Survive a machine change
Browse files- scripts/launch.py +8 -2
- scripts/pipeline.sh +25 -1
- scripts/supervise.py +25 -6
scripts/launch.py
CHANGED
|
@@ -58,11 +58,17 @@ def get_uv(root):
|
|
| 58 |
|
| 59 |
|
| 60 |
def alive(pidfile):
|
|
|
|
|
|
|
| 61 |
try:
|
| 62 |
-
|
| 63 |
-
|
| 64 |
except (OSError, ValueError):
|
| 65 |
return False
|
|
|
|
|
|
|
|
|
|
|
|
|
| 66 |
|
| 67 |
|
| 68 |
def status(root):
|
|
|
|
| 58 |
|
| 59 |
|
| 60 |
def alive(pidfile):
|
| 61 |
+
"""True only if the recorded process exists and is our supervisor (a pid file left over
|
| 62 |
+
from an earlier machine can name an unrelated process)."""
|
| 63 |
try:
|
| 64 |
+
pid = int(pidfile.read_text())
|
| 65 |
+
os.kill(pid, 0)
|
| 66 |
except (OSError, ValueError):
|
| 67 |
return False
|
| 68 |
+
try:
|
| 69 |
+
return b"supervise.py" in Path(f"/proc/{pid}/cmdline").read_bytes()
|
| 70 |
+
except OSError:
|
| 71 |
+
return False
|
| 72 |
|
| 73 |
|
| 74 |
def status(root):
|
scripts/pipeline.sh
CHANGED
|
@@ -72,7 +72,7 @@ s_env() {
|
|
| 72 |
for d in "$UV_PYTHON_INSTALL_DIR"/cpython-3.12*; do
|
| 73 |
if [ -d "$d" ] && [ ! -x "$d/bin/python3.12" ]; then echo "removing incomplete $d"; rm -rf "$UV_PYTHON_INSTALL_DIR"; break; fi
|
| 74 |
done
|
| 75 |
-
if ! "$PY" -c "import sys" >/dev/null 2>&1; then rm -rf "$VENV"; "$UV" venv --python 3.12 "$VENV"; fi
|
| 76 |
"$UV" pip install --python "$PY" "vllm==0.31.0" accelerate datasets
|
| 77 |
"$UV" cache clean >/dev/null 2>&1 || true
|
| 78 |
"$PY" - <<'EOF'
|
|
@@ -194,6 +194,30 @@ s_bench_target() { SUB="" bench --tag target_only --target-only --what batch,lon
|
|
| 194 |
log "pipeline start: run $(cat "$STATE/run_id"), code $(cat "$REPO/.commit" 2>/dev/null | cut -c1-12), training run $RUN_NAME"
|
| 195 |
# a vLLM server left behind by an interrupted run would still hold the GPU
|
| 196 |
if pkill -f -- "--served-model-name target" 2>/dev/null; then log "stopped a leftover vLLM server"; sleep 20; fi
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 197 |
stage env 1 s_env
|
| 198 |
stage download 1 s_download
|
| 199 |
stage baseline 1 s_baseline
|
|
|
|
| 72 |
for d in "$UV_PYTHON_INSTALL_DIR"/cpython-3.12*; do
|
| 73 |
if [ -d "$d" ] && [ ! -x "$d/bin/python3.12" ]; then echo "removing incomplete $d"; rm -rf "$UV_PYTHON_INSTALL_DIR"; break; fi
|
| 74 |
done
|
| 75 |
+
if ! "$PY" -c "import sys" >/dev/null 2>&1; then rm -rf "$VENV" "$UV_CACHE_DIR"; "$UV" venv --python 3.12 "$VENV"; fi
|
| 76 |
"$UV" pip install --python "$PY" "vllm==0.31.0" accelerate datasets
|
| 77 |
"$UV" cache clean >/dev/null 2>&1 || true
|
| 78 |
"$PY" - <<'EOF'
|
|
|
|
| 194 |
log "pipeline start: run $(cat "$STATE/run_id"), code $(cat "$REPO/.commit" 2>/dev/null | cut -c1-12), training run $RUN_NAME"
|
| 195 |
# a vLLM server left behind by an interrupted run would still hold the GPU
|
| 196 |
if pkill -f -- "--served-model-name target" 2>/dev/null; then log "stopped a leftover vLLM server"; sleep 20; fi
|
| 197 |
