vllm-ascend vnpu v1

vllm_ascend/patch/platform/patch_executor.py

@@ -0,0 +1,44 @@
+import time
+
+from vllm.executor.executor_base import logger, ExecutorBase
+
+
+original_init = ExecutorBase.__init__
+def init(self, *args, **kwargs):
+    original_init(self, *args, **kwargs)
+    self.is_offloaded = False
+
+
+def offload_vram(self):
+    if self.is_offloaded:
+        logger.warning("Executor is already offloaded.")
+        return
+    time_before_offload = time.perf_counter()
+    self.collective_rpc("offload_vram")
+    time_after_offload = time.perf_counter()
+
+    self.is_offloaded = True
+    logger.info(f"Offloading VRAM costs {time_after_offload - time_before_offload:.6f} seconds.")
+
+
+def reload_vram(self) -> bool:
+    if not self.is_offloaded:
+        logger.warning("Executor is not offloaded.")
+        return True
+
+    time_before_reload = time.perf_counter()
+    prev_is_self = self.collective_rpc("reload_vram")
+    time_after_reload = time.perf_counter()
+    self.is_offloaded = False
+    logger.info(f"Reloading VRAM costs {time_after_reload - time_before_reload:.6f} seconds.")
+    return prev_is_self
+
+
+def determine_available_memory_idle_offload_mode(self) -> int:
+    return self.collective_rpc("determine_available_memory_idle_offload_mode")
+
+
+ExecutorBase.__init__ = init
+ExecutorBase.offload_vram = offload_vram
+ExecutorBase.reload_vram = reload_vram
+ExecutorBase.determine_available_memory_idle_offload_mode = determine_available_memory_idle_offload_mode