File size: 10,023 Bytes
4ddfe4d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 | #!/usr/bin/env python3
"""Snapshot of project progress for the live progress artifact (pushed into its db as progress/current)."""
import glob, json, os, re, subprocess, time, datetime as dt
os.chdir(os.path.expanduser("~/local-reasoning-research"))
R = "results/raw"
def jl(path):
try:
return [json.loads(l) for l in open(path)]
except FileNotFoundError:
return []
def running(pattern):
return subprocess.run(["pgrep", "-f", pattern], capture_output=True).returncode == 0
def current_local():
log = open(f"{R}/logs/queue_quality.log").read() if os.path.exists(f"{R}/logs/queue_quality.log") else ""
m = re.findall(r"^== (\S+) (\S+) ", log, re.M)
return m[-1] if m else None
QUEUES = []
cur = current_local()
for label, model in (("super-nvfp4", "Nemotron 3.5 Super (NVFP4)"), ("qwen38fn-q4", "Qwen3.8-Flash-Next (Q4)"), ("bolt-iq2", "Nemotron 3.5 Bolt (IQ2)")):
for task, total in (("aime", 15), ("lcb", 10)):
rows = jl(f"{R}/quality/{label}/{task}.jsonl")
active = cur == (label, task) and running("queue_quality.sh")
QUEUES.append(dict(group="Local quality (RTX PRO 4000)", model=model, task="AIME 2026" if task == "aime" else "LiveCodeBench v6 hard",
done=len(rows), total=total, correct=sum(r["ok"] for r in rows),
length_capped=sum(r.get("finish") == "length" for r in rows),
status="running" if active else ("done" if len(rows) >= total else "queued"),
last_s=round(rows[-1]["wall_s"]) if rows else None))
for d, model, pat in (("opus-5-5", "Claude Opus 5.5 (subscription)", "queue_cloud.sh opus"), ("minimax-m3", "MiniMax-M3 (subscription)", "queue_cloud.sh minimax")):
alive = running(pat)
for task, total in (("aime", 15), ("lcb", 10)):
rows = jl(f"{R}/quality/{d}/{task}.jsonl")
st = "done" if len(rows) >= total else ("running" if alive else "stopped")
QUEUES.append(dict(group="Cloud reference", model=model, task="AIME 2026" if task == "aime" else "LiveCodeBench v6 hard",
done=len(rows), total=total, correct=sum(r["ok"] for r in rows), length_capped=0, status=st,
last_s=round(rows[-1]["wall_s"]) if rows else None))
lc = sorted(glob.glob(f"{R}/cloud/{d}/v2_262144_s*.json"))
QUEUES.append(dict(group="Cloud reference", model=model, task="Long context 262K (v2 prompts)", done=len(lc), total=3,
correct=sum(json.load(open(f))["scores"]["_score"] == 1.0 for f in lc), length_capped=0,
status="done" if len(lc) >= 3 else ("running" if alive else "stopped"), last_s=None))
# Phase 2: 64K coding reruns of capped items + 10-run stability suite
P2 = running("queue_phase3.sh") or running("queue_super1m.sh")
p2log = open(f"{R}/logs/queue_phase3.log").read() if os.path.exists(f"{R}/logs/queue_phase3.log") else ""
p2cur = (re.findall(r"^== (\S+) (\S+) ", p2log, re.M) or [None])[-1]
for label, model, n in (("qwen38fn-q4", "Qwen3.8-Flash-Next (Q4)", 1),):
rows = jl(f"{R}/quality/{label}/lcb64k.jsonl")
st = "done" if len(rows) >= n else ("running" if P2 and p2cur == (label, "lcb64k") else ("queued" if P2 else "stopped"))
QUEUES.append(dict(group="Coding reruns at 64K (dropped after MiniMax went 0/4; only 1 Qwen item run)", model=model, task="LiveCodeBench capped items",
done=len(rows), total=n, correct=sum(r["ok"] for r in rows), length_capped=sum(r.get("finish") == "length" for r in rows),
status=st, last_s=None))
rows = jl(f"{R}/quality/minimax-m3/lcb64k.jsonl")
QUEUES.append(dict(group="Coding reruns at 64K (dropped after MiniMax went 0/4; only 1 Qwen item run)", model="MiniMax-M3 (subscription)", task="LiveCodeBench capped items",
done=len(rows), total=4, correct=sum(r["ok"] for r in rows), length_capped=0,
status="done" if len(rows) >= 4 else ("running" if running("lcb64k") else "stopped"), last_s=None))
stab = sorted(glob.glob(f"{R}/STABILITY-v2-*/*_s*_*[0-9].json"))
st_ok = sum(json.load(open(f))["scores"]["_score"] == 1.0 for f in stab)
QUEUES.append(dict(group="Phase 2: 256K stability qualification (brief 4.4)", model="Qwen3.8-Flash-Next (Q4, YaRN x2)",
task="10 consecutive runs, >=262,144 input tokens", done=len(stab), total=10, correct=st_ok, length_capped=0,
status="done" if len(stab) >= 10 else ("running" if P2 and p2cur and p2cur[0] == "qwen" else ("queued" if P2 else "stopped")), last_s=None))
for label, model, tok, n in (("qwen38fn-q4-yarn4-544k-q8kv", "Qwen3.8-Flash-Next (YaRN x4, q8 KV)", "512K", 2),
("super-nvfp4-1m-q8kv", "Nemotron 3.5 Super (native 1M, q8 KV)", "512K + 1M", 2)):
fs = sorted(glob.glob(f"{R}/{label}/*_s*_*[0-9].json"))
QUEUES.append(dict(group="Context extension (512K / 1M)", model=model, task=f"{tok} long-context runs", done=len(fs), total=n,
