969 B
969 B
paths
| paths | |
|---|---|
|
Mutate ScheduleBatch fields out of place
Never mutate a ScheduleBatch direct field in place — no .extend()/.append(),
+=/-=/|=, tensor .add_()/.fill_(), or slice/index assignment. Build the
new value and rebind the field:
# Bad
self.reqs.extend(other.reqs)
self.seq_lens.add_(1)
self.extend_lens[i] -= encoder_len
# Good
self.reqs = self.reqs + other.reqs
self.seq_lens = self.seq_lens + 1
lens = self.extend_lens[:]; lens[i] -= encoder_len; self.extend_lens = lens # loop a copy, rebind once
Why: copy() snapshots and the overlap scheduler's queued references rely on old
objects staying frozen (the prerequisite for the per-step immutable ScheduleBatch
refactor); derived values like seq_lens_sum only recompute safely when fields
are rebound.
Out of scope (mutation intended): penalizer buffers, CUDA graph static buffers,
ForwardBatch fields, attention-backend metadata.