matt/ollama37
1
0
Fork 0
mirror of https://github.com/dogkeeper886/ollama37.git synced 2025-12-18 11:47:07 +00:00
Code Issues Packages Projects Releases Wiki Activity

llama: update vendored code to commit 46e3556 (#8308)

Browse Source
This commit is contained in:
Jeffrey Morgan
2025-01-08 11:22:01 -08:00
committed by GitHub
parent 57f038ec7b
commit 1deafd8254
305 changed files with 16048 additions and 12926 deletions
Download Patch File Download Diff File Expand all files Collapse all files

34
llama/patches/0014-llama-Ensure-KV-cache-is-fully-defragmented.patch
View File

@@ -19,12 +19,12 @@ multiple batches of processing until everything is complete.
1 file changed, 46 insertions(+), 53 deletions(-)
diff --git a/src/llama.cpp b/src/llama.cpp
index 4778a9ed..654e32bc 100644
index bac66c24..c95da45d 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -3025,6 +3025,13 @@ struct llama_kv_cache {
}
};
@@ -3536,6 +3536,13 @@ static struct ggml_tensor * llm_build_rwkv6_channel_mix(
return ggml_mul(ctx, r, llm_build_lora_mm(lctx, ctx, layer->channel_mix_value, k));
}
+// block of KV slots to move when defragging
+struct llama_kv_defrag_move {
@@ -33,10 +33,10 @@ index 4778a9ed..654e32bc 100644
+ uint32_t len;
+};
+
struct llama_control_vector {
std::vector<struct ggml_tensor *> tensors; // per layer
std::vector<ggml_context_ptr> ctxs;
@@ -10802,35 +10809,23 @@ struct llm_build_context {
struct llm_build_context {
const llama_model & model;
llama_context & lctx;
@@ -3712,35 +3719,23 @@ struct llm_build_context {
return gf;
}
@@ -78,7 +78,7 @@ index 4778a9ed..654e32bc 100644
ggml_tensor * view_v_src;
ggml_tensor * view_v_dst;
@@ -10838,31 +10833,29 @@ struct llm_build_context {
@@ -3748,31 +3743,29 @@ struct llm_build_context {
if (flash_attn) {
// NOTE: the V cache is not transposed when using flash attention
view_v_src = ggml_view_2d(ctx0, kv_self.v_l[il],
@@ -118,7 +118,7 @@ index 4778a9ed..654e32bc 100644
}
//LLAMA_LOG_INFO("gf->n_nodes = %d\n", gf->n_nodes);
@@ -17325,7 +17318,7 @@ struct llm_build_context {
@@ -10856,7 +10849,7 @@ struct llm_build_context {
}
};
@@ -127,7 +127,7 @@ index 4778a9ed..654e32bc 100644
llama_ubatch dummy = {};
dummy.equal_seqs = true;
@@ -17335,7 +17328,7 @@ static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const
@@ -10866,7 +10859,7 @@ static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const
llm.init();
@@ -136,7 +136,7 @@ index 4778a9ed..654e32bc 100644
llm.free();
@@ -18351,7 +18344,12 @@ static int llama_decode_internal(
@@ -11329,7 +11322,12 @@ static int llama_decode_internal(
kv_self.head = 0;
}
@@ -150,7 +150,7 @@ index 4778a9ed..654e32bc 100644
if (!slot) {
return 1;
}
@@ -18756,8 +18754,8 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) {
@@ -11735,8 +11733,8 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) {
//const int64_t t_start = ggml_time_us();
@@ -161,7 +161,7 @@ index 4778a9ed..654e32bc 100644
// each move requires 6*n_layer tensors (see build_defrag)
// - source view, destination view, copy operation
@@ -18821,19 +18819,11 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) {
@@ -11800,19 +11798,11 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) {
// are we moving a continuous block of memory?
bool cont = false;
@@ -181,7 +181,7 @@ index 4778a9ed..654e32bc 100644
cont = false;
continue;
}
@@ -18849,8 +18839,10 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) {
@@ -11828,8 +11818,10 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) {
kv_self.head = n_used;
if (!cont) {
@@ -193,7 +193,7 @@ index 4778a9ed..654e32bc 100644
}
nf++;
@@ -18860,22 +18852,16 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) {
@@ -11839,22 +11831,16 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) {
}
}
@@ -218,7 +218,7 @@ index 4778a9ed..654e32bc 100644
#if 0
// CPU defrag
@@ -18950,11 +18936,18 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) {
@@ -11929,11 +11915,18 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) {
#else
// ggml_graph defrag