".1f")])
+                chart = base.mark_bar(color=primary_color, cornerRadiusEnd=4) + base.mark_text(
+                    align="left", dx=4, fontWeight="bold").encode(text=alt.Text("tok/s:Q", format=".1f"))
+                st.altair_chart(chart.properties(height=50 + 45 * len(bars)), width="stretch")
+
+        with st.expander(t("label.all_runs")):
+            show_table(t, df.round(3))
+
+        d1, d2, _ = st.columns([1, 1, 5])
+        csv = df.rename(columns=lambda c: t(f"col.{c}")).to_csv(index=False).encode("utf-8")
+        d1.download_button(t("label.download"), csv,
+                           file_name=f"tokenrace_{datetime.now():%Y%m%d_%H%M%S}.csv",
+                           mime="text/csv", help=t("help.download"), width="stretch")
+        if d2.button(t("label.clear"), help=t("help.clear"), width="stretch"):
+            ss.history = []
+            st.rerun()
+
+
+# ---------------------------------------------------------------------------
+# Pagina
+# ---------------------------------------------------------------------------
+
+def main():
+    st.set_page_config(page_title="TokenRace", page_icon=str(ROOT / "assets" / "logo" / "tokenrace-favicon.png"),
+                       layout="wide")
+    defaults = load_defaults()
+    ui, measure = defaults["ui"], defaults["measure"]
+    servers = load_servers()
+    css = (ROOT / "assets" / "style.css").read_text(encoding="utf-8")
+    st.markdown(f"<style>:root {{--tr-primary: {ui['primary_color']};}}\n{css}</style>",
+                unsafe_allow_html=True)
+
+    ss = st.session_state
+    ss.setdefault("history", [])
+    ss.setdefault("models", {})
+
+    # Lingua: scelta nell'intestazione e ricordata nell'URL (?lang=it).
+    # Se l'URL cambia da fuori (link, ricarica), ha la precedenza sulla scelta in memoria.
+    url_lang = pick_language(ui["languages"], ui["default_language"])
+    if url_lang != ss.get("applied_lang"):
+        ss.lang_choice = url_lang
+    lang = ss.get("lang_choice") or ss.get("applied_lang") or url_lang
+    t = Translator(lang, ui["fallback_language"])
+    render_header(t, len(servers), ui["languages"], lang)
+    st.query_params["lang"] = lang
+    ss.applied_lang = lang
+
+    all_prompts = load_prompts()
+    target = defaults["long_prompt"]["target_chars"]
+    prompt_texts = all_prompts.get(lang, all_prompts[ui["fallback_language"]])
+    presets = bench.build_prompts(prompt_texts, target)
+    ss.setdefault("preset", "short")
+    # Alla prima apertura, o se cambia la lingua e il prompt non è stato modificato,
+    # mostra il prompt pronto nella lingua scelta.
+    old_lang = ss.get("prompt_lang")
+    if "prompt" not in ss:
+        ss.prompt = presets[ss.preset]
+    elif old_lang and old_lang != lang and old_lang in all_prompts:
+        if ss.prompt == bench.build_prompts(all_prompts[old_lang], target)[ss.preset]:
+            ss.prompt = presets[ss.preset]
+    ss.prompt_lang = lang
+
+    server, base_url, api_key, model = render_server_form(t, servers, measure)
+    prompt, params, start = render_prompt_form(t, server, presets, defaults, model)
+
+    left, right = st.columns([3, 2])
+    with left.container(border=True):
