""" gui/pages/models.py โ€” ๐ŸŽ Models Reads config/models.json live on every render โ€” no hardcoding. Schema: { "cloud": { "linear": {...}, "agentic": {...} }, "local": { "linear": {...}, "agentic": {...} } } """ import json import streamlit as st import pandas as pd import plotly.express as px from pathlib import Path from gui.config import PROJECT_ROOT, WF_COLORS from gui.db import q_safe from gui.helpers import fl # โ”€โ”€ Theme helpers โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€ def _is_dark() -> bool: return st.session_state.get("theme", "dark") == "dark" def _tok() -> dict: if _is_dark(): return dict( bg0="#0d1117", bg1="#111827", bg2="#1f2937", bg3="#374151", t1="#f1f5f9", t2="#94a3b8", t3="#475569", brd="#1f2937", brd2="#374151", accent="#3b82f6", ) return dict( bg0="#f3f4f6", bg1="#ffffff", bg2="#f9fafb", bg3="#d1d5db", t1="#1f2937", t2="#4b5563", t3="#6b7280", brd="#d1d5db", brd2="#9ca3af", accent="#3b82f6", ) # โ”€โ”€ Load models.json โ€” no caching so changes show immediately โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€ def _load_models() -> dict: for candidate in [ PROJECT_ROOT / "config" / "models.json", PROJECT_ROOT / "models.json", ]: if candidate.exists(): try: return json.loads(candidate.read_text()) except Exception as e: return {"_error": str(e), "_path": str(candidate)} return {"_missing": True} # โ”€โ”€ DB stats keyed by (model_name, workflow_type) โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€ def _load_stats() -> dict: stats = {} try: df, _ = q_safe(""" SELECT e.model_name, r.workflow_type, COUNT(*) AS runs, ROUND(AVG(r.total_energy_uj)/1e6, 4) AS avg_j, ROUND(AVG(r.dynamic_energy_uj)/1e6, 4) AS avg_dyn_j, ROUND(AVG(r.duration_ns)/1e9, 3) AS avg_dur_s, ROUND(AVG(r.total_tokens), 1) AS avg_tokens, ROUND(AVG(CASE WHEN r.total_tokens > 0 THEN r.total_energy_uj / r.total_tokens END) / 1e3, 4) AS avg_mj_tok, ROUND(AVG(r.ipc), 3) AS avg_ipc, ROUND(AVG(r.carbon_g)*1000, 4) AS avg_carbon_mg FROM runs r JOIN experiments e ON r.exp_id = e.exp_id WHERE e.model_name IS NOT NULL GROUP BY e.model_name, r.workflow_type """) if not df.empty: for _, row in df.iterrows(): stats[(str(row["model_name"]), str(row["workflow_type"]))] = row.to_dict() except Exception: pass return stats def _find_stat(stats: dict, spec: dict, wf: str) -> dict | None: """Match a spec to a stats row. Tries model_id then name.""" for field in ("model_id", "name"): v = spec.get(field) if v: hit = stats.get((str(v), wf)) if hit: return hit return None # โ”€โ”€ Spec block (linear or agentic half of a card) โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€ def _spec_block(spec: dict, wf: str, wf_stats: dict | None, t: dict) -> str: bg2, bg3 = t["bg2"], t["bg3"] t1, t2, t3 = t["t1"], t["t2"], t["t3"] brd = t["brd"] wf_clr = "#16a34a" if wf == "linear" else "#ef4444" name = spec.get("name", "โ€”") mid = spec.get("model_id", "") ep = spec.get("api_endpoint", "") mpath = spec.get("model_path", "") mtok = spec.get("max_tokens", "โ€”") temp = spec.get("temperature", "โ€”") tools_s = spec.get("tools_supported", False) tools = spec.get("tools", []) def _row(k, v, vc=None): vc = vc or t2 return ( f"
" f"{k}" f"{str(v) if v not in (None,'') else 'โ€”'}
" ) html = ( f"
" f"
{wf}
" f"
" f"{name}
" ) if mid: html += _row("Model ID", mid) if ep: html += _row("Endpoint", ep.replace("https://","")[:38]) if mpath: html += _row("Model file", Path(mpath).name) html += _row("Max tokens", mtok) html += _row("Temperature", temp) if tools_s and tools: html += _row("Tools", "โœ“ " + ", ".join(tools), "#16a34a") else: html += _row("Tools", "โœ“ enabled" if tools_s else "โœ— disabled", "#16a34a" if tools_s else t3) # Measured energy section if wf_stats: avg_j = wf_stats.get("avg_j", 0) or 0 avg_mj = wf_stats.get("avg_mj_tok", 0) or 0 avg_dur = wf_stats.get("avg_dur_s", 0) or 0 avg_tok = wf_stats.get("avg_tokens", 0) or 0 runs = int(wf_stats.get("runs", 0) or 0) html += ( f"
" f"
" f"Measured ยท {runs} runs ยท {avg_tok:.0f} avg tokens
" f"
" ) for lbl, val, unit in [ ("Energy", avg_j, "J"), ("mJ/token", avg_mj, "mJ"), ("Duration", avg_dur, "s"), ]: html += ( f"
