frank-default-small: ncpu-moe=32(8GPU, ~51.1 t/s gen)
This commit is contained in:
@@ -31,8 +31,8 @@ NGRAM_MOD_N_MIN=48
|
|||||||
NGRAM_MOD_N_MAX=64
|
NGRAM_MOD_N_MAX=64
|
||||||
NGRAM_MOD_N_MATCH=24
|
NGRAM_MOD_N_MATCH=24
|
||||||
|
|
||||||
# Expert offload: 35 of 40 layers keep MoE weights on CPU; 5 on GPU.
|
# Expert offload: 32 of 40 layers keep MoE weights on CPU; 8 on GPU.
|
||||||
N_CPU_MOE=35
|
N_CPU_MOE=32
|
||||||
N_GPU_LAYERS=all
|
N_GPU_LAYERS=all
|
||||||
|
|
||||||
# KV cache on GPU, both K and V at f16.
|
# KV cache on GPU, both K and V at f16.
|
||||||
|
|||||||
@@ -36,7 +36,7 @@ fi
|
|||||||
: "${CACHE_TYPE_K:=f16}"
|
: "${CACHE_TYPE_K:=f16}"
|
||||||
: "${CACHE_TYPE_V:=f16}"
|
: "${CACHE_TYPE_V:=f16}"
|
||||||
: "${VERBOSITY:=4}"
|
: "${VERBOSITY:=4}"
|
||||||
: "${N_CPU_MOE:=35}"
|
: "${N_CPU_MOE:=32}"
|
||||||
: "${N_GPU_LAYERS:=all}"
|
: "${N_GPU_LAYERS:=all}"
|
||||||
: "${BATCH_SIZE:=4096}"
|
: "${BATCH_SIZE:=4096}"
|
||||||
: "${UBATCH_SIZE:=2048}"
|
: "${UBATCH_SIZE:=2048}"
|
||||||
|
|||||||
Reference in New Issue
Block a user