kernel-env / train /dashboard.py
Shreyansh327's picture
Publish OpenEnv kernel environment and bounded GRPO proof artifacts
e313085 verified
Raw History Blame Contribute Delete
20.7 kB
"""Live, loopback-only viewer for a small_run.py experiment; no extra GPU required.
python dashboard.py --run-dir ../runs/proof-20261019/results --remote-dir /runs/proof-20261019-7b \
--container ta-... --log /tmp/kernel-proof-train.log --billing /tmp/kernel-proof-billing-progress.json
"""
import argparse
import json
import re
import subprocess
import threading
import time
from collections import defaultdict
from html import escape
from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
from pathlib import Path
from baseline_eval import summarize
from grpo_kernel import episode_return
from report import CSS, episode_html
# Read artifacts only: never import or execute generated kernels in the viewer.
READ = r'''
import json,sys
from pathlib import Path
p=Path(sys.argv[1])
def read(name, lines=False):
f=p/name
if not f.exists(): return [] if lines else {}
if not lines: return json.loads(f.read_text())
out=[]
for line in f.read_text().splitlines():
try: out.append(json.loads(line))
except ValueError: pass
return out
states=sorted((p/'adapter').glob('checkpoint-*/trainer_state.json'),key=lambda x:int(x.parent.name.split('-')[-1]))
print(json.dumps({'manifest':read('manifest.json'),'provenance':read('provenance.json'),
'before':read('before.jsonl',True),'train':read('adapter/rollouts-rank0.jsonl',True),
'after':read('after.jsonl',True),'comparison':read('comparison.json'),
'trainer':json.loads(states[-1].read_text()) if states else {}}))
'''
def view(snapshot, log_text="", cost=None):
manifest = snapshot.get("manifest", {})
events = []
for line in log_text.splitlines():
try:
event = json.loads(line)
if isinstance(event, dict) and event.get("event") == "step":
events.append(event)
except ValueError:
pass
goal = manifest.get("steps", 16)
progress = max([int(n) for n in re.findall(r"(\d+)/" + str(goal) + r"\s*\[", log_text)]
+ [snapshot.get("trainer", {}).get("global_step", 0)])
train = snapshot.get("train", [])
by_step = defaultdict(list)
for ep in train:
by_step[ep["step"] + 1].append(ep)
metrics = {int(m["step"]): m for m in snapshot.get("trainer", {}).get("log_history", []) if "step" in m}
steps = []
for step, eps in sorted(by_step.items()):
s = summarize(eps)
steps.append({"step": step, "episodes": len(eps), "correct": s["correct_pct"],
"task": escape(", ".join(sorted({e["task_id"].split("/")[-1] for e in eps}))),
"first": s["correct_turn1_pct"], "mean_return": sum(
episode_return([t["reward"] for t in e["turns"]], "discounted", 0.4)
for e in eps) / len(eps), "updated": step <= progress,
"grad_norm": metrics.get(step, {}).get("grad_norm"),
**{k: metrics.get(step, {}).get(k) for k in ("loss", "epoch", "learning_rate", "entropy")}})
before, after = snapshot.get("before", []), snapshot.get("after", [])
phase = "Complete" if snapshot.get("comparison") else "Final evaluation" if after else (
"Training" if train else "Baseline")
if not snapshot.get("comparison") and "PHASE after:" in log_text:
phase = "Final evaluation"
groups = {}
for name, ids in manifest.get("groups", {}).items():
groups[name] = {label: summarize([e for e in eps if e["task_id"] in ids])
for label, eps in [("before", before), ("after", after)]}
sections = {}
for key, eps in [("before", before), ("train", train), ("after", after)]:
parts = []
for i, ep in enumerate(eps):
label = f'Step {ep["step"] + 1} · episode {i + 1}' if key == "train" else f'Episode {i + 1}'
body = episode_html(ep).replace('<details class="ep">', f'<details class="ep" id="{key}-{i}">', 1)
for j in range(body.count('<details class="inner">')):
body = body.replace('<details class="inner">', f'<details class="inner" id="{key}-{i}-detail-{j}">', 1)
parts.append(f'<article><p class="note">{escape(label)}</p>{body}</article>')
