biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

Optimize model execution with CUDA graph (#1926)

Browse Source 
Co-authored-by: Chen Shen <scv119@gmail.com>
Co-authored-by: Antoni Baum <antoni.baum@protonmail.com>
This commit is contained in:
Woosuk Kwon
2023-12-16 21:12:08 -08:00
committed by GitHub
parent eed74a558f
commit 37ca558103
34 changed files with 557 additions and 254 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
vllm/model_executor/models/bloom.py
View File

@@ -118,14 +118,12 @@ class BloomAttention(nn.Module):
hidden_states: torch.Tensor,
kv_cache: KVCache,
input_metadata: InputMetadata,
cache_event: Optional[torch.cuda.Event],
) -> torch.Tensor:
del position_ids # Unused.
qkv, _ = self.query_key_value(hidden_states)
q, k, v = qkv.chunk(chunks=3, dim=-1)
k_cache, v_cache = kv_cache
attn_output = self.attn(q, k, v, k_cache, v_cache, input_metadata,
cache_event)
attn_output = self.attn(q, k, v, k_cache, v_cache, input_metadata)
output, _ = self.dense(attn_output)
return output
@@ -184,7 +182,6 @@ class BloomBlock(nn.Module):
hidden_states: torch.Tensor,
kv_cache: KVCache,
input_metadata: InputMetadata,
cache_event: Optional[torch.cuda.Event],
) -> torch.Tensor:
# Layer norm at the beginning of the transformer layer.
layernorm_output = self.input_layernorm(hidden_states)
@@ -201,7 +198,6 @@ class BloomBlock(nn.Module):
hidden_states=layernorm_output,
kv_cache=kv_cache,
input_metadata=input_metadata,
cache_event=cache_event,
)
attention_output = attention_output + residual
layernorm_output = self.post_attention_layernorm(attention_output)
@@ -250,19 +246,16 @@ class BloomModel(nn.Module):
position_ids: torch.Tensor,
kv_caches: List[KVCache],
input_metadata: InputMetadata,
cache_events: Optional[List[torch.cuda.Event]],
) -> torch.Tensor:
hidden_states = self.word_embeddings(input_ids)
hidden_states = self.word_embeddings_layernorm(hidden_states)
for i in range(len(self.h)):
cache_event = None if cache_events is None else cache_events[i]
layer = self.h[i]
hidden_states = layer(
position_ids,
hidden_states,
kv_caches[i],
input_metadata,
cache_event,
)
hidden_states = self.ln_f(hidden_states)
return hidden_states
@@ -288,10 +281,9 @@ class BloomForCausalLM(nn.Module):
positions: torch.Tensor,
kv_caches: List[KVCache],
input_metadata: InputMetadata,
cache_events: Optional[List[torch.cuda.Event]],
) -> torch.Tensor:
hidden_states = self.transformer(input_ids, positions, kv_caches,
input_metadata, cache_events)
input_metadata)
return hidden_states
def sample(