[diffusion][kernel] avoid 4D scale-shift autotuning (#36521)

python/sglang/kernels/ops/diffusion/README.md

@@ -90,6 +90,7 @@ Several norms look interchangeable and are not. Start here.
 
 | Entry point | Backend | Contract | Applies to |
 |---|---|---|---|
+| `fuse_scale_shift_kernel` | Triton | close | contiguous BLC; scalar/row/token modulation plus causal-video `[B, F, 1, C]`, using a static capped power-of-two tile to avoid request-time autotuning |
 | `fused_rmsnorm_scale_shift_bitexact` | Triton | bit-exact vs flashinfer CuTe RMSNorm + aten modulate | bf16, contiguous rows, `H == 64 * threads_per_row` |
 | `fused_scale_residual_rmsnorm_scale_shift_bitexact` | Triton | bit-exact, incl. the preceding residual-gate add | as above |
 | `fused_layernorm_modulate` | Triton | bit-exact vs aten `vectorized_layer_norm` | bf16, `N % 4 == 0`, 16B-aligned |

python/sglang/kernels/ops/diffusion/modulate/scale_shift_triton.py

@@ -272,16 +272,6 @@ def _fused_residual_layernorm_scale_shift_gate_select01_kernel(
     tl.store(gate_row_ptr + cols, gate, mask=mask)
 
 
-@triton.autotune(
-    configs=[
-        triton.Config({"BLOCK_N": 64}, num_warps=2),
-        triton.Config({"BLOCK_N": 128}, num_warps=4),
-        triton.Config({"BLOCK_N": 256}, num_warps=4),
-        triton.Config({"BLOCK_N": 512}, num_warps=4),
-        triton.Config({"BLOCK_N": 1024}