Skip to content

llama : allocate GLM_DSA indexer cache only in "full" indexer layers #26474

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
fairydreaming merged 2 commits into from
Aug 3, 2026
+11 −7
Merged

llama : allocate GLM_DSA indexer cache only in "full" indexer layers #26474

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
bda1f55
llama : allocate indexer cache only in "full" indexer layers
sszymczy Aug 2, 2026
32326ca
Merge remote-tracking branch 'upstream/master' into glm-lid-cache-filter
sszymczy Aug 3, 2026
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-kv-cache-dsa.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,7 +23,8 @@ llama_kv_cache_dsa::llama_kv_cache_dsa(
uint32_t n_pad,
uint32_t n_swa,
llama_swa_type swa_type,
const layer_filter_cb & filter,
const layer_filter_cb & filter_mla,
const layer_filter_cb & filter_lid,
const layer_reuse_cb & reuse) :
hparams_lid(model.hparams), n_stream(unified ? 1 : n_seq_max) {

Expand All @@ -32,7 +33,7 @@ llama_kv_cache_dsa::llama_kv_cache_dsa(
kv_mla = std::make_unique<llama_kv_cache>(
model, model.hparams, type_k, type_v,
v_trans, offload, unified, kv_size, n_seq_max, n_pad,
n_swa, swa_type, nullptr, filter, reuse, nullptr);
n_swa, swa_type, nullptr, filter_mla, reuse, nullptr);

// we use llama_kv_cache for caching indexer keys
// by hand-tweaking some hparams we fool it to create
Expand All @@ -49,7 +50,7 @@ llama_kv_cache_dsa::llama_kv_cache_dsa(
kv_lid = std::make_unique<llama_kv_cache>(
model, hparams_lid, type_k, type_v,
v_trans, offload, unified, kv_size, n_seq_max, n_pad,
n_swa, swa_type, nullptr, filter, reuse, nullptr);
n_swa, swa_type, nullptr, filter_lid, reuse, nullptr);
}

void llama_kv_cache_dsa::clear(bool data) {
Expand Down
3 changes: 2 additions & 1 deletion src/llama-kv-cache-dsa.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,7 +26,8 @@ class llama_kv_cache_dsa : public llama_memory_i {
uint32_t n_pad,
uint32_t n_swa,
llama_swa_type swa_type,
const layer_filter_cb & filter,
const layer_filter_cb & filter_mla,
const layer_filter_cb & filter_lid,
const layer_reuse_cb & reuse);

~llama_kv_cache_dsa() = default;
Expand Down
8 changes: 5 additions & 3 deletions src/llama-model.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2101,10 +2101,11 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
} else {
// Main context: DSA cache for the trunk layers only - the nextn
// layer(s) are never attended by the trunk graph.
llama_kv_cache::layer_filter_cb filter = nullptr;
llama_kv_cache::layer_filter_cb filter_mla = nullptr;
if (hparams.n_layer_nextn > 0) {
filter = [&](uint32_t il) { return il < hparams.n_layer(); };
filter_mla = [&](uint32_t il) { return il < hparams.n_layer(); };
}
llama_kv_cache::layer_filter_cb filter_lid = [&](uint32_t il) { return il < hparams.n_layer() && (arch != LLM_ARCH_GLM_DSA || hparams.is_indexer_full(il)); };

res = new llama_kv_cache_dsa(
*this,
Expand All @@ -2118,7 +2119,8 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
1,
hparams.n_swa,
hparams.swa_type,
filter,
filter_mla,
filter_lid,
nullptr);
}
} break;
Expand Down
Loading