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

vLLM: Apply attention optimizations for selective batching (#9758)

Browse source 
* fuse_rope for prefil

* apply kv_cache optimizations

* apply fast_decoding_path

* Re-enable kv_cache optimizations for prefill

* reduce KV_CACHE_ALLOC_BLOCK for selective_batching
This commit is contained in:
Guancheng Fu 2023-12-25 10:29:31 +08:00 committed by GitHub
parent ed8ed76d4f
commit daf536fb2d
2 changed files with 126 additions and 85 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

105
python/llm/src/bigdl/llm/transformers/models/llama.py
View file

@ -317,6 +317,8 @@ def llama_attention_selective_batching_forward_4_31(
padding_mask: Optional[torch.LongTensor] = None, padding_mask: Optional[torch.LongTensor] = None,
**kwargs, **kwargs,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
# Minimize this value to reduce memory allocation.
KV_CACHE_ALLOC_BLOCK_LENGTH = 64
bsz, q_len, _ = hidden_states.size() bsz, q_len, _ = hidden_states.size()
device = hidden_states.device device = hidden_states.device
# for flash attention # for flash attention
@ -334,35 +336,48 @@ def llama_attention_selective_batching_forward_4_31(
attention_dtype = original_dtype attention_dtype = original_dtype
# TODO: decoding fast path # TODO: decoding fast path
# use_fuse_rope = should_use_fuse_rope(self, hidden_states, position_ids) use_fuse_rope = should_use_fuse_rope(self, hidden_states, position_ids)
# enough_kv_room = is_enough_kv_cache_room(past_key_value[0]) enough_kv_room = past_key_value is not None and is_enough_kv_cache_room_4_31(past_key_value[0])
# is_q4_0 = self.q_proj.qtype == SYM_INT4 is_q4_0 = self.q_proj.qtype == SYM_INT4
# no_tp = not self.config.pretraining_tp > 1 no_tp = not self.config.pretraining_tp > 1
# decoding_fast_path = (no_tp and is_q4_0 and use_fuse_rope and decoding_fast_path = (no_tp and is_q4_0 and use_fuse_rope and
# enough_kv_room and bsz * q_len == 1) bsz * q_len == 1)
updated_past_key_values = []
# single batch decoding fast path # single batch decoding fast path
# forward_qkv takes will perform QKV projection, rotary position embedding # forward_qkv takes will perform QKV projection, rotary position embedding
# and save the key/value states to cache, then return query states and the # and save the key/value states to cache, then return query states and the
# extended key/value cache # extended key/value cache
# if decoding_fast_path: if decoding_fast_path:
# hidden_states = hidden_states.view(1, -1) past_k = past_key_value[0][0]
# kv_seq_len = past_key_value[0].shape[-2] past_v = past_key_value[0][1]
# cache_k = past_key_value[0] kv_seq_len = past_k.shape[-2]
# cache_v = past_key_value[1] if not enough_kv_room:
# import linear_q4_0 new_cache_k, new_cache_v = extend_kv_cache(1,
# query_states, key_states, value_states = linear_q4_0.forward_qkv(hidden_states, self.num_key_value_heads, # Support GQA
# self.q_proj.weight, self.head_dim,
# self.k_proj.weight, kv_seq_len,
# self.v_proj.weight, kv_seq_len +
# position_ids, KV_CACHE_ALLOC_BLOCK_LENGTH,
# cache_k, cache_v, dtype=past_k.dtype,
# self.q_proj.weight.qtype, device=device)
# kv_seq_len, new_cache_k[:] = past_k
# self.head_dim) new_cache_v[:] = past_v
# kv_seq_len += 1 past_k = new_cache_k
past_v = new_cache_v
# else: hidden_states = hidden_states.view(1, -1)
import linear_q4_0
query_states, key_states, value_states = linear_q4_0.forward_qkv(hidden_states,
self.q_proj.weight,
self.k_proj.weight,
self.v_proj.weight,
position_ids,
past_k, past_v,
self.q_proj.weight.qtype,
kv_seq_len,
self.head_dim)
kv_seq_len += 1
else:
if self.config.pretraining_tp > 1: if self.config.pretraining_tp > 1:
invalidInputError(False, f"vLLM: config.pretraining_tp > 1 not supported yet") invalidInputError(False, f"vLLM: config.pretraining_tp > 1 not supported yet")
else: else:
@ -381,24 +396,44 @@ def llama_attention_selective_batching_forward_4_31(
if past_key_value is not None: if past_key_value is not None:
kv_seq_len += max(kv_pair[0].shape[-2] for kv_pair in past_key_value) kv_seq_len += max(kv_pair[0].shape[-2] for kv_pair in past_key_value)
# TODO: fuse_rope if use_fuse_rope:
query_states, key_states = apply_rotary_pos_emb_no_cache_xpu(query_states,
key_states,
position_ids,
"llama")
else:
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len) cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, query_states, key_states = apply_rotary_pos_emb(query_states, key_states,
cos, sin, position_ids, "llama") cos, sin, position_ids, "llama")
updated_past_key_values = []
if past_key_value is not None: if past_key_value is not None:
batched_attention_output = [] batched_attention_output = []
# print(f"type of attention_mask is {type(attention_mask)}") # print(f"type of attention_mask is {type(attention_mask)}")
for batch in range(bsz): for batch in range(bsz):
enough_kv_room = is_enough_kv_cache_room_4_31(past_key_value[batch])
past_k, past_v = past_key_value[batch] past_k, past_v = past_key_value[batch]
current_kv_len = past_k.shape[-2] + 1 current_kv_len = past_k.shape[-2] + 1
if not enough_kv_room:
# allocate new
new_cache_k, new_cache_v = extend_kv_cache(1,
self.num_key_value_heads,
self.head_dim,
past_k.size(2),
current_kv_len +
KV_CACHE_ALLOC_BLOCK_LENGTH,
dtype=past_k.dtype,
device=device)
new_cache_k[:] = past_k
new_cache_v[:] = past_v
past_k = new_cache_k
past_v = new_cache_v
current_key_states = torch.cat([past_k, current_key_states = key_states[batch: batch + 1, :, :, :]
key_states[batch: batch + 1, :, :, :]], dim=2) current_value_states = value_states[batch: batch + 1, :, :, :]
current_value_states = torch.cat([past_v, current_key_states, current_value_states = append_kv_cache(past_k,
value_states[batch: batch + 1, :, :, :]], dim=2) past_v,
current_key_states,
current_value_states)
updated_past_key_values.append((current_key_states, current_value_states)) updated_past_key_values.append((current_key_states, current_value_states))
current_key_states = repeat_kv(current_key_states, self.num_key_value_groups) current_key_states = repeat_kv(current_key_states, self.num_key_value_groups)
@ -434,8 +469,8 @@ def llama_attention_selective_batching_forward_4_31(
attn_output = self.o_proj(attn_output) attn_output = self.o_proj(attn_output)
return attn_output, None, updated_past_key_values return attn_output, None, updated_past_key_values
# TODO: Assume always use_cache # Assume always use_cache
# print(f"prefill with batch size {bsz}") # prefill or decoding fast path
for batch in range(bsz): for batch in range(bsz):
updated_past_key_values.append((key_states[batch: batch + 1, :, :, :], updated_past_key_values.append((key_states[batch: batch + 1, :, :, :],
value_states[batch: batch+1, :, :, :])) value_states[batch: batch+1, :, :, :]))
@ -445,6 +480,10 @@ def llama_attention_selective_batching_forward_4_31(
dtype=attention_dtype) dtype=attention_dtype)
value_states = repeat_kv(value_states, self.num_key_value_groups).to(device, value_states = repeat_kv(value_states, self.num_key_value_groups).to(device,
dtype=attention_dtype) dtype=attention_dtype)
# Can also happens for decoding fast path
if isinstance(attention_mask, list):
# For decoding fast path
attention_mask = attention_mask[0]
attn_output, attn_weights = native_sdp(query_states, attn_output, attn_weights = native_sdp(query_states,
key_states, key_states,
value_states, value_states,