sections[key] = "".join(parts) or '<p class="note">No episodes yet. This updates automatically.</p>'
return {"phase": phase, "updated": time.time(), "steps": steps, "completed": progress, "goal": goal,
"kernel_evals": len(events) if events else sum(len(e["turns"]) for k in ("before", "train", "after") for e in snapshot.get(k, [])),
"activity": (f'{events[-1]["task_id"]} · attempt {events[-1]["turn"]} · {events[-1]["stage"]}' if events else "Collecting the first submissions…"),
"cost": cost, "before": summarize(before), "after": summarize(after), "sections": sections,
"comparison": snapshot.get("comparison"), "model": manifest.get("model", "Loading…"),
"groups": groups, "metrics": metrics, "warning": None}
PAGE = r'''<!doctype html><html lang="en"><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
<title>Kernel Lab · Live experiment</title><style>__CSS__
body{margin:0}main{max-width:1120px;gap:28px}.top{display:flex;justify-content:space-between;align-items:start;gap:16px}
.live{color:var(--accent);font-size:13px}.stats{justify-content:space-between}.stat{min-width:120px}
.pipeline{display:grid;grid-template-columns:repeat(5,1fr);gap:8px}.phase{padding:12px;border-top:3px solid var(--line);font-size:13px}
.phase.active{border-color:var(--accent);background:var(--surface)}.phase.done{border-color:var(--ok)}
.panel{background:var(--surface);border:1px solid var(--line);padding:20px;border-radius:8px;min-width:0}
.stepgrid{display:grid;grid-template-columns:repeat(16,1fr);gap:5px}.cell{background:var(--code);padding:8px 0;text-align:center;font:12px var(--mono);border-radius:3px;border-bottom:3px solid var(--line)}
.cell.solved{background:var(--add);border-color:var(--ok)}.cell.partial{border-color:var(--warn)}
button,select,input{font:inherit;border:1px solid var(--line);background:var(--surface);color:var(--fg);padding:8px 12px;border-radius:5px}
button{cursor:pointer}button[aria-selected="true"]{color:var(--accent);border-color:var(--accent)}
.toolbar{display:flex;flex-wrap:wrap;gap:8px;align-items:center}.toolbar input{margin-left:auto;max-width:240px}
#episodes{display:grid;gap:14px}article .note{margin:0 0 5px}svg{width:100%;height:155px}#warning{color:var(--warn)}
.small{font-size:12px;color:var(--muted)}a{color:var(--accent)}.compare td{padding:10px 8px}.progress{height:5px;background:var(--code);margin-top:6px}.progress>div{height:100%;background:var(--accent)}
.traces{display:grid;grid-template-columns:repeat(auto-fit,minmax(260px,1fr));gap:16px}.traces svg{height:190px}.traces h3{font-size:14px;margin:0}.trace-label{display:flex;justify-content:space-between;align-items:center;gap:6px}.trace-label select{max-width:170px;font-size:12px;padding:4px}
@media(max-width:650px){.two{grid-template-columns:1fr}.stepgrid{grid-template-columns:repeat(8,1fr)}.pipeline{grid-template-columns:repeat(5,minmax(0,1fr))}.phase{padding:8px 2px;font-size:11px}.stats{gap:14px}.stat b{font-size:1.4rem}}
</style><main><header class="top"><div><div class="eyebrow">Kernel Lab / OpenEnv × GRPO</div><h1>A small run. Every step visible.</h1><p class="sub" id="model">Connecting to experiment…</p></div><div><div class="live" id="live" role="status">● Connecting</div><div class="small" id="stamp"></div></div></header>
<div class="pipeline" id="pipeline"></div>
<div class="stats"><div class="stat"><b id="completed">–</b><span>Training updates</span></div><div class="stat"><b id="evals">–</b><span>Kernel submissions</span></div><div class="stat"><b id="base">–</b><span>Baseline correct</span></div><div class="stat"><b id="after">Pending</b><span>After RL, correct</span></div><div class="stat"><b id="cost">–</b><span>Reported spend / $50 cap</span></div></div><p class="small" id="activity" role="status"></p>
<p class="small" id="warning" role="status"></p>
