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

25
vllm/model_executor/models/chatglm.py
View File

@@ -230,6 +230,7 @@ class GLMAttention(nn.Module):
config: ChatGLMConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
):
super().__init__()
self.hidden_size = config.hidden_size
@@ -285,7 +286,8 @@ class GLMAttention(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 forward(
self,
@@ -364,6 +366,7 @@ class GLMBlock(nn.Module):
config: ChatGLMConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
):
super().__init__()
self.apply_residual_connection_post_layernorm = (
@@ -377,7 +380,10 @@ class GLMBlock(nn.Module):
eps=config.layernorm_epsilon)
# Self attention.
self.self_attention = GLMAttention(config, cache_config, quant_config)
self.self_attention = GLMAttention(config,
cache_config,
quant_config,
prefix=f"{prefix}.self_attention")
self.hidden_dropout = config.hidden_dropout
# Layernorm on the attention output
@@ -446,7 +452,8 @@ class GLMTransformer(nn.Module):
# Transformer layers.
self.start_layer, self.end_layer, self.layers = make_layers(
self.num_layers,
lambda prefix: GLMBlock(config, cache_config, quant_config),
lambda prefix: GLMBlock(
config, cache_config, quant_config, prefix=prefix),
prefix=f"{prefix}.layers",
)
@@ -500,16 +507,22 @@ class ChatGLMModel(nn.Module):
self.num_layers = config.num_layers
self.multi_query_group_num = config.multi_query_group_num
self.kv_channels = config.kv_channels
self.encoder = GLMTransformer(config, cache_config, quant_config)
self.encoder = GLMTransformer(config,
cache_config,
quant_config,
prefix=f"{prefix}.encoder")
self.output_layer = ParallelLMHead(config.padded_vocab_size,
config.hidden_size,
quant_config=quant_config)
quant_config=quant_config,
prefix=f"{prefix}.output_layer")
vision_config_flag = getattr(config, 'vision_config', None)
if vision_config_flag is not None:
self.vision_config = Namespace(**config.vision_config)
self.vision = EVA2CLIPModel(self.config, quant_config)
self.vision = EVA2CLIPModel(self.config,
quant_config,
prefix=f"{prefix}.vision")
else:
self.vision = None