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

api/langfuse_client.py

@@ -55,9 +55,14 @@ def span(
     as_type: str = "span",
     input: Any = None,
     metadata: dict[str, Any] | None = None,
+    model: str | None = None,
 ) -> Iterator[Any]:
     """Open a Langfuse observation; yields the span object or a no-op.
 
+    Pass `model="<id>"` when `as_type="generation"` so Langfuse can label
+    the trace and compute cost when it knows the model's pricing. Token
+    usage is reported by the body via `span.update(usage_details=...)`.
+
     Exceptions raised by the *body* of the `with` block propagate up
     untouched. Only Langfuse's own setup failures fall through to a no-op
     span — we never want to mask a real error.
@@ -66,10 +71,16 @@ def span(
     if lf is None:
         yield _NullSpan()
         return
+    kwargs: dict[str, Any] = {
+        "name": name,
+        "as_type": as_type,
+        "input": input,
+        "metadata": metadata or {},
+    }
+    if model is not None:
+        kwargs["model"] = model
     try:
-        cm = lf.start_as_current_observation(
-            name=name, as_type=as_type, input=input, metadata=metadata or {}
-        )
+        cm = lf.start_as_current_observation(**kwargs)
     except Exception as e:
         logger.warning("langfuse setup for %s failed: %s", name, e)
         yield _NullSpan()