correct=sum(json.load(open(f))["scores"]["_score"] == 1.0 for f in fs), length_capped=0,
status="done" if len(fs) >= n else ("running" if (label.startswith("super") and running("queue_super1m.sh")) or (P2 and p2cur and label.startswith(p2cur[0][:5])) else ("queued" if P2 else "stopped")), last_s=None))
# Qwen focus phase
FQ = running("queue_qwen_focus.sh")
for label, task, n, pat in (("CANON-qwen38fn-q4-288k", "Q4: 3 shared 262K prompts (same as Opus/MiniMax)", 3, "v2_*.json"),
("CANON-qwen38fn-iq3s-288k", "3-bit IQ3_S: same 3 shared 262K prompts", 3, "v2_*.json")):
fs = sorted(f for f in glob.glob(f"{R}/{label}/{pat}") if "monitor" not in f)
QUEUES.append(dict(group="Qwen focus: quantization + same-prompt comparison", model="Qwen3.8-Flash-Next", task=task, done=len(fs), total=n,
correct=sum(json.load(open(f))["scores"]["_score"] == 1.0 for f in fs), length_capped=0,
status="done" if len(fs) >= n else ("running" if FQ else "stopped"), last_s=None))
rows = jl(f"{R}/quality/qwen38fn-iq3s/aime.jsonl")
QUEUES.append(dict(group="Qwen focus: quantization + same-prompt comparison", model="Qwen3.8-Flash-Next", task="3-bit IQ3_S: AIME 2026 (Q4 scored 14/15)",
done=len(rows), total=15, correct=sum(r["ok"] for r in rows), length_capped=sum(r.get("finish") == "length" for r in rows),
status="done" if len(rows) >= 15 else ("running" if FQ else "stopped"), last_s=None))
sp = sorted(glob.glob(f"{R}/speed/*.json"))
QUEUES.append(dict(group="Qwen focus: speed options", model="Qwen3.8-Flash-Next", task="baseline / MTP drafting / 3-bit (GPU-expert option dropped: OOM)",
done=len(sp), total=3, correct=len(sp), length_capped=0, status="done" if len(sp) >= 3 else ("running" if FQ else "stopped"), last_s=None))
ag = sorted(glob.glob(f"{R}/agentic/*.json"))
QUEUES.append(dict(group="Claude Code integration (agentic, local model)", model="Qwen3.8-Flash-Next (Q4) via Claude Code",
task="easy task (2 bugs, hinted) x3 + hard task (3 bugs, no hints) x3", done=len(ag), total=6,
correct=sum(json.load(open(f)).get("tests_pass") and json.load(open(f)).get("tests_untouched") for f in ag), length_capped=0,
status="done" if len(ag) >= 6 else ("running" if running("queue_agentic.sh") else "stopped"), last_s=None))
# waiting on a usage-limit reset?
waiting = None
lp = f"{R}/logs/queue_cloud_opus.log"
if os.path.exists(lp):
w = [l for l in open(lp) if "[usage-limit]" in l]
if w and running("queue_cloud.sh opus"):
waiting = w[-1].strip()[:200]
LONG = []
for f in sorted(glob.glob(f"{R}/*/*_s*_*[0-9].json")):
if "/quality/" in f: continue
d = json.load(open(f)); s = d["scores"]
n = sum(1 for k in s if not k.startswith("_"))
LONG.append(dict(model=d["label"], tokens=d["prompt_tokens_server"], seed=d["seed"], suite=d.get("suite", "easy"),
score=sum(v["ok"] for k, v in s.items() if not k.startswith("_")), of=n,
prefill_min=round(d["prefill_s"] / 60, 1), gen_tps=round(d["gen_tps"] or 0, 1)))
for f in sorted(glob.glob(f"{R}/cloud/*/*.json")):
d = json.load(open(f)); s = d["scores"]
LONG.append(dict(model=d["backend"] + (" (v2)" if "v2_" in f else ""), tokens=d.get("usage", {}).get("input"), seed=d["seed"],
suite=d.get("suite"), score=sum(v["ok"] for k, v in s.items() if not k.startswith("_")),
of=sum(1 for k in s if not k.startswith("_")), prefill_min=None, gen_tps=None))
PHASES = [
("Environment discovery", "done"), ("Model research & HGA", "done"), ("Candidate triage", "done"),
("Hardware feasibility", "done"), ("Benchmark harness", "done"),
("256K qualification runs (64K→128K→256K)", "done"), ("Quality suite: AIME + coding", "done"),
("Cloud comparison (Opus, MiniMax)", "done"), ("Coding reruns at 64K budget", "done"), ("10-run 256K stability suite", "done"),
("512K / 1M extension", "done"), ("Qwen focus: speed + 3-bit quantization", "done"),
("Fine-tuning assessment", "done"), ("Deployment + Claude Code integration", "done"), ("Final report", "done"),
]
snap = dict(updated_at=dt.datetime.now(dt.timezone.utc).isoformat(timespec="seconds"),
gpu=subprocess.run(["nvidia-smi", "-i", "GPU-4e9ed5ea-d6d0-183f-9335-fd87f65d3f16",
"--query-gpu=memory.used,utilization.gpu,temperature.gpu,power.draw", "--format=csv,noheader,nounits"],
capture_output=True, text=True).stdout.strip(),
queues=QUEUES, longctx=LONG, phases=[dict(name=n, status=s) for n, s in PHASES], waiting=waiting)
out = "/tmp/claude-1000/-home-adam/aee8f88e-e0a7-4d76-9aed-0c7c432b0b53/scratchpad/progress.json"
json.dump(snap, open(out, "w"))
print(out, sum(q["done"] for q in QUEUES), "/", sum(q["total"] for q in QUEUES))
|