[diffusion] fix: use dynamic LoRA for LTX2 original stage-two (#25591)

python/sglang/multimodal_gen/runtime/pipelines/ltx_2_pipeline.py

@@ -806,6 +806,13 @@ class LTX2TwoStagePipeline(_BaseLTX2Pipeline):
             server_args.pipeline_config.vae_config.arch_config
         )
 
+    def _should_merge_lora_for_phase(self, phase: str) -> bool:
+        if phase == "stage2" and self._ltx2_residency.mode == "original":
+            # original mode reuses one DiT for both phases; dynamic LoRA avoids
+            # request-time merge/unmerge without keeping another DiT resident
+            return False
+        return self._should_merge_stage2_distilled_lora(self.server_args)
+
     def initialize_pipeline(self, server_args: ServerArgs):
         super().initialize_pipeline(server_args)
         server_args.component_paths = _resolve_ltx2_two_stage_component_paths(
@@ -967,16 +974,15 @@ class LTX2TwoStagePipeline(_BaseLTX2Pipeline):
                 strength=lora_strengths,
             )
             if phase == "stage2":
-                # Official LTX-2.3 two-stage builds stage 2 with distilled LoRA fused
+                # premerged modes keep official LTX-2.3 fused stage-2 LoRA; original
-                # into the transformer weights. Legacy LTX-2 should keep the
+                # avoids single-DiT request-time merge/unmerge with dynamic LoRA
-                # preexisting unmerged behavior to avoid regressing stage 2 quality.
+                set_lora_kwargs["merge_weights"] = self._should_merge_lora_for_phase(
-                set_lora_kwargs["merge_weights"] = (
+                    phase
-                    self._should_merge_stage2_distilled_lora(self.server_args)
                 )
             elif phase == "stage1" and self.pipeline_name == "LTX2TwoStageHQPipeline":
                 # Official HQ also builds stage 1 with distilled LoRA fused.
-                set_lora_kwargs["merge_weights"] = (
+                set_lora_kwargs["merge_weights"] = self._should_merge_lora_for_phase(
-                    self._should_merge_stage2_distilled_lora(self.server_args)
+                    phase
                 )
             self.set_lora(
                 **set_lora_kwargs,

python/sglang/multimodal_gen/test/server/perf_baselines.json

@@ -1362,7 +1362,7 @@
                 "1": 115.0,
                 "2": 130.0
             },
-            "expected_e2e_ms": 2684.91,
+            "expected_e2e_ms": 3150.0,
             "expected_avg_denoise_ms": 95.0,
             "expected_median_denoise_ms": 120.0,
             "estimated_full_test_time_s": 125.2
@@ -2477,7 +2477,7 @@
                 "32": 184.4
             },
             "expected_e2e_ms": 18039.38,
-            "expected_avg_denoise_ms": 183.75,
+            "expected_avg_denoise_ms": 271.95,
             "expected_median_denoise_ms": 179.26,
             "estimated_full_test_time_s": 160.0
         },
@@ -2487,7 +2487,7 @@
                 "TextEncodingStage": 1554.35,
                 "LTX2TextConnectorStage": 27.64,
                 "LTX2HalveResolutionStage": 0.05,
-                "LTX2LoRASwitchStage": 114.37,
+                "LTX2LoRASwitchStage": 70.0,
                 "LTX2SigmaPreparationStage": 0.4,
                 "TimestepPreparationStage": 23.51,
                 "LTX2AVLatentPreparationStage": 0.15,
@@ -2531,9 +2531,9 @@
                 "29": 237.2,
                 "30": 127.08,
                 "31": 123.1,
-                "32": 128.06
+                "32": 209.80
             },
-            "expected_e2e_ms": 17660.77,
+            "expected_e2e_ms": 12000.0,
             "expected_avg_denoise_ms": 238.8,
             "expected_median_denoise_ms": 246.85,
             "estimated_full_test_time_s": 170.0