mirror of
https://github.com/huggingface/text-generation-inference.git
synced 2025-04-19 22:02:06 +00:00
fix Qwen VL break in intel platform (#3002)
* fix Qwen VL break in intel platform Signed-off-by: Wang, Yi A <yi.a.wang@intel.com> * could use PositionRotaryEmbedding impl so rocm and ipex could all work Signed-off-by: Wang, Yi A <yi.a.wang@intel.com> --------- Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
This commit is contained in:
parent
b86c3947ab
commit
76bcb4948d
@ -570,22 +570,6 @@ class RotaryPositionEmbeddingMultimodalSections(PositionRotaryEmbedding):
|
||||
.to(inv_freq.device)
|
||||
)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
query: torch.Tensor,
|
||||
key: torch.Tensor,
|
||||
cos: torch.Tensor,
|
||||
sin: torch.Tensor,
|
||||
):
|
||||
# rotate half the sequence length
|
||||
rot = cos.shape[-1] // 2
|
||||
q2 = torch.cat([-query[..., rot:], query[..., :rot]], dim=-1)
|
||||
k2 = torch.cat([-key[..., rot:], key[..., :rot]], dim=-1)
|
||||
|
||||
# apply the rotation
|
||||
rotary_emb.apply_rotary(query, q2, cos, sin, query, q2, True)
|
||||
rotary_emb.apply_rotary(key, k2, cos, sin, key, k2, True)
|
||||
|
||||
def _update_cos_sin_cache(
|
||||
self, dtype: torch.dtype, device: torch.device, seqlen: int
|
||||
):
|
||||
@ -614,7 +598,4 @@ class RotaryPositionEmbeddingMultimodalSections(PositionRotaryEmbedding):
|
||||
|
||||
cos = self._cos_cached[position_ids].gather(1, self._sections[:slen])
|
||||
sin = self._sin_cached[position_ids].gather(1, self._sections[:slen])
|
||||
|
||||
cos = torch.cat([cos, cos], dim=-1)
|
||||
sin = torch.cat([sin, sin], dim=-1)
|
||||
return cos, sin
|
||||
|
Loading…
Reference in New Issue
Block a user