+
# a machine change can keep the "done" markers but lose what the stages produced: check each
|
| 198 |
+
# finished stage's output and run the stage again when it is gone
|
| 199 |
+
check_done() { # check_done <stage> <command ...>
|
| 200 |
+
local name=$1; shift
|
| 201 |
+
if is_done "$name" && ! "$@" >/dev/null 2>&1; then rm -f "$STATE/$name.done"; log "redo $name (its output is missing)"; fi
|
| 202 |
+
}
|
| 203 |
+
has_files() { compgen -G "$1" >/dev/null; }
|
| 204 |
+
shards_ok() { # every shard listed by its metadata file is present and non-empty
|
| 205 |
+
local d=$1 j
|
| 206 |
+
has_files "$d/*.json" || return 1
|
| 207 |
+
for j in "$d"/*.json; do [ -s "${j%.json}.tok.npy" ] && [ -s "${j%.json}.hid.npy" ] || return 1; done
|
| 208 |
+
}
|
| 209 |
+
check_done env "$PY" -c "import vllm, torch, transformers"
|
| 210 |
+
check_done download "$PY" -c "import json, os, sys; d = '$TARGET'; f = set(json.load(open(d + '/model.safetensors.index.json'))['weight_map'].values()); sys.exit(0 if os.path.isfile('$KIT/contract.json') and all(os.path.getsize(os.path.join(d, x)) > 0 for x in f) else 1)"
|
| 211 |
+
check_done baseline test -s "$BASELINE/model.safetensors"
|
| 212 |
+
check_done probe test -s "$PROBE/model.safetensors"
|
| 213 |
+
check_done prompts has_files "$DATA/prompts/*.jsonl"
|
| 214 |
+
check_done smoke_fp8 test -s "$RESULTS/smoke_fp8.json"
|
| 215 |
+
check_done generate has_files "$DATA/gen/*.jsonl"
|
| 216 |
+
check_done extract shards_ok "$DATA/train"
|
| 217 |
+
check_done extract shards_ok "$DATA/dev"
|
| 218 |
+
check_done train test -s "$OUTRUN/sub_bf16/model.safetensors"
|
| 219 |
+
check_done eval_main test -s "$RESULTS/$(new_tag).json"
|
| 220 |
+
check_done eval_baseline test -s "$RESULTS/eval_baseline.json"
|
| 221 |
stage env 1 s_env
|
| 222 |
stage download 1 s_download
|
| 223 |
stage baseline 1 s_baseline
|
scripts/supervise.py
CHANGED
|
@@ -48,8 +48,16 @@ def remote_sha():
|
|
| 48 |
return m.group(1) if m else None
|
| 49 |
|
| 50 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 51 |
def install(sha):
|
| 52 |
dest = CODE / sha[:12]
|
|
|
|
|
|
|
| 53 |
if not dest.exists():
|
| 54 |
data = urllib.request.urlopen(f"https://codeload.github.com/{SOURCE}/tar.gz/{sha}", timeout=300).read()
|
| 55 |
tmp = CODE / f".{sha[:12]}.tmp"
|
|
@@ -118,9 +126,12 @@ def stop(proc):
|
|
| 118 |
|
| 119 |
def sync():
|
| 120 |
if VENV_PY.exists() and (ROOT / "repo" / "scripts" / "sync.py").exists():
|
| 121 |
-
|
| 122 |
-
|
| 123 |
-
|
|
|
|
|
|
|
|
|
|
| 124 |
|
| 125 |
|
| 126 |
def start_pipeline(extra):
|
|
@@ -141,7 +152,7 @@ def main():
|
|
| 141 |
if os.environ.get("RUN_ID") and not (STATE / "run_id").exists():
|
| 142 |
(STATE / "run_id").write_text(os.environ["RUN_ID"] + "\n")
|
| 143 |
head = ROOT / "repo" / ".commit"
|
| 144 |
-
current = head.read_text().strip() if
|
| 145 |
ran = (STATE / "pipeline_commit").read_text().strip() if (STATE / "pipeline_commit").exists() else ""
