Skip to content

memory: Hybrid context shift #17009

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 5 commits into from
Nov 10, 2025
Merged

memory: Hybrid context shift #17009

Show file tree
Hide file tree
Changes from 4 commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
4fb572c
feat(memory): Only fail partial erasure of recurrent tail
gabe-l-hart Nov 4, 2025
24aff7f
fix(main): Check the output of seq_rm for prefix matching
gabe-l-hart Nov 4, 2025
3b59021
fix(memory): Fix condition for partial erasure failure if p0 > pos
gabe-l-hart Nov 4, 2025
b8f9c56
style: Fix extra parens
gabe-l-hart Nov 5, 2025
104d8ac
fix(main.cpp): Set n_matching_session_tokens to 0 on cache clear
gabe-l-hart Nov 6, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 4 additions & 3 deletions src/llama-memory-recurrent.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -151,7 +151,8 @@ bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
p1 = std::numeric_limits<llama_pos>::max();
}

// models like Mamba or RWKV can't have a state partially erased
// models like Mamba or RWKV can't have a state partially erased at the end
// of the sequence because their state isn't preserved for previous tokens
if (seq_id >= (int64_t) size) {
// could be fatal
return false;
Expand All @@ -160,8 +161,8 @@ bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
int32_t & tail_id = cells[seq_id].tail;
if (tail_id >= 0) {
const auto & cell = cells[tail_id];
// partial intersection is invalid
if ((0 < p0 && p0 < cell.pos) || (0 < p1 && p1 <= cell.pos)) {
// partial intersection is invalid if it includes the final pos
if (0 < p0 && p0 <= cell.pos && p1 > cell.pos) {
//printf("[DEBUG] inside `llama_memory_recurrent::seq_rm`: partial intersection is invalid, so returning false\n");
return false;
}
Expand Down
5 changes: 4 additions & 1 deletion tools/main/main.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -354,7 +354,10 @@ int main(int argc, char ** argv) {
}

// remove any "future" tokens that we might have inherited from the previous session
llama_memory_seq_rm(mem, -1, n_matching_session_tokens, -1);
if (!llama_memory_seq_rm(mem, -1, n_matching_session_tokens, -1)) {
LOG_INF("%s: unable to resuse common prefix\n", __func__);
llama_memory_seq_rm(mem, -1, -1, -1);
Comment on lines 357 to 360
Copy link
Collaborator

@compilade compilade Nov 5, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm wondering if this will correctly trigger reevaluation of the context afterwards.

n_matching_session_tokens seems to be used mostly for cosmetic purposes except when it's equal to embd_inp.size(), so maybe it could be set to 0.

But I think there might be something else necessary here. (For example, should session_tokens be cleared?)

(This is mostly relevant when reloading a saved session, and not necessarily during context shifting, though)

Copy link
Collaborator Author

@gabe-l-hart gabe-l-hart Nov 5, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hm, yeah, I'm still a little murky on the logic going on here. I'll look more carefully and see what I can find.

Copy link

@leok7v leok7v Nov 6, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Sorry folks,
I am new to merge management of commits.
b8f9c56923ab34bf24d4fe5476efe9b02736918d
Don't know who and when presses merge button for the commit after all checks are passed?

Copy link
Collaborator Author

@gabe-l-hart gabe-l-hart Nov 6, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think we're holding on this until I resolve the question in this thread. I'll try to get to that today!

Copy link
Collaborator Author

@gabe-l-hart gabe-l-hart Nov 6, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ok, I see what you're getting at here with the clause below that recalculates the cached logits for the final token. I think the right thing to do here is to set n_matching_session_tokens to 0 since we're essentially saying there are no matching tokens anymore.

}
}

LOG_DBG("recalculate the cached logits (check): embd_inp.size() %zu, n_matching_session_tokens %zu, embd_inp.size() %zu, session_tokens.size() %zu\n",
Expand Down
Loading