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:
Woosuk Kwon
2023-12-16 21:12:08 -08:00
committed by GitHub
co-authored by Chen Shen Antoni Baum
parent eed74a558f
commit 37ca558103
34 changed files with 557 additions and 254 deletions
+4 -1
View File
@@ -21,12 +21,14 @@ class InputMetadata:
max_context_len: Optional[int],
context_lens: Optional[torch.Tensor],
block_tables: Optional[torch.Tensor],
use_cuda_graph: bool,
) -> None:
self.prompt_lens = prompt_lens
self.max_context_len = max_context_len
self.slot_mapping = slot_mapping
self.context_lens = context_lens
self.block_tables = block_tables
self.use_cuda_graph = use_cuda_graph
self.is_prompt = len(prompt_lens) > 0
# Set during the execution of the first attention op.
@@ -39,4 +41,5 @@ class InputMetadata:
f"max_context_len={self.max_context_len}, "
f"slot_mapping={self.slot_mapping}, "
f"context_lens={self.context_lens}, "
f"block_tables={self.block_tables})")
f"block_tables={self.block_tables}, "
f"use_cuda_graph={self.use_cuda_graph})")