+        st.markdown(f"##### {t('section.response')}")
+        resp_box = st.container(height=ui["response_height"], border=False).empty()
+    with right.container(border=True):
+        h1, h2 = st.columns([3, 2], vertical_alignment="center")
+        h1.markdown(f"##### {t('section.live')}")
+        status_box = h2.empty()
+        live_box = st.empty()
+        reps_box = st.empty()
+
+    if start and model:
+        ctx = {"server": server, "base_url": base_url, "api_key": api_key, "model": model,
+               "preset": ss.preset, "lang": lang, "params": params}
+        run_test(t, ctx, prompt, prompt_texts["run_id"], measure, resp_box, status_box, live_box,
+                 reps_box)
+    else:
+        resp_box.caption(t("msg.response_placeholder"))
+        live_box.caption(t("msg.live_placeholder"))
+
+    render_results(t, ui["primary_color"])
+    repo = ui["repo_url"]
+    st.markdown(f'<div class="tr-footer">TokenRace · {t("app.footer")} · '
+                f'<a href="{repo}" target="_blank">{repo.removeprefix("https://")}</a></div>',
+                unsafe_allow_html=True)
+
+
+if __name__ == "__main__":
+    main()

tokenrace/_data/assets/logo/tokenrace-favicon.svg

@@ -0,0 +1,14 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 64 64">
+<defs><pattern id="c2" width="10" height="10" patternUnits="userSpaceOnUse">
+<rect width="10" height="10" fill="#FFFFFF"/><rect width="5" height="5" fill="#1E1E1E"/><rect x="5" y="5" width="5" height="5" fill="#1E1E1E"/></pattern>
+<clipPath id="cc"><circle cx="31" cy="33" r="27"/></clipPath></defs>
+<circle cx="33" cy="33.5" r="30.5" fill="#FFFFFF"/>
+<circle cx="34" cy="34" r="27" fill="#B93A0C" stroke="#2A1A12" stroke-width="4"/>
+<circle cx="31" cy="33" r="27" fill="#FF5A1F"/>
+<g clip-path="url(#cc)"><path d="M0 16 Q31 6 64 16 L64 26 Q31 16 0 26 Z" fill="url(#c2)" stroke="#2A1A12" stroke-width="3"/></g>
+<circle cx="31" cy="33" r="27" fill="none" stroke="#2A1A12" stroke-width="4"/>
+<ellipse cx="24" cy="35" rx="5" ry="6.5" fill="#FFFFFF" stroke="#2A1A12" stroke-width="2.5"/>
+<ellipse cx="39" cy="35" rx="5" ry="6.5" fill="#FFFFFF" stroke="#2A1A12" stroke-width="2.5"/>
+<circle cx="26" cy="36" r="3" fill="#2A1A12"/><circle cx="41" cy="36" r="3" fill="#2A1A12"/>
+<path d="M22 46 Q31 55 41 45 Z" fill="#2A1A12" stroke="#2A1A12" stroke-width="2.5" stroke-linejoin="round"/>
+</svg>

tokenrace/_data/assets/logo/tokenrace-icon.svg

@@ -0,0 +1,47 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="-6 28 250 206">
+<defs>
+  <filter id="sticker" x="-15%" y="-15%" width="130%" height="130%">
+    <feMorphology in="SourceAlpha" operator="dilate" radius="6" result="d"/>
+    <feFlood flood-color="#FFFFFF"/><feComposite in2="d" operator="in" result="w"/>
+    <feMerge><feMergeNode in="w"/><feMergeNode in="SourceGraphic"/></feMerge>
+  </filter>
+  <clipPath id="edge"><circle cx="153" cy="133" r="80"/></clipPath>
+  <pattern id="chk" width="12" height="12" patternUnits="userSpaceOnUse">
+    <rect width="12" height="12" fill="#FFFFFF"/>
+    <rect width="6" height="6" fill="#1E1E1E"/><rect x="6" y="6" width="6" height="6" fill="#1E1E1E"/>
+  </pattern>
