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/qwen.py
View File

@@ -112,14 +112,12 @@ class QWenAttention(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.chunk(chunks=3, dim=-1)
q, k = self.rotary_emb(positions, q, k)
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.c_proj(attn_output)
return output
@@ -156,7 +154,6 @@ class QWenBlock(nn.Module):
hidden_states: torch.Tensor,
kv_cache: KVCache,
input_metadata: InputMetadata,
cache_event: Optional[torch.cuda.Event],
residual: Optional[torch.Tensor],
) -> Tuple[torch.Tensor, torch.Tensor]:
# Self Attention
@@ -170,7 +167,6 @@ class QWenBlock(nn.Module):
hidden_states=hidden_states,
kv_cache=kv_cache,
input_metadata=input_metadata,
cache_event=cache_event,
)
# Fully Connected
@@ -206,19 +202,16 @@ class QWenModel(nn.Module):
positions: torch.Tensor,
kv_caches: List[KVCache],
input_metadata: InputMetadata,
cache_events: Optional[List[torch.cuda.Event]],
) -> torch.Tensor:
hidden_states = self.wte(input_ids)
residual = None
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, residual = layer(
positions,
hidden_states,
kv_caches[i],
input_metadata,
cache_event,
residual,
)
hidden_states, _ = self.ln_f(hidden_states, residual)
@@ -245,10 +238,9 @@ class QWenLMHeadModel(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(