vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

llama : cont

Browse source
This commit is contained in:
Georgi Gerganov 2024-02-26 18:24:55 +02:00
parent 4e35db1a81
commit 547ddae5f7
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
3 changed files with 11 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

9
common/common.cpp
View file

@ -335,6 +335,12 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
break; break;
} }
params.yarn_beta_slow = std::stof(argv[i]); params.yarn_beta_slow = std::stof(argv[i]);
} else if (arg == "--defrag-thold" || arg == "-dt") {
if (++i >= argc) {
invalid_param = true;
break;
}
params.defrag_thold = std::stof(argv[i]);
} else if (arg == "--samplers") { } else if (arg == "--samplers") {
if (++i >= argc) { if (++i >= argc) {
invalid_param = true; invalid_param = true;
@ -1004,6 +1010,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
printf(" --yarn-attn-factor N YaRN: scale sqrt(t) or attention magnitude (default: 1.0)\n"); printf(" --yarn-attn-factor N YaRN: scale sqrt(t) or attention magnitude (default: 1.0)\n");
printf(" --yarn-beta-slow N YaRN: high correction dim or alpha (default: %.1f)\n", params.yarn_beta_slow); printf(" --yarn-beta-slow N YaRN: high correction dim or alpha (default: %.1f)\n", params.yarn_beta_slow);
printf(" --yarn-beta-fast N YaRN: low correction dim or beta (default: %.1f)\n", params.yarn_beta_fast); printf(" --yarn-beta-fast N YaRN: low correction dim or beta (default: %.1f)\n", params.yarn_beta_fast);
printf(" -dt N, --defrag-thold N\n");
printf(" KV cache defragmentation threshold (default: %.1f, < 0 - disabled)\n", params.defrag_thold);
printf(" --ignore-eos ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n"); printf(" --ignore-eos ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
printf(" --no-penalize-nl do not penalize newline token\n"); printf(" --no-penalize-nl do not penalize newline token\n");
printf(" --temp N temperature (default: %.1f)\n", (double)sparams.temp); printf(" --temp N temperature (default: %.1f)\n", (double)sparams.temp);
@ -1285,6 +1293,7 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
cparams.yarn_beta_fast = params.yarn_beta_fast; cparams.yarn_beta_fast = params.yarn_beta_fast;
cparams.yarn_beta_slow = params.yarn_beta_slow; cparams.yarn_beta_slow = params.yarn_beta_slow;
cparams.yarn_orig_ctx = params.yarn_orig_ctx; cparams.yarn_orig_ctx = params.yarn_orig_ctx;
cparams.defrag_thold = params.defrag_thold;
cparams.offload_kqv = !params.no_kv_offload; cparams.offload_kqv = !params.no_kv_offload;
cparams.type_k = kv_cache_type_from_str(params.cache_type_k); cparams.type_k = kv_cache_type_from_str(params.cache_type_k);

1
common/common.h
View file

@ -75,6 +75,7 @@ struct gpt_params {
float yarn_beta_fast = 32.0f; // YaRN low correction dim float yarn_beta_fast = 32.0f; // YaRN low correction dim
float yarn_beta_slow = 1.0f; // YaRN high correction dim float yarn_beta_slow = 1.0f; // YaRN high correction dim
int32_t yarn_orig_ctx = 0; // YaRN original context length int32_t yarn_orig_ctx = 0; // YaRN original context length
float defrag_thold = -1.0f; // KV cache defragmentation threshold
int32_t rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED; int32_t rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
ggml_numa_strategy numa = GGML_NUMA_STRATEGY_DISABLED; ggml_numa_strategy numa = GGML_NUMA_STRATEGY_DISABLED;

2
llama.cpp
View file

@ -8008,7 +8008,7 @@ static int llama_decode_internal(
} }
// decide if we need to defrag the kv cache // decide if we need to defrag the kv cache
{ if (cparams.defrag_thold >= 0.0f) {
const float fragmentation = kv_self.n >= 128 ? 1.0f - float(kv_self.used + n_tokens)/float(kv_self.n) : 0.0f; const float fragmentation = kv_self.n >= 128 ? 1.0f - float(kv_self.used + n_tokens)/float(kv_self.n) : 0.0f;
// queue defragmentation for next llama_kv_cache_update // queue defragmentation for next llama_kv_cache_update