diff --git a/pyproject.toml b/pyproject.toml
index 06099e0..0a4a9f4 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -34,7 +34,7 @@ serve = ["fastapi>=0.110", "uvicorn>=0.29"]
 vllm = ["fastapi>=0.110", "uvicorn>=0.29", "vllm>=0.12"]  # decision-model support: vLLM nightly from 2026-10-15
 fla = ["flash-linear-attention>=0.5"]  # fast Gated DeltaNet kernels for Qwen3.5/3.8 backbones
 data = ["datasets>=3"]
-dev = ["pytest>=8", "ruff>=0.6", "mypy>=1.11", "fastapi>=0.110", "httpx>=0.27"]
+dev = ["pytest>=8", "ruff>=0.6", "mypy>=1.11", "fastapi>=0.110", "uvicorn>=0.29", "httpx>=0.27"]
 
 [project.scripts]
 decisionformer = "decisionformer.cli:main"
diff --git a/src/decisionformer/serve.py b/src/decisionformer/serve.py
index 7f934ba..00899d8 100644
--- a/src/decisionformer/serve.py
+++ b/src/decisionformer/serve.py
@@ -90,7 +90,7 @@ def load_decider(model: str, *, backend: str = "transformers", device: str | Non
 
 def serve(model: str, *, host: str = "127.0.0.1", port: int = 8000, device: str | None = None,
           backend: str = "transformers", api_key_env: str = "DECISIONFORMER_API_KEY", **load_options: Any) -> None:
+    decider = load_decider(model, backend=backend, device=device, **load_options)  # model errors first
     import uvicorn
 
-    decider = load_decider(model, backend=backend, device=device, **load_options)
     uvicorn.run(create_app(decider, api_key=os.environ.get(api_key_env)), host=host, port=port)
diff --git a/tests/conftest.py b/tests/conftest.py
index 6ebdcfc..732f033 100644
--- a/tests/conftest.py
+++ b/tests/conftest.py
@@ -212,7 +212,8 @@ def data_dir(tmp_path_factory) -> Path:
     "qwen3_moe": {"num_experts": 4, "num_experts_per_tok": 2, "moe_intermediate_size": 32, "decoder_sparse_step": 1},
     "olmoe": {"num_experts": 4, "num_experts_per_tok": 2},
     "gpt_oss": {**_MOE, "layer_types": ["sliding_attention", "full_attention"], "sliding_window": 64},
-    "deepseek_v3": {"n_routed_experts": 4, "num_experts_per_tok": 2, "moe_intermediate_size": 32,
+    # multi-head latent attention has as many key/value heads as query heads
+    "deepseek_v3": {"n_routed_experts": 4, "num_experts_per_tok": 2, "moe_intermediate_size": 32, "num_key_value_heads": 4,
                     "first_k_dense_replace": 1, "kv_lora_rank": 16, "q_lora_rank": 16, "qk_rope_head_dim": 8,
                     "qk_nope_head_dim": 8, "v_head_dim": 8, "n_group": 1, "topk_group": 1},
     # hybrid linear attention (Gated DeltaNet) + full attention
@@ -236,8 +237,9 @@ def build_zoo_model(root: Path, model_type: str) -> Path:
     tokenizer = make_tokenizer("encoder" if encoder else "decoder")
     kwargs = dict(_TINY, vocab_size=len(tokenizer), **(ZOO_ENCODERS if encoder else ZOO_DECODERS)[model_type])
     if not encoder:
-        kwargs.update(num_key_value_heads=2, eos_token_id=SPECIALS.index("<|end|>"), bos_token_id=None,
-                      tie_word_embeddings=True, **({} if model_type == "falcon" else {"head_dim": 8}))
+        kwargs = {"num_key_value_heads": 2, **kwargs}
+        kwargs.update(eos_token_id=SPECIALS.index("<|end|>"), bos_token_id=None, tie_word_embeddings=True,
+                      **({} if model_type == "falcon" else {"head_dim": 8}))
     torch.manual_seed(0)
     config = AutoConfig.for_model(model_type, **kwargs)
     if getattr(config, "is_encoder_decoder", False):