changes based on torchtitan

example_trainer/vllm_patching/__init__.py

@@ -0,0 +1,37 @@
+"""
+vLLM Patching Module - Enables shared memory weight updates.
+
+This module patches vLLM's GPUModelRunner to:
+1. Call share_memory_() on model weights after loading
+2. Spawn a daemon process that receives NCCL weight updates from trainers
+3. Enable real-time weight synchronization without restarting vLLM
+
+Usage:
+    # Import this BEFORE importing vllm
+    from example_trainer.vllm_patching import apply_patches
+    apply_patches()
+    
+    # Then import vllm normally
+    from vllm import AsyncLLM
+"""
+
+from .patched_gpu_runner import PatchedGPUModelRunner, apply_patches
+from .weight_updater import weight_updater_process
+from .distributed_utils import (
+    init_process_group,
+    broadcast_object_list,
+    get_inference_urls,
+    get_json_data,
+)
+
+__all__ = [
+    "PatchedGPUModelRunner",
+    "apply_patches",
+    "weight_updater_process",
+    "init_process_group",
+    "broadcast_object_list",
+    "get_inference_urls",
+    "get_json_data",
+]
+
+