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

ggml : reduce hash table reset cost (#8698)

Browse source 
* ggml : reduce hash table reset cost

* fix unreachable code warnings after GGML_ASSERT(false)

* GGML_ASSERT(false) -> GGML_ABORT("fatal error")

* GGML_ABORT use format string
This commit is contained in:
slaren 2024-07-27 04:41:55 +02:00 • committed by GitHub
parent 01245f5b16
commit 2b1f616b20
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
46 changed files with 851 additions and 754 deletions
Download patch file Download diff file Expand all files Collapse all files

6
ggml/src/ggml-cuda/fattn-common.cuh
View file

@ -564,7 +564,7 @@ static void on_no_fattn_vec_case(const int D) {
fprintf(stderr, "Unsupported KV type combination for head_size 64.\n");
fprintf(stderr, "By default only f16 KV cache is supported.\n");
fprintf(stderr, "Compile with GGML_CUDA_FA_ALL_QUANTS for V cache quantization support.\n");
GGML_ASSERT(false);
GGML_ABORT("fatal error");
} else if (D == 128) {
fprintf(stderr, "Unsupported KV type combination for head_size 128.\n");
fprintf(stderr, "Supported combinations:\n");
@ -572,11 +572,11 @@ static void on_no_fattn_vec_case(const int D) {
fprintf(stderr, " - K == q8_0, V == q8_0, 8.50 BPV\n");
fprintf(stderr, " - K == f16, V == f16, 16.00 BPV\n");
fprintf(stderr, "Compile with GGML_CUDA_FA_ALL_QUANTS for all combinations of q4_0, q4_1, q5_0, q5_1, q8_0, and f16.\n");
GGML_ASSERT(false);
GGML_ABORT("fatal error");
} else {
fprintf(stderr, "Unsupported KV type combination for head_size 256.\n");
fprintf(stderr, "Only f16 is supported.\n");
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
}