" f"
{lbl}
" f"
{val:.3f}
" f"
{unit}
" f"
" ) html += "
" else: html += ( f"
" f"Not yet measured
" ) html += "
" return html # โ”€โ”€ Full group card (cloud / local) โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€ def _group_card(gkey: str, gcfg: dict, stats: dict, t: dict) -> str: bg1, bg2, bg3 = t["bg1"], t["bg2"], t["bg3"] t1, t2, t3 = t["t1"], t["t2"], t["t3"] brd = t["brd"] clr = {"cloud": "#38bdf8", "local": "#22c55e"}.get(gkey, "#7c3aed") icon = {"cloud": "โ˜", "local": "๐Ÿ’ป" }.get(gkey, "โš™") lin_spec = gcfg.get("linear", {}) age_spec = gcfg.get("agentic", {}) lin_stats = _find_stat(stats, lin_spec, "linear") age_stats = _find_stat(stats, age_spec, "agentic") html = ( f"
" # header f"
" f"{icon}" f"" f"{gkey.title()} provider" f"{gkey}" f"
" ) html += _spec_block(lin_spec, "linear", lin_stats, t) html += _spec_block(age_spec, "agentic", age_stats, t) # Tax multiple if both measured if lin_stats and age_stats: lj = lin_stats.get("avg_j") or 0 aj = age_stats.get("avg_j") or 0 if lj > 0: mult = aj / lj mc = "#ef4444" if mult >= 2 else "#d97706" if mult >= 1.2 else "#16a34a" html += ( f"
" f"
" f"
Orchestration tax
" f"
agentic รท linear
" f"
" f"{mult:.2f}ร—" f"
" ) html += "
" return html # โ”€โ”€ Main render โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€ def render(ctx: dict = None): t = _tok() bg1 = t["bg1"]; bg2 = t["bg2"]; bg3 = t["bg3"] t1 = t["t1"]; t2 = t["t2"]; t3 = t["t3"] brd = t["brd"]; brd2 = t["brd2"]; accent = t["accent"] # โ”€โ”€ Page header โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€ h_col, btn_col = st.columns([7, 1]) with h_col: st.markdown( f"
" f"๐ŸŽ Models" f"Apple-to-Apple energy comparison" f"
", unsafe_allow_html=True) with btn_col: if st.button("โ†บ Reload", key="models_reload"): st.rerun() # No cache โ€” file is always re-read # โ”€โ”€ Load config live โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€ cfg = _load_models() if cfg.get("_missing"): st.markdown( f"
" f"config/models.json not found.
" f"Expected at: " f"" f"{PROJECT_ROOT}/config/models.json" f"
", unsafe_allow_html=True) return if cfg.get("_error"): st.error(f"Error reading models.json: {cfg['_error']}") return # Show config path + mod time cfg_path = PROJECT_ROOT / "config" / "models.json" _mtime = "โ€”" try: import datetime as _dt _mtime = _dt.datetime.fromtimestamp( cfg_path.stat().st_mtime).strftime("%Y-%m-%d %H:%M:%S") except Exception: pass groups = [k for k in cfg if not k.startswith("_") and isinstance(cfg[k], dict)] st.markdown( f"
" f"Live from " f"" f"config/models.json" f" ยท modified {_mtime}" f" ยท {len(groups)} provider group{'s' if len(groups) != 1 else ''}: " f"{', '.join(groups)}
", unsafe_allow_html=True) # โ”€โ”€ Model config cards โ€” one per provider group โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€ stats = _load_stats() cols = st.columns(max(len(groups), 1)) for i, gkey in enumerate(groups): with cols[i]: st.markdown(_group_card(gkey, cfg[gkey], stats, t), unsafe_allow_html=True) st.markdown( f"
", unsafe_allow_html=True) # โ”€โ”€ Apple-to-Apple charts โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€ st.markdown( f"
" f"Measured comparison โ€” all runs
", unsafe_allow_html=True) _full_df = pd.DataFrame() try: _full_df, _ = q_safe(""" SELECT e.model_name, e.provider, r.workflow_type, e.task_name, COUNT(*) AS runs, ROUND(AVG(r.total_energy_uj)/1e6, 4) AS avg_energy_j, ROUND(AVG(r.dynamic_energy_uj)/1e6, 4) AS avg_dynamic_j, ROUND(AVG(r.duration_ns)/1e9, 3) AS avg_duration_s, ROUND(AVG(r.total_tokens), 1) AS avg_tokens, ROUND(AVG(CASE WHEN r.total_tokens > 0 THEN r.total_energy_uj / r.total_tokens END) / 1e3, 4) AS avg_mj_per_token, ROUND(AVG(r.ipc), 3) AS avg_ipc, ROUND(AVG(r.carbon_g)*1000, 4) AS avg_carbon_mg, ROUND(AVG(r.water_ml), 4) AS avg_water_ml FROM runs r JOIN experiments e ON r.exp_id = e.exp_id WHERE e.model_name IS NOT NULL GROUP BY e.model_name, e.provider, r.workflow_type, e.task_name ORDER BY e.provider, e.model_name, r.workflow_type """) except Exception: pass if _full_df.empty: st.markdown( f"
" f"No run data yet โ€” run experiments to see charts here.
