BurnyCoder's picture
Publish the verified paper-review protocol before model inference
387a4cf verified
Raw History Blame Contribute Delete
12.7 kB
"""Run, resume and publish the registered pilot, paired confirmation and midpoint fork."""
from __future__ import annotations
from contextlib import contextmanager
import os
import subprocess
import tempfile
import time
import zipfile
import httpx
from .data import dump, load
from .domain import Domain
from .endpoints import freeze, roster, score
from .inference import Broker, MODEL, MODEL_DIGEST, now
from .reporting import report
from .sandbox import Worker
from .search import Study, run_name
ARMS = ("editable", "native", "uniform", "handcrafted")
HF = "BurnyCoder/rqgm-editable-parent-selection"
def preflight(project):
"""Lock the largest registered GPU-resident context using only a synthetic input."""
path = project / "results" / "runtime.json"
if path.exists():
return load(path)
attempts = []
for ctx in (16384, 8192, 4096):
b = Broker(project, "preflight-" + str(ctx), ctx)
try:
b.ask(
"hardware/synthetic",
"preflight",
"Return OK.",
"Synthetic GPU test.",
8,
)
processes = b.client.get("/api/ps").json()["models"]
process = next(m for m in processes if m["digest"] == MODEL_DIGEST)
attempts.append(
dict(
ctx=ctx,
size=process["size"],
size_vram=process["size_vram"],
compute=b.metrics(),
)
)
if (
process["size_vram"] >= process["size"]
and process.get("context_length", ctx) == ctx
):
hardware = subprocess.check_output(
[
"nvidia-smi",
"--query-gpu=name,memory.total,driver_version",
"--format=csv,noheader",
],
text=True,
).strip()
runtime = dict(
model=MODEL,
digest=MODEL_DIGEST,
ollama=b.version,
ctx=ctx,
gpu=hardware,
attempts=attempts,
registered=now(),
)
dump(path, runtime)
return runtime
except (httpx.HTTPError, RuntimeError, KeyError, StopIteration) as exc:
attempts.append(dict(ctx=ctx, error=str(exc)))
finally:
b.close()
dump(project / ".private" / "failed-preflight.json", attempts)
raise RuntimeError(
"Registered contexts did not achieve full GPU residency; see the preflight log"
)
@contextmanager
def domain_for(project, name):
"""Open and close the per-run compute ledger on all execution paths."""
runtime = load(project / "results" / "runtime.json")
broker = Broker(project, name, runtime["ctx"])
try:
if runtime["ollama"] != broker.version or runtime["digest"] != MODEL_DIGEST:
raise ValueError("The frozen inference backend changed")
yield Domain(project, broker, Worker())
finally:
broker.close()
def search(project, stage, seed, arm, budget, stop_at=None):
"""Use the same user-facing operation from the CLI and the multi-run campaign."""
name = run_name(stage, seed, arm)
mid = (
project
/ ".private"
/ "runs"
/ run_name("confirm", 1, "editable")
/ "midpoint.json"
if arm == "fork-native"
else None
)
if stage == "confirm":
registered = load(project / "results" / "budget.json")["budget"]
if (
budget != registered
or (seed, arm) not in roster()
or (project / "results" / "freeze.json").exists()
):
raise ValueError(
"Confirmatory configuration is outside its registered cohort"
)
elif stage != "pilot" or seed != 0 or budget != 64 or arm not in ARMS:
raise ValueError(
"Pilot uses seed zero, the four registered arms and 64 outcomes"
)
with domain_for(project, name) as d:
return search_result(
Study(project, d, stage, seed, arm, budget, mid).run(stop_at)
)
def search_result(study):
return dict(
name=study.name,
n=study.n,
budget=study.info["budget"],
epoch=study.core.epoch_vector,
)
def evaluate(project, split, name):
with domain_for(project, "eval-" + split + "-" + name) as d:
out = score(project, d, split, name)
return dict(
name=name, split=split, papers=len(out["rows"]), complete=out["complete"]
)
def budget_decision(project, deadline, remaining=None):
"""Price all future search and endpoint runs from the slowest pilot timings only."""
path = project / "results" / "budget.json"
if path.exists():
return load(path)
searches = [
load(project / "results" / "runs" / (run_name("pilot", 0, a) + ".json"))
for a in ARMS
]
eps = [
load(project / "results" / "evaluations" / ("pilot-" + r["name"] + ".json"))
for r in searches
]
eps.append(load(project / "results" / "evaluations" / "pilot-root.json"))
if not all(r["complete"] for r in [*searches, *eps]):
raise ValueError("Incomplete pilot block")
per_eval = max(r["elapsed"] / r["n"] for r in searches)
per_paper = max(e["elapsed"] / len(e["rows"]) for e in eps)
available = deadline - time.time() if remaining is None else remaining
search_equivalents = len(roster()) - 0.5
endpoints = len(roster()) + 1 + 2 * 8
test_items = load(project / "results" / "split-manifest.json")["splits"]["test"][
"count"
]
reserve = per_paper * test_items * endpoints
projections = {
str(b): (per_eval * b * search_equivalents + reserve) / 3600
for b in (64, 128, 256)
}
fits = [b for b in (64, 128, 256) if projections[str(b)] * 3600 <= available]
decision = dict(
budget=max(fits, default=64),
deadline=deadline,
available_hours=available / 3600,
chosen_at=now(),
pilot_seconds_per_search_eval=per_eval,
pilot_seconds_per_item=per_paper,
search_equivalents=search_equivalents,
planned_endpoints=endpoints,
heldout_reserve_seconds=reserve,
projections_hours=projections,
all_projected_to_fit=bool(fits),
)
dump(path, decision)
return decision
def command(project, *args):
"""Run publication tools with argument arrays and checked exit statuses."""
return subprocess.run(list(args), cwd=project, text=True, check=True)
def publish(project, label):
"""Publish only tracked source plus committed public artifacts to both destinations."""
report(project)
branch = subprocess.check_output(
["git", "branch", "--show-current"], cwd=project, text=True
).strip()
if branch != "main":
raise RuntimeError(
"Merge the implementation PR before publishing empirical artifacts"
)
command(project, "git", "add", "--", "README.md", "reports", "docs", "results")
changed = subprocess.run(
["git", "diff", "--cached", "--quiet"], cwd=project
).returncode
if changed == 1:
command(project, "git", "commit", "-m", label)
elif changed != 0:
raise RuntimeError("Unable to check staged artifacts")
command(project, "git", "push", "origin", "main")
rev = subprocess.check_output(
["git", "rev-parse", "HEAD"], cwd=project, text=True
).strip()
with tempfile.TemporaryDirectory(dir=project / ".private") as temp:
from pathlib import Path
folder = Path(temp)
archive = folder / "study.zip"
command(
project, "git", "archive", "--format=zip", "--output", str(archive), rev
)
with zipfile.ZipFile(archive) as source:
source.extractall(folder)
archive.unlink()
body = (folder / "README.md").read_text(encoding="utf-8")
(folder / "README.md").write_text(
"---\nbase_model: Qwen/Qwen3.5-4B\nlibrary_name: rqgm-parent-study\ntags:\n- paper-review\n- evolutionary-search\n---\n\n"
+ body
+ "\n\nGitHub source revision "
+ rev
+ ". The model artifacts are the executable workspaces and selectors in each run archive.\n",
encoding="utf-8",
)
command(
project,
"hf",
"upload",
HF,
str(folder),
".",
"--type",
"model",
"--commit-message",
label,
"--quiet",
)
return rev
def campaign(project, publish_results):
"""Keep the deadline, completed milestones and resumption state local to one study."""
ledger = project / ".private" / "campaign.json"
state = (
load(ledger)
if ledger.exists()
else dict(started=now(), deadline=time.time() + 7 * 86400, done=[])
)
dump(ledger, state)
def step(tag, fn):
if tag in state["done"]:
return
print(f"[{now()}] Starting {tag}", flush=True)
fn()
if publish_results:
publish(project, "Record " + tag)
else:
report(project)
state["done"].append(tag)
dump(ledger, state)
print(f"[{now()}] Completed {tag}", flush=True)
step("GPU preflight", lambda: preflight(project))
step("pilot root evaluation", lambda: evaluate(project, "pilot", "root"))
for arm in ARMS:
name = run_name("pilot", 0, arm)
step(name + " search", lambda a=arm: search(project, "pilot", 0, a, 64))
step(name + " evaluation", lambda n=name: evaluate(project, "pilot", n))
step("timing-selected budget", lambda: budget_decision(project, state["deadline"]))
chosen = load(project / "results" / "budget.json")
if not (project / "results" / "freeze.json").exists():
groups = [[(seed, "editable"), (seed, "native")] for seed in range(1, 9)]
groups.extend([[item] for item in roster()[16:]])
for group in groups:
outstanding = [
(s, a)
for s, a in group
if run_name("confirm", s, a) + " search" not in state["done"]
]
if not outstanding:
continue
completed = [
(s, a)
for s, a in roster()
if run_name("confirm", s, a) + " search" in state["done"]
]
planned = completed + outstanding
endpoints = (
1 + len(planned) + sum(a in ("editable", "native") for s, a in planned)
)
paper_count = load(project / "results" / "split-manifest.json")["splits"][
"test"
]["count"]
reserve = endpoints * paper_count * chosen["pilot_seconds_per_item"]
cost = (
sum(0.5 if a == "fork-native" else 1 for s, a in outstanding)
* chosen["budget"]
* chosen["pilot_seconds_per_search_eval"]
)
# Always try the first primary pair. Thereafter protect time for already-earned endpoints.
if completed and time.time() + reserve + cost > state["deadline"]:
dump(
project / "results" / "time-stop.json",
dict(
at=now(),
next_group=outstanding,
projected_search_seconds=cost,
reserved_endpoint_seconds=reserve,
remaining_seconds=state["deadline"] - time.time(),
),
)
break
for seed, arm in outstanding:
name = run_name("confirm", seed, arm)
step(
name + " search",
lambda a=arm, s=seed: search(
project, "confirm", s, a, chosen["budget"]
),
)
incomplete = (
project / "results" / "time-stop.json"
).exists() or time.time() >= state["deadline"]
step("joint reviewer freeze", lambda: freeze(project, allow_incomplete=incomplete))
manifest = load(project / "results" / "freeze.json")
for name in manifest["entries"]:
step("sealed " + name, lambda n=name: evaluate(project, "test", n))
return dict(completed=state["done"], missing=manifest["missing"])