+  <radialGradient id="face" cx="0.38" cy="0.32" r="0.8">
+    <stop offset="0" stop-color="#FF8A4F"/><stop offset="0.55" stop-color="#FF5A1F"/><stop offset="1" stop-color="#E5470F"/>
+  </radialGradient>
+  <clipPath id="coin"><circle cx="140" cy="128" r="80"/></clipPath>
+  <clipPath id="mouth"><path d="M114 150 Q141 186 170 147 Z"/></clipPath>
+</defs><g filter="url(#sticker)"><g transform="rotate(4 140 140)">
+<g stroke-linecap="round" fill="none">
+  <path d="M4 150 H34" stroke="#FF5A1F" stroke-width="9"/>
+  <path d="M14 174 H52" stroke="#2A1A12" stroke-width="9"/>
+  <path d="M8 198 H58" stroke="#FFA27A" stroke-width="9"/>
+</g>
+<g stroke="#2A1A12" stroke-width="5" stroke-linejoin="round">
+  <path d="M70 86 C50 76 34 84 14 72 L20 94 C38 100 52 96 70 102 Z" fill="url(#chk)"/>
+  <path d="M70 100 C54 108 40 118 22 116 L32 132 C48 130 58 120 72 112 Z" fill="url(#chk)"/>
+</g>
+<circle cx="153" cy="133" r="80" fill="#B93A0C" stroke="#2A1A12" stroke-width="7"/>
+<g clip-path="url(#edge)" stroke="#2A1A12" stroke-width="3.5" stroke-opacity="0.8"><path d="M164.8 66.0 L167.6 50.3"/><path d="M174.0 68.3 L179.0 53.1"/><path d="M182.8 71.9 L189.8 57.5"/><path d="M191.0 76.6 L200.0 63.4"/><path d="M198.5 82.5 L209.2 70.6"/><path d="M205.1 89.3 L217.3 79.0"/><path d="M210.7 97.0 L224.2 88.5"/><path d="M215.1 105.3 L229.7 98.8"/><path d="M218.4 114.3 L233.7 109.8"/><path d="M220.3 123.5 L236.2 121.3"/><path d="M221.0 133.0 L237.0 133.0"/><path d="M220.3 142.5 L236.2 144.7"/><path d="M218.4 151.7 L233.7 156.2"/><path d="M215.1 160.7 L229.7 167.2"/><path d="M210.7 169.0 L224.2 177.5"/><path d="M205.1 176.7 L217.3 187.0"/><path d="M198.5 183.5 L209.2 195.4"/><path d="M191.0 189.4 L200.0 202.6"/><path d="M182.8 194.1 L189.8 208.5"/><path d="M174.0 197.7 L179.0 212.9"/><path d="M164.8 200.0 L167.6 215.7"/><path d="M155.4 201.0 L155.9 216.9"/><path d="M145.9 200.6 L144.2 216.5"/><path d="M136.5 199.0 L132.7 214.5"/><path d="M127.5 196.0 L121.5 210.9"/><path d="M119.0 191.9 L111.0 205.7"/><path d="M111.1 186.6 L101.3 199.2"/><path d="M104.1 180.2 L92.6 191.4"/><path d="M98.0 173.0 L85.0 182.4"/><path d="M93.0 164.9 L78.8 172.4"/><path d="M89.1 156.3 L74.1 161.7"/><path d="M86.5 147.1 L70.8 150.5"/><path d="M85.2 137.7 L69.2 138.9"/><path d="M85.2 128.3 L69.2 127.1"/></g>
+<circle cx="140" cy="128" r="80" fill="url(#face)"/>
+<circle cx="140" cy="128" r="68" fill="none" stroke="#FFFFFF" stroke-opacity="0.28" stroke-width="4"/>
+<path d="M82 118 A60 60 0 0 1 100 84" fill="none" stroke="#FFFFFF" stroke-opacity="0.55" stroke-width="7" stroke-linecap="round"/><g clip-path="url(#coin)">
+  <path d="M50 86 Q140 60 230 86 L230 106 Q140 80 50 106 Z" fill="url(#chk)" stroke="#2A1A12" stroke-width="5"/>
+</g><circle cx="140" cy="128" r="80" fill="none" stroke="#2A1A12" stroke-width="7"/>
+<g stroke="#2A1A12" stroke-linecap="round">
+  <path d="M106 106 Q117 100 129 106" stroke-width="7" fill="none"/><path d="M153 106 Q165 100 177 106" stroke-width="7" fill="none"/>
+</g>
+<g stroke="#2A1A12" stroke-width="5">
+  <ellipse cx="118" cy="128" rx="13" ry="17" fill="#FFFFFF"/>
+  <ellipse cx="164" cy="128" rx="13" ry="17" fill="#FFFFFF"/>
+</g>
