chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

[Bug fix] Fix Gemma 2 and fix Gemma 3 multimodal with bs > 1 on NPU (#9871)

Browse Source 
Co-authored-by: Maksim <makcum888e@mail.ru>
This commit is contained in:
ssshinigami
2025-09-08 11:19:40 +03:00
committed by GitHub
parent ee21817c6b
commit 5dd8c6444b
3 changed files with 11 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
python/sglang/srt/layers/logits_processor.py
View File

@@ -46,10 +46,12 @@ from sglang.srt.model_executor.forward_batch_info import (
ForwardBatch,
ForwardMode,
)
from sglang.srt.utils import dump_to_file, use_intel_amx_backend
from sglang.srt.utils import dump_to_file, is_npu, use_intel_amx_backend
logger = logging.getLogger(__name__)
_is_npu = is_npu()
@dataclasses.dataclass
class LogitsProcessorOutput:
@@ -517,7 +519,12 @@ class LogitsProcessor(nn.Module):
logits = logits[:, : self.config.vocab_size].float()
if self.final_logit_softcapping:
fused_softcap(logits, self.final_logit_softcapping)
if not _is_npu:
fused_softcap(logits, self.final_logit_softcapping)
else:
logits = self.final_logit_softcapping * torch.tanh(
logits / self.final_logit_softcapping
)
return logits