Fix Extra VRAM Usage in PPTRT (PaddlePaddle#76296)

Limerances · web-flow · commit 7bc4bc487eb5 · 2025-11-13T11:22:24.000+08:00
diff --git a/python/paddle/tensorrt/export.py b/python/paddle/tensorrt/export.py
@@ -390,6 +390,8 @@ def convert_to_trt(program, trt_config, scope):
             scope=scope,
         )
 
+        paddle.device.empty_cache()
+
         # specify certain operators to be excluded from entering TensorRT
         if trt_config.disable_ops:
             forbid_op_lower_trt(program, trt_config.disable_ops)

Original file line number	Diff line number	Diff line change
`@@ -390,6 +390,8 @@ def convert_to_trt(program, trt_config, scope):`
`390`	`390`	`scope=scope,`
`391`	`391`	`)`
`392`	`392`
	`393`	`+ paddle.device.empty_cache()`
	`394`	`+`
`393`	`395`	`# specify certain operators to be excluded from entering TensorRT`
`394`	`396`	`if trt_config.disable_ops:`
`395`	`397`	`forbid_op_lower_trt(program, trt_config.disable_ops)`