[ROCm][CI] Fix v1/logits_processors failure on ROCm (#29927)

Signed-off-by: Micah Williamson <micah.williamson@amd.com>

tests/v1/logits_processors/test_custom_offline.py

 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import random
-import sys
 from typing import Any
 
 import pytest
@@ -10,7 +9,6 @@ from tests.utils import create_new_process_for_each_test
 from tests.v1.logits_processors.utils import (
     DUMMY_LOGITPROC_ARG,
     DUMMY_LOGITPROC_FQCN,
-    DUMMY_LOGITPROC_MODULE,
     MAX_TOKENS,
     MODEL_NAME,
     POOLING_MODEL_NAME,
@@ -18,7 +16,6 @@ from tests.v1.logits_processors.utils import (
     CustomLogitprocSource,
     DummyLogitsProcessor,
     WrappedPerReqLogitsProcessor,
-    dummy_module,
     prompts,
 )
 from tests.v1.logits_processors.utils import entry_points as fake_entry_points
@@ -162,8 +159,6 @@ def test_custom_logitsprocs(monkeypatch, logitproc_source: CustomLogitprocSource
     kwargs: dict[str, list[str | type[LogitsProcessor]]] = {}
     if logitproc_source == CustomLogitprocSource.LOGITPROC_SOURCE_FQCN:
         # Scenario: load logitproc based on fully-qualified class name (FQCN)
-        # Inject dummy module which defines logitproc
-        sys.modules[DUMMY_LOGITPROC_MODULE] = dummy_module
         kwargs["logits_processors"] = [DUMMY_LOGITPROC_FQCN]
     elif logitproc_source == CustomLogitprocSource.LOGITPROC_SOURCE_CLASS:
         # Scenario: load logitproc from provided class object

tests/v1/logits_processors/test_custom_online.py

@@ -14,11 +14,9 @@ from tests.utils import RemoteOpenAIServerCustom, create_new_process_for_each_te
 from tests.v1.logits_processors.utils import (
     DUMMY_LOGITPROC_ARG,
     DUMMY_LOGITPROC_FQCN,
-    DUMMY_LOGITPROC_MODULE,
     MAX_TOKENS,
     MODEL_NAME,
     TEMP_GREEDY,
-    dummy_module,
     prompts,
 )
 from tests.v1.logits_processors.utils import entry_points as fake_entry_points
@@ -47,20 +45,14 @@ def _server_with_logitproc_entrypoint(
     main.main()
 
 
-def _server_with_logitproc_module(
+def _server_with_logitproc_fqcn(
     env_dict: dict[str, str] | None,
     model: str,
     vllm_serve_args: list[str],
 ) -> None:
     """Start vLLM server, inject module with dummy logitproc"""
-
-    # Patch `modules` to inject dummy logitproc module
     from vllm.entrypoints.cli import main
 
-    sys.modules[DUMMY_LOGITPROC_MODULE] = dummy_module
-
-    # fork is required for workers to see entrypoint patch
-    os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "fork"
     if env_dict is not None:
         os.environ.update(env_dict)
 
@@ -99,7 +91,7 @@ def server(default_server_args, request, monkeypatch):
     if request.param:
         # Launch server, append FQCN argument, inject dummy logitproc module
         args = default_server_args + request.param
-        _server_fxn = _server_with_logitproc_module
+        _server_fxn = _server_with_logitproc_fqcn
     else:
         # Launch server, inject dummy logitproc entrypoint
         args = default_server_args

tests/v1/logits_processors/utils.py

@@ -27,7 +27,7 @@ DUMMY_LOGITPROC_ARG = "target_token"
 TEMP_GREEDY = 0.0
 MAX_TOKENS = 20
 DUMMY_LOGITPROC_ENTRYPOINT = "dummy_logitproc"
-DUMMY_LOGITPROC_MODULE = "DummyModule"
+DUMMY_LOGITPROC_MODULE = "tests.v1.logits_processors.utils"
 DUMMY_LOGITPROC_FQCN = f"{DUMMY_LOGITPROC_MODULE}:DummyLogitsProcessor"