|
| 146 |
# force: (re)start the pipeline once on every supervisor start - finished stages are skipped,
|
| 147 |
# and a machine restart in the middle of a stage resumes there.
|
|
@@ -183,8 +194,16 @@ def main():
|
|
| 183 |
log(f"pipeline exited with code {proc.returncode}")
|
| 184 |
reported = True
|
| 185 |
sync()
|
| 186 |
-
if current and
|
| 187 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 188 |
ran, force, reported, last_sync = current, False, False, time.time()
|
| 189 |
(STATE / "pipeline_commit").write_text(current + "\n")
|
| 190 |
log(f"pipeline started (pid {proc.pid}) at {current[:12]}")
|
|
|
|
| 48 |
return m.group(1) if m else None
|
| 49 |
|
| 50 |
|
| 51 |
+
def complete(code_dir):
|
| 52 |
+
"""A code snapshot is usable only with its commit stamp and the pipeline script (folders
|
| 53 |
+
can survive a machine change without their files)."""
|
| 54 |
+
return (code_dir / ".commit").is_file() and (code_dir / "scripts" / "pipeline.sh").is_file()
|
| 55 |
+
|
| 56 |
+
|
| 57 |
def install(sha):
|
| 58 |
dest = CODE / sha[:12]
|
| 59 |
+
if dest.exists() and not complete(dest):
|
| 60 |
+
shutil.rmtree(dest, ignore_errors=True)
|
| 61 |
if not dest.exists():
|
| 62 |
data = urllib.request.urlopen(f"https://codeload.github.com/{SOURCE}/tar.gz/{sha}", timeout=300).read()
|
| 63 |
tmp = CODE / f".{sha[:12]}.tmp"
|
|
|
|
| 126 |
|
| 127 |
def sync():
|
| 128 |
if VENV_PY.exists() and (ROOT / "repo" / "scripts" / "sync.py").exists():
|
| 129 |
+
try:
|
| 130 |
+
with open(LOGS / "sync.log", "a") as f:
|
| 131 |
+
subprocess.run([str(VENV_PY), str(ROOT / "repo" / "scripts" / "sync.py")], stdout=f,
|
| 132 |
+
stderr=subprocess.STDOUT, env=dict(tool_env(), EMC_ROOT=str(ROOT)), timeout=1200)
|
| 133 |
+
except Exception as e: # noqa: BLE001 - an upload problem never stops the supervisor
|
| 134 |
+
log(f"log push failed: {type(e).__name__}: {e}")
|
| 135 |
|
| 136 |
|
| 137 |
def start_pipeline(extra):
|
|
|
|
| 152 |
if os.environ.get("RUN_ID") and not (STATE / "run_id").exists():
|
| 153 |
(STATE / "run_id").write_text(os.environ["RUN_ID"] + "\n")
|
| 154 |
head = ROOT / "repo" / ".commit"
|
| 155 |
+
current = head.read_text().strip() if complete(ROOT / "repo") else ""
|
| 156 |
ran = (STATE / "pipeline_commit").read_text().strip() if (STATE / "pipeline_commit").exists() else ""
|
| 157 |
# force: (re)start the pipeline once on every supervisor start - finished stages are skipped,
|
| 158 |
# and a machine restart in the middle of a stage resumes there.
|
|
|
|
| 194 |
log(f"pipeline exited with code {proc.returncode}")
|
| 195 |
reported = True
|
| 196 |
sync()
|
| 197 |
+
if current and not complete(ROOT / "repo"):
|
| 198 |
+
current = "" # snapshot lost its files: fetch it again next minute
|
| 199 |
+
log("code snapshot incomplete; reinstalling")
|
| 200 |
+
elif current and (current != ran or force):
|
| 201 |
+
try:
|
| 202 |
+
proc = start_pipeline(env)
|
| 203 |
+
except Exception as e: # noqa: BLE001 - retried next minute
|
| 204 |
+
log(f"pipeline start failed: {type(e).__name__}: {e}")
|
| 205 |
+
time.sleep(60)
|
| 206 |
+
continue
|
| 207 |
ran, force, reported, last_sync = current, False, False, time.time()
|
| 208 |
(STATE / "pipeline_commit").write_text(current + "\n")
|
| 209 |
log(f"pipeline started (pid {proc.pid}) at {current[:12]}")
|