langfuse generation spans with token usage; switch llm to vllm (qwen2.5-coder-7b on mcrndeb)

api/runtime/runner.py

@@ -128,7 +128,7 @@ async def _synthesize_node(state: RunState) -> dict[str, Any]:
         for f in findings
     )
 
-    with lf.span("synthesize", as_type="generation", input={"findings": len(findings)}) as span:
+    with lf.span("synthesize", input={"findings": len(findings)}) as span:
         answer = chat(
             system=load("synthesize.system"),
             user=render(
@@ -137,6 +137,7 @@ async def _synthesize_node(state: RunState) -> dict[str, Any]:
                 findings_block=findings_block,
             ),
             max_tokens=512,
+            span_name="synthesize.gen",
         )
         span.update(output={"answer": answer})
     return {"answer": answer}