4
python/llm/src/bigdl/llm/vllm/model_executor/models/bigdl_llama.py
View file

@ -196,7 +196,8 @@ class BigDLLlamaForCausalLM(BigDLModelForCausalLM):
if enable_vllm_se_batching: if enable_vllm_se_batching:
attention_mask = [torch.tensor(x, device=self.device).unsqueeze(0) attention_mask = [torch.tensor(x, device=self.device).unsqueeze(0)
for x in decoding_attention_mask_list] for x in decoding_attention_mask_list]
position_ids = torch.tensor(decoding_position_ids).long().unsqueeze(-1) position_ids = torch.tensor(decoding_position_ids, device=self.device).long()
position_ids = position_ids.unsqueeze(-1)
else: else:
attention_mask = torch.tensor(decoding_attention_mask_list, device=self.device) attention_mask = torch.tensor(decoding_attention_mask_list, device=self.device)
position_ids = None position_ids = None
@ -214,6 +215,7 @@ class BigDLLlamaForCausalLM(BigDLModelForCausalLM):
if enable_vllm_se_batching: if enable_vllm_se_batching:
position_ids = attention_mask.long().cumsum(-1) - 1 position_ids = attention_mask.long().cumsum(-1) - 1
position_ids.masked_fill_(attention_mask == 0, 1) position_ids.masked_fill_(attention_mask == 0, 1)
position_ids.to(self.device)
else: else:
position_ids = None position_ids = None
kwargs = { kwargs = {