shivank21's picture
Add files using upload-large-folder tool
cb49561 verified
Raw History Blame Contribute Delete
13.4 kB
import json
import math
import os
import select
import signal
import statistics
import struct
import subprocess
import sys
import time
import numpy as np
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
import graph_spec
_HDR = struct.Struct("<4sQ")
N_WORKERS = 8
N_CPUS = 8
WARMUPS = 3
REPS = 15
BUILD_TIMEOUT_SEC = 20.0
CALL_TIMEOUT_SEC = 120.0
ABS_TOL = 1e-3
REL_TOL = 1e-3
AGENT_UID = 1001
class WorkerError(RuntimeError):
"""Any failure of a child: crash, protocol violation, timeout, bad output."""
def _uid_of(pid):
try:
with open("/proc/%d/status" % pid, "rb") as fh:
for line in fh:
if line.startswith(b"Uid:"):
return int(line.split()[1])
except (OSError, IndexError, ValueError):
pass
return None
def _tasks_stopped(pids):
for pid in pids:
try:
tids = os.listdir("/proc/%d/task" % pid)
except OSError:
continue
for tid in tids:
try:
with open("/proc/%d/task/%s/stat" % (pid, tid), "rb") as fh:
state = fh.read().rsplit(b")", 1)[1].split()[0]
except (OSError, IndexError):
continue
if state not in (b"T", b"Z", b"X"):
return False
return True
class Child:
"""A worker subprocess, SIGSTOPped whenever it is not the one being timed."""
def __init__(self, kind, rundir, python, launch_prefix=(), env=None,
popen_kwargs=None):
self.kind = kind
argv = list(launch_prefix) + [python, "-I", "-B",
os.path.join(rundir, "worker.py"), kind]
self.proc = subprocess.Popen(
argv, stdin=subprocess.PIPE, stdout=subprocess.PIPE,
stderr=subprocess.PIPE, cwd=rundir, env=env, close_fds=True,
start_new_session=True, **(popen_kwargs or {}))
self.rfd = self.proc.stdout.fileno()
self.wfd = self.proc.stdin.fileno()
self.stopped = False
self._pids = None
self._expect(b"HI", BUILD_TIMEOUT_SEC)
def _send(self, cmd, payload=b""):
buf = _HDR.pack(cmd.ljust(4, b"\0"), len(payload)) + payload
off = 0
deadline = time.monotonic() + CALL_TIMEOUT_SEC
while off < len(buf):
if time.monotonic() > deadline:
raise WorkerError("%s: timed out sending a request" % self.kind)
_, w, _ = select.select([], [self.wfd], [], 1.0)
if not w:
continue
try:
off += os.write(self.wfd, buf[off:off + (1 << 20)])
except OSError as exc:
raise WorkerError("%s: pipe closed while sending (%s)" % (self.kind, exc))
def _read_exactly(self, n, deadline):
out = bytearray()
while len(out) < n:
remaining = deadline - time.monotonic()
if remaining <= 0:
raise WorkerError("%s: timed out waiting for a response" % self.kind)
r, _, _ = select.select([self.rfd], [], [], min(1.0, remaining))
if not r:
continue
chunk = os.read(self.rfd, n - len(out))
if not chunk:
raise WorkerError("%s: exited before responding (rc=%s)"
% (self.kind, self.proc.poll()))
out += chunk
return bytes(out)
def _expect(self, cmd, timeout):
deadline = time.monotonic() + timeout
got, ln = _HDR.unpack(self._read_exactly(_HDR.size, deadline))
payload = self._read_exactly(ln, deadline) if ln else b""
got = got.rstrip(b"\0")
if got != cmd:
raise WorkerError("%s: expected %r, got %r" % (self.kind, cmd, got))
return payload
def _signal(self, sig):
try:
os.killpg(self.proc.pid, sig)
except OSError:
pass
def tree_pids(self, refresh=False):
"""The child's own process tree, including anything it started itself."""
if self._pids is not None and not refresh:
return self._pids
seen, stack = [], [self.proc.pid]
while stack:
pid = stack.pop()
seen.append(pid)
try:
for tid in os.listdir("/proc/%d/task" % pid):
with open("/proc/%d/task/%s/children" % (pid, tid)) as fh:
stack.extend(int(v) for v in fh.read().split())
except (OSError, ValueError):
pass
self._pids = seen
return seen
def freeze(self):
"""SIGSTOP, then wait until every thread has really stopped.
