Skip to content

Slice returned routed_experts at routed_experts_start_len - #45

Draft
nightlessbaron wants to merge 2 commits into
prodfrom
feat/routed-experts-delta
Draft

Slice returned routed_experts at routed_experts_start_len#45
nightlessbaron wants to merge 2 commits into
prodfrom
feat/routed-experts-delta

Conversation

@nightlessbaron

Copy link
Copy Markdown

Plumbs the existing routed_experts_start_len request field through chat completions: the tokenizer manager slices the returned expert tensor at the offset, so multi-turn clients that already hold the prefix rows receive only the rows for newly processed tokens. Default 0 keeps the full echo unchanged.

Motivation: with 61 layers × topk 8 × int32 the expert echo is 1952 B/token; echoing the full context every turn made TITO responses O(context) (~200MB at 16k ctx for ~110-token turns) and saturated the gateway NIC. Client-side delta merge lands in the companion miles PR.

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