pullrequests Search Results · bucket language:JavaScript language:JavaScript language:HTML language:JavaScript language:Python
Filter by
997k results
... token-classification forward.
Sequence variants are bounded:
- Buckets are multiples of the largest attention window, up to the position limit.
- Each bucket gets a per-layer attention working ...
feature
P1
roadmap-v3
... is presented).
Problem
HPU decode bucketing pads decode batches to fixed bucket sizes: real requests come first; pad rows ride along as
zero-length query segments with sentinel state indices. The ...
... correctly returns 0, but the intensity line used a truthy check (suds ? bucket : 2). Zero took the
missing-rating default, energy 2.
Fix and validation
Use suds == null so 0 goes through the existing 1–5 ...
Fixes the production boot crash undefined is not an object (evaluating n.concreteDark.push ) in v6.9 landmark ruin
archetypes by initializing the concreteDark geometry bucket in all three landmark ...
... . And c=16 steps average 5.7 rows, so every batch size up to 8
needs a graph.
What changed
Stacked on #1099 with #1101 picked first (it brings #1100).
- The 1344x768 bucket captured and its VAE ...
area/api
area/benchmark
area/conductor
area/distributed
area/engine
area/graph
area/infra
area/integrations
area/model-shared
area/observability
area/other
area/resources
area/rust
area/transport
area/utils
area/worker
model/many
size/XXL
Relevant issues
Closes #50586
Changes
_ensure_litellm_metadata (unified guardrail): a request that already has a metadata bucket (chat completions) gets the
user_api_key_* fields there through ...