CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-09 10:17:06 +00:00
Code Issues Packages Projects Releases Wiki Activity

context : fix n_ctx_per_seq computation

Browse Source
This commit is contained in:
Georgi Gerganov
2025-10-23 14:51:26 +03:00
parent a42fb77147
commit 492f628c58
2 changed files with 7 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
src/llama-context.cpp
View File

@@ -112,11 +112,9 @@ llama_context::llama_context(
} }
} }
const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max;
LLAMA_LOG_INFO("%s: n_seq_max = %u\n", __func__, cparams.n_seq_max); LLAMA_LOG_INFO("%s: n_seq_max = %u\n", __func__, cparams.n_seq_max);
LLAMA_LOG_INFO("%s: n_ctx = %u\n", __func__, cparams.n_ctx); LLAMA_LOG_INFO("%s: n_ctx = %u\n", __func__, cparams.n_ctx);
LLAMA_LOG_INFO("%s: n_ctx_per_seq = %u\n", __func__, n_ctx_per_seq); LLAMA_LOG_INFO("%s: n_ctx_per_seq = %u\n", __func__, n_ctx_per_seq());
LLAMA_LOG_INFO("%s: n_batch = %u\n", __func__, cparams.n_batch); LLAMA_LOG_INFO("%s: n_batch = %u\n", __func__, cparams.n_batch);
LLAMA_LOG_INFO("%s: n_ubatch = %u\n", __func__, cparams.n_ubatch); LLAMA_LOG_INFO("%s: n_ubatch = %u\n", __func__, cparams.n_ubatch);
LLAMA_LOG_INFO("%s: causal_attn = %d\n", __func__, cparams.causal_attn); LLAMA_LOG_INFO("%s: causal_attn = %d\n", __func__, cparams.causal_attn);
@@ -125,14 +123,14 @@ llama_context::llama_context(
LLAMA_LOG_INFO("%s: freq_base = %.1f\n", __func__, cparams.rope_freq_base); LLAMA_LOG_INFO("%s: freq_base = %.1f\n", __func__, cparams.rope_freq_base);
LLAMA_LOG_INFO("%s: freq_scale = %g\n", __func__, cparams.rope_freq_scale); LLAMA_LOG_INFO("%s: freq_scale = %g\n", __func__, cparams.rope_freq_scale);
if (n_ctx_per_seq < hparams.n_ctx_train) { if (n_ctx_per_seq() < hparams.n_ctx_train) {
LLAMA_LOG_WARN("%s: n_ctx_per_seq (%u) < n_ctx_train (%u) -- the full capacity of the model will not be utilized\n", LLAMA_LOG_WARN("%s: n_ctx_per_seq (%u) < n_ctx_train (%u) -- the full capacity of the model will not be utilized\n",
__func__, n_ctx_per_seq, hparams.n_ctx_train); __func__, n_ctx_per_seq(), hparams.n_ctx_train);
} }
if (n_ctx_per_seq > hparams.n_ctx_train) { if (n_ctx_per_seq() > hparams.n_ctx_train) {
LLAMA_LOG_WARN("%s: n_ctx_per_seq (%u) > n_ctx_train (%u) -- possible training context overflow\n", LLAMA_LOG_WARN("%s: n_ctx_per_seq (%u) > n_ctx_train (%u) -- possible training context overflow\n",
__func__, n_ctx_per_seq, hparams.n_ctx_train); __func__, n_ctx_per_seq(), hparams.n_ctx_train);
} }
if (!hparams.vocab_only) { if (!hparams.vocab_only) {
@@ -454,7 +452,7 @@ uint32_t llama_context::n_ctx() const {
} }
uint32_t llama_context::n_ctx_per_seq() const { uint32_t llama_context::n_ctx_per_seq() const {
return cparams.n_ctx / cparams.n_seq_max; return cparams.kv_unified ? cparams.n_ctx : cparams.n_ctx / cparams.n_seq_max;
} }
uint32_t llama_context::n_batch() const { uint32_t llama_context::n_batch() const {

2
src/llama-model.cpp
View File

@@ -6712,7 +6712,7 @@ float llama_model::get_rope_freq_scale(const llama_cparams & cparams, int il) co
} }
ggml_tensor * llama_model::get_rope_factors(const llama_cparams & cparams, int il) const { ggml_tensor * llama_model::get_rope_factors(const llama_cparams & cparams, int il) const {
const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max; const uint32_t n_ctx_per_seq = cparams.kv_unified ? cparams.n_ctx : cparams.n_ctx / cparams.n_seq_max;
// choose long/short freq factors based on the context size // choose long/short freq factors based on the context size
if (layers[il].rope_freqs != nullptr) { if (layers[il].rope_freqs != nullptr) {