ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-01-26 17:20:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

Hadamard transforms for K-cache - CPU only (#1033)

Browse Source 
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2025-12-04 06:51:11 +01:00
committed by GitHub
parent 0581f90c0f
commit 18fdd80eaf
13 changed files with 155 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
common/common.cpp
View File

@@ -1394,6 +1394,10 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
params.merge_qkv = true;
return true;
}
if (arg == "-khad" || arg == "--k-cache-hadamard") {
params.k_cache_hadamard = true;
return true;
}
if (arg == "--numa") {
CHECK_ARG
std::string value(argv[i]);
@@ -2074,6 +2078,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
options.push_back({ "*", "-gr, --graph-reuse", "enable graph reuse (default: %s)", params.graph_reuse ? "enabled" : "disabled" });
options.push_back({ "*", "-ser, --smart-expert-reduction", "experts reduction (default: %d,%g)", params.min_experts, params.thresh_experts});
options.push_back({ "*", "-mqkv, --merge-qkv,", "merge Q,K,V (default: %d)", params.merge_qkv});
options.push_back({ "*", "-khad, --k-cache-hadamard,", "Use Hadamard transform for K-cache (default: %d)", params.k_cache_hadamard});
options.push_back({ "*", "-vq, --validate-quants", "validate quantized data while loading the model (default: %d)", params.validate_quants});
options.push_back({ "*", "-p, --prompt PROMPT", "prompt to start generation with\n"
"in conversation mode, this will be used as system prompt\n"
@@ -3063,9 +3068,11 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
cparams.fused_mmad = params.fused_mmad;
cparams.rope_cache = params.rope_cache;
cparams.graph_reuse = params.graph_reuse;
cparams.k_cache_hadamard = params.k_cache_hadamard;
cparams.min_experts = params.min_experts;
cparams.thresh_experts = params.thresh_experts;
cparams.only_active_experts = params.only_active_exps;
cparams.k_cache_hadamard = params.k_cache_hadamard;
cparams.type_k = kv_cache_type_from_str(params.cache_type_k);
cparams.type_v = kv_cache_type_from_str(params.cache_type_v);
@@ -4209,6 +4216,7 @@ void yaml_dump_non_result_info(FILE * stream, const gpt_params & params, const l
fprintf(stream, "fused_mmad: %s # default: true\n", params.fused_mmad ? "true" : "false");
fprintf(stream, "rope_cache: %s # default: false\n", params.rope_cache ? "true" : "false");
fprintf(stream, "graph_reuse: %s # default: false\n", params.graph_reuse ? "true" : "false");
fprintf(stream, "k_cache_hadamard: %s # default: false\n", params.k_cache_hadamard ? "true" : "false");
fprintf(stream, "ser: %d,%g # defaulr: -1,0\n", params.min_experts, params.thresh_experts);
fprintf(stream, "temp: %f # default: 0.8\n", sparams.temp);

1
common/common.h
View File

@@ -276,6 +276,7 @@ struct gpt_params {
bool validate_quants = false; // if true, check for NaNs while loading the model
bool only_active_exps = true; // if true, offload only active experts (relevant only for hybrid CPU/GPU)
bool merge_qkv = false; // if true, merge separate Q, K, V tensors into a single, contiguous tensor
bool k_cache_hadamard = false; // if true, use Hadamard transform for the K-cache (only makes sense with quantized cache)
std::string cache_type_k = "f16"; // KV cache data type for the K
std::string cache_type_v = "f16"; // KV cache data type for the V