pin: kernels/phase4_session.py (RepoMissing resume branch, disk gate, PHASE4_PREP_ONLY)
Browse files- kernels/phase4_session.py +29 -4
kernels/phase4_session.py
CHANGED
|
@@ -47,6 +47,7 @@ PLANNING_TOK_PER_S = float(os.environ.get("PLANNING_TOK_PER_S", "7732"))
|
|
| 47 |
SESSION_GPU_HOURS = float(os.environ.get("SESSION_GPU_HOURS", "11.0"))
|
| 48 |
QUOTA_LEFT_HOURS = float(os.environ.get("QUOTA_LEFT_HOURS", "30.0"))
|
| 49 |
RESERVE_HOURS = float(os.environ.get("RESERVE_HOURS", "0.6")) # startup, val eval, the last push
|
|
|
|
| 50 |
|
| 51 |
|
| 52 |
def sh(argv, label, timeout=None, env=None):
|
|
@@ -141,22 +142,38 @@ def main():
|
|
| 141 |
"shard_dataset.py": ("train/shard_dataset.py",
|
| 142 |
"4111832e79445fb0ffce6a73a02da4c97255d6c7886762682e04a417f2695a78"),
|
| 143 |
"hubckpt.py": ("train/hubckpt.py",
|
| 144 |
-
"
|
| 145 |
"train_ounce100m.py": ("train/train_ounce100m.py",
|
| 146 |
-
"
|
| 147 |
})
|
| 148 |
|
| 149 |
import ounce100m_credentials
|
| 150 |
print("creds:", json.dumps(ounce100m_credentials.install(verify=True)), flush=True)
|
| 151 |
rc, out = sh(["bash", "-c", "nvidia-smi --query-gpu=name,memory.used,memory.total "
|
| 152 |
-
"--format=csv,noheader; df -
|
| 153 |
"env")
|
|
|
|
|
|
|
| 154 |
free_gb = None
|
| 155 |
for line in out.splitlines():
|
| 156 |
if "overlay" in line or line.startswith("/dev/"):
|
| 157 |
parts = line.split()
|
| 158 |
if len(parts) >= 4:
|
| 159 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 160 |
|
| 161 |
# The Hub pointer is the only truth about where the run is (never local state -- §3.13).
|
| 162 |
import hubckpt
|
|
@@ -209,6 +226,14 @@ def main():
|
|
| 209 |
"burn quota on work that cannot be resumed from. Raise SESSION_GPU_HOURS or wait for "
|
| 210 |
"quota.", flush=True)
|
| 211 |
raise SystemExit(8)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 212 |
|
| 213 |
argv = [sys.executable, "-u", "train_ounce100m.py",
|
| 214 |
"--root", f"{WORK}/mixroot", "--out", f"{WORK}/run",
|
|
|
|
| 47 |
SESSION_GPU_HOURS = float(os.environ.get("SESSION_GPU_HOURS", "11.0"))
|
| 48 |
QUOTA_LEFT_HOURS = float(os.environ.get("QUOTA_LEFT_HOURS", "30.0"))
|
| 49 |
RESERVE_HOURS = float(os.environ.get("RESERVE_HOURS", "0.6")) # startup, val eval, the last push
|
| 50 |
+
MIN_FREE_GB = float(os.environ.get("MIN_FREE_GB", "8")) # mix 2.3 GB + two checkpoints + slack
|
| 51 |
|
| 52 |
|
| 53 |
def sh(argv, label, timeout=None, env=None):
|
|
|
|
| 142 |
"shard_dataset.py": ("train/shard_dataset.py",
|
| 143 |
"4111832e79445fb0ffce6a73a02da4c97255d6c7886762682e04a417f2695a78"),
|
| 144 |
"hubckpt.py": ("train/hubckpt.py",
|
| 145 |
+
"c8c958417dc48db92f8be5c4b505581c746be9cc10110a358230ac47c480d5bf"),
|
| 146 |
"train_ounce100m.py": ("train/train_ounce100m.py",
|
| 147 |
+
"487edc1be54f03b5f1a312901580cc3e1504c412322ed074ceff869febbfb8bc"),
|
| 148 |
})
|
| 149 |
|
| 150 |
import ounce100m_credentials
|
| 151 |
print("creds:", json.dumps(ounce100m_credentials.install(verify=True)), flush=True)
|
| 152 |
rc, out = sh(["bash", "-c", "nvidia-smi --query-gpu=name,memory.used,memory.total "
|
| 153 |
+
"--format=csv,noheader; df -k /kaggle/working | tail -1; free -g | head -2"],
|
| 154 |
"env")
|
| 155 |
+
# `df -k` not `-h`: the human-readable form prints "318G"/"500M"/"1.5T" and a float() on those raises,
|
| 156 |
+
# which would take down the session before training even starts.
|
| 157 |
free_gb = None
|
| 158 |
for line in out.splitlines():
|
| 159 |
if "overlay" in line or line.startswith("/dev/"):
|
| 160 |
parts = line.split()
|
| 161 |
if len(parts) >= 4:
|
| 162 |
+
try:
|
| 163 |
+
free_gb = float(parts[3]) / 1048576.0
|
| 164 |
+
except ValueError:
|
| 165 |
+
free_gb = None
|
| 166 |
+
if free_gb is None:
|
| 167 |
+
print("VERDICT REFUSED_TO_START: could not read the free space on /kaggle/working from "
|
| 168 |
+
"df output -- refusing to guess whether the mix and two checkpoints will fit", flush=True)
|
| 169 |
+
raise SystemExit(10)
|
| 170 |
+
if free_gb < MIN_FREE_GB:
|
| 171 |
+
print(f"VERDICT REFUSED_TO_START: {free_gb:.1f} GB free on /kaggle/working, this run needs at "
|
| 172 |
+
f"least {MIN_FREE_GB} GB (2.3 GB mix + ~1.7 GB per checkpoint + the unpacked resume dir). "
|
| 173 |
+
"A Kaggle GPU instance starts with ~20 GB, so a low number here means something is left "
|
| 174 |
+
"over from a previous session on this box.", flush=True)
|
| 175 |
+
raise SystemExit(10)
|
| 176 |
+
print(f"disk: {free_gb:.1f} GB free", flush=True)
|
| 177 |
|
| 178 |
# The Hub pointer is the only truth about where the run is (never local state -- §3.13).
|
| 179 |
import hubckpt
|
|
|
|
| 226 |
"burn quota on work that cannot be resumed from. Raise SESSION_GPU_HOURS or wait for "
|
| 227 |
"quota.", flush=True)
|
| 228 |
raise SystemExit(8)
|
| 229 |
+
if os.environ.get("PHASE4_PREP_ONLY") == "1":
|
| 230 |
+
# Everything a session can get wrong before it bills GPU time has now been checked: the pinned
|
| 231 |
+
# code hashes, the credentials resolve, the checkpoint pointer is readable, the published mix is
|
| 232 |
+
# downloadable and large enough, and the plan lands on a checkpoint boundary. A CPU instance can
|
| 233 |
+
# run all of that for free (E-031's arithmetic is exactly the kind of thing to find there).
|
| 234 |
+
print("VERDICT PREP_ONLY_OK stop_after_steps", pl["stop_after_steps"],
|
| 235 |
+
"intervals", pl["intervals_this_session"], flush=True)
|
| 236 |
+
return
|
| 237 |
|
| 238 |
argv = [sys.executable, "-u", "train_ounce100m.py",
|
| 239 |
"--root", f"{WORK}/mixroot", "--out", f"{WORK}/run",
|