Optimize model execution with CUDA graph (#1926)
Co-authored-by: Chen Shen <scv119@gmail.com> Co-authored-by: Antoni Baum <antoni.baum@protonmail.com>
This commit is contained in:
co-authored by
Chen Shen
Antoni Baum
parent
eed74a558f
commit
37ca558103
@@ -94,14 +94,12 @@ class GPTJAttention(nn.Module):
|
||||
hidden_states: torch.Tensor,
|
||||
kv_cache: KVCache,
|
||||
input_metadata: InputMetadata,
|
||||
cache_event: Optional[torch.cuda.Event],
|
||||
) -> torch.Tensor:
|
||||
qkv, _ = self.qkv_proj(hidden_states)
|
||||
q, k, v = qkv.chunk(chunks=3, dim=-1)
|
||||
q, k = self.rotary_emb(position_ids, 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)
|
||||
attn_output, _ = self.out_proj(attn_output)
|
||||
return attn_output
|
||||
|
||||
@@ -156,7 +154,6 @@ class GPTJBlock(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)
|
||||
@@ -165,7 +162,6 @@ class GPTJBlock(nn.Module):
|
||||
hidden_states=hidden_states,
|
||||
kv_cache=kv_cache,
|
||||
input_metadata=input_metadata,
|
||||
cache_event=cache_event,
|
||||
)
|
||||
mlp_output = self.mlp(hidden_states)
|
||||
hidden_states = attn_output + mlp_output + residual
|
||||
@@ -196,18 +192,15 @@ class GPTJModel(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.wte(input_ids)
|
||||
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
|
||||
@@ -238,10 +231,9 @@ class GPTJForCausalLM(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(
|
||||
|
||||
Reference in New Issue
Block a user