Skip to content

FreeToken brings datacenter-scale model serving to your desktop. Run massive models locally, fast and efficiently.

fix(checkpoint): FTW truncated shard raises OSError instead of silent garbage load (#25)

Lancerbd372b630a02
diff --git a/python/freetoken/checkpoint/ftw.py b/python/freetoken/checkpoint/ftw.py
index 9d6df684..a365e3d5 100644
--- a/python/freetoken/checkpoint/ftw.py
+++ b/python/freetoken/checkpoint/ftw.py
@@ -69,18 +69,19 @@ def layer_bank_entry_name(bank_name: str, layer_id: int) -> str:
def _pread_into(fd: int, mv: memoryview, offset: int) -> None:
"""POSIX positional read into ``mv`` at ``offset``, looping over any short preadv.
- preadv may return short (a signal, or the EOF-adjacent tail); the bare call this
- replaces ignored the return value, so a short read silently left the tail of the
- destination unfilled -- garbage bytes in the middle of a weight tensor with no
- error anywhere. Resuming at the running offset stays O_DIRECT-legal: the writer
- pads every tensor to ALIGN and cuts shards at ALIGN boundaries, so direct-IO
- short reads land on block boundaries."""
+ preadv may return short (a signal, or the EOF-adjacent tail); the loop resumes
+ at the running offset, which stays O_DIRECT-legal: the writer pads every tensor
+ to ALIGN and cuts shards at ALIGN boundaries, so direct-IO short reads land on
+ block boundaries. EOF before the buffer is filled raises ``OSError`` — a
+ truncated shard must not silently load garbage weights."""
done = 0
total = len(mv)
while done < total:
n = os.preadv(fd, [mv[done:]], offset + done)
if n == 0:
- break # EOF
+ raise OSError(
+ f"unexpected EOF reading FTW: got {done}/{total} bytes at offset {offset}"
+ )
done += n
@@ -318,11 +319,20 @@ class FTWReader:
if self._direct:
def rd(job):
file, fo, do, ln = job
- _pread_into(self._fd(file), dest[do:do + ln], fo)
+ try:
+ _pread_into(self._fd(file), dest[do:do + ln], fo)
+ except OSError as e:
+ raise OSError(f"shard {file}: {e}") from e
else:
def rd(job):
file, fo, do, ln = job
- dest[do:do + ln] = self._map(file)[fo:fo + ln] # memcpy from the mapping
+ mv = self._map(file)
+ if fo + ln > len(mv):
+ raise OSError(
+ f"unexpected EOF reading FTW: shard {file} has "
+ f"{len(mv)} bytes, need {ln} at offset {fo}"
+ )
+ dest[do:do + ln] = mv[fo:fo + ln]
if len(jobs) <= 1:
for j in jobs: