ayo/ipex-llm
1
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

use new fp16 sdp in llama and mistral (#10734)

Browse source
This commit is contained in:
Yishuo Wang 2024-04-12 10:49:02 +08:00 committed by GitHub
parent 019293e1b9
commit 8086554d33
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 42 additions and 36 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

28
python/llm/src/ipex_llm/transformers/models/llama.py
View file

@ -647,12 +647,11 @@ def llama_attention_forward_4_31_original(
past_key_value = (key_states, value_states) if use_cache else None
if not self.training and not hidden_states.requires_grad and \
use_flash_attention(query_states, key_states, attention_mask):
# 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)
if not self.training and not hidden_states.requires_grad and \
use_flash_attention(query_states, key_states, attention_mask):
attn_output = F.scaled_dot_product_attention(query_states.to(device, dtype=torch.float16),
key_states.to(device, dtype=torch.float16),
value_states.to(device, dtype=torch.float16),
@ -660,13 +659,14 @@ def llama_attention_forward_4_31_original(
attn_weights = None
elif not self.training and not hidden_states.requires_grad and \
use_esimd_sdp(q_len, key_states.shape[2], self.head_dim, query_states, attention_mask):
import linear_fp16_esimd
attn_output = linear_fp16_esimd.sdp_forward(query_states,
key_states,
value_states)
import linear_q4_0
attn_output = linear_q4_0.sdp_fp16(query_states, key_states, value_states, attention_mask)
attn_output = attn_output.view(query_states.shape)
attn_weights = None
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)
# otherwise, use native attention
attn_output, attn_weights = native_sdp(query_states, key_states, value_states,
attention_mask,
@ -1305,12 +1305,11 @@ def llama_attention_forward_4_36_original(
past_key_value.key_cache[self.layer_idx] = key_states
past_key_value.value_cache[self.layer_idx] = value_states
if not self.training and not hidden_states.requires_grad and \
use_flash_attention(query_states, key_states, attention_mask):
# 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)
if not self.training and not hidden_states.requires_grad and \
use_flash_attention(query_states, key_states, attention_mask):
# now only use flash attention for first token
attn_output = F.scaled_dot_product_attention(query_states.to(device, dtype=torch.float16),
key_states.to(device, dtype=torch.float16),
@ -1319,13 +1318,14 @@ def llama_attention_forward_4_36_original(
attn_weights = None
elif not self.training and not hidden_states.requires_grad and \
use_esimd_sdp(q_len, key_states.shape[2], self.head_dim, query_states):
import linear_fp16_esimd
attn_output = linear_fp16_esimd.sdp_forward(query_states,
key_states,
value_states)
import linear_q4_0
attn_output = linear_q4_0.sdp_fp16(query_states, key_states, value_states, attention_mask)
attn_output = attn_output.view(query_states.shape)
attn_weights = None
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)
# otherwise, use native attention
attn_output, attn_weights = native_sdp(query_states, key_states, value_states,
attention_mask,

30
python/llm/src/ipex_llm/transformers/models/mistral.py
View file

@ -495,13 +495,12 @@ def mistral_attention_forward_original(
else:
attention_dtype = original_dtype
if fsdp_flag:
# repeat k/v heads if n_kv_heads < n_heads
key_states = repeat_kv(key_states, self.num_key_value_groups).to(device,
dtype=attention_dtype)
value_states = repeat_kv(value_states, self.num_key_value_groups).to(device,
dtype=attention_dtype)
if fsdp_flag:
attn_output = F.scaled_dot_product_attention(query_states.to(dtype=attention_dtype),
key_states,
value_states,
@ -510,15 +509,19 @@ def mistral_attention_forward_original(
attn_output = attn_output.transpose(1, 2).contiguous()
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
elif use_esimd_sdp(q_len, key_states.shape[2], self.head_dim, query_states):
import linear_fp16_esimd
attn_output = linear_fp16_esimd.sdp_forward(query_states,
key_states,
value_states)
# new fp16 sdp doesn't require repeat_kv
import linear_q4_0
attn_output = linear_q4_0.sdp_fp16(query_states, key_states, value_states, attention_mask)
attn_output = attn_output.view(query_states.shape)
attn_weights = None
attn_output = attn_output.transpose(1, 2).contiguous()
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
else:
# repeat k/v heads if n_kv_heads < n_heads
key_states = repeat_kv(key_states, self.num_key_value_groups).to(device,
dtype=attention_dtype)
value_states = repeat_kv(value_states, self.num_key_value_groups).to(device,
dtype=attention_dtype)
attn_output, attn_weights = compute_attn_outputs_weights(query_states,
key_states,
value_states,
@ -885,13 +888,12 @@ def mistral_attention_forward_4_36_original(
else:
attention_dtype = original_dtype
if fsdp_flag:
# repeat k/v heads if n_kv_heads < n_heads
key_states = repeat_kv(key_states, self.num_key_value_groups).to(device,
dtype=attention_dtype)
value_states = repeat_kv(value_states, self.num_key_value_groups).to(device,
dtype=attention_dtype)
if fsdp_flag:
attn_output = F.scaled_dot_product_attention(query_states.to(dtype=attention_dtype),
key_states,
value_states,
@ -900,15 +902,19 @@ def mistral_attention_forward_4_36_original(
attn_output = attn_output.transpose(1, 2).contiguous()
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
elif use_esimd_sdp(q_len, key_states.shape[2], self.head_dim, query_states):
import linear_fp16_esimd
attn_output = linear_fp16_esimd.sdp_forward(query_states,
key_states,
value_states)
# new fp16 sdp doesn't require repeat_kv
import linear_q4_0
attn_output = linear_q4_0.sdp_fp16(query_states, key_states, value_states, attention_mask)
attn_output = attn_output.view(query_states.shape)
attn_weights = None
attn_output = attn_output.transpose(1, 2).contiguous()
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
else:
# repeat k/v heads if n_kv_heads < n_heads
key_states = repeat_kv(key_states, self.num_key_value_groups).to(device,
dtype=attention_dtype)
value_states = repeat_kv(value_states, self.num_key_value_groups).to(device,
dtype=attention_dtype)
attn_output, attn_weights = compute_attn_outputs_weights(query_states,
key_states,
value_states,