Returning early would let this child's threads -- BLAS worker threads busy-spin
for a while after the last call returns -- drain into the other child's timed
region, which would bias whichever executor is measured second.
"""
if not self.stopped:
self._signal(signal.SIGSTOP)
self.stopped = True
deadline = time.monotonic() + 2.0
while not _tasks_stopped(self.tree_pids()):
if time.monotonic() > deadline:
break
time.sleep(0.0002)
def thaw(self):
if self.stopped:
self._signal(signal.SIGCONT)
self.stopped = False
self._send(b"PING")
self._expect(b"PONG", CALL_TIMEOUT_SEC)
def build(self, seed):
t0 = time.perf_counter()
self._send(b"BLD", str(int(seed)).encode())
self._expect(b"RDY", BUILD_TIMEOUT_SEC)
return time.perf_counter() - t0
def call(self, xb):
"""Round-trip one input. Returns (seconds, header, raw output bytes)."""
t0 = time.perf_counter()
self._send(b"CALL", xb)
payload = self._expect(b"OUT", CALL_TIMEOUT_SEC)
dt = time.perf_counter() - t0
head, _, body = payload.partition(b"|")
return dt, head.decode("utf-8", "replace"), body
def stderr_tail(self, limit=2000):
self._signal(signal.SIGCONT)
self._signal(signal.SIGKILL)
try:
return self.proc.stderr.read().decode("utf-8", "replace")[-limit:]
except Exception:
return ""
def close(self):
self._signal(signal.SIGCONT)
self.stopped = False
try:
self._send(b"EXIT")
self.proc.wait(timeout=10)
except Exception:
self._signal(signal.SIGKILL)
try:
self.proc.wait(timeout=10)
except subprocess.TimeoutExpired:
pass
for stream in (self.proc.stdin, self.proc.stdout, self.proc.stderr):
try:
stream.close()
except Exception:
pass
def _quiesce(active):
"""Stop every unprivileged process that is not part of `active`'s tree.
An executor that detaches a helper into its own session would otherwise keep burning
cycles while the other executor is on the clock. Only the privileged harness can do
this, so the developer-side harness (which runs unprivileged) skips it; the minimum
statistic is what protects a measurement there.
"""
if os.geteuid() != 0:
return 0
keep_pids = set(active.tree_pids(refresh=True))
strays = 0
try:
pids = [int(p) for p in os.listdir("/proc") if p.isdigit()]
except OSError:
return 0
for pid in pids:
if pid in keep_pids or _uid_of(pid) != AGENT_UID:
continue
try:
os.kill(pid, signal.SIGSTOP)
strays += 1
except OSError:
pass
return strays
def _check_output(head, body, spec):
"""Validate a returned tensor. Returns the array or raises WorkerError."""
T, d = spec["T"], spec["d_model"]
try:
ndim, dtype, dims = head.split(",")
dims = tuple(int(v) for v in dims.split("x")) if dims else ()
except ValueError:
raise WorkerError("unparseable output header %r" % head)
if int(ndim) != 2 or dims != (T, d):
raise WorkerError("output shape %r, expected (%d, %d)" % (dims, T, d))
if dtype != "float32":
raise WorkerError("output dtype %r, expected float32" % dtype)
if len(body) != T * d * 4:
raise WorkerError("output payload is %d bytes, expected %d"
% (len(body), T * d * 4))
return np.frombuffer(body, dtype=np.float32).reshape(T, d)
def _errors(y, y_ref):
a = y.astype(np.float64)
b = y_ref.astype(np.float64)
if not np.all(np.isfinite(a)):
return float("inf"), float("inf")
diff = np.abs(a - b)
max_abs = float(diff.max()) if diff.size else 0.0
denom = float(np.linalg.norm(b))
rel = float(np.linalg.norm(diff) / denom) if denom > 0 else max_abs
return max_abs, rel
def measure_instance(seed, ref, sub, warmups=WARMUPS, reps=REPS):
"""Time one instance. Returns a record; raises WorkerError on any child failure."""
