diff --git a/python/sglang/test/kits/eval_accuracy_kit.py b/python/sglang/test/kits/eval_accuracy_kit.py index 04afac8cbe..93d481b8a3 100644 --- a/python/sglang/test/kits/eval_accuracy_kit.py +++ b/python/sglang/test/kits/eval_accuracy_kit.py @@ -94,6 +94,7 @@ def _run_sgl_eval( chat_template_kwargs: Optional[dict] = None, max_tokens: Optional[int] = None, temperature: Optional[float] = None, + top_p: Optional[float] = None, accept_length_thres: Optional[float] = None, ): """Shared sgl-eval driver for the MMLU sanity gate and the GSM8K ``sgl_eval`` backend. @@ -118,6 +119,7 @@ def _run_sgl_eval( num_threads=num_threads, max_tokens=max_tokens, temperature=temperature, + top_p=top_p, chat_template_kwargs=( chat_template_kwargs if chat_template_kwargs is not None @@ -176,7 +178,7 @@ class GSM8KMixin: gsm8k_backend: str = "run_eval" # "run_eval" | "sgl_eval" gsm8k_thinking: bool = False # sgl_eval backend gsm8k_max_tokens: Optional[int] = None # sgl_eval backend - # None keeps run_eval's greedy default; set both to route the run through + # None keeps the backend's default; set both to route the run through # the sampling path. gsm8k_temperature: Optional[float] = None gsm8k_top_p: Optional[float] = None @@ -200,6 +202,8 @@ class GSM8KMixin: num_threads=self.gsm8k_num_threads, thinking=self.gsm8k_thinking, max_tokens=self.gsm8k_max_tokens, + temperature=self.gsm8k_temperature, + top_p=self.gsm8k_top_p, accept_length_thres=self.gsm8k_accept_length_thres, ) else: diff --git a/test/registered/e2e/models/test_glm53_flash_b200.py b/test/registered/e2e/models/test_glm53_flash_b200.py index 99ae8c230d..0084843c3c 100644 --- a/test/registered/e2e/models/test_glm53_flash_b200.py +++ b/test/registered/e2e/models/test_glm53_flash_b200.py @@ -75,12 +75,14 @@ class TestGLM53FlashB200LowLatency( GSM8KMixin, _GLM53FlashB200Base, ): - gsm8k_score_threshold = 0.93 - # Match the established DSA+MTP accuracy workload. The generic 200-question, - # 5-shot defaults leave a single question worth 0.5 percentage points and - # make this tight quality floor unnecessarily sensitive to kernel numerics. - gsm8k_num_examples = 500 - gsm8k_num_shots = 20 + gsm8k_score_threshold = 0.96 + gsm8k_num_examples = 1319 + gsm8k_num_threads = 256 + gsm8k_backend = "sgl_eval" + gsm8k_thinking = True + gsm8k_max_tokens = 8192 + gsm8k_temperature = 1.0 + gsm8k_top_p = 0.95 accept_length_thres = 4.0 bs_1_speed_thres = 250 server_args = [ @@ -100,9 +102,14 @@ class TestGLM53FlashB200HighThroughput( GSM8KMixin, _GLM53FlashB200Base, ): - gsm8k_score_threshold = 0.93 - gsm8k_num_examples = 500 - gsm8k_num_shots = 20 + gsm8k_score_threshold = 0.96 + gsm8k_num_examples = 1319 + gsm8k_num_threads = 256 + gsm8k_backend = "sgl_eval" + gsm8k_thinking = True + gsm8k_max_tokens = 8192 + gsm8k_temperature = 1.0 + gsm8k_top_p = 0.95 server_args = [ *COMMON_SERVER_ARGS, "--attn-dp-size", @@ -115,9 +122,14 @@ class TestGLM53FlashB200DFlash2( GSM8KMixin, _GLM53FlashB200Base, ): - gsm8k_score_threshold = 0.93 - gsm8k_num_examples = 500 - gsm8k_num_shots = 20 + gsm8k_score_threshold = 0.96 + gsm8k_num_examples = 1319 + gsm8k_num_threads = 256 + gsm8k_backend = "sgl_eval" + gsm8k_thinking = True + gsm8k_max_tokens = 8192 + gsm8k_temperature = 1.0 + gsm8k_top_p = 0.95 server_args = [ *COMMON_SERVER_ARGS, "--speculative-algorithm", diff --git a/test/registered/e2e/models/test_glm53_flash_h200.py b/test/registered/e2e/models/test_glm53_flash_h200.py index 1f9e0747fa..10d7cab65b 100644 --- a/test/registered/e2e/models/test_glm53_flash_h200.py +++ b/test/registered/e2e/models/test_glm53_flash_h200.py @@ -74,12 +74,14 @@ class TestGLM53FlashH200LowLatency( GSM8KMixin, _GLM53FlashH200Base, ): - gsm8k_score_threshold = 0.93 - # Match the established DSA+MTP accuracy workload. The generic 200-question, - # 5-shot defaults leave a single question worth 0.5 percentage points and - # make this tight quality floor unnecessarily sensitive to kernel numerics. - gsm8k_num_examples = 500 - gsm8k_num_shots = 20 + gsm8k_score_threshold = 0.96 + gsm8k_num_examples = 1319 + gsm8k_num_threads = 256 + gsm8k_backend = "sgl_eval" + gsm8k_thinking = True + gsm8k_max_tokens = 8192 + gsm8k_temperature = 1.0 + gsm8k_top_p = 0.95 accept_length_thres = 4.0 bs_1_speed_thres = 200 server_args = [ @@ -99,9 +101,14 @@ class TestGLM53FlashH200HighThroughput( GSM8KMixin, _GLM53FlashH200Base, ): - gsm8k_score_threshold = 0.93 - gsm8k_num_examples = 500 - gsm8k_num_shots = 20 + gsm8k_score_threshold = 0.96 + gsm8k_num_examples = 1319 + gsm8k_num_threads = 256 + gsm8k_backend = "sgl_eval" + gsm8k_thinking = True + gsm8k_max_tokens = 8192 + gsm8k_temperature = 1.0 + gsm8k_top_p = 0.95 server_args = [ *COMMON_SERVER_ARGS, "--attn-dp-size",