CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-06 09:46:50 +00:00
Code Issues Packages Projects Releases Wiki Activity

memory : remove KV cache size padding (#16812)

Browse Source 
* memory : remove KV cache size padding

* cont : restore padding for n_kv tensor shape

* server : use slot context size instead of training context size

* server : simplify context limit logic
This commit is contained in:
Georgi Gerganov
2025-10-28 20:19:44 +02:00
committed by GitHub
parent a8ca18b4b8
commit 85a7d8677b
6 changed files with 14 additions and 54 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
src/llama-model.cpp
View File

@@ -19641,7 +19641,7 @@ struct llm_build_apertus : public llm_graph_context {
}
};
llama_memory_i * llama_model::create_memory(const llama_memory_params & params, llama_cparams & cparams) const {
llama_memory_i * llama_model::create_memory(const llama_memory_params & params, const llama_cparams & cparams) const {
llama_memory_i * res;
switch (arch) {
@@ -19692,17 +19692,13 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
};
}
const auto padding = llama_kv_cache::get_padding(cparams);
cparams.n_ctx = GGML_PAD(cparams.n_ctx, padding);
res = new llama_memory_hybrid(
/* model */ *this,
/* attn_type_k */ params.type_k,
/* attn_type_v */ params.type_v,
/* attn_v_trans */ !cparams.flash_attn,
/* attn_kv_size */ cparams.n_ctx,
/* attn_n_pad */ padding,
/* attn_n_pad */ 1,
/* attn_n_swa */ hparams.n_swa,
/* attn_swa_type */ hparams.swa_type,
/* recurrent_type_k */ GGML_TYPE_F32,
@@ -19714,23 +19710,12 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
/* filter_attn */ std::move(filter_attn),
/* filter_recr */ std::move(filter_recr));
} else {
const auto padding = llama_kv_cache::get_padding(cparams);
uint32_t n_ctx_per_stream = cparams.n_ctx;
if (!cparams.kv_unified) {
n_ctx_per_stream = (cparams.n_ctx + cparams.n_seq_max - 1)/cparams.n_seq_max;
n_ctx_per_stream = GGML_PAD(n_ctx_per_stream, padding);
cparams.n_ctx = n_ctx_per_stream*cparams.n_seq_max;
} else {
n_ctx_per_stream = GGML_PAD(n_ctx_per_stream, padding);
cparams.n_ctx = n_ctx_per_stream;
}
LLAMA_LOG_DEBUG("%s: n_ctx = %u (padded)\n", __func__, cparams.n_ctx);
llama_memory_i::layer_reuse_cb reuse = nullptr;
if (arch == LLM_ARCH_GEMMA3N) {
@@ -19757,7 +19742,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
n_ctx_per_stream,
cparams.n_seq_max,
cparams.n_ubatch,
padding,
1,
nullptr,
reuse);
} else {
@@ -19772,7 +19757,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
cparams.kv_unified,
n_ctx_per_stream,
cparams.n_seq_max,
padding,
1,
hparams.n_swa,
hparams.swa_type,
nullptr,