spec = graph_spec.sample_spec(seed)
xbs = [np.ascontiguousarray(x).tobytes()
for x in graph_spec.build_inputs(spec, warmups + reps)]
ref.freeze()
sub.thaw()
build_sub = sub.build(seed)
sub.freeze()
ref.thaw()
build_ref = ref.build(seed)
ref.freeze()
times = {id(ref): [], id(sub): []}
max_abs = 0.0
max_rel = 0.0
strays = 0
for i, xb in enumerate(xbs):
first, second = (ref, sub) if i % 2 == 0 else (sub, ref)
got = {}
for child in (first, second):
other = sub if child is ref else ref
other.freeze()
strays += _quiesce(child)
child.thaw()
dt, head, body = child.call(xb)
child.freeze()
got[id(child)] = (dt, head, body)
y_ref = _check_output(got[id(ref)][1], got[id(ref)][2], spec)
y_sub = _check_output(got[id(sub)][1], got[id(sub)][2], spec)
a, r = _errors(y_sub, y_ref)
max_abs = max(max_abs, a)
max_rel = max(max_rel, r)
if i >= warmups:
times[id(ref)].append(got[id(ref)][0])
times[id(sub)].append(got[id(sub)][0])
ref_t, sub_t = times[id(ref)], times[id(sub)]
ref_best, sub_best = min(ref_t), min(sub_t)
equivalent = bool(max_abs <= ABS_TOL and max_rel <= REL_TOL)
return {
"seed": int(seed),
"shape": {k: spec[k] for k in
("d_model", "d_ff", "n_layers", "n_heads", "n_kv_heads", "T")},
"ref_best_s": ref_best,
"sub_best_s": sub_best,
"ref_median_s": statistics.median(ref_t),
"sub_median_s": statistics.median(sub_t),
"ref_jitter": statistics.median(ref_t) / ref_best - 1.0,
"sub_jitter": statistics.median(sub_t) / sub_best - 1.0,
"build_ref_s": build_ref,
"build_sub_s": build_sub,
"strays_stopped": strays,
"max_abs_err": max_abs,
"rel_fro_err": max_rel,
"equivalent": equivalent,
"ratio": (ref_best / sub_best) if (equivalent and sub_best > 0) else 0.0,
}
def child_env():
"""The environment both children get. BLAS gets the full worker budget."""
return {
"PATH": "/usr/local/bin:/usr/local/sbin:/usr/sbin:/usr/bin:/sbin:/bin",
"HOME": "/tmp",
"LANG": "C.UTF-8",
"PYTHONDONTWRITEBYTECODE": "1",
"PYTHONHASHSEED": "0",
"OPENBLAS_NUM_THREADS": str(N_WORKERS),
"OMP_NUM_THREADS": str(N_WORKERS),
"MKL_NUM_THREADS": str(N_WORKERS),
"NUMEXPR_NUM_THREADS": str(N_WORKERS),
"EXEC_N_WORKERS": str(N_WORKERS),
"EXEC_N_CPUS": str(N_CPUS),
}
def measure(seeds, rundir, python=sys.executable, launch_prefix=(),
popen_kwargs=None):
"""Measure every instance and aggregate. Returns the metric record."""
env = child_env()
ref = sub = None
per_instance = []
failure = None
stderr_tail = ""
try:
ref = Child("ref", rundir, python, (), env, None)
sub = Child("sub", rundir, python, launch_prefix, env, popen_kwargs)
for seed in seeds:
per_instance.append(measure_instance(int(seed), ref, sub))
except WorkerError as exc:
failure = str(exc)
except Exception as exc:
failure = "%s: %s" % (type(exc).__name__, exc)
finally:
if failure is not None and sub is not None:
stderr_tail = sub.stderr_tail()
for child in (ref, sub):
if child is not None:
try:
child.close()
except Exception:
pass
n_ok = sum(1 for r in per_instance if r["equivalent"])
valid = bool(failure is None and len(per_instance) == len(seeds) == n_ok)
if valid:
metric = math.exp(
sum(math.log(r["ratio"]) for r in per_instance) / len(per_instance))
else:
metric = 0.0
return {
"metric_name": "geometric_mean_speedup",
"metric": metric,
"valid": valid,
"failure": failure,
"stderr_tail": stderr_tail,
"n_instances": len(seeds),
"n_equivalent": n_ok,
"per_instance": per_instance,
}
def dump(record, path):
with open(path, "w") as fh:
json.dump(record, fh, indent=2, sort_keys=True)