Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-26 14:20:31 +01:00
Code Issues Packages Projects Releases Wiki Activity

server : fix default draft model parameters (#10586)

Browse Source 
* server : force F16 KV cache for the draft model

ggml-ci

* server : fix draft params

ggml-ci

* server : various params fixes

ggml-ci
This commit is contained in:
Georgi Gerganov 2024-12-03 11:20:00 +02:00 committed by GitHub
parent 642330ac7c
commit 70b98fadbc
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 14 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
examples/server/server.cpp
View File

@ -696,8 +696,9 @@ struct server_context {
params_dft.devices = params_base.speculative.devices; params_dft.devices = params_base.speculative.devices;
params_dft.model = params_base.speculative.model; params_dft.model = params_base.speculative.model;
params_dft.n_ctx = params_base.speculative.n_ctx; params_dft.n_ctx = params_base.speculative.n_ctx == 0 ? params_base.n_ctx / params_base.n_parallel : params_base.speculative.n_ctx;
params_dft.n_gpu_layers = params_base.speculative.n_gpu_layers; params_dft.n_gpu_layers = params_base.speculative.n_gpu_layers;
params_dft.n_parallel = 1;
common_init_result llama_init_dft = common_init_from_params(params_dft); common_init_result llama_init_dft = common_init_from_params(params_dft);
@ -717,8 +718,14 @@ struct server_context {
return false; return false;
} }
cparams_dft = common_context_params_to_llama(params_base); const int n_ctx_dft = llama_n_ctx(llama_init_dft.context);
cparams_dft.n_batch = llama_n_ctx(llama_init_dft.context);
cparams_dft = common_context_params_to_llama(params_dft);
cparams_dft.n_batch = n_ctx_dft;
// force F16 KV cache for the draft model for extra performance
cparams_dft.type_k = GGML_TYPE_F16;
cparams_dft.type_v = GGML_TYPE_F16;
// the context is not needed - we will create one for each slot // the context is not needed - we will create one for each slot
llama_free(llama_init_dft.context); llama_free(llama_init_dft.context);
@ -2322,6 +2329,10 @@ struct server_context {
continue; continue;
} }
if (slot.state != SLOT_STATE_GENERATING) {
continue;
}
llama_token id = slot.sampled; llama_token id = slot.sampled;
struct common_speculative_params params_spec; struct common_speculative_params params_spec;