Deprecate global_server_args_dict (#11331)
This commit is contained in:
@@ -45,7 +45,7 @@ from sglang.srt.layers.linear import (
|
||||
)
|
||||
from sglang.srt.layers.quantization import QuantizationConfig
|
||||
from sglang.srt.layers.rotary_embedding import apply_rotary_pos_emb
|
||||
from sglang.srt.managers.schedule_batch import global_server_args_dict
|
||||
from sglang.srt.server_args import get_global_server_args
|
||||
from sglang.srt.utils import add_prefix
|
||||
|
||||
ROTARY_EMBED_CLASSES = {
|
||||
@@ -468,7 +468,7 @@ class VisionAttention(nn.Module):
|
||||
_passed_backend = qkv_backend
|
||||
qkv_backend = self._determine_attention_backend(_passed_backend)
|
||||
if (
|
||||
global_server_args_dict["mm_attention_backend"] is None
|
||||
get_global_server_args().mm_attention_backend is None
|
||||
and _passed_backend is None
|
||||
):
|
||||
print_info_once(f"Multimodal attention backend not set. Use {qkv_backend}.")
|
||||
@@ -528,7 +528,7 @@ class VisionAttention(nn.Module):
|
||||
- CUDA: "triton_attn"
|
||||
- Non-CUDA: "sdpa"
|
||||
"""
|
||||
override_backend = global_server_args_dict["mm_attention_backend"]
|
||||
override_backend = get_global_server_args().mm_attention_backend
|
||||
if override_backend is not None:
|
||||
backend = override_backend
|
||||
elif passed_backend is not None:
|
||||
|
||||
Reference in New Issue
Block a user