Enable benchmarked two-slot medium profile
This commit is contained in:
+7
-1
@@ -40,7 +40,7 @@ UNCENSORED_UBATCH_SIZE=128
|
||||
EXPERIMENTAL_CONTEXT=76800
|
||||
FAST_GPU_DEVICES=0,1
|
||||
MEDIUM_GPU_DEVICES=0,1
|
||||
MEDIUM_TENSOR_SPLIT=90,10
|
||||
MEDIUM_TENSOR_SPLIT=85,15
|
||||
LARGE_GPU_DEVICES=0,1
|
||||
LARGE_TENSOR_SPLIT=86,14
|
||||
ULTRA_GPU_DEVICES=0,1
|
||||
@@ -51,3 +51,9 @@ UNCENSORED_MTP_MAX=2
|
||||
EXPERIMENTAL_GPU_DEVICES=0
|
||||
LLAMA_THREADS=6
|
||||
LLAMA_THREADS_BATCH=6
|
||||
FAST_PARALLEL_SLOTS=1
|
||||
MEDIUM_PARALLEL_SLOTS=2
|
||||
LARGE_PARALLEL_SLOTS=1
|
||||
ULTRA_PARALLEL_SLOTS=1
|
||||
UNCENSORED_PARALLEL_SLOTS=1
|
||||
EXPERIMENTAL_PARALLEL_SLOTS=1
|
||||
|
||||
Reference in New Issue
Block a user