+<circle cx="123" cy="131" r="7.5" fill="#2A1A12"/><circle cx="169" cy="131" r="7.5" fill="#2A1A12"/>
+<circle cx="126" cy="127" r="2.6" fill="#FFFFFF"/><circle cx="172" cy="127" r="2.6" fill="#FFFFFF"/>
+<ellipse cx="98" cy="154" rx="9" ry="6" fill="#FFB4A0" opacity="0.75"/>
+<ellipse cx="186" cy="152" rx="9" ry="6" fill="#FFB4A0" opacity="0.75"/>
+<path d="M114 150 Q141 186 170 147 Z" fill="#2A1A12" stroke="#2A1A12" stroke-width="5" stroke-linejoin="round"/>
+<g clip-path="url(#mouth)"><ellipse cx="145" cy="172" rx="14" ry="8" fill="#FF8FA3"/></g></g></g></svg>

tokenrace/_data/assets/logo/tokenrace-mascot.svg

@@ -0,0 +1,57 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="-8 -4 272 264">
+<defs>
+  <filter id="sticker" x="-15%" y="-15%" width="130%" height="130%">
+    <feMorphology in="SourceAlpha" operator="dilate" radius="6" result="d"/>
+    <feFlood flood-color="#FFFFFF"/><feComposite in2="d" operator="in" result="w"/>
+    <feMerge><feMergeNode in="w"/><feMergeNode in="SourceGraphic"/></feMerge>
+  </filter>
+  <clipPath id="edge"><circle cx="153" cy="133" r="80"/></clipPath>
+  <pattern id="chk" width="12" height="12" patternUnits="userSpaceOnUse">
+    <rect width="12" height="12" fill="#FFFFFF"/>
+    <rect width="6" height="6" fill="#1E1E1E"/><rect x="6" y="6" width="6" height="6" fill="#1E1E1E"/>
+  </pattern>
+  <radialGradient id="face" cx="0.38" cy="0.32" r="0.8">
+    <stop offset="0" stop-color="#FF8A4F"/><stop offset="0.55" stop-color="#FF5A1F"/><stop offset="1" stop-color="#E5470F"/>
+  </radialGradient>
+  <clipPath id="coin"><circle cx="140" cy="128" r="80"/></clipPath>
+  <clipPath id="mouth"><path d="M114 150 Q141 186 170 147 Z"/></clipPath>
+</defs><g filter="url(#sticker)"><g transform="rotate(6 140 140)">
+<g stroke-linecap="round" fill="none">
+  <path d="M4 150 H34" stroke="#FF5A1F" stroke-width="9"/>
+  <path d="M14 174 H52" stroke="#2A1A12" stroke-width="9"/>
+  <path d="M8 198 H58" stroke="#FFA27A" stroke-width="9"/>
+</g>
+<g stroke="#2A1A12" stroke-width="5" stroke-linejoin="round">
+  <path d="M70 86 C50 76 34 84 14 72 L20 94 C38 100 52 96 70 102 Z" fill="url(#chk)"/>
+  <path d="M70 100 C54 108 40 118 22 116 L32 132 C48 130 58 120 72 112 Z" fill="url(#chk)"/>
+</g>
+<g stroke="#2A1A12" stroke-linecap="round" stroke-linejoin="round" fill="none" stroke-width="9">
+  <path d="M74 150 Q54 158 46 142"/>
+  <path d="M118 200 Q104 214 86 204"/>
+</g>
+<circle cx="45" cy="135" r="11" fill="#FFFFFF" stroke="#2A1A12" stroke-width="5"/>
+<g transform="translate(78 196) rotate(35)" stroke="#2A1A12" stroke-width="5" stroke-linejoin="round"><path d="M-17 7 Q-17 -10 0 -10 Q21 -8 23 5 Q21 11 -17 7 Z" fill="#FFFFFF"/><path d="M-12 3 H18" stroke="#FF5A1F" stroke-width="4"/></g>
+<path d="M160 204 Q172 218 174 234" stroke="#2A1A12" stroke-width="9" stroke-linecap="round" fill="none"/>
+<g transform="translate(180 238) rotate(0)" stroke="#2A1A12" stroke-width="5" stroke-linejoin="round"><path d="M-17 7 Q-17 -10 0 -10 Q21 -8 23 5 Q21 11 -17 7 Z" fill="#FFFFFF"/><path d="M-12 3 H18" stroke="#FF5A1F" stroke-width="4"/></g>
+<circle cx="153" cy="133" r="80" fill="#B93A0C" stroke="#2A1A12" stroke-width="7"/>
