fix: adding teacache.params back to sampling params as intended (#20665)

python/sglang/multimodal_gen/configs/sample/sampling_params.py

@@ -156,6 +156,9 @@ class SamplingParams:
 
     # TeaCache parameters
     enable_teacache: bool = False
+    teacache_params: Any = (
+        None  # TeaCacheParams or WanTeaCacheParams, set by model-specific subclass
+    )
 
     # Profiling
     profile: bool = False

python/sglang/multimodal_gen/configs/sample/wan.py

@@ -212,6 +212,11 @@ class Wan2_2_Base_SamplingParams(SamplingParams):
         "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走"
     )
 
+    # TODO(Wan2.2): TeaCache coefficients need to be calibrated for Wan2.2 by
+    # profiling L1 distances across timesteps. Until then, teacache_params is None
+    # and enable_teacache will be accepted but silently no-op.
+    # Consider using Cache-DiT (SGLANG_CACHE_DIT_ENABLED=1) as an alternative.
+
 
 @dataclass
 class Wan2_2_TI2V_5B_SamplingParam(Wan2_2_Base_SamplingParams):

python/sglang/multimodal_gen/test/server/testcase_configs.py

@@ -768,6 +768,23 @@ TWO_GPU_CASES_A = [
             prompt=T2V_PROMPT,
         ),
     ),
+    # TeaCache smoke test for Wan2.2 T2V A14B — verifies enable_teacache=True
+    # doesn't crash. Perf check disabled because Wan2.2-specific TeaCache
+    # coefficients are not yet calibrated (teacache_params=None, so no speedup).
+    DiffusionTestCase(
+        "wan2_2_t2v_a14b_teacache_2gpu",
+        DiffusionServerArgs(
+            model_path=DEFAULT_WAN_2_2_T2V_A14B_MODEL_NAME_FOR_TEST,
+            modality="video",
+            custom_validator="video",
+            num_gpus=2,
+        ),
+        DiffusionSamplingParams(
+            prompt=T2V_PROMPT,
+            extras={"enable_teacache": True},
+        ),
+        run_perf_check=False,
+    ),
     # LoRA test case for transformer_2 support
     DiffusionTestCase(
         "wan2_2_t2v_a14b_lora_2gpu",