Fix bench latency benchmark (#1225)

.github/workflows/e2e-test.yml

@@ -38,6 +38,11 @@ jobs:
           cd test/srt
           python3 -m unittest test_serving_throughput.TestServingThroughput.test_default
 
+      - name: Benchmark Serving Latency
+        timeout-minutes: 10
+        run: |
+          python3 -m sglang.bench_latency --model meta-llama/Meta-Llama-3.1-8B-Instruct --batch-size 1 --input 128 --output 8
+
       - name: Benchmark Serving Throughput (w/o RadixAttention)
         timeout-minutes: 10
         run: |

python/sglang/bench_latency.py

@@ -200,16 +200,14 @@ def extend(reqs, model_runner):
         tree_cache=None,
     )
     batch.prepare_for_extend(model_runner.model_config.vocab_size)
-    output = model_runner.forward(batch, ForwardMode.EXTEND)
-    next_token_ids = batch.sample(output.next_token_logits)
-    return next_token_ids, output.next_token_logits, batch
+    sample_output, logits_output = model_runner.forward(batch, ForwardMode.EXTEND)
+    return sample_output.batch_next_token_ids, logits_output.next_token_logits, batch
 
 
 def decode(input_token_ids, batch, model_runner):
     batch.prepare_for_decode(input_token_ids.cpu().numpy())
-    output = model_runner.forward(batch, ForwardMode.DECODE)
-    next_token_ids = batch.sample(output.next_token_logits)
-    return next_token_ids, output.next_token_logits
+    sample_output, logits_output = model_runner.forward(batch, ForwardMode.DECODE)
+    return sample_output.batch_next_token_ids, logits_output.next_token_logits
 
 
 @torch.inference_mode()