[core] LLM.collective_rpc interface and RLHF example (#12084)

Signed-off-by: youkaichao <youkaichao@gmail.com>

vllm/plugins/__init__.py

@@ -1,9 +1,6 @@
 import logging
-import os
 from typing import Callable, Dict
 
-import torch
-
 import vllm.envs as envs
 
 logger = logging.getLogger(__name__)
@@ -50,34 +47,6 @@ def load_general_plugins():
     processes. They should be designed in a way that they can be loaded
     multiple times without causing issues.
     """
-
-    # all processes created by vllm will load plugins,
-    # and here we can inject some common environment variables
-    # for all processes.
-
-    # see https://github.com/vllm-project/vllm/issues/10480
-    os.environ['TORCHINDUCTOR_COMPILE_THREADS'] = '1'
-    # see https://github.com/vllm-project/vllm/issues/10619
-    torch._inductor.config.compile_threads = 1
-
-    from vllm.platforms import current_platform
-
-    if current_platform.is_xpu():
-        # see https://github.com/pytorch/pytorch/blob/43c5f59/torch/_dynamo/config.py#L158
-        torch._dynamo.config.disable = True
-    if current_platform.is_hpu():
-        # NOTE(kzawora): PT HPU lazy backend (PT_HPU_LAZY_MODE = 1)
-        # does not support torch.compile
-        # Eager backend (PT_HPU_LAZY_MODE = 0) must be selected for
-        # torch.compile support
-        is_lazy = os.environ.get('PT_HPU_LAZY_MODE', '1') == '1'
-        if is_lazy:
-            torch._dynamo.config.disable = True
-            # NOTE(kzawora) multi-HPU inference with HPUGraphs (lazy-only)
-            # requires enabling lazy collectives
-            # see https://docs.habana.ai/en/latest/PyTorch/Inference_on_PyTorch/Inference_Using_HPU_Graphs.html # noqa: E501
-            os.environ['PT_HPU_ENABLE_LAZY_COLLECTIVES'] = 'true'
-
     global plugins_loaded
     if plugins_loaded:
         return