Eager resolution for vllm.current_platform when running Pathways (#1150)

Signed-off-by: Richard Liu <ricliu@google.com>

Author: richardsliu

tpu_inference/__init__.py

@@ -12,9 +12,29 @@
     logger.info("Running vLLM on TPU via Pathways proxy.")
     # Must run pathwaysutils.initialize() before any JAX operations
     try:
+        import traceback
+
         import pathwaysutils
+        import vllm
+        from vllm.platforms import (resolve_current_platform_cls_qualname,
+                                    resolve_obj_by_qualname)
         pathwaysutils.initialize()
         logger.info("Module pathwaysutils is imported.")
+
+        # Pathways requires eager resolution of vllm.current_platform instead of
+        # lazy resolution in the normal code path. Since this part involves
+        # global topology discovery across multiple hosts, the platform
+        # resolution must happen before other components are loaded.
+        logger.info("Eagerly resolving vLLM current_platform for Pathways.")
+        platform_cls_qualname = resolve_current_platform_cls_qualname()
+        resolved_platform_instance = resolve_obj_by_qualname(
+            platform_cls_qualname)()
+        vllm.platforms._current_platform = resolved_platform_instance
+        vllm.platforms._init_trace = "".join(traceback.format_stack())
+        logger.info(
+            f"vLLM platform resolved to: {resolved_platform_instance.__class__.__name__}"
+        )
+
     except Exception as e:
         logger.error(
             f"Error occurred while importing pathwaysutils or logging TPU info: {e}"