[Diffusion] Fuse FLUX.2 NVFP4 FC1, SwiGLU, and FC2 quantization (#37096)

Co-authored-by: Cursor <cursoragent@cursor.com>

test/registered/kernels/ops/diffusion/test_model_fast_paths.py

@@ -82,6 +82,7 @@ from sglang.multimodal_gen.runtime.models.dits.flux import (
     _flux_norm_modulate,
 )
 from sglang.multimodal_gen.runtime.models.dits.flux_2 import (
+    _can_use_nvfp4_swiglu_quant_fusion,
     _flux2_norm_modulate,
     _flux2_swiglu,
 )
@@ -121,6 +122,7 @@ from sglang.multimodal_gen.runtime.models.vaes.wan_vae_cuda_opt import (
     VaeFastPathGate,
 )
 from sglang.multimodal_gen.runtime.models.vaes.wanvae import WanRMS_norm
+from sglang.multimodal_gen.runtime.platforms.interface import DeviceCapability
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
@@ -342,6 +344,11 @@ class TestFlux2EagerFusions(CustomTestCase):
         self.assertFalse(flux2._FLUX2_SWIGLU.disabled)
         self.assertEqual(len(flux2._FLUX2_SWIGLU_SIGS), 2)
 
+    def test_nvfp4_swiglu_quant_fusion_is_sm103_only(self):
+        self.assertFalse(_can_use_nvfp4_swiglu_quant_fusion(DeviceCapability(10, 0)))
+        self.assertTrue(_can_use_nvfp4_swiglu_quant_fusion(DeviceCapability(10, 3)))
+        self.assertFalse(_can_use_nvfp4_swiglu_quant_fusion(DeviceCapability(12, 0)))
+
     def test_fp16_preserves_reference_path(self):
         x = torch.randn(1, 17, 512, device="cuda", dtype=torch.float16)
         expected = F.silu(x[..., :256]) * x[..., 256:]

test/registered/kernels/ops/diffusion/test_sites.py

@@ -35,11 +35,15 @@ from sglang.kernels.ops.diffusion import (
     QualityGatedFusion,
     can_use_ln_modulate,
     flashinfer_rmsnorm_diagnostic_hint,
+    flux2_nvfp4_swiglu_quant_active,
     fused_ln_modulate,
     fused_ln_modulate_active,
+    mark_flux2_nvfp4_swiglu_quant_site,
     mark_fused_ln_modulate_site,
+    mount_flux2_nvfp4_swiglu_quant,
     mount_fused_ln_modulate,
     tensors_equal,
+    unmount_flux2_nvfp4_swiglu_quant,
     unmount_fused_ln_modulate,
 )
 from sglang.test.ci.ci_register import register_cpu_ci, register_cuda_ci
@@ -92,6 +96,18 @@ def test_quality_gate_rejection_is_all_or_nothing():
     assert not fusion.mount(nn.Module())
 
 
+def test_flux2_nvfp4_swiglu_quant_is_disabled_until_quality_gate_mounts():
+    site = nn.Module()
+    root = nn.ModuleList([site])
+    mark_flux2_nvfp4_swiglu_quant_site(site)
+
+    assert not flux2_nvfp4_swiglu_quant_active(site)
+    assert mount_flux2_nvfp4_swiglu_quant(root)
+    assert flux2_nvfp4_swiglu_quant_active(site)
+    unmount_flux2_nvfp4_swiglu_quant(root)
+    assert not flux2_nvfp4_swiglu_quant_active(site)
+
+
 def test_qwen_image_added_qkv_site_is_request_scoped():
     site = nn.Module()
     site.to_added_qkv = nn.Module()