+<g clip-path="url(#edge)" stroke="#2A1A12" stroke-width="3.5" stroke-opacity="0.8"><path d="M164.8 66.0 L167.6 50.3"/><path d="M174.0 68.3 L179.0 53.1"/><path d="M182.8 71.9 L189.8 57.5"/><path d="M191.0 76.6 L200.0 63.4"/><path d="M198.5 82.5 L209.2 70.6"/><path d="M205.1 89.3 L217.3 79.0"/><path d="M210.7 97.0 L224.2 88.5"/><path d="M215.1 105.3 L229.7 98.8"/><path d="M218.4 114.3 L233.7 109.8"/><path d="M220.3 123.5 L236.2 121.3"/><path d="M221.0 133.0 L237.0 133.0"/><path d="M220.3 142.5 L236.2 144.7"/><path d="M218.4 151.7 L233.7 156.2"/><path d="M215.1 160.7 L229.7 167.2"/><path d="M210.7 169.0 L224.2 177.5"/><path d="M205.1 176.7 L217.3 187.0"/><path d="M198.5 183.5 L209.2 195.4"/><path d="M191.0 189.4 L200.0 202.6"/><path d="M182.8 194.1 L189.8 208.5"/><path d="M174.0 197.7 L179.0 212.9"/><path d="M164.8 200.0 L167.6 215.7"/><path d="M155.4 201.0 L155.9 216.9"/><path d="M145.9 200.6 L144.2 216.5"/><path d="M136.5 199.0 L132.7 214.5"/><path d="M127.5 196.0 L121.5 210.9"/><path d="M119.0 191.9 L111.0 205.7"/><path d="M111.1 186.6 L101.3 199.2"/><path d="M104.1 180.2 L92.6 191.4"/><path d="M98.0 173.0 L85.0 182.4"/><path d="M93.0 164.9 L78.8 172.4"/><path d="M89.1 156.3 L74.1 161.7"/><path d="M86.5 147.1 L70.8 150.5"/><path d="M85.2 137.7 L69.2 138.9"/><path d="M85.2 128.3 L69.2 127.1"/></g>
+<circle cx="140" cy="128" r="80" fill="url(#face)"/>
+<circle cx="140" cy="128" r="68" fill="none" stroke="#FFFFFF" stroke-opacity="0.28" stroke-width="4"/>
+<path d="M82 118 A60 60 0 0 1 100 84" fill="none" stroke="#FFFFFF" stroke-opacity="0.55" stroke-width="7" stroke-linecap="round"/><g clip-path="url(#coin)">
+  <path d="M50 86 Q140 60 230 86 L230 106 Q140 80 50 106 Z" fill="url(#chk)" stroke="#2A1A12" stroke-width="5"/>
+</g><circle cx="140" cy="128" r="80" fill="none" stroke="#2A1A12" stroke-width="7"/>
+<g stroke="#2A1A12" stroke-linecap="round">
+  <path d="M106 106 Q117 100 129 106" stroke-width="7" fill="none"/><path d="M153 106 Q165 100 177 106" stroke-width="7" fill="none"/>
+</g>
+<g stroke="#2A1A12" stroke-width="5">
+  <ellipse cx="118" cy="128" rx="13" ry="17" fill="#FFFFFF"/>
+  <ellipse cx="164" cy="128" rx="13" ry="17" fill="#FFFFFF"/>
+</g>
+<circle cx="123" cy="131" r="7.5" fill="#2A1A12"/><circle cx="169" cy="131" r="7.5" fill="#2A1A12"/>
+<circle cx="126" cy="127" r="2.6" fill="#FFFFFF"/><circle cx="172" cy="127" r="2.6" fill="#FFFFFF"/>
+<ellipse cx="98" cy="154" rx="9" ry="6" fill="#FFB4A0" opacity="0.75"/>
+<ellipse cx="186" cy="152" rx="9" ry="6" fill="#FFB4A0" opacity="0.75"/>
+<path d="M114 150 Q141 186 170 147 Z" fill="#2A1A12" stroke="#2A1A12" stroke-width="5" stroke-linejoin="round"/>
+<g clip-path="url(#mouth)"><ellipse cx="145" cy="172" rx="14" ry="8" fill="#FF8FA3"/></g>
+<path d="M214 148 Q240 152 242 126" stroke="#2A1A12" stroke-width="9" stroke-linecap="round" fill="none"/>
+<circle cx="242" cy="119" r="11" fill="#FFFFFF" stroke="#2A1A12" stroke-width="5"/></g></g></svg>

tokenrace/_data/assets/style.css

@@ -0,0 +1,37 @@
+/* Stile di TokenRace. Il colore principale arriva da config/defaults.toml
+   come variabile --tr-primary. */
+
+.block-container {padding-top: 2rem; padding-bottom: 1rem; max-width: 1400px;}
+h5 {margin: 0 0 0.6rem 0 !important; padding: 0 !important;}
