minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

[AMD] Fix Handle missing rope_theta in get_rope_config for Grok-1 (#21518)

Browse Source
This commit is contained in:
Michael
2026-03-31 10:58:12 -07:00
committed by GitHub
parent 91048b2a8e
commit 8d919bbd44
1 changed files with 4 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/grok.py
+4 -2
View File
@@ -61,7 +61,6 @@ from sglang.srt.model_executor.forward_batch_info import ForwardBatch
from sglang.srt.model_loader.loader import DefaultModelLoader from sglang.srt.model_loader.loader import DefaultModelLoader
from sglang.srt.model_loader.weight_utils import default_weight_loader from sglang.srt.model_loader.weight_utils import default_weight_loader
from sglang.srt.utils import add_prefix, is_npu from sglang.srt.utils import add_prefix, is_npu
from sglang.srt.utils.hf_transformers_utils import get_rope_config
_is_npu = is_npu() _is_npu = is_npu()
@@ -478,7 +477,10 @@ class Grok1DecoderLayer(nn.Module):
self.layer_id = layer_id self.layer_id = layer_id
self.alt_stream = alt_stream or torch.cuda.Stream() self.alt_stream = alt_stream or torch.cuda.Stream()
rope_theta, _ = get_rope_config(config) rope_theta = getattr(config, "rope_theta", None)
if rope_theta is None:
rope_params = getattr(config, "rope_parameters", None)
rope_theta = rope_params["rope_theta"] if rope_params else 10000
self.self_attn = Grok1Attention( self.self_attn = Grok1Attention(
config=config, config=config,
hidden_size=self.hidden_size, hidden_size=self.hidden_size,