Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-03 19:42:14 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'upstream' into concedo_experimental

Browse Source 
# Conflicts:
#	docs/backend/SYCL.md
#	ggml/src/ggml-cpu/ggml-cpu.c
#	ggml/src/ggml-opencl/CMakeLists.txt
#	ggml/src/ggml-opencl/ggml-opencl.cpp
#	ggml/src/ggml-opencl/kernels/cvt.cl
#	ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32_ns.cl
#	ggml/src/ggml-opencl/kernels/gemv_moe_q4_k_f32_ns.cl
#	ggml/src/ggml-sycl/backend.hpp
#	ggml/src/ggml-sycl/common.hpp
#	ggml/src/ggml-sycl/dmmv.cpp
#	ggml/src/ggml-sycl/ggml-sycl.cpp
#	scripts/sync_vendor.py
#	tests/CMakeLists.txt
#	tests/test-alloc.cpp
#	tests/test-backend-ops.cpp
#	tests/test-chat.cpp
#	tests/test-gguf.cpp
#	tests/test-save-load-state.cpp
This commit is contained in:
Concedo
2026-07-13 20:53:37 +08:00
parent 1f9dd9c398 f4253ef965
commit fa21872f97
76 changed files with 1399 additions and 1110 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-graph.cpp
+3 -3
View File
@@ -843,7 +843,7 @@ static void dsv4_build_comp_inputs(
GGML_ASSERT(n_stream > 0);
GGML_ASSERT(n_tokens%n_stream == 0);
inp.kq_mask = ggml_new_tensor_4d(ctx, cparams.flash_attn && strcmp(name, "lid") != 0 ? GGML_TYPE_F16 : GGML_TYPE_F32, plan.n_kv, n_tokens/n_stream, 1, n_stream);
inp.kq_mask = ggml_new_tensor_4d(ctx, (strcmp(name, "lid") != 0 && cparams.flash_attn) || (strcmp(name, "lid") == 0 && cparams.fused_lid) ? GGML_TYPE_F16 : GGML_TYPE_F32, plan.n_kv, n_tokens/n_stream, 1, n_stream);
ggml_set_input(inp.kq_mask);
ggml_set_name(inp.kq_mask, (std::string("dsv4_") + name + "_kq_mask").c_str());
}
@@ -3026,9 +3026,9 @@ llm_graph_input_attn_k_dsa * llm_graph_context::build_attn_inp_k_dsa() const {
{
inp->self_k_idxs_lid = mctx_cur->get_lid()->build_input_k_idxs(ctx0, ubatch);
// ensure F32 mask
// ensure that mask type matches fused lightning indexer use (requires f16 mask)
auto cparams_copy = cparams;
cparams_copy.flash_attn = false;
cparams_copy.flash_attn = cparams.fused_lid;
inp->self_kq_mask_lid = build_attn_inp_kq_mask(ctx0, mctx_cur->get_lid(), ubatch, cparams_copy);
inp->self_kq_mask_lid_cnv = inp->self_kq_mask_lid;