[Fix] Correct dense FP8 Marlin bias ordering (#35020)

python/sglang/srt/layers/quantization/marlin_utils_fp8.py

@@ -186,10 +186,12 @@ def prepare_fp8_layer_for_marlin(
     marlin_scales = fp8_fused_exponent_bias_into_scales(marlin_scales)
     layer.weight_scale = torch.nn.Parameter(marlin_scales, requires_grad=False)
 
+    # The dense FP8 Marlin wrapper adds bias after the kernel returns, so the
+    # bias must remain in logical output-channel order. Only scales need the
+    # Marlin tile permutation.
     if hasattr(layer, "bias") and layer.bias is not None:
         assert layer.bias.shape == (part_size_n,)
-        bias = marlin_permute_bias(layer.bias)
+        layer.bias = torch.nn.Parameter(layer.bias.detach(), requires_grad=False)
-        layer.bias = torch.nn.Parameter(bias, requires_grad=False)
 
 
 def prepare_moe_fp8_layer_for_marlin(

test/registered/unit/layers/quantization/test_marlin_utils_fp8.py

@@ -0,0 +1,54 @@
+"""Tests for FP8 Marlin utilities."""
+
+import unittest
+from unittest.mock import patch
+
+import torch
+
+from sglang.srt.layers.quantization import marlin_utils_fp8
+from sglang.test.ci.ci_register import register_cpu_ci
+from sglang.test.test_utils import CustomTestCase
+
+register_cpu_ci(est_time=5, suite="base-a-test-cpu")
+
+
+class TestFp8MarlinBias(CustomTestCase):
+    def test_dense_bias_remains_in_logical_output_order(self):
+        size_k = 32
+        size_n = 32
+
+        layer = torch.nn.Module()
+        layer.input_size_per_partition = size_k
+        layer.output_size_per_partition = size_n
+        layer.orig_dtype = torch.float16
+        layer.weight_block_size = None
+        layer.weight = torch.nn.Parameter(
+            torch.zeros((size_k, size_n), dtype=torch.float8_e4m3fn),
+            requires_grad=False,
+        )
+        layer.weight_scale = torch.nn.Parameter(
+            torch.ones((size_n,), dtype=torch.float32), requires_grad=False
+        )
+        original_bias = torch.arange(size_n, dtype=torch.float16)
+        layer.bias = torch.nn.Parameter(original_bias.clone(), requires_grad=False)
+
+        with (
+            patch.object(
+                marlin_utils_fp8,
+                "marlin_make_workspace",
+                return_value=torch.empty(0, dtype=torch.int32),
+            ),
+            patch.object(
+                marlin_utils_fp8,
+                "gptq_marlin_repack",
+                return_value=torch.empty(0, dtype=torch.int32),
+                create=True,
+            ),
+        ):
+            marlin_utils_fp8.prepare_fp8_layer_for_marlin(layer)
+
+        torch.testing.assert_close(layer.bias, original_bias)
+
+
+if __name__ == "__main__":
+    unittest.main(verbosity=3)