From f33b83b4ccae1613317cded8ae64b1f9f866a8d4 Mon Sep 17 00:00:00 2001
From: WenhaoZhang <42087078+niehen6174@users.noreply.github.com>
Date: Mon, 17 Aug 2026 18:47:31 +0800
Subject: [PATCH] [diffusion] fix: fix h3 swap peft SwiGLU lora_B halves when
 loading FFN Lora (#34940)

---
 .../runtime/pipelines_core/lora_pipeline.py     | 17 +++++++++++++++++
 1 file changed, 17 insertions(+)

diff --git a/python/sglang/multimodal_gen/runtime/pipelines_core/lora_pipeline.py b/python/sglang/multimodal_gen/runtime/pipelines_core/lora_pipeline.py
index 4b4ab0851..300dd649e 100644
--- a/python/sglang/multimodal_gen/runtime/pipelines_core/lora_pipeline.py
+++ b/python/sglang/multimodal_gen/runtime/pipelines_core/lora_pipeline.py
@@ -39,6 +39,22 @@ os.environ["TOKENIZERS_PARALLELISM"] = "false"
 logger = init_logger(__name__)
 
 
+def _swap_peft_swiglu_fc1_lora_b(
+    source_name: str, target_name: str, weight: torch.Tensor
+) -> torch.Tensor:
+    # Only the PEFT -> native H3 FFN rewrite: ff.net.0.proj [value; gate]
+    # onto mlp.fc1 [gate; value]. Native fused mlp.fc1 and other models'
+    # ff.net.0.proj (e.g. Flux) must not match.
+    if (
+        weight.dim() != 2
+        or ".ff.net.0.proj.lora_B" not in source_name
+        or not target_name.endswith(".mlp.fc1.lora_B")
+    ):
+        return weight
+    value, gate = weight.chunk(2, dim=0)
+    return torch.cat([gate, value], dim=0)
+
+
 class LoRAPipeline(ComposedPipelineBase):
     """
     Pipeline that supports injecting LoRA adapters into the diffusion transformer.
@@ -775,6 +791,7 @@ class LoRAPipeline(ComposedPipelineBase):
                 else:
                     continue
 
+            weight = _swap_peft_swiglu_fc1_lora_b(name, target_name, weight)
             if target_name in self.lora_adapters[lora_nickname]:
                 raise ValueError(
                     f"Dit target weight name {target_name} already exists in lora_adapters[{lora_nickname}]"