Relax GSM8K thresholds for the GLM-5.2 DSA-MTP variants (#38725)

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

python/sglang/test/server_fixtures/dsa_mtp_fixture.py

@@ -32,8 +32,9 @@ from sglang.test.test_utils import (
 class DsaMtpEvalConfigDefaults:
     """Eval thresholds & params shared across DSA-MTP regression variants."""
 
-    # GSM8KMixin defaults.
-    gsm8k_accuracy_thres = 0.935
+    # GSM8KMixin defaults. `gsm8k_accuracy_thres` is measured on the FP8
+    # variants; lower-scoring quantizations override it per variant.
+    gsm8k_accuracy_thres = 0.925
     gsm8k_accept_length_thres = 3.7
     gsm8k_num_questions = 500
     gsm8k_num_threads = 500

test/registered/e2e/models/test_dsa_glm52_nvfp4_dp_mtp.py

@@ -24,6 +24,7 @@ class TestGLM52NVFP4DPMTP(
     mem_fraction_static = 0.88
     enable_dp_attention = True
     bs_1_speed_thres = 180
+    gsm8k_accuracy_thres = 0.92
     extra_server_args = [
         "--moe-runner-backend",
         "flashinfer_trtllm",

test/registered/e2e/models/test_dsa_glm52_nvfp4_tp_mtp.py

@@ -22,6 +22,7 @@ class TestGLM52NVFP4TPMTP(
     tp_size = 4
     mem_fraction_static = 0.8
     bs_1_speed_thres = 280
+    gsm8k_accuracy_thres = 0.92
     extra_server_args = [
         "--moe-runner-backend",
         "flashinfer_trtllm",