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

14
vllm/model_executor/models/commandr.py
View File

@@ -120,6 +120,7 @@ class CohereAttention(nn.Module):
config: CohereConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
):
super().__init__()
tp_size = get_tensor_model_parallel_world_size()
@@ -175,7 +176,8 @@ class CohereAttention(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")
if self.use_qk_norm:
self.q_norm = LayerNorm(param_shape=(self.num_heads,
self.head_dim),
@@ -215,13 +217,15 @@ class CohereDecoderLayer(nn.Module):
def __init__(self,
config: CohereConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None):
quant_config: Optional[QuantizationConfig] = None,
prefix: str = ""):
super().__init__()
self.hidden_size = config.hidden_size
self.self_attn = CohereAttention(config,
cache_config,
quant_config=quant_config)
quant_config=quant_config,
prefix=f"{prefix}.self_attn")
self.mlp = CohereMLP(config, quant_config=quant_config)
self.input_layernorm = LayerNorm(param_shape=(config.hidden_size),
@@ -271,8 +275,8 @@ class CohereModel(nn.Module):
config.hidden_size)
self.start_layer, self.end_layer, self.layers = make_layers(
config.num_hidden_layers,
lambda prefix: CohereDecoderLayer(config, cache_config,
quant_config),
lambda prefix: CohereDecoderLayer(
config, cache_config, quant_config, prefix=prefix),
prefix=f"{prefix}.layers")
self.norm = LayerNorm(param_shape=(config.hidden_size),
eps=config.layer_norm_eps)