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

8
vllm/model_executor/models/jamba.py
View File

@@ -102,7 +102,8 @@ class JambaMambaDecoderLayer(nn.Module):
config: JambaConfig,
layer_idx: int,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None) -> None:
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "") -> None:
super().__init__()
self.config = config
self.mamba = MambaMixer(hidden_size= config.hidden_size,
@@ -157,6 +158,7 @@ class JambaAttentionDecoderLayer(nn.Module):
layer_idx: int,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
) -> None:
super().__init__()
self.hidden_size = config.hidden_size
@@ -198,6 +200,7 @@ class JambaAttentionDecoderLayer(nn.Module):
self.scaling,
num_kv_heads=self.num_kv_heads,
cache_config=cache_config,
prefix=f"{prefix}.attn",
)
num_experts = config.layers_num_experts[layer_idx]
@@ -287,7 +290,8 @@ class JambaModel(nn.Module):
layer_class(config,
layer_idx=i,
cache_config=cache_config,
quant_config=quant_config))
quant_config=quant_config,
prefix=f"{prefix}.layers.{i}"))
self.layers = nn.ModuleList(decoder_layers)
self.final_layernorm = RMSNorm(config.hidden_size,
eps=config.rms_norm_eps)