vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-12 09:59:41 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 3

Merge branch 'upstream' into concedo_experimental

Browse source 
# Conflicts:
#	Makefile
#	README.md
#	common/CMakeLists.txt
#	common/common.cpp
#	common/common.h
#	examples/embedding/embedding.cpp
#	examples/imatrix/imatrix.cpp
#	examples/infill/infill.cpp
#	examples/parallel/parallel.cpp
#	examples/perplexity/perplexity.cpp
#	examples/rpc/README.md
#	examples/save-load-state/save-load-state.cpp
#	examples/server/README.md
#	examples/speculative/speculative.cpp
#	tests/test-sampling.cpp
This commit is contained in:
Concedo 2024-09-10 16:39:23 +08:00
commit a947558e0e
45 changed files with 2628 additions and 6743 deletions
Download patch file Download diff file Expand all files Collapse all files

4
src/llama.cpp
View file

@ -9317,7 +9317,7 @@ static struct ggml_tensor * llm_build_copy_mask_state(
// FIXME: zero-out NANs?
states = ggml_mul(ctx, states, state_mask);
// copy states which won't be changed further (between n_seqs and n_rs)
// copy states which won't be changed further (between n_seqs and n_kv)
ggml_build_forward_expand(graph,
ggml_cpy(ctx,
ggml_view_1d(ctx, states, n_state*(n_kv - n_seqs), n_seqs*n_state*ggml_element_size(states)),
@ -17607,6 +17607,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
quantize &= name.find("time_mix_first.weight") == std::string::npos;
quantize &= name.find("time_mix_w1.weight") == std::string::npos;
quantize &= name.find("time_mix_w2.weight") == std::string::npos;
quantize &= name.find("time_mix_decay_w1.weight") == std::string::npos;
quantize &= name.find("time_mix_decay_w2.weight") == std::string::npos;
// do not quantize relative position bias (T5)
quantize &= name.find("attn_rel_b.weight") == std::string::npos;