[diffusion] CI: tighten selected perf baselines (#25510)

python/sglang/multimodal_gen/runtime/server_args.py

@@ -1201,6 +1201,7 @@ class ServerArgs(DisaggArgsMixin):
             help="The number of warmup steps to perform for each resolution.",
         )
 
+        # layerwise offload
         parser.add_argument(
             "--dit-cpu-offload",
             action=StoreBoolean,
@@ -1233,11 +1234,8 @@ class ServerArgs(DisaggArgsMixin):
             default=ServerArgs.dit_offload_prefetch_size,
             help="The size of prefetch for dit-layerwise-offload. If the value is between 0.0 and 1.0, it is treated as a ratio of the total number of layers. If the value is >= 1, it is treated as the absolute number of layers. 0.0 means prefetch 1 layer (lowest memory). Values above 0.5 might have peak memory close to no offload but worse performance.",
         )
-        parser.add_argument(
+
-            "--use-fsdp-inference",
+        # offload flags
-            action=StoreBoolean,
-            help="Use FSDP inference to shard DiT weights across GPUs. For single-GPU memory pressure, prefer CPU or layerwise offload.",
-        )
         parser.add_argument(
             "--text-encoder-cpu-offload",
             action=StoreBoolean,
@@ -1253,6 +1251,12 @@ class ServerArgs(DisaggArgsMixin):
             action=StoreBoolean,
             help="Use CPU offload for VAE. Enable if run out of memory.",
         )
+
+        parser.add_argument(
+            "--use-fsdp-inference",
+            action=StoreBoolean,
+            help="Use FSDP inference to shard DiT weights across GPUs. For single-GPU memory pressure, prefer CPU or layerwise offload.",
+        )
         parser.add_argument(
             "--pin-cpu-memory",
             action=StoreBoolean,
@@ -1278,6 +1282,7 @@ class ServerArgs(DisaggArgsMixin):
             help="Disable autocast for denoising loop and vae decoding in pipeline sampling",
         )
 
+        # quantization
         parser.add_argument(
             "--quantization",
             type=str,

python/sglang/multimodal_gen/test/server/ascend/perf_baselines_npu.json

@@ -147,7 +147,7 @@
                 "per_frame_generation": null
             },
             "denoise_step_ms": {
-                "0": 101.91,
+                "0": 330.0,
                 "1": 303.37,
                 "2": 548.94,
                 "3": 542.02,
@@ -200,7 +200,8 @@
             },
             "expected_e2e_ms": 29360.56,
             "expected_avg_denoise_ms": 524.67,
-            "expected_median_denoise_ms": 537.54
+            "expected_median_denoise_ms": 537.54,
+            "estimated_full_test_time_s": 157.8
         },
        "wan2_2_t2v_14b_w8a8_8npu": {
             "stages_ms": {