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/gpt_bigcode.py
View File

@@ -52,6 +52,7 @@ class GPTBigCodeAttention(nn.Module):
config: GPTBigCodeConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
):
super().__init__()
self.hidden_size = config.hidden_size
@@ -92,7 +93,8 @@ class GPTBigCodeAttention(nn.Module):
scale=self.scale,
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,
@@ -151,6 +153,7 @@ class GPTBigCodeBlock(nn.Module):
config: GPTBigCodeConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
):
super().__init__()
hidden_size = config.hidden_size
@@ -158,7 +161,10 @@ class GPTBigCodeBlock(nn.Module):
hidden_size)
self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
self.attn = GPTBigCodeAttention(config, cache_config, quant_config)
self.attn = GPTBigCodeAttention(config,
cache_config,
quant_config,
prefix=f"{prefix}.attn")
self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
self.mlp = GPTBigMLP(inner_dim, config, quant_config)
@@ -210,7 +216,8 @@ class GPTBigCodeModel(nn.Module):
self.wpe = nn.Embedding(config.max_position_embeddings, self.embed_dim)
self.start_layer, self.end_layer, self.h = make_layers(
config.num_hidden_layers,
lambda prefix: GPTBigCodeBlock(config, cache_config, quant_config),
lambda prefix: GPTBigCodeBlock(
config, cache_config, quant_config, prefix=prefix),
prefix=f"{prefix}.h",
)
self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)