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

11
vllm/model_executor/models/chameleon.py
View File

@@ -223,6 +223,7 @@ class ChameleonAttention(nn.Module):
quant_config: Optional[QuantizationConfig] = None,
bias: bool = False,
cache_config: Optional[CacheConfig] = None,
prefix: str = "",
) -> None:
super().__init__()
self.hidden_size = hidden_size
@@ -276,7 +277,8 @@ class ChameleonAttention(nn.Module):
self.scaling,
num_kv_heads=self.num_kv_heads,
cache_config=cache_config,
quant_config=quant_config)
quant_config=quant_config,
prefix=f"{prefix}.attn")
def _apply_qk_norm(self, q: torch.Tensor,
k: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
@@ -313,6 +315,7 @@ class ChameleonDecoderLayer(nn.Module):
config: ChameleonConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
) -> None:
super().__init__()
self.hidden_size = config.hidden_size
@@ -336,6 +339,7 @@ class ChameleonDecoderLayer(nn.Module):
quant_config=quant_config,
bias=False,
cache_config=cache_config,
prefix=f"{prefix}.self_attn",
)
self.mlp = ChameleonMLP(
hidden_size=self.hidden_size,
@@ -386,6 +390,7 @@ class ChameleonSwinDecoderLayer(nn.Module):
config: ChameleonConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
) -> None:
super().__init__()
self.hidden_size = config.hidden_size
@@ -409,6 +414,7 @@ class ChameleonSwinDecoderLayer(nn.Module):
quant_config=quant_config,
bias=False,
cache_config=cache_config,
prefix=f"{prefix}.self_attn",
)
self.mlp = ChameleonMLP(
hidden_size=self.hidden_size,
@@ -855,7 +861,8 @@ class ChameleonModel(nn.Module):
config.num_hidden_layers,
lambda prefix: decoder_layer(config=config,
cache_config=cache_config,
quant_config=quant_config),
quant_config=quant_config,
prefix=prefix),
prefix=f"{prefix}.layers",
)