refactor(qwen3.5): rely on the vLLM 0.24 branch contract

Author: CherryLemon

README.md

@@ -92,11 +92,13 @@ In theory, vllm-plugin-FL can support all models available in vLLM, as long as n
     pip install --no-build-isolation -e .
     ```
 
-### vLLM 0.24 runtime compatibility
+### Qwen3.5 text-only causal model support
 
-For vLLM 0.24.x, the plugin registers Qwen3.5 text-only causal model
-support at process startup without modifying the vLLM installation.
-It also routes `vllm._custom_ops.moe_sum` to the FlagGems implementation when
+The plugin supports Qwen3.5 text-only causal models. It installs the required
+config, model, and checkpoint compatibility at process startup without
+modifying the vLLM installation.
+
+For MoE variants, it also routes `vllm._custom_ops.moe_sum` to FlagGems when
 `USE_FLAGGEMS=1` and `moe_sum` is not excluded by the FlagOS
 whitelist/blacklist. This supports arbitrary MoE top-k with FP32 accumulation
 and handles empty-token batches without rebuilding or replacing vLLM's

…/unit_tests/patches/test_moe_sum_v024.py tests/unit_tests/patches/test_moe_sum.pytests/unit_tests/patches/test_moe_sum_v024.py renamed to tests/unit_tests/patches/test_moe_sum.py tests/unit_tests/patches/test_moe_sum_v024.py renamed to tests/unit_tests/patches/test_moe_sum.py

@@ -1,6 +1,6 @@
 from types import ModuleType
 
-from vllm_fl.patches import moe_sum_v024
+from vllm_fl.patches import moe_sum
 
 
 class _FakeTensor:
@@ -21,18 +21,17 @@ def test_moe_sum_patch_routes_nonempty_and_guards_empty(monkeypatch):
     ops = ModuleType("fake_vllm_custom_ops")
     ops.moe_sum = lambda input, output: calls.append(("original", input, output))
 
-    monkeypatch.setattr(moe_sum_v024, "is_vllm_024", lambda: True)
     monkeypatch.setattr(
-        moe_sum_v024, "use_flaggems_op", lambda op_name: op_name == "moe_sum"
+        moe_sum, "use_flaggems_op", lambda op_name: op_name == "moe_sum"
     )
     monkeypatch.setattr(
-        moe_sum_v024,
+        moe_sum,
         "_flag_gems_moe_sum",
         lambda input, output: calls.append(("flag_gems", input, output)),
     )
 
-    assert moe_sum_v024.patch_vllm_moe_sum(ops) is True
-    assert moe_sum_v024.patch_vllm_moe_sum(ops) is False
+    assert moe_sum.patch_vllm_moe_sum(ops) is True
+    assert moe_sum.patch_vllm_moe_sum(ops) is False
 
     nonempty_input = _FakeTensor(24)
     nonempty_output = _FakeTensor(8)
@@ -48,15 +47,14 @@ def test_moe_sum_patch_uses_stride_safe_fallback(monkeypatch):
     ops = ModuleType("fake_vllm_custom_ops")
     ops.moe_sum = lambda input, output: None
 
-    monkeypatch.setattr(moe_sum_v024, "is_vllm_024", lambda: True)
-    monkeypatch.setattr(moe_sum_v024, "use_flaggems_op", lambda op_name: True)
+    monkeypatch.setattr(moe_sum, "use_flaggems_op", lambda op_name: True)
     monkeypatch.setattr(
-        moe_sum_v024,
+        moe_sum,
         "_torch_moe_sum",
         lambda input, output: calls.append((input, output)),
     )
 
-    assert moe_sum_v024.patch_vllm_moe_sum(ops) is True
+    assert moe_sum.patch_vllm_moe_sum(ops) is True
     noncontiguous_input = _FakeTensor(24, hidden_stride=2)
     output = _FakeTensor(8)
     ops.moe_sum(noncontiguous_input, output)
@@ -69,8 +67,7 @@ def test_moe_sum_patch_respects_flaggems_disable(monkeypatch):
     original = lambda input, output: None
     ops.moe_sum = original
 
-    monkeypatch.setattr(moe_sum_v024, "is_vllm_024", lambda: True)
-    monkeypatch.setattr(moe_sum_v024, "use_flaggems_op", lambda op_name: False)
+    monkeypatch.setattr(moe_sum, "use_flaggems_op", lambda op_name: False)
 
-    assert moe_sum_v024.patch_vllm_moe_sum(ops) is False
+    assert moe_sum.patch_vllm_moe_sum(ops) is False
     assert ops.moe_sum is original

