[Bugfix] Add contiguous call inside rope kernel wrapper (#17091)
Signed-off-by: 苏政渊 <suzhengyuan@moonshot.cn> Co-authored-by: 苏政渊 <suzhengyuan@moonshot.cn>
This commit is contained in:
co-authored by
苏政渊
parent
165cb56329
commit
17eb306fcc
@@ -938,8 +938,7 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
|
||||
decode_ql_nope, decode_q_pe = \
|
||||
self._q_proj_and_k_up_proj(decode_hs_or_q_c)
|
||||
decode_q_pe[...], decode_k_pe[...] = self.rotary_emb(
|
||||
attn_metadata.decode.input_positions, decode_q_pe.contiguous(),
|
||||
decode_k_pe)
|
||||
attn_metadata.decode.input_positions, decode_q_pe, decode_k_pe)
|
||||
|
||||
if has_prefill:
|
||||
assert attn_metadata.prefill is not None
|
||||
@@ -948,8 +947,8 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
|
||||
prefill_q_pe = prefill_q[..., self.qk_nope_head_dim:]
|
||||
|
||||
prefill_q_pe[...], prefill_k_pe[...] = self.rotary_emb(
|
||||
attn_metadata.prefill.input_positions,
|
||||
prefill_q_pe.contiguous(), prefill_k_pe)
|
||||
attn_metadata.prefill.input_positions, prefill_q_pe,
|
||||
prefill_k_pe)
|
||||
|
||||
# write the latent and rope to kv cache
|
||||
if kv_cache.numel() > 0:
|
||||
|
||||
Reference in New Issue
Block a user