<section><div class="top"><div><h2>Optimizer traces</h2><p class="sub">Recorded metrics for each update. Hover over a point for its exact value and epoch.</p></div><a href="/metrics.json" download="training-metrics.json">Download all metrics ↓</a></div><div class="traces"><div class="panel"><h3>Reward / discounted episode return</h3><svg id="rewardChart" viewBox="0 0 350 190" role="img" aria-label="Reward by optimizer step"></svg></div><div class="panel"><h3>GRPO policy loss</h3><svg id="lossChart" viewBox="0 0 350 190" role="img" aria-label="GRPO loss by optimizer step"></svg></div><div class="panel"><div class="trace-label"><h3>More traces</h3><select id="metric" aria-label="Additional training metric"><option value="learning_rate">Learning rate</option><option value="grad_norm">Gradient norm</option><option value="entropy">Policy entropy</option><option value="reward_std">Reward standard deviation</option><option value="epoch">Epoch</option><option value="completions/mean_length">Completion tokens</option><option value="clip_ratio/region_mean">Clipped-token ratio</option><option value="step_time">Seconds per step</option><option value="num_tokens">Tokens processed</option></select></div><svg id="metricChart" viewBox="0 0 350 190" role="img" aria-label="Selected training metric by optimizer step"></svg></div></div><p class="small">X axis: optimizer step. Groups use different training tasks. A lower policy loss or higher training reward alone does not establish better held-out kernels.</p></section>
<div class="two"><section class="panel"><h2>How training is progressing</h2><p class="note">Each step samples four episodes of two attempts. Green = all four solved; amber = some or none. A filled cell means the optimizer update completed.</p><div class="stepgrid" id="stepgrid"></div><svg id="chart" viewBox="0 0 640 155" role="img" aria-label="Correct training episodes by optimizer step"></svg><div class="small">Bar height: correct episodes in that training group. This is training performance, not held-out progress.</div></section>
<section class="panel"><h2>The comparison that matters</h2><p class="note">Same tasks, model family, GPU and greedy decoding. Each task gets two attempts.</p><table class="compare"><thead><tr><th>Task group</th><th>Base</th><th>After RL</th></tr></thead><tbody id="comparison"></tbody></table><p class="small">Four training tasks, three reserved transfer tasks, three official held-out tasks. A ten-task pilot cannot establish broad generalization.</p></section></div>
<section><h2>Training steps</h2><div class="scroll"><table><thead><tr><th>Step</th><th>Epoch</th><th>Task</th><th>State</th><th>Correct episodes</th><th>Turn 1 correct</th><th>Mean return</th><th>Loss</th><th>Learning rate</th><th>Gradient norm</th></tr></thead><tbody id="steps"></tbody></table></div><p class="small">Groups can contain different tasks, so bar heights are not a learning curve on a fixed task. Trainer traces sync from saved checkpoint history. Final evaluation measures improvement.</p></section>
<section><div class="top"><div><h2>Inspect every attempt</h2><p class="sub">Open an episode for code, feedback, runtime and a diff between turns.</p></div><button id="refresh">Refresh now</button></div><div class="toolbar" role="tablist" aria-label="Experiment phase"><button data-phase="before" role="tab">Baseline</button><button data-phase="train" role="tab" aria-selected="true">Training</button><button data-phase="after" role="tab">After RL</button><input id="search" aria-label="Filter episodes by task name" placeholder="Filter by task name…"></div><div id="episodes"></div></section>
<footer class="small">Private live view on this computer · refreshes every 10 seconds · cost reporting can lag · <a href="https://app.notion.com/p/3fcecae370058136a9a5ef9dcc0457a1" target="_blank" rel="noopener">Build notes and findings ↗</a></footer></main>
<script>
