wylab/llama.cpp
0
0
Fork 1
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-14 18:02:52 +02:00
Code Issues Packages Projects Releases Wiki Activity

server : add comment

Browse Source 
[no ci]
This commit is contained in:
Georgi Gerganov
2026-05-08 08:50:23 +03:00
parent 12c7cfbe83
commit 233d1aee69
1 changed files with 12 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/server/server-context.cpp
+12 -2
View File
@@ -2968,7 +2968,17 @@ private:
}
// TODO: avoid restoring the draft context and re-evaluating the drafted tokens when not needed [TAG_SPEC_AVOID_DRAFT_REEVAL]
// for methods that require the target embeddings, I think we have to re-evaluate the draft tokens?
// for now, always re-evaluate for simplicity
// ref: https://github.com/ggml-org/llama.cpp/pull/22728#issuecomment-4400925384
//
// | spec type | need re-eval |
// | --- | --- |
// | draft model | no |
// | MTP (std) | yes |
// | MTP Gemma4 | no |
// | Eagle3 | yes |
// | DFlash | yes? |
//
if (ctx_drft) {
// TODO: update as needed for MTP, Eagle3, etc.
const bool need_tgt_embd = false;
@@ -2978,7 +2988,7 @@ private:
}
// the logic here varies depending on the speculative decoding method
// - some draft contexts require emebeddings from the target context, others don't
// - some draft contexts require embeddings from the target context, others don't
// - some draft contexts involve an encoder step to transform the target embeddings to draft embeddings
// TODO: extract this in a function ?
{