…/unit_tests/patches/test_qwen3_5_v024.py …/unit_tests/patches/test_qwen3_5_text.pytests/unit_tests/patches/test_qwen3_5_v024.py renamed to tests/unit_tests/patches/test_qwen3_5_text.py tests/unit_tests/patches/test_qwen3_5_v024.py renamed to tests/unit_tests/patches/test_qwen3_5_text.py

@@ -1,6 +1,6 @@
 from types import SimpleNamespace
 
-from vllm_fl.patches import qwen3_5_v024 as compat
+from vllm_fl.patches import qwen3_5_text as compat
 
 
 def test_text_config_convertor_remaps_conditional_architecture():
@@ -65,15 +65,14 @@ def test_apply_registers_only_plugin_owned_lazy_models(monkeypatch):
             architecture, model
         )
     )
-    monkeypatch.setattr(compat, "is_vllm_024", lambda: True)
     monkeypatch.setattr(transformers_config, "_CONFIG_REGISTRY", {})
     monkeypatch.setattr(model_config, "MODELS_CONFIG_MAP", {})
     monkeypatch.setattr(model_arch_config_convertor, "MODEL_ARCH_CONFIG_CONVERTORS", {})
     monkeypatch.setattr(model_registry, "_TEXT_GENERATION_MODELS", {})
     monkeypatch.setattr(model_registry, "_VLLM_MODELS", {})
     monkeypatch.setattr(model_registry, "ModelRegistry", fake_registry)
 
-    assert compat.apply_qwen3_5_v024_patches() is True
+    assert compat.apply_qwen3_5_text_patches() is True
 
     assert registered == {
         "Qwen3_5ForCausalLM": ("vllm_fl.models.qwen3_5:Qwen3_5ForCausalLM"),

vllm_fl/__init__.py

@@ -138,17 +138,13 @@ def register_router():
 
 def register_model():
     """Register FL-specific models not yet upstream."""
-    # Keep all vLLM 0.24 compatibility changes plugin-owned.  These hooks are
-    # idempotent because general plugins are loaded independently in spawned
-    # model-inspection and worker processes.
-    from vllm_fl.patches._version import is_vllm_024
+    # General plugins are loaded independently in spawned model-inspection and
+    # worker processes, so all runtime compatibility hooks must be idempotent.
+    from vllm_fl.patches.moe_sum import patch_vllm_moe_sum
+    from vllm_fl.patches.qwen3_5_text import apply_qwen3_5_text_patches
 
-    if is_vllm_024():
-        from vllm_fl.patches.qwen3_5_v024 import apply_qwen3_5_v024_patches
-        from vllm_fl.patches.moe_sum_v024 import patch_vllm_moe_sum
-
-        apply_qwen3_5_v024_patches()
-        patch_vllm_moe_sum()
+    apply_qwen3_5_text_patches()
+    patch_vllm_moe_sum()
 
     _register_flagcx_connector()
 

vllm_fl/models/qwen3_5.py

@@ -1,10 +1,10 @@
 # Copyright (c) 2026 BAAI. All rights reserved.
-"""Runtime compatibility shim for Qwen3.5 text-only models on vLLM 0.24.
+"""Runtime compatibility shim for Qwen3.5 text-only causal models.
 
 The model implementation remains the one shipped by vLLM. Importing this
 module adds the hybrid-model metadata, cache helpers, and VL checkpoint prefix
-mapping that the v0.24 text-only classes are missing, then re-exports those
-upstream classes for lazy registration by the FL plugin.
+mapping that the upstream text-only classes are missing, then re-exports those
+classes for lazy registration by the FL plugin.
 """
 
 from vllm.model_executor.layers.mamba.mamba_utils import (

vllm_fl/patches/_version.py

@@ -1,11 +1,10 @@
 # Copyright (c) 2026 BAAI. All rights reserved.
-"""Route vLLM 0.24 MoE reduction through FlagGems at runtime."""
+"""Route vLLM MoE reduction through FlagGems at runtime."""
 
