Simplify eagle tests and TP sync in grammar backend (#4066)

2025-03-04 13:40:40 -08:00
parent 03b0364f76
commit 77a3954bf7
14 changed files with 122 additions and 126 deletions
--- a/test/srt/test_gguf.py
+++ b/test/srt/test_gguf.py
@@ -15,7 +15,7 @@ class TestGGUF(unittest.TestCase):
            filename="qwen2-1_5b-instruct-q4_k_m.gguf",
        )

-        engine = sgl.Engine(model_path=model_path, random_seed=42)
+        engine = sgl.Engine(model_path=model_path, random_seed=42, cuda_graph_max_bs=2)
        outputs = engine.generate(prompt, sampling_params)["text"]
        engine.shutdown()