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

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
+15 -11
View File
@@ -117,6 +117,7 @@ class Phi3SmallSelfAttention(nn.Module):
layer_idx: int,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
) -> None:
super().__init__()
self.layer_idx = layer_idx
@@ -214,15 +215,14 @@ class Phi3SmallSelfAttention(nn.Module):
"homo_head": self.homo_heads
}
self.attn = Attention(
self.num_heads_per_partition,
self.head_dim,
self.scale,
num_kv_heads=self.num_kv_heads_per_partion,
cache_config=cache_config,
quant_config=quant_config,
blocksparse_params=bs_params,
)
self.attn = Attention(self.num_heads_per_partition,
self.head_dim,
self.scale,
num_kv_heads=self.num_kv_heads_per_partion,
cache_config=cache_config,
quant_config=quant_config,
blocksparse_params=bs_params,
prefix=f"{prefix}.attn")
def forward(
self,
@@ -259,13 +259,15 @@ class Phi3SmallDecoderLayer(nn.Module):
layer_idx: int,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
):
super().__init__()
self.hidden_size = config.hidden_size
self.self_attn = Phi3SmallSelfAttention(config,
layer_idx,
cache_config=cache_config,
quant_config=quant_config)
quant_config=quant_config,
prefix=f"{prefix}.self_attn")
self.mlp = Phi3SmallMLP(config, quant_config)
self.input_layernorm = nn.LayerNorm(config.hidden_size,
@@ -315,7 +317,9 @@ class Phi3SmallModel(nn.Module):
config.num_hidden_layers,
lambda prefix: Phi3SmallDecoderLayer(config,
int(prefix.split('.')[-1]),
cache_config, quant_config),
cache_config,
quant_config,
prefix=prefix),
prefix=f"{prefix}.layers")
self.final_layernorm = nn.LayerNorm(config.hidden_size,