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

13
vllm/model_executor/models/gpt_bigcode.py
View File

@@ -95,7 +95,6 @@ class GPTBigCodeAttention(nn.Module):
hidden_states: torch.Tensor,
kv_cache: KVCache,
input_metadata: InputMetadata,
cache_event: Optional[torch.cuda.Event],
) -> torch.Tensor:
qkv, _ = self.c_attn(hidden_states)
q, k, v = qkv.split(
@@ -107,7 +106,7 @@ class GPTBigCodeAttention(nn.Module):
)
key_cache, value_cache = kv_cache
attn_output = self.attn(q, k, v, key_cache, value_cache,
input_metadata, cache_event)
input_metadata)
attn_output, _ = self.c_proj(attn_output)
return attn_output
@@ -167,7 +166,6 @@ class GPTBigCodeBlock(nn.Module):
hidden_states: torch.Tensor,
kv_cache: KVCache,
input_metadata: InputMetadata,
cache_event: Optional[torch.cuda.Event],
) -> torch.Tensor:
residual = hidden_states
hidden_states = self.ln_1(hidden_states)
@@ -175,7 +173,6 @@ class GPTBigCodeBlock(nn.Module):
hidden_states=hidden_states,
kv_cache=kv_cache,
input_metadata=input_metadata,
cache_event=cache_event,
)
# residual connection
hidden_states = attn_output + residual
@@ -215,17 +212,14 @@ class GPTBigCodeModel(nn.Module):
position_ids: torch.Tensor,
kv_caches: List[KVCache],
input_metadata: InputMetadata,
cache_events: Optional[List[torch.cuda.Event]],
) -> torch.Tensor:
inputs_embeds = self.wte(input_ids)
position_embeds = self.wpe(position_ids)
hidden_states = inputs_embeds + position_embeds
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(hidden_states, kv_caches[i], input_metadata,
cache_event)
hidden_states = layer(hidden_states, kv_caches[i], input_metadata)
hidden_states = self.ln_f(hidden_states)
return hidden_states
@@ -251,10 +245,9 @@ class GPTBigCodeForCausalLM(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(