Support W&B 0.28 and vLLM 0.24 DSV4 LoRA

pyproject.toml

@@ -29,7 +29,7 @@ backend = [
     "accelerate==1.7.0",
     "awscli>=1.38.1",
     "setuptools>=78.1.0",
-    "wandb==0.25.0",
+    "wandb==0.28.0",
     "transformers==5.2.0",
     "duckdb>=1.0.0",
     "pyarrow>=15.0.0",
@@ -162,6 +162,7 @@ conflicts = [
     ],
 ]
 override-dependencies = [
+    "click==8.2.0",
     "flashinfer-python==0.6.8.post1",
     "megatron-core==0.17.0",
     "numpy<2",

tests/unit/test_dsv4_vllm_runtime_patches.py

@@ -25,6 +25,31 @@ def _load_dsv4_patches_module():
     return module
 
 
+def test_dsv4_lora_support_declares_vllm_024_manager_protocol(monkeypatch) -> None:
+    patches = _load_dsv4_patches_module()
+
+    class FakeDeepseekV4ForCausalLM:
+        pass
+
+    manager_patches: list[type] = []
+    monkeypatch.setattr(
+        patches,
+        "_import_dsv4_model_module",
+        lambda: SimpleNamespace(DeepseekV4ForCausalLM=FakeDeepseekV4ForCausalLM),
+    )
+    monkeypatch.setattr(
+        patches,
+        "_patch_dsv4_lora_manager_indexer_skip",
+        manager_patches.append,
+    )
+
+    patches.patch_dsv4_lora_support()
+
+    assert getattr(FakeDeepseekV4ForCausalLM, "supports_lora") is True
+    assert getattr(FakeDeepseekV4ForCausalLM, "lora_manager") is None
+    assert manager_patches == [FakeDeepseekV4ForCausalLM]
+
+
 def test_dsv4_compressor_helper_uses_punica_metadata_without_full_batch_lora(
     monkeypatch,
 ) -> None:

vllm_runtime/src/art_vllm_runtime/dsv4_patches.py

@@ -239,6 +239,7 @@ def patch_dsv4_lora_support() -> None:
     }
     model_cls.is_3d_moe_weight = False
     model_cls.is_non_gated_moe = False
+    model_cls.lora_manager = None
     model_cls.lora_skip_prefixes = ["mtp", "indexer"]
     model_cls._art_dsv4_lora_patched = True
     _patch_dsv4_lora_manager_indexer_skip(model_cls)