Refactor Attention (#1840)

This commit is contained in:
Woosuk Kwon
2023-11-29 15:37:31 -08:00
committed by GitHub
parent 0229c386c5
commit a9e4574261
16 changed files with 354 additions and 492 deletions
+11 -10
View File
@@ -24,11 +24,12 @@ from transformers import GPTJConfig
from vllm.model_executor.input_metadata import InputMetadata
from vllm.model_executor.layers.activation import get_act_fn
from vllm.model_executor.layers.attention import PagedAttentionWithRoPE
from vllm.model_executor.layers.attention import PagedAttention
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
LinearMethodBase,
QKVParallelLinear,
RowParallelLinear)
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.sampler import Sampler
from vllm.model_executor.layers.vocab_parallel_embedding import (
VocabParallelEmbedding, ParallelLMHead)
@@ -77,15 +78,14 @@ class GPTJAttention(nn.Module):
rope_theta = getattr(config, "rope_theta", 10000)
max_position_embeddings = getattr(config, "max_position_embeddings",
8192)
self.attn = PagedAttentionWithRoPE(
self.num_heads,
self.rotary_emb = get_rope(
self.head_size,
scaling,
config.rotary_dim,
base=rope_theta,
rotary_dim=config.rotary_dim,
max_position=max_position_embeddings,
is_neox_style=False)
self.warmup = False
base=rope_theta,
is_neox_style=False,
)
self.attn = PagedAttention(self.num_heads, self.head_size, scaling)
def forward(
self,
@@ -97,9 +97,10 @@ class GPTJAttention(nn.Module):
) -> torch.Tensor:
qkv, _ = self.qkv_proj(hidden_states)
q, k, v = qkv.chunk(chunks=3, dim=-1)
q, k = self.rotary_emb(position_ids, q, k)
k_cache, v_cache = kv_cache
attn_output = self.attn(position_ids, q, k, v, k_cache, v_cache,
input_metadata, cache_event)
attn_output = self.attn(q, k, v, k_cache, v_cache, input_metadata,
cache_event)
attn_output, _ = self.out_proj(attn_output)
return attn_output