This PR is a refactoring of model runner, to decouple it from the cla…

…sses specifically designed for GPU. Signed-off-by: Shanshan Shen <467638484@qq.com>
vllm-project · Feb 5, 2025 · 8b45b23 · 8b45b23
1 parent d5e7756
commit 8b45b23
Show file tree

Hide file tree

Showing 4 changed files with 1,104 additions and 328 deletions.
diff --git a/.gitignore b/.gitignore
@@ -192,3 +192,4 @@ cython_debug/
 # PyPI configuration file
 .pypirc
 
+kernel_meta/
diff --git a/vllm_ascend/attention.py b/vllm_ascend/attention.py
@@ -321,6 +321,13 @@ class AscendMetadataBuilder(CommonMetadataBuilder[AscendMetadata]):
 
     _metadata_cls = AscendMetadata
 
+    def __init__(self, input_builder: "ModelInputForNPUBuilder"):
+        self.input_builder = input_builder
+        self.runner = input_builder.runner
+        self.sliding_window = input_builder.sliding_window
+        self.block_size = input_builder.block_size
+        self.prepare()
+
     def compute_npu_slot_indices(self, is_profile_run, slot_indices, seq_id,
                                  seq_len, context_len, start_idx, block_size,
                                  block_tables, max_query_len):
Original file line number	Diff line number	Diff line change
Expand Up		@@ -192,3 +192,4 @@ cython_debug/
		# PyPI configuration file
		.pypirc

		kernel_meta/