[Perf] RMSNorm uncontiguous support, 1.2~3.1x kernel performance improvement (#49750)
Signed-off-by: yewentao256 <zhyanwentao@126.com> Signed-off-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
This commit is contained in:
@@ -271,11 +271,11 @@ class Step3p5Attention(nn.Module):
|
||||
q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
|
||||
# Add qk-norm inline similar to Qwen3 MOE attention
|
||||
q_by_head = q.view(*q.shape[:-1], q.shape[-1] // self.head_dim, self.head_dim)
|
||||
q_by_head = self.q_norm(q_by_head.contiguous())
|
||||
q_by_head = self.q_norm(q_by_head)
|
||||
q = q_by_head.view(q.shape)
|
||||
|
||||
k_by_head = k.view(*k.shape[:-1], k.shape[-1] // self.head_dim, self.head_dim)
|
||||
k_by_head = self.k_norm(k_by_head.contiguous())
|
||||
k_by_head = self.k_norm(k_by_head)
|
||||
k = k_by_head.view(k.shape)
|
||||
if self.use_rope:
|
||||
q, k = self.rotary_emb(positions, q, k)
|
||||
|
||||
Reference in New Issue
Block a user