Skip to content

Metal: add _gs params for shared-expert and fused layer decode#589

Closed
RDouglasSharp wants to merge 1 commit into
JustVugg:devfrom
RDouglasSharp:pr-shared-gs
Closed

Metal: add _gs params for shared-expert and fused layer decode#589
RDouglasSharp wants to merge 1 commit into
JustVugg:devfrom
RDouglasSharp:pr-shared-gs

Conversation

@RDouglasSharp

Copy link
Copy Markdown
Contributor

Extend coli_metal_layer_decode with per-weight group-size (gs) parameters for all attention weights (qa/qb/kva/kvb/o) and shared-expert weights (shg/shu/shd). Update bind_gemv, AttnW, encode_attention, and the three shared-expert bind_gemv calls to forward gs. Update the layer_forward_rows call site in colibri.c.

Also includes the common bind_gemv/AttnW/enccode_attention changes from PR #588 for self-contained compilation; the two patches will merge cleanly.

Extend coli_metal_layer_decode with per-weight group-size (gs) params
for all attention weights (qa/qb/kva/kvb/o) and shared-expert weights
(shg/shu/shd). Update bind_gemv, AttnW struct, encode_attention, and
all shared-expert bind_gemv calls to forward gs. Update the colibri.c
layer_forward_rows call site to supply the new args.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant