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

13
vllm/model_executor/models/mpt.py
View File

@@ -50,6 +50,7 @@ class MPTAttention(nn.Module):
config: MPTConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
):
super().__init__()
self.d_model = config.d_model
@@ -115,7 +116,8 @@ class MPTAttention(nn.Module):
alibi_slopes=alibi_slopes,
num_kv_heads=self.num_kv_heads,
cache_config=cache_config,
quant_config=quant_config)
quant_config=quant_config,
prefix=f"{prefix}.attn")
def forward(
self,
@@ -176,11 +178,15 @@ class MPTBlock(nn.Module):
config: MPTConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
):
super().__init__()
hidden_size = config.d_model
self.norm_1 = nn.LayerNorm(hidden_size)
self.attn = MPTAttention(config, cache_config, quant_config)
self.attn = MPTAttention(config,
cache_config,
quant_config,
prefix=f"{prefix}.attn")
self.norm_2 = nn.LayerNorm(hidden_size)
self.ffn = MPTMLP(config, quant_config)
@@ -224,7 +230,8 @@ class MPTModel(nn.Module):
)
self.start_layer, self.end_layer, self.blocks = make_layers(
config.n_layers,
lambda prefix: MPTBlock(config, cache_config, quant_config),
lambda prefix: MPTBlock(
config, cache_config, quant_config, prefix=prefix),
prefix=f"{prefix}.blocks")
self.norm_f = nn.LayerNorm(config.d_model)
if config.no_bias: