🐛 fix prometheus metric labels (#27)

This fixes a miss where I had seen usages of `.labels` `**`a dictionary
into kwargs, and I accidentally passed a raw dictionary as a value
instead of using keyword arguments 🤦. This caused metrics to show eg.
`method="{'method':'prefill'}"` instead of `method=prefill`

Signed-off-by: Joe Runde <[email protected]>

vllm/tgis_utils/metrics.py

@@ -71,7 +71,7 @@ def observe_queue_time(self, engine_output: RequestOutput):
             engine_output.metrics.time_in_queue)
 
     def count_request_failure(self, reason: FailureReasonLabel):
-        self.tgi_request_failure.labels({"err": reason}).inc(1)
+        self.tgi_request_failure.labels(err=reason).inc(1)
 
 
 class TGISStatLogger(StatLogger):
@@ -120,13 +120,11 @@ def log(self, stats: Stats) -> None:
         self.tgi_batch_current_size.set(stats.num_running_sys)
 
         for ttft in stats.time_to_first_tokens_iter:
-            self.tgi_batch_inference_duration.labels({
-                "method": "prefill"
-            }).observe(ttft)
+            self.tgi_batch_inference_duration.labels(
+                method="prefill").observe(ttft)
         for tpot in stats.time_per_output_tokens_iter:
-            self.tgi_batch_inference_duration.labels({
-                "method": "next_token"
-            }).observe(tpot)
+            self.tgi_batch_inference_duration.labels(
+                method="next_token").observe(tpot)
 
         for input_len in stats.num_prompt_tokens_requests:
             self.tgi_request_input_length.observe(input_len)