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

Remove unnecessary bf16 assert in rotate_activation (#24686)

Browse Source
This commit is contained in:
Yuhao Yang
2026-05-09 05:00:52 +08:00
committed by GitHub
parent f30d1d0b0a
commit 09912fd89d
2 changed files with 1 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/attention/dsv4/compressor.py
+1 -1
View File
@@ -104,7 +104,7 @@ class CompressorBackendMixin:
freqs_cis_cache, freqs_cis_cache,
plan, plan,
) )
return rotate_activation(kv_compressed.bfloat16()) if rotate else kv_compressed return rotate_activation(kv_compressed) if rotate else kv_compressed
def forward_core_compressor( def forward_core_compressor(
self, self,
python/sglang/srt/layers/attention/nsa/nsa_indexer.py
-1
View File
@@ -148,7 +148,6 @@ class BaseIndexerMetadata(ABC):
def rotate_activation(x: torch.Tensor) -> torch.Tensor: def rotate_activation(x: torch.Tensor) -> torch.Tensor:
assert x.dtype == torch.bfloat16
# from sgl_kernel import hadamard_transform # from sgl_kernel import hadamard_transform
if _is_hip: if _is_hip:
from fast_hadamard_transform import hadamard_transform from fast_hadamard_transform import hadamard_transform