let current=null,phase='train',lastHtml='';const $=id=>document.getElementById(id);const n=x=>x==null?'–':x.toFixed(2);const score=s=>s&&s.episodes?Math.round(s.correct_pct*s.episodes/100)+' / '+s.episodes:'Pending';
const numeric=x=>x==null?'–':Math.abs(x)>0&&Math.abs(x)<0.001?x.toExponential(2):x.toFixed(4);
function trace(id,key,color){let rows=Object.values(current.metrics).filter(m=>typeof m[key]==='number'&&Number.isFinite(m[key])).sort((a,b)=>a.step-b.step);if(!rows.length){$(id).innerHTML='<text x="20" y="90" fill="currentColor" font-size="12">Waiting for trainer metrics…</text>';return}let lo=Math.min(...rows.map(r=>r[key])),hi=Math.max(...rows.map(r=>r[key]));let span=hi-lo||Math.max(Math.abs(hi)*0.1,0.001);lo-=span*.12;hi+=span*.12;let x=s=>48+(s-1)*285/Math.max(current.goal-1,1),y=v=>145-(v-lo)*125/(hi-lo);let axes='';for(let i=0;i<3;i++){let v=lo+(hi-lo)*i/2,yy=y(v);axes+='<line x1="48" x2="335" y1="'+yy+'" y2="'+yy+'" stroke="#7c8b82" stroke-opacity=".25"/><text x="2" y="'+(yy+4)+'" fill="currentColor" font-size="10">'+numeric(v)+'</text>'}let points=rows.map(r=>x(r.step)+','+y(r[key])).join(' ');$(id).innerHTML=axes+'<polyline points="'+points+'" fill="none" stroke="'+color+'" stroke-width="2.5"/>'+rows.map(r=>'<circle tabindex="0" cx="'+x(r.step)+'" cy="'+y(r[key])+'" r="4" fill="'+color+'"><title>Step '+r.step+' · epoch '+n(r.epoch)+' · '+key+': '+r[key]+'</title></circle>').join('')+'<text x="48" y="169" fill="currentColor" font-size="11">1</text><text x="320" y="169" fill="currentColor" font-size="11">'+current.goal+'</text><text x="145" y="184" fill="currentColor" font-size="10">Optimizer step</text>'}
function traces(){if(!current)return;trace('rewardChart','reward','#4fc2ae');trace('lossChart','loss','#b29bf7');trace('metricChart',$('metric').value,'#e8ae58')}
function episodes(){if(!current)return;let h=current.sections[phase];if(h!==lastHtml){let opened=[...document.querySelectorAll('details[open][id]')].map(e=>e.id);$('episodes').innerHTML=h;opened.forEach(id=>{let e=$(id);if(e)e.open=true});lastHtml=h}let q=$('search').value.toLowerCase();document.querySelectorAll('#episodes article').forEach(e=>e.hidden=!e.querySelector('.task').textContent.toLowerCase().includes(q));document.querySelectorAll('[data-phase]').forEach(e=>e.setAttribute('aria-selected',e.dataset.phase===phase))}
function render(d){current=d;$('model').textContent=d.model+' · LoRA rank 16 · 2× H100 · 2-turn episodes';$('live').textContent='● '+d.phase;$('stamp').textContent='Synced '+new Date(d.updated*1000).toLocaleTimeString();$('completed').textContent=d.completed+' / '+d.goal;$('evals').textContent=d.kernel_evals;$('base').textContent=score(d.before);$('after').textContent=score(d.after);$('cost').textContent=d.cost==null?'Pending':'$'+d.cost.toFixed(2)+' / $50';$('warning').textContent=d.warning||'';$('activity').textContent='Latest evaluator feedback: '+d.activity;
let p=['GPU checks','Baseline','GRPO training','Final evaluation','Results'];let idx=d.phase==='Complete'?4:d.phase==='Final evaluation'?3:d.phase==='Training'?2:1;$('pipeline').innerHTML=p.map((s,i)=>'<div class="phase '+(i<idx?'done':i===idx?'active':'')+'">'+(i<idx?'✓ ':i===idx?'● ':'')+s+'</div>').join('');
let by=new Map(d.steps.map(s=>[s.step,s]));$('stepgrid').innerHTML=Array.from({length:d.goal},(_,i)=>{let s=by.get(i+1);return '<div class="cell '+(s&&s.updated?(s.correct===100?'solved':'partial'):'')+'" title="Step '+(i+1)+(s?': '+s.correct+'% correct'+(s.updated?', update complete':', collecting rollout'):', pending')+'">'+(i+1)+'</div>'}).join('');
let chart='<line x1="25" y1="125" x2="635" y2="125" stroke="#aaa"/><text x="0" y="20" fill="currentColor" font-size="11">100%</text><text x="0" y="126" fill="currentColor" font-size="11">0</text>';$('chart').innerHTML=chart+d.steps.map(s=>{let x=35+(s.step-1)*36,w=24,h=s.correct;return '<rect x="'+x+'" y="'+(125-h)+'" width="'+w+'" height="'+h+'" fill="'+(s.updated?'#0d6b5f':'#aab8b1')+'"/><text x="'+(x+8)+'" y="145" fill="currentColor" font-size="10">'+s.step+'</text>'}).join('');
