use sdp fp8 causal kernel (#11023)
This commit is contained in:
parent
c34f85e7d0
commit
fad1dbaf60
1 changed files with 24 additions and 23 deletions
|
|
@ -137,22 +137,23 @@ def attention_forward(
|
|||
key_states, value_states = past_key_value.update(key_states, value_states,
|
||||
self.layer_idx, None)
|
||||
|
||||
if (isinstance(past_key_value, DynamicFp8Cache) and
|
||||
use_sdp_fp8(q_len, kv_seq_len, query_states)):
|
||||
import linear_q4_0
|
||||
attn_output = linear_q4_0.sdp_fp8(query_states, key_states, value_states, attention_mask)
|
||||
elif (isinstance(past_key_value, DynamicNormalCache) and
|
||||
use_sdp(q_len, kv_seq_len, self.head_dim, query_states)):
|
||||
if use_sdp(q_len, kv_seq_len, self.head_dim, query_states):
|
||||
import linear_q4_0
|
||||
if isinstance(past_key_value, DynamicFp8Cache):
|
||||
attn_output = linear_q4_0.sdp_fp8(query_states, key_states, value_states,
|
||||
attention_mask)
|
||||
else:
|
||||
attn_output = linear_q4_0.sdp(query_states, key_states, value_states, attention_mask)
|
||||
elif use_sdp_causal(q_len, kv_seq_len, query_states, self.training):
|
||||
import linear_q4_0
|
||||
if isinstance(past_key_value, DynamicFp8Cache):
|
||||
attn_output = linear_q4_0.sdp_fp8_causal(query_states, key_states, value_states)
|
||||
else:
|
||||
attn_output = linear_q4_0.sdp_causal(query_states, key_states, value_states)
|
||||
else:
|
||||
if isinstance(past_key_value, DynamicFp8Cache):
|
||||
key_states, value_states = restore_fp8_kv_cache(key_states, value_states,
|
||||
query_states.dtype)
|
||||
if use_sdp_causal(q_len, kv_seq_len, query_states, self.training):
|
||||
import linear_q4_0
|
||||
attn_output = linear_q4_0.sdp_causal(query_states, key_states, value_states)
|
||||
else:
|
||||
# repeat k/v heads if n_kv_heads < n_heads
|
||||
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
||||
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
||||
|
|
|
|||
Loading…
Reference in a new issue