+[data-testid="stVerticalBlock"] {gap: 0.5rem;}
+[data-testid="stWidgetLabel"] p {font-size: 0.8rem;}
+.stTextInput input, .stNumberInput input, .stTextArea textarea,
+[data-baseweb="select"] div {font-size: 0.85rem;}
+[data-testid="stCaptionContainer"] {margin-top: -0.3rem;}
+[data-testid="stMetricValue"] {font-size: 1.5rem;}
+[data-testid="stMetricLabel"] p {font-size: 0.75rem;}
+
+/* Intestazione */
+.tr-hero {display: flex; align-items: baseline; gap: 0.8rem; flex-wrap: wrap;
+          margin-bottom: 0.6rem;}
+.tr-hero h1 {font-size: 2rem; margin: 0; padding: 0; display: flex; align-items: center;}
+.tr-hero h1 img {height: 2.6rem; margin-right: 0.45rem;}
+.tr-hero h1 span {color: var(--tr-primary);}
+.tr-hero p {margin: 0; opacity: 0.65; font-size: 0.95rem;}
+.tr-pill {display: inline-block; padding: 0.1rem 0.6rem; border-radius: 999px;
+          background: color-mix(in srgb, var(--tr-primary) 12%, transparent);
+          color: var(--tr-primary); font-size: 0.75rem; font-weight: 600;}
+
+/* Piè di pagina */
+.tr-footer {text-align: center; opacity: 0.55; font-size: 0.8rem; margin-top: 1.5rem;}
+
+/* Stato della ripetizione, in alto a destra del riquadro delle misure */
+.tr-status-wrap {text-align: right; margin-bottom: 1rem;}
+.tr-status {display: inline-flex; align-items: center; gap: 0.4rem; padding: 0.15rem 0.7rem;
+            border-radius: 999px; font-size: 0.75rem; font-weight: 600; white-space: nowrap;}
+.tr-dot {width: 0.5rem; height: 0.5rem; border-radius: 50%; background: currentColor;}
+.tr-status-running {color: var(--tr-primary);
+                    background: color-mix(in srgb, var(--tr-primary) 12%, transparent);}
+.tr-status-running .tr-dot {animation: tr-pulse 1s ease-in-out infinite;}
+.tr-status-done {color: #1a9f53; background: color-mix(in srgb, #1a9f53 12%, transparent);}
+@keyframes tr-pulse {0%, 100% {opacity: 1;} 50% {opacity: 0.25;}}

tokenrace/_data/config/defaults.toml

@@ -0,0 +1,27 @@
+# Valori predefiniti di TokenRace. Modificali liberamente.
+
+[params]
+# Valori iniziali dei campi del test (si possono cambiare anche dall'interfaccia)
+temperature = 0.0
+seed = 42
+max_tokens = 512
+repetitions = 3
+
+[measure]
+ram_interval_s = 0.5       # ogni quanto campionare la RAM di sistema
+ui_refresh_s = 0.1         # ogni quanto aggiornare risposta e misure durante lo streaming
+connect_timeout_s = 5      # attesa massima per collegarsi al server
+read_timeout_s = 600       # attesa massima per la risposta
+models_timeout_s = 5       # attesa massima per GET /models
+
+[long_prompt]
+target_chars = 32000       # lunghezza del contesto dei prompt "Lungo" e "Cache" (≈ 8.000 token)
+
+[ui]
+languages = ["en", "it"]   # lingue disponibili: un file locales/<codice>.toml per ciascuna
+default_language = "en"    # lingua all'apertura (si cambia dal selettore o con ?lang=it)
+fallback_language = "en"   # usata per i testi che mancano in una traduzione
+response_height = 430      # altezza in pixel del riquadro della risposta
+# Il colore principale va cambiato in due posti: qui (per lo stile) e in .streamlit/config.toml
+primary_color = "#FF5A1F"
+repo_url = "https://github.com/ernext/tokenrace"