tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-29 12:35:16 +00:00
Code Issues Packages Projects Releases Wiki Activity

CUDA: optimize FA for GQA + large batches (#12014)

Browse Source
This commit is contained in:
Johannes Gäßler
2025-02-22 12:20:17 +01:00
committed by GitHub
parent 335eb04a91
commit 5fa07c2f93
32 changed files with 940 additions and 411 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
ggml/src/ggml-cuda/template-instances/generate_cu_files.py
View File

@ -18,7 +18,7 @@ SOURCE_FATTN_MMA_START = """// This file has been autogenerated by generate_cu_f
"""
SOURCE_FATTN_MMA_CASE = "DECL_FATTN_MMA_F16_CASE({head_size}, {cols_per_block});\n"
SOURCE_FATTN_MMA_CASE = "DECL_FATTN_MMA_F16_CASE({head_size}, {ncols1}, {ncols2});\n"
TYPES_MMQ = [
"GGML_TYPE_Q4_0", "GGML_TYPE_Q4_1", "GGML_TYPE_Q5_0", "GGML_TYPE_Q5_1", "GGML_TYPE_Q8_0",
@ -57,12 +57,18 @@ for vkq_size in [16, 32]:
with open(f"fattn-vec-f{vkq_size}-instance-hs{head_size}-{get_short_name(type_k)}-{get_short_name(type_v)}.cu", "w") as f:
f.write(SOURCE_FATTN_VEC.format(vkq_size=vkq_size, head_size=head_size, type_k=type_k, type_v=type_v))
for cols_per_block in [8, 16, 32, 64]:
with open(f"fattn-mma-f16-instance-cpb{cols_per_block}.cu", "w") as f:
f.write(SOURCE_FATTN_MMA_START)
for ncols in [8, 16, 32, 64, 128]:
for ncols2 in [1, 2, 4, 8]:
ncols1 = ncols // ncols2
if ncols == 128:
continue # Too much register pressure.
with open(f"fattn-mma-f16-instance-ncols1_{ncols1}-ncols2_{ncols2}.cu", "w") as f:
f.write(SOURCE_FATTN_MMA_START)
for head_size in [64, 80, 96, 112, 128, 256]:
f.write(SOURCE_FATTN_MMA_CASE.format(cols_per_block=cols_per_block, head_size=head_size))
for head_size in [64, 80, 96, 112, 128, 256]:
if ncols == 128 and head_size == 256:
continue # Needs too much shared memory.
f.write(SOURCE_FATTN_MMA_CASE.format(ncols1=ncols1, ncols2=ncols2, head_size=head_size))
for type in TYPES_MMQ:
with open(f"mmq-instance-{get_short_name(type)}.cu", "w") as f: