tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-26 11:45:21 +00:00
Code Issues Packages Projects Releases Wiki Activity

kv-cache : add SWA support (#13194)

Browse Source 
* kv-cache : prepare for SWA

ggml-ci

* kv-cache : initial iSWA implementation

ggml-ci

* kv-cache : rework error recovery logic

ggml-ci

* models : fix Phi-3 SWA parameters

ggml-ci

* model : adjust Granite to rope factor changes

ggml-ci

* server : check if context can do shifts

ggml-ci

* iswa : for now, always enable shifts (experiment)

ggml-ci

* kv-cache : simplify SWA logic

ggml-ci

* kv-cache : apply defrag when we fail to find slots for the batch

ggml-ci

* llama : update docs about llama_decode

ggml-ci

* kv-cache : update warning logs when no space for the batch is available

ggml-ci

* llama : add llama_kv_self_seq_pos_min()

* kv-cache : keep track of partial SWA computes and print warnings

* server : disallow use cases involving partial SWA context

ggml-ci

* llama : add param to control SWA cache size

ggml-ci

* minor : clean-up

ggml-ci
This commit is contained in:
Georgi Gerganov
2025-05-20 08:05:46 +03:00
committed by GitHub
parent f0adb80bf7
commit e298d2fbd0
15 changed files with 1426 additions and 650 deletions
Download Patch File Download Diff File Expand all files Collapse all files

28
include/llama.h
View File

@ -361,10 +361,11 @@ extern "C" {
// Keep the booleans together and at the end of the struct to avoid misalignment during copy-by-value.
bool embeddings; // if true, extract embeddings (together with logits)
bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
bool flash_attn; // whether to use flash attention [EXPERIMENTAL]
bool no_perf; // whether to measure performance timings
bool op_offload; // whether to offload host tensor operations to device
bool offload_kqv; // offload the KQV ops (including the KV cache) to GPU
bool flash_attn; // use flash attention [EXPERIMENTAL]
bool no_perf; // measure performance timings
bool op_offload; // offload host tensor operations to device
bool swa_full; // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
};
// model quantization parameters
@ -730,10 +731,18 @@ extern "C" {
llama_pos p1,
int d);
// Returns the smallest position present in the KV cache for the specified sequence
// This is typically non-zero only for SWA caches
// Return -1 if the sequence is empty
LLAMA_API llama_pos llama_kv_self_seq_pos_min(
struct llama_context * ctx,
llama_seq_id seq_id);
// Returns the largest position present in the KV cache for the specified sequence
// Return -1 if the sequence is empty
LLAMA_API llama_pos llama_kv_self_seq_pos_max(
struct llama_context * ctx,
llama_seq_id seq_id);
llama_seq_id seq_id);
// Defragment the KV cache
// This will be applied:
@ -943,9 +952,12 @@ extern "C" {
// Requires KV cache.
// For encode-decoder contexts, processes the batch using the decoder.
// Positive return values does not mean a fatal error, but rather a warning.
// 0 - success
// 1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
// < 0 - error. the KV cache state is restored to the state before this call
// Upon non-zero return values, the KV cache state is restored to the state before this call
// 0 - success
// 1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
// 2 - aborted
// -1 - invalid input batch
// < -1 - error
LLAMA_API int32_t llama_decode(
struct llama_context * ctx,
struct llama_batch batch);