Support hidden_dim % 4 == 0 in per_token_quant_fp8 (#12883)

2025-11-10 17:13:14 +08:00
parent 7bffc5dc25
commit 05559a4a90
2 changed files with 19 additions and 4 deletions
@@ -36,7 +36,7 @@ def sglang_per_token_quant_fp8(

@pytest.mark.parametrize(
    "num_tokens,hidden_dim",
-    list(itertools.product([128, 256, 512], [512, 1368, 2048, 4096])),
+    list(itertools.product([128, 256, 512], [512, 1076, 1368, 2048, 4096])),
 )
 def test_per_token_quant_compare_implementations(
    num_tokens: int,