Fix repetition penalty (#3139)

.github/workflows/pr-test.yml

@@ -29,7 +29,7 @@ concurrency:
 jobs:
 
   unit-test-frontend:
-    if: github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request'
+    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && github.event.pull_request.draft == false
     runs-on: 1-gpu-runner
     steps:
       - name: Checkout code
@@ -50,7 +50,7 @@ jobs:
           python3 run_suite.py --suite per-commit
 
   unit-test-backend-1-gpu:
-    if: github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request'
+    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && github.event.pull_request.draft == false
     runs-on: 1-gpu-runner
     strategy:
       matrix:
@@ -77,7 +77,7 @@ jobs:
 
 
   unit-test-backend-2-gpu:
-    if: github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request'
+    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && github.event.pull_request.draft == false
     runs-on: 2-gpu-runner
     steps:
       - name: Checkout code
@@ -114,7 +114,7 @@ jobs:
           python3 test_moe_ep.py
 
   performance-test-1-gpu-part-1:
-    if: github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request'
+    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && github.event.pull_request.draft == false
     runs-on: 1-gpu-runner
     steps:
       - name: Checkout code
@@ -158,7 +158,7 @@ jobs:
 
 
   performance-test-1-gpu-part-2:
-    if: github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request'
+    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && github.event.pull_request.draft == false
     runs-on: 1-gpu-runner
     steps:
       - name: Checkout code
@@ -189,7 +189,7 @@ jobs:
           python3 -m unittest test_bench_serving.TestBenchServing.test_offline_throughput_default_fp8
 
   performance-test-2-gpu:
-    if: github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request'
+    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && github.event.pull_request.draft == false
     runs-on: 2-gpu-runner
     steps:
       - name: Checkout code
@@ -227,7 +227,7 @@ jobs:
 
 
   accuracy-test-1-gpu:
-    if: github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request'
+    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && github.event.pull_request.draft == false
     runs-on: 1-gpu-runner
     steps:
       - name: Checkout code
@@ -251,7 +251,7 @@ jobs:
 
 
   accuracy-test-2-gpu:
-    if: github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request'
+    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && github.event.pull_request.draft == false
     runs-on: 2-gpu-runner
     steps:
       - name: Checkout code

python/sglang/srt/sampling/penaltylib/penalizers/repetition_penalty.py

@@ -67,6 +67,7 @@ class BatchedRepetitionPenalizer(_BatchedPenalizer):
 
     def _apply(self, logits: torch.Tensor) -> torch.Tensor:
         apply_scaling_penalties(logits, self.cumulated_repetition_penalties)
+        return logits
 
     def _filter(self, indices_to_keep: List[int], indices_tensor_to_keep: torch.Tensor):
         self.repetition_penalties = self.repetition_penalties[indices_tensor_to_keep]