Relax flaky B200 GSM8K accuracy thresholds (#20304)

Co-authored-by: Alison Shao <alisonshao@Mac.attlocal.net>
This commit is contained in:
Alison Shao
2026-03-11 12:35:21 -07:00
committed by GitHub
parent d093e70067
commit 7b44bc923e
2 changed files with 5 additions and 5 deletions

View File

@@ -70,7 +70,7 @@ class TestDeepseekV3FP4(CustomTestCase):
f"### test_gsm8k (deepseek-v3-fp4)\n" f'{metrics["accuracy"]=:.3f}\n'
)
self.assertGreater(metrics["accuracy"], 0.935)
self.assertGreater(metrics["accuracy"], 0.93)
def test_bs_1_speed(self):
args = BenchArgs(port=int(self.base_url.split(":")[-1]), max_new_tokens=2048)
@@ -136,7 +136,7 @@ class TestDeepseekV3FP4PiecewiseCudaGraph(CustomTestCase):
f"### test_gsm8k (deepseek-v3-fp4)\n" f'{metrics["accuracy"]=:.3f}\n'
)
self.assertGreater(metrics["accuracy"], 0.935)
self.assertGreater(metrics["accuracy"], 0.93)
def test_bs_1_speed(self):
args = BenchArgs(port=int(self.base_url.split(":")[-1]), max_new_tokens=2048)
@@ -206,7 +206,7 @@ class TestDeepseekV3FP4CutlassMoE(CustomTestCase):
f"### test_gsm8k (deepseek-v3-fp4-cutlass-moe)\n"
f'{metrics["accuracy"]=:.3f}\n'
)
self.assertGreater(metrics["accuracy"], 0.935)
self.assertGreater(metrics["accuracy"], 0.93)
class TestDeepseekV3FP4SymmetricMemory(CustomTestCase):
@@ -260,7 +260,7 @@ class TestDeepseekV3FP4SymmetricMemory(CustomTestCase):
f"### test_gsm8k (deepseek-v3-fp4)\n" f'{metrics["accuracy"]=:.3f}\n'
)
self.assertGreater(metrics["accuracy"], 0.935)
self.assertGreater(metrics["accuracy"], 0.93)
if __name__ == "__main__":

View File

@@ -85,7 +85,7 @@ class TestEagleDPAttnServerSmall(CustomTestCase):
def test_a_gsm8k(self):
metrics, avg_spec_accept_length = test_gsm8k(self.base_url)
self.assertGreater(metrics["accuracy"], 0.64)
self.assertGreater(metrics["accuracy"], 0.62)
self.assertGreater(avg_spec_accept_length, 2.7)