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/qwen.py
View File

@@ -442,6 +442,7 @@ class QWenAttention(nn.Module):
rope_scaling: Optional[Dict[str, Any]] = None,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
):
super().__init__()
self.hidden_size = hidden_size
@@ -478,7 +479,8 @@ class QWenAttention(nn.Module):
self.head_dim,
self.scaling,
cache_config=cache_config,
quant_config=quant_config)
quant_config=quant_config,
prefix=f"{prefix}.attn")
def forward(
self,
@@ -502,6 +504,7 @@ class QWenBlock(nn.Module):
config: PretrainedConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
):
super().__init__()
self.ln_1 = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
@@ -514,7 +517,8 @@ class QWenBlock(nn.Module):
rope_theta=rope_theta,
rope_scaling=rope_scaling,
cache_config=cache_config,
quant_config=quant_config)
quant_config=quant_config,
prefix=f"{prefix}.attn")
self.ln_2 = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
@@ -568,7 +572,8 @@ class QWenModel(nn.Module):
)
self.start_layer, self.end_layer, self.h = make_layers(
config.num_hidden_layers,
lambda prefix: QWenBlock(config, cache_config, quant_config),
lambda prefix: QWenBlock(
config, cache_config, quant_config, prefix=prefix),
prefix=f"{prefix}.h")
self.ln_f = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
self.make_empty_intermediate_tensors = (