[XPU]feat: add DeepSeek-V4 XPU attention decode path (#42953)
Signed-off-by: Ma Jian <jian1.ma@intel.com>
This commit is contained in:
@@ -368,7 +368,7 @@ class DeepseekCompressor(nn.Module):
|
||||
fp8_scale=fp8_scale,
|
||||
)
|
||||
else:
|
||||
# Indexer path (head_dim == 128) or AMD: triton, legacy UE8M0 only.
|
||||
# Indexer path (head_dim == 128) or non-CUDA GPUs (AMD, XPU, etc.).
|
||||
compress_norm_rope_store_fn = compress_norm_rope_store_triton
|
||||
extra_kwargs = {}
|
||||
|
||||
|
||||
Reference in New Issue
Block a user