$('steps').innerHTML=d.steps.map(s=>'<tr><td>'+s.step+'</td><td>'+n(s.epoch)+'</td><td>'+s.task+'</td><td>'+(s.updated?'✓ Updated':'Rollout collected')+'</td><td>'+Math.round(s.correct*s.episodes/100)+' / '+s.episodes+'</td><td>'+s.first.toFixed(0)+'%</td><td>'+n(s.mean_return)+'</td><td>'+numeric(s.loss)+'</td><td>'+numeric(s.learning_rate)+'</td><td>'+numeric(s.grad_norm)+'</td></tr>').join('')||'<tr><td colspan="10">Waiting for the first rollout group.</td></tr>';
let groups=[['training_tasks','Training tasks'],['reserved_transfer','Reserved transfer'],['official_heldout','Official held-out']];$('comparison').innerHTML=groups.map(([k,label])=>{let c=d.groups[k];return '<tr><td>'+label+'</td><td>'+(c?score(c.before):'Pending')+'</td><td>'+(c?score(c.after):'Pending')+'</td></tr>'}).join('');episodes();traces()}
async function refresh(){try{let r=await fetch('/state',{cache:'no-store'});if(!r.ok)throw Error('Connection failed');render(await r.json())}catch(e){$('warning').textContent='Live connection unavailable. Last successful snapshot remains visible.'}}
document.querySelectorAll('[data-phase]').forEach(e=>e.onclick=()=>{phase=e.dataset.phase;lastHtml='';episodes()});$('search').oninput=episodes;$('metric').onchange=traces;$('refresh').onclick=refresh;refresh();setInterval(refresh,10000);
</script></html>'''
def main():
p = argparse.ArgumentParser()
p.add_argument("--run-dir", type=Path, required=True)
p.add_argument("--remote-dir", required=True)
p.add_argument("--container", default="")
p.add_argument("--log", type=Path, required=True)
p.add_argument("--billing", type=Path, required=True)
p.add_argument("--app-id", action="append", default=[])
p.add_argument("--port", type=int, default=8099)
args = p.parse_args()
args.run_dir.mkdir(parents=True, exist_ok=True)
cache = args.run_dir / "live-snapshot.json"
snapshot = json.loads(cache.read_text()) if cache.exists() else {}
state = view(snapshot)
synced_at = cache.stat().st_mtime if cache.exists() else time.time()
def collect():
nonlocal snapshot, state, synced_at
while True:
warning = None
try:
if (args.run_dir / "comparison.json").exists():
r = subprocess.run(["python3", "-c", READ, str(args.run_dir)], capture_output=True, text=True, check=True)
else:
r = subprocess.run(["modal", "container", "exec", args.container, "--", "python", "-c", READ,
args.remote_dir], capture_output=True, text=True, check=True, timeout=25)
snapshot = json.loads(r.stdout)
cache.write_text(json.dumps(snapshot))
synced_at = time.time()
except (OSError, ValueError, subprocess.SubprocessError):
warning = "Cloud sync unavailable; showing the last snapshot. Final artifacts will replace it when downloaded."
cost = None
if args.billing.exists():
try:
rows = json.loads(args.billing.read_text())
cost = sum(float(r["Cost"]) for r in rows if r["Object ID"] in args.app_id)
except ValueError:
cost = state.get("cost")
state = view(snapshot, args.log.read_text() if args.log.exists() else "", cost)
state["updated"] = synced_at
state["warning"] = warning
time.sleep(20)
class Handler(BaseHTTPRequestHandler):
def do_GET(self):
if self.path not in ("/", "/state", "/metrics.json"):
self.send_error(404)
return
body = (json.dumps(state["metrics"] if self.path == "/metrics.json" else state).encode()
if self.path != "/" else PAGE.replace("__CSS__", CSS).encode())
self.send_response(200)
self.send_header("Content-Type", "application/json" if self.path != "/" else "text/html; charset=utf-8")
self.send_header("Cache-Control", "no-store")
self.send_header("Content-Length", str(len(body)))
self.end_headers()
self.wfile.write(body)
def log_message(self, *_):
pass
threading.Thread(target=collect, daemon=True).start()
print(f"Kernel Lab: http://127.0.0.1:{args.port}", flush=True)
ThreadingHTTPServer(("127.0.0.1", args.port), Handler).serve_forever()
if __name__ == "__main__":
main()