Teak mem fraction (#20)
This commit is contained in:
@@ -278,7 +278,7 @@ class ModelRunner:
|
||||
load_format=self.load_format,
|
||||
revision=None,
|
||||
)
|
||||
self.model = model
|
||||
self.model = model.eval()
|
||||
|
||||
def profile_max_num_token(self, total_gpu_memory):
|
||||
available_gpu_memory = get_available_gpu_memory(
|
||||
|
||||
@@ -26,10 +26,14 @@ class ServerArgs:
|
||||
if self.tokenizer_path is None:
|
||||
self.tokenizer_path = self.model_path
|
||||
if self.mem_fraction_static is None:
|
||||
if self.tp_size > 1:
|
||||
self.mem_fraction_static = 0.8
|
||||
if self.tp_size >= 8:
|
||||
self.mem_fraction_static = 0.80
|
||||
elif self.tp_size >= 4:
|
||||
self.mem_fraction_static = 0.82
|
||||
elif self.tp_size >= 2:
|
||||
self.mem_fraction_static = 0.85
|
||||
else:
|
||||
self.mem_fraction_static = 0.9
|
||||
self.mem_fraction_static = 0.90
|
||||
|
||||
@staticmethod
|
||||
def add_cli_args(parser: argparse.ArgumentParser):
|
||||
|
||||
Reference in New Issue
Block a user