Support DeepGEMM for standard MoE dispatch (#33128)

Co-authored-by: Sam Li <lsam@nvidia.com>

test/registered/unit/layers/quantization/test_deepgemm_ue8m0_requant.py

@@ -5,13 +5,14 @@ from sglang.test.ci.ci_register import register_cpu_ci
 register_cpu_ci(est_time=5, suite="base-a-test-cpu")
 
 import unittest
-from unittest.mock import patch
+from unittest.mock import call, patch
 
 import torch
 from compressed_tensors.quantization import QuantizationStrategy
 
 import sglang.srt.layers.quantization.fp8_utils as fp8_utils
 from sglang.srt.layers import deep_gemm_wrapper
+from sglang.srt.layers.quantization import fp8 as fp8_quant
 from sglang.srt.layers.quantization.compressed_tensors.schemes.compressed_tensors_w8a8_fp8 import (
     CompressedTensorsW8A8Fp8,
 )
@@ -162,6 +163,60 @@ class TestDeepGemmUE8M0Requant(CustomTestCase):
         self.assertTrue(layer.weight_scale.format_ue8m0)
         requant.assert_called_once()
 
+    def test_fp8_moe_requants_standard_layer_for_deepgemm(self):
+        method = fp8_quant.Fp8MoEMethod.__new__(fp8_quant.Fp8MoEMethod)
+        method.convert_mxfp8_to_block = False
+        method.use_mxfp8 = False
+        method.is_fp4_expert = False
+        method.dequant_fp4_to_fp8 = False
+        method.quant_config = unittest.mock.Mock(weight_block_size=BLOCK_SIZE)
+
+        layer = torch.nn.Module()
+        layer.w13_weight, layer.w13_weight_scale_inv = _make_params()
+        layer.w2_weight, layer.w2_weight_scale_inv = _make_params()
+
+        def _mark_ue8m0(weight, weight_scale, *args, **kwargs):
+            weight_scale.format_ue8m0 = True
+            return True
+
+        with patch.multiple(
+            fp8_quant,
+            _is_cpu=False,
+            _is_fp8_fnuz=False,
+            _use_aiter=False,
+        ), patch.object(
+            method, "is_deepgemm_moe_runner_backend_enabled", return_value=True
+        ), patch.object(
+            fp8_quant,
+            "requant_block_scale_ue8m0_for_deepgemm",
+            side_effect=_mark_ue8m0,
+        ) as requant:
+            method.process_weights_after_loading_block_quant(layer)
+
+        self.assertEqual(
+            requant.call_args_list,
+            [
+                call(
+                    layer.w13_weight,
+                    layer.w13_weight_scale_inv,
+                    BLOCK_SIZE,
+                    use_deepgemm_runner=True,
+                    output_dtype=torch.bfloat16,
+                    weight_shape=layer.w13_weight.shape[-2:],
+                ),
+                call(
+                    layer.w2_weight,
+                    layer.w2_weight_scale_inv,
+                    BLOCK_SIZE,
+                    use_deepgemm_runner=True,
+                    output_dtype=torch.bfloat16,
+                    weight_shape=layer.w2_weight.shape[-2:],
+                ),
+            ],
+        )
+        self.assertTrue(layer.w13_weight_scale_inv.format_ue8m0)
+        self.assertTrue(layer.w2_weight_scale_inv.format_ue8m0)
+
 
 if __name__ == "__main__":
     unittest.main(verbosity=3)