", unsafe_allow_html=True) return # Task filter _, fc = st.columns([4, 1]) with fc: _tasks = sorted(_full_df.task_name.dropna().unique().tolist()) _sel = st.selectbox("Task", ["all"] + _tasks, key="models_task") _filt = (_full_df if _sel == "all" else _full_df[_full_df.task_name == _sel]).copy() _filt["model_wf"] = ( _filt["model_name"].astype(str) + " ยท " + _filt["workflow_type"].astype(str)) _pl = dict( paper_bgcolor=bg1, plot_bgcolor=bg2, font=dict(color=t1, size=10), margin=dict(t=20, b=70, l=50, r=10), legend=dict(bgcolor=bg1, bordercolor=brd, borderwidth=0.5, font=dict(color=t2)), height=260) for (ch_l, ch_r), (col_l, col_r), (lbl_l, lbl_r) in [ (("avg_energy_j", "avg_mj_per_token"), st.columns(2), ("Energy (J) โ€” model ร— workflow", "mJ / token โ€” model ร— workflow")), (("avg_duration_s", "avg_carbon_mg"), st.columns(2), ("Avg duration (s)", "Carbon footprint (mg COโ‚‚)")), ]: for col, metric, lbl in [(ch_l, col_l, lbl_l), (ch_r, col_r, lbl_r)]: pass # loop structure โ€” see below # Draw charts explicitly (avoid loop variable confusion) ch1, ch2 = st.columns(2) for col_ctx, metric, lbl in [ (ch1, "avg_energy_j", "Energy (J) โ€” model ร— workflow"), (ch2, "avg_mj_per_token","mJ / token โ€” model ร— workflow"), ]: with col_ctx: st.markdown( f"
{lbl}
", unsafe_allow_html=True) fig = px.bar( _filt.groupby(["model_wf","workflow_type"])[metric] .mean().reset_index(), x="model_wf", y=metric, color="workflow_type", barmode="group", color_discrete_map=WF_COLORS, labels={metric: lbl, "model_wf": ""}) fig.update_xaxes(tickangle=30) fig.update_layout(**_pl) st.plotly_chart(fl(fig), use_container_width=True) ch3, ch4 = st.columns(2) for col_ctx, metric, lbl in [ (ch3, "avg_duration_s", "Avg duration (s)"), (ch4, "avg_carbon_mg", "Carbon footprint (mg COโ‚‚)"), ]: with col_ctx: st.markdown( f"
{lbl}
", unsafe_allow_html=True) fig = px.bar( _filt.groupby(["model_wf","workflow_type"])[metric] .mean().reset_index(), x="model_wf", y=metric, color="workflow_type", barmode="group", color_discrete_map=WF_COLORS, labels={metric: lbl, "model_wf": ""}) fig.update_xaxes(tickangle=30) fig.update_layout(**_pl) st.plotly_chart(fl(fig), use_container_width=True) # โ”€โ”€ Full matrix โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€ st.markdown( f"
", unsafe_allow_html=True) st.markdown( f"
Full comparison matrix
", unsafe_allow_html=True) _show = [c for c in [ "model_name","provider","workflow_type","task_name","runs", "avg_energy_j","avg_dynamic_j","avg_duration_s","avg_tokens", "avg_mj_per_token","avg_ipc","avg_carbon_mg","avg_water_ml"] if c in _filt.columns] st.dataframe(_filt[_show].round(4), use_container_width=True, hide_index=True)