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
@@ -110,14 +110,12 @@ class InternLMAttention(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(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.o_proj(attn_output)
|
||||
return output
|
||||
|
||||
@@ -160,7 +158,6 @@ class InternLMDecoderLayer(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
|
||||
@@ -175,7 +172,6 @@ class InternLMDecoderLayer(nn.Module):
|
||||
hidden_states=hidden_states,
|
||||
kv_cache=kv_cache,
|
||||
input_metadata=input_metadata,
|
||||
cache_event=cache_event,
|
||||
)
|
||||
|
||||
# Fully Connected
|
||||
@@ -214,19 +210,16 @@ class InternLMModel(nn.Module):
|
||||
positions: torch.Tensor,
|
||||
kv_caches: List[KVCache],
|
||||
input_metadata: InputMetadata,
|
||||
cache_events: Optional[List[torch.cuda.Event]],
|
||||
) -> torch.Tensor:
|
||||
hidden_states = self.embed_tokens(input_ids)
|
||||
residual = None
|
||||
for i in range(len(self.layers)):
|
||||
cache_event = None if cache_events is None else cache_events[i]
|
||||
layer = self.layers[i]
|
||||
hidden_states, residual = layer(
|
||||
positions,
|
||||
hidden_states,
|
||||
kv_caches[i],
|
||||
input_metadata,
|
||||
cache_event,
|
||||
residual,
|
||||
)
|
||||
hidden_states, _ = self.norm(hidden_states, residual)
|
||||
@@ -253,10 +246,9 @@ class InternLMForCausalLM(nn.Module):
|
||||
positions: torch.Tensor,
|
||||
kv_caches: List[KVCache],
|
||||
input_metadata: InputMetadata,
|
||||
cache_events: Optional[List[torch.cuda.Event]],
|
||||
) -> torch.Tensor:
|
||||
hidden_states = self.model(input_ids, positions, kv_caches,
|
||||
input_metadata, cache_events)
|
||||
input_metadata)
|
||||
return hidden_states
|
||||
|
||||
def sample(
|
||||
|
||||
Reference in New Issue
Block a user