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

17
vllm/model_executor/models/phi.py
View File

@@ -69,7 +69,8 @@ class PhiAttention(nn.Module):
def __init__(self,
config: PhiConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None):
quant_config: Optional[QuantizationConfig] = None,
prefix: str = ""):
super().__init__()
self.total_num_heads = config.num_attention_heads
self.hidden_size = config.hidden_size
@@ -116,7 +117,8 @@ class PhiAttention(nn.Module):
self.head_size,
scaling,
cache_config=cache_config,
quant_config=quant_config)
quant_config=quant_config,
prefix=f"{prefix}.attn")
def forward(
self,
@@ -167,11 +169,15 @@ class PhiLayer(nn.Module):
def __init__(self,
config: PhiConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None):
quant_config: Optional[QuantizationConfig] = None,
prefix: str = ""):
super().__init__()
self.input_layernorm = nn.LayerNorm(config.hidden_size,
eps=config.layer_norm_eps)
self.self_attn = PhiAttention(config, cache_config, quant_config)
self.self_attn = PhiAttention(config,
cache_config,
quant_config,
prefix=f"{prefix}.self_attn")
self.mlp = PhiMLP(config, quant_config)
def forward(
@@ -210,7 +216,8 @@ class PhiModel(nn.Module):
config.hidden_size)
self.start_layer, self.end_layer, self.layers = make_layers(
config.num_hidden_layers,
lambda prefix: PhiLayer(config, cache_config, quant_config),
lambda prefix: PhiLayer(
config, cache_config, quant_config, prefix=prefix),
prefix=f"{prefix}.layers")
self.final_layernorm = nn.LayerNorm(config.hidden_size,
eps=config.layer_norm_eps)