-from functools import wraps
 import logging
+from functools import wraps
 from types import ModuleType
 
-from vllm_fl.patches._version import is_vllm_024
 from vllm_fl.utils import use_flaggems_op
 
 logger = logging.getLogger(__name__)
@@ -30,7 +29,7 @@ def patch_vllm_moe_sum(ops_module: ModuleType | None = None) -> bool:
     every call.  The replacement therefore covers Triton/FP8 experts while
     leaving the vLLM wheel and its compiled extensions untouched.
     """
-    if not is_vllm_024() or not use_flaggems_op("moe_sum"):
+    if not use_flaggems_op("moe_sum"):
         return False
 
     if ops_module is None:
@@ -56,7 +55,7 @@ def moe_sum_flagos(input, output):
     moe_sum_flagos._vllm_fl_moe_sum_patch = True
     moe_sum_flagos._vllm_fl_original = original
     ops_module.moe_sum = moe_sum_flagos
-    logger.info("Monkey-patched vLLM 0.24 moe_sum -> FlagGems runtime operator")
+    logger.info("Monkey-patched vLLM moe_sum -> FlagGems runtime operator")
     return True
 
 

vllm_fl/patches/moe_sum_v024.py vllm_fl/patches/moe_sum.pyvllm_fl/patches/moe_sum_v024.py renamed to vllm_fl/patches/moe_sum.py

@@ -1,8 +1,8 @@
 # Copyright (c) 2026 BAAI. All rights reserved.
-"""Install Qwen3.5 text-only compatibility at vLLM runtime.
+"""Install Qwen3.5 text-only causal model support at vLLM runtime.
 
 The compatibility layer mirrors the text-only config, model, and checkpoint
-handling added after vLLM 0.24 without modifying the vLLM installation.
+handling missing from upstream vLLM without modifying the vLLM installation.
 """
 
 import logging
@@ -13,7 +13,6 @@
 from vllm.transformers_utils.model_arch_config_convertor import (
     ModelArchConfigConvertorBase,
 )
-from vllm_fl.patches._version import is_vllm_024
 
 logger = logging.getLogger(__name__)
 
@@ -44,9 +43,9 @@ def get_architectures(self) -> list[str]:
                 _CONDITIONAL_TO_CAUSAL.get(arch, arch) for arch in architectures
             ]
 
-        # vLLM 0.24 consults hf_config.architectures again in the runtime model
-        # loader, after ModelArchitectureConfig has been built.  Keep both
-        # views synchronized so it cannot fall back to a stale VL architecture.
+        # The runtime model loader consults hf_config.architectures again after
+        # ModelArchitectureConfig has been built. Keep both views synchronized
+        # so it cannot fall back to a stale VL architecture.
         if normalized != architectures:
             self.hf_config.architectures = normalized.copy()
         return normalized
@@ -66,19 +65,20 @@ def verify_and_update_config(vllm_config) -> None:
             rope_parameters.pop("mrope_interleaved", None)
 
 
-def apply_qwen3_5_v024_patches() -> bool:
-    """Register the runtime shims required by pristine vLLM 0.24.x.
+def apply_qwen3_5_text_patches() -> bool:
+    """Register Qwen3.5 text-only causal model support.
 
-    Returns ``True`` when the compatibility path applies to this vLLM release.
-    Repeated calls are safe.
+    Repeated calls are safe. Returns ``True`` after installing the runtime
+    registrations.
     """
-    if not is_vllm_024():
-        return False
-
-    from vllm.model_executor.models import config as model_config
-    from vllm.model_executor.models import registry as model_registry
-    from vllm.transformers_utils import config as transformers_config
-    from vllm.transformers_utils import model_arch_config_convertor
+    from vllm.model_executor.models import (
+        config as model_config,
+        registry as model_registry,
+    )
+    from vllm.transformers_utils import (
+        config as transformers_config,
+        model_arch_config_convertor,
+    )
 
     config_registry = transformers_config._CONFIG_REGISTRY
     config_registry.setdefault("qwen3_5_text", "Qwen3_5TextConfig")
@@ -98,20 +98,17 @@ def apply_qwen3_5_v024_patches() -> bool:
         model_registry._TEXT_GENERATION_MODELS.setdefault(
             architecture, ("qwen3_5", class_name)
         )
-        model_registry._VLLM_MODELS.setdefault(
-            architecture, ("qwen3_5", class_name)
-        )
+        model_registry._VLLM_MODELS.setdefault(architecture, ("qwen3_5", class_name))
         model_registry.ModelRegistry.register_model(
             architecture, f"vllm_fl.models.qwen3_5:{class_name}"
         )
 
-    logger.info("Installed vLLM 0.24 Qwen3.5 text-only runtime compatibility")
+    logger.info("Installed Qwen3.5 text-only causal model runtime compatibility")
     return True
 
 
 __all__ = [
     "Qwen3_5ForCausalLMConfig",
     "Qwen3_5TextModelArchConfigConvertor",
-    "apply_qwen3_5_v024_patches",
-    "is_vllm_024",
+    "apply_qwen3_5_text_patches",
 ]