load punica wrapper obj dynamically

Signed-off-by: Shanshan Shen <[email protected]>

vllm/lora/punica_wrapper/punica_selector.py

@@ -1,7 +1,13 @@
 from vllm.platforms import current_platform
+from vllm.utils import print_info_once, resolve_obj_by_qualname
 
 from .punica_base import PunicaWrapperBase
 
 
 def get_punica_wrapper(*args, **kwargs) -> PunicaWrapperBase:
-    return current_platform.get_punica_wrapper(*args, **kwargs)
+    punica_wrapper_qualname = current_platform.get_punica_wrapper()
+    punica_wrapper_cls = resolve_obj_by_qualname(punica_wrapper_qualname)
+    punica_wrapper = punica_wrapper_cls(*args, **kwargs)
+    assert punica_wrapper is not None
+    print_info_once("Using " + punica_wrapper_qualname.rsplit(".", 1)[1] + ".")
+    return punica_wrapper

vllm/platforms/cuda.py

@@ -15,9 +15,6 @@
 import vllm._C  # noqa
 import vllm.envs as envs
 from vllm.logger import init_logger
-from vllm.lora.punica_wrapper.punica_base import PunicaWrapperBase
-from vllm.lora.punica_wrapper.punica_gpu import PunicaWrapperGPU
-from vllm.utils import print_info_once
 
 from .interface import DeviceCapability, Platform, PlatformEnum
 
@@ -145,9 +142,8 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
             cache_config.block_size = 16
 
     @classmethod
-    def get_punica_wrapper(cls, *args, **kwargs) -> PunicaWrapperBase:
-        print_info_once("Using PunicaWrapperGPU.")
-        return PunicaWrapperGPU(*args, **kwargs)
+    def get_punica_wrapper(cls) -> str:
+        return "vllm.lora.punica_wrapper.punica_gpu.PunicaWrapperGPU"
 
 
 # NVML utils

vllm/platforms/hpu.py

@@ -3,9 +3,6 @@
 import torch
 
 from vllm.logger import init_logger
-from vllm.lora.punica_wrapper.punica_base import PunicaWrapperBase
-from vllm.lora.punica_wrapper.punica_hpu import PunicaWrapperHPU
-from vllm.utils import print_info_once
 
 from .interface import Platform, PlatformEnum, _Backend
 
@@ -63,6 +60,5 @@ def is_pin_memory_available(cls):
         return False
 
     @classmethod
-    def get_punica_wrapper(cls, *args, **kwargs) -> PunicaWrapperBase:
-        print_info_once("Using PunicaWrapperHPU.")
-        return PunicaWrapperHPU(*args, **kwargs)
+    def get_punica_wrapper(cls) -> str:
+        return "vllm.lora.punica_wrapper.punica_hpu.PunicaWrapperHPU"

vllm/platforms/interface.py

@@ -8,7 +8,6 @@
 import torch
 
 from vllm.logger import init_logger
-from vllm.lora.punica_wrapper.punica_base import PunicaWrapperBase
 
 if TYPE_CHECKING:
     from vllm.config import VllmConfig
@@ -240,7 +239,7 @@ def is_pin_memory_available(cls) -> bool:
         return True
 
     @classmethod
-    def get_punica_wrapper(cls, *args, **kwargs) -> PunicaWrapperBase:
+    def get_punica_wrapper(cls) -> str:
         """
         Return the punica wrapper for current platform.
         """

vllm/platforms/rocm.py

@@ -6,9 +6,6 @@
 
 import vllm.envs as envs
 from vllm.logger import init_logger
-from vllm.lora.punica_wrapper.punica_base import PunicaWrapperBase
-from vllm.lora.punica_wrapper.punica_gpu import PunicaWrapperGPU
-from vllm.utils import print_info_once
 
 from .interface import DeviceCapability, Platform, PlatformEnum, _Backend
 
@@ -115,6 +112,5 @@ def verify_quantization(cls, quant: str) -> None:
         envs.VLLM_USE_TRITON_AWQ = True
 
     @classmethod
-    def get_punica_wrapper(cls, *args, **kwargs) -> PunicaWrapperBase:
-        print_info_once("Using PunicaWrapperGPU.")
-        return PunicaWrapperGPU(*args, **kwargs)
+    def get_punica_wrapper(cls) -> str:
+        return "vllm.lora.punica_wrapper.punica_gpu.PunicaWrapperGPU"