[Bugfix] Fix HPU multiprocessing executor (vllm-project#12167)

Signed-off-by: Konrad Zawora <[email protected]>

vllm/config.py

@@ -1293,7 +1293,7 @@ def __post_init__(self) -> None:
                 raise ValueError(f"worker-use-ray can't be used with "
                                  f"distributed executor backend "
                                  f"'{self.distributed_executor_backend}'.")
-        ray_only_devices = ["tpu", "hpu"]
+        ray_only_devices = ["tpu"]
         from vllm.platforms import current_platform
         if (current_platform.device_type in ray_only_devices
                 and self.world_size > 1):

vllm/engine/arg_utils.py

@@ -397,7 +397,7 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
             'or equal to the number of GPUs available, "mp" will be used to '
             'keep processing on a single host. Otherwise, this will default '
             'to "ray" if Ray is installed and fail otherwise. Note that tpu '
-            'and hpu only support Ray for distributed inference.')
+            'only supports Ray for distributed inference.')
 
         parser.add_argument(
             '--worker-use-ray',

vllm/platforms/hpu.py

@@ -1,7 +1,9 @@
+import os
 from typing import TYPE_CHECKING, Optional
 
 import torch
 
+from vllm import envs
 from vllm.logger import init_logger
 
 from .interface import Platform, PlatformEnum, _Backend
@@ -58,6 +60,22 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
         cache_config = vllm_config.cache_config
         if cache_config and cache_config.block_size is None:
             cache_config.block_size = 128
+        if (parallel_config.distributed_executor_backend == 'mp'
+                and envs.VLLM_WORKER_MULTIPROC_METHOD == 'fork'):
+            if os.environ.get("VLLM_WORKER_MULTIPROC_METHOD",
+                              None) is not None:
+                logger.warning("On HPU, VLLM_WORKER_MULTIPROC_METHOD=fork "
+                               "might cause application hangs on exit. Using "
+                               "VLLM_WORKER_MULTIPROC_METHOD=fork anyway, "
+                               "as it was explicitly requested.")
+            else:
+                logger.warning(
+                    "On HPU, VLLM_WORKER_MULTIPROC_METHOD=fork "
+                    "might cause application hangs on exit. Setting "
+                    "VLLM_WORKER_MULTIPROC_METHOD to 'spawn'. "
+                    "To override that behavior, please set "
+                    "VLLM_WORKER_MULTIPROC_METHOD=fork explicitly.")
+                os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
 
     @classmethod
     def is_pin_memory_available(cls):

vllm/worker/hpu_worker.py

@@ -130,7 +130,6 @@ def execute_model(
         self,
         execute_model_req: Optional[ExecuteModelRequest] = None,
     ) -> Optional[List[SamplerOutput]]:
-        assert execute_model_req is not None
         # VLLM_HPU_LOG_STEP_GRAPH_COMPILATION     - will log graph compilations per engine step, only when there was any - highly recommended to use alongside PT_HPU_METRICS_GC_DETAILS! # noqa:E501
         # VLLM_HPU_LOG_STEP_GRAPH_COMPILATION_ALL - will log graph compilations per engine step, always, even if there were none # noqa:E501
         # VLLM_HPU_LOG_STEP_CPU_FALLBACKS         - will log cpu fallbacks per engine step, only when there was any # noqa:E501
@@ -144,7 +143,8 @@ def execute_model(
             'VLLM_HPU_LOG_STEP_CPU_FALLBACKS_ALL', '0') != '0'
         log_cpu_fallbacks = os.environ.get('VLLM_HPU_LOG_STEP_CPU_FALLBACKS',
                                            '0') != '0' or log_cpu_fallbacks_all
-        if log_graph_compilation or log_cpu_fallbacks:
+        if (log_graph_compilation or log_cpu_fallbacks) and \
+            execute_model_req is not None:
             from habana_frameworks.torch.hpu.metrics import metric_localcontext
             seq_group_metadata_list = execute_model_req.seq_group_metadata_list
             is_prompt = any([