Disable guided decoding to resolve chat completion errors

The vLLM server was returning 500 errors on chat requests due to
guided decoding (outlines) import issues. Since basic chat doesn't
require structured generation, disabled guided decoding entirely
and removed airportsdata dependency.

app.py

@@ -27,6 +27,7 @@ VLLM_ARGS = [
     "--gpu-memory-utilization", "0.90",
     "--trust-remote-code",
     "--disable-log-requests",                # reduce log noise
+    "--disable-guided-decoding",             # skip guided decoding (outlines) to avoid import issues
 ]
 if "AWQ" in MODEL_ID.upper():
     VLLM_ARGS += ["--quantization", "awq_marlin"]  # faster AWQ kernel if available

requirements.txt

@@ -9,6 +9,3 @@ vllm==0.6.3.post1
 torch==2.4.0
 transformers>=4.44
 accelerate>=0.30
-
-# Required for vLLM guided decoding (even if not actively used)
-airportsdata>=20240400