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

14
vllm/model_executor/models/bloom.py
View File

@@ -78,6 +78,7 @@ class BloomAttention(nn.Module):
config: BloomConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
):
super().__init__()
self.hidden_size = config.hidden_size
@@ -116,7 +117,8 @@ class BloomAttention(nn.Module):
scaling,
alibi_slopes=alibi_slopes,
cache_config=cache_config,
quant_config=quant_config)
quant_config=quant_config,
prefix=f"{prefix}.attn")
def forward(
self,
@@ -168,14 +170,17 @@ class BloomBlock(nn.Module):
config: BloomConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
):
super().__init__()
hidden_size = config.hidden_size
self.input_layernorm = nn.LayerNorm(hidden_size,
eps=config.layer_norm_epsilon)
self.self_attention = BloomAttention(config, cache_config,
quant_config)
self.self_attention = BloomAttention(config,
cache_config,
quant_config,
prefix=f"{prefix}.self_attention")
self.post_attention_layernorm = nn.LayerNorm(
hidden_size, eps=config.layer_norm_epsilon)
self.mlp = BloomMLP(config, quant_config)
@@ -242,7 +247,8 @@ class BloomModel(nn.Module):
# Transformer blocks
self.start_layer, self.end_layer, self.h = make_layers(
config.num_hidden_layers,
lambda prefix: BloomBlock(config, cache_config, quant_config),
lambda prefix: BloomBlock(
config, cache_config, quant_config, prefix=prefix),
prefix=f"{prefix}.h")
# Final Layer Norm