Skip to content

Commit f0cb485

Browse files
gabe-l-hartgabe-l-hart
committed
fix: Use per-layer sizing everywhere in kv caches
Branch: GraniteFour Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
1 parent e00e525 commit f0cb485

File tree

2 files changed

+16
-16
lines changed

2 files changed

+16
-16
lines changed

‎src/llama-kv-cache-recurrent.cpp

Lines changed: 8 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -69,8 +69,8 @@ llama_kv_cache_recurrent::llama_kv_cache_recurrent(
6969
continue;
7070
}
7171

72-
const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i) + hparams.n_embd_k_s();
73-
const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i) + hparams.n_embd_v_s();
72+
const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i) + hparams.n_embd_k_s(i);
73+
const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i) + hparams.n_embd_v_s(i);
7474

7575
const char * dev_name = "CPU";
7676

@@ -771,7 +771,7 @@ void llama_kv_cache_recurrent::state_write_data(llama_io_write_i & io, const std
771771
// Iterate and write all the keys first, each row is a cell
772772
// Get whole range at a time
773773
for (uint32_t il = 0; il < n_layer; ++il) {
774-
const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s();
774+
const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s(il);
775775

776776
// Write key type
777777
const int32_t k_type_i = (int32_t)k_l[il]->type;
@@ -791,7 +791,7 @@ void llama_kv_cache_recurrent::state_write_data(llama_io_write_i & io, const std
791791

792792
if (!v_trans) {
793793
for (uint32_t il = 0; il < n_layer; ++il) {
794-
const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
794+
const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(il);
795795

796796
// Write value type
797797
const int32_t v_type_i = (int32_t)v_l[il]->type;
@@ -812,7 +812,7 @@ void llama_kv_cache_recurrent::state_write_data(llama_io_write_i & io, const std
812812
// When v is transposed, we also need the element size and get the element ranges from each row
813813
const uint32_t kv_size = size;
814814
for (uint32_t il = 0; il < n_layer; ++il) {
815-
const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
815+
const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(il);
816816

817817
// Write value type
818818
const int32_t v_type_i = (int32_t)v_l[il]->type;
@@ -959,7 +959,7 @@ bool llama_kv_cache_recurrent::state_read_data(llama_io_read_i & io, uint32_t ce
959959

960960
// For each layer, read the keys for each cell, one row is one cell, read as one contiguous block
961961
for (uint32_t il = 0; il < n_layer; ++il) {
962-
const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s();
962+
const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s(il);
963963

964964
// Read type of key
965965
int32_t k_type_i_ref;
@@ -987,7 +987,7 @@ bool llama_kv_cache_recurrent::state_read_data(llama_io_read_i & io, uint32_t ce
987987

988988
if (!v_trans) {
989989
for (uint32_t il = 0; il < n_layer; ++il) {
990-
const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
990+
const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(il);
991991

992992
// Read type of value
993993
int32_t v_type_i_ref;
@@ -1015,7 +1015,7 @@ bool llama_kv_cache_recurrent::state_read_data(llama_io_read_i & io, uint32_t ce
10151015
} else {
10161016
// For each layer, read the values for each cell (transposed)
10171017
for (uint32_t il = 0; il < n_layer; ++il) {
1018-
const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
1018+
const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(il);
10191019

10201020
// Read type of value
10211021
int32_t v_type_i_ref;

‎src/llama-kv-cache-unified.cpp

Lines changed: 8 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -68,8 +68,8 @@ llama_kv_cache_unified::llama_kv_cache_unified(
6868
continue;
6969
}
7070

71-
const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s();
72-
const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
71+
const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s(il);
72+
const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(il);
7373

7474
const char * dev_name = "CPU";
7575

@@ -1369,7 +1369,7 @@ void llama_kv_cache_unified::state_write_data(llama_io_write_i & io, const std::
13691369
for (const auto & layer : layers) {
13701370
const uint32_t il = layer.il;
13711371

1372-
const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s();
1372+
const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s(il);
13731373

13741374
// Write key type
13751375
const int32_t k_type_i = (int32_t)layer.k->type;
@@ -1391,7 +1391,7 @@ void llama_kv_cache_unified::state_write_data(llama_io_write_i & io, const std::
13911391
for (const auto & layer : layers) {
13921392
const uint32_t il = layer.il;
13931393

1394-
const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
1394+
const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(il);
13951395

13961396
// Write value type
13971397
const int32_t v_type_i = (int32_t)layer.v->type;
@@ -1415,7 +1415,7 @@ void llama_kv_cache_unified::state_write_data(llama_io_write_i & io, const std::
14151415
for (const auto & layer : layers) {
14161416
const uint32_t il = layer.il;
14171417

1418-
const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
1418+
const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(il);
14191419

14201420
// Write value type
14211421
const int32_t v_type_i = (int32_t)layer.v->type;
@@ -1558,7 +1558,7 @@ bool llama_kv_cache_unified::state_read_data(llama_io_read_i & io, uint32_t cell
15581558
for (const auto & layer : layers) {
15591559
const uint32_t il = layer.il;
15601560

1561-
const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s();
1561+
const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s(il);
15621562

15631563
// Read type of key
15641564
int32_t k_type_i_ref;
@@ -1588,7 +1588,7 @@ bool llama_kv_cache_unified::state_read_data(llama_io_read_i & io, uint32_t cell
15881588
for (const auto & layer : layers) {
15891589
const uint32_t il = layer.il;
15901590

1591-
const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
1591+
const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(il);
15921592

15931593
// Read type of value
15941594
int32_t v_type_i_ref;
@@ -1618,7 +1618,7 @@ bool llama_kv_cache_unified::state_read_data(llama_io_read_i & io, uint32_t cell
16181618
for (const auto & layer : layers) {
16191619
const uint32_t il = layer.il;
16201620

1621-
const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
1621+
const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(il);
16221622

16231623
// Read type of value
16241624
int32_t v_type_i_ref;

0 commit comments

Comments
 (0)