pullrequests Search Results · bucket language:JavaScript language:JavaScript language:JavaScript language:Python language:Python
Filter by
904k results
The multi-observation mapmaker s per-bucket accumulation loop is replaced by a single map_reduce call on a HostEnsemble
of the bucket s slots.
What changed
_accumulate_bucket builds a HostEnsemble ...
... .*, which SGLang remaps only after each bucket arrives. Miles therefore split wq_a + wkv or
compressor.wkv + wgate across update calls, and SGLang asserted on a half-filled fusion cache.
Validation ...
Draft until it passes on both trials and rsWVPRunner v3.2.4 (final) is published. This branch builds now: its files come
from the v3.2.4-rc1 pre-release, not the bucket.
Why
Web VPython 3.2.3 ...
... (sliding-window): the window is position-based, so a short state would lose real context past (L_max − W).
v1 solution is length-bucketing: group only requests whose state_ids tokenise to equal length ...
... :
- It kept a list of timestamps for each tenant:operation in a module-level singleton.
- Buckets were never evicted, so rotating tenant IDs grew memory without bound.
- Limits were hard-coded ...