biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[torch.compile] support all attention backends (#10558)

Browse Source 
Signed-off-by: youkaichao <youkaichao@gmail.com>
This commit is contained in:
youkaichao
2024-11-22 14:04:42 -08:00
committed by GitHub
parent db100c5cde
commit eebad39f26
77 changed files with 876 additions and 648 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
vllm/model_executor/models/baichuan.py
View File

@@ -116,6 +116,7 @@ class BaiChuanAttention(nn.Module):
max_position_embeddings: int = 8192,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
):
super().__init__()
self.hidden_size = hidden_size
@@ -158,7 +159,8 @@ class BaiChuanAttention(nn.Module):
self.head_dim,
scaling,
alibi_slopes=alibi_slopes,
quant_config=quant_config)
quant_config=quant_config,
prefix=f"{prefix}.attn")
else:
self.rotary_emb = get_rope(
self.head_dim,
@@ -171,7 +173,8 @@ class BaiChuanAttention(nn.Module):
self.head_dim,
self.scaling,
cache_config=cache_config,
quant_config=quant_config)
quant_config=quant_config,
prefix=f"{prefix}.attn")
def forward(
self,
@@ -195,7 +198,8 @@ class BaiChuanDecoderLayer(nn.Module):
config: PretrainedConfig,
position_embedding: str,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None):
quant_config: Optional[QuantizationConfig] = None,
prefix: str = ""):
super().__init__()
self.hidden_size = config.hidden_size
rope_theta = getattr(config, "rope_theta", 10000)
@@ -209,6 +213,7 @@ class BaiChuanDecoderLayer(nn.Module):
max_position_embeddings=max_position_embeddings,
cache_config=cache_config,
quant_config=quant_config,
prefix=f"{prefix}.self_attn",
)
self.mlp = BaiChuanMLP(
hidden_size=self.hidden_size,
@@ -275,8 +280,11 @@ class BaiChuanModel(nn.Module):
)
self.start_layer, self.end_layer, self.layers = make_layers(
config.num_hidden_layers,
lambda prefix: BaiChuanDecoderLayer(config, position_embedding,
cache_config, quant_config),
lambda prefix: BaiChuanDecoderLayer(config,
position_embedding,
cache_config,
quant_config,
prefix=prefix),
prefix=f"{prefix}.layers",
)
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)