-
Notifications
You must be signed in to change notification settings - Fork 10
Pull requests: vcruz305/vllm-exl3
Author
Label
Projects
Milestones
Reviews
Assignee
Sort
Pull requests list
fix(loader): don't pre-filter speculative-draft expert weights with the main-stack window (+9-20% decode)
#35
opened Sep 20, 2026 by
fattchris
Contributor
Loading…
fix(moe): default the native codebook gate to MCG instead of refusing (fixes 103 failing tests on main)
#34
opened Sep 20, 2026 by
fattchris
Contributor
Loading…
feat(moe): fixed-shape padded MoE with zero host sync — capturable in CUDA graphs
#33
opened Sep 20, 2026 by
fattchris
Contributor
Loading…
fix(moe): expert codebook as a build-time parameter (P2B_CB) + fail-closed check
#32
opened Sep 20, 2026 by
fattchris
Contributor
Loading…
feat: multi-K fused MoE kernel (p2b_fused_moe_mk) — per-expert K tables, one cooperative launch
#31
opened Sep 16, 2026 by
fattchris
Contributor
Loading…
perf: GB10 load-path — pinned bounce, POPULATE_READ, batched markers, prescan cache
#28
opened Sep 16, 2026 by
fattchris
Contributor
Loading…
ProTip!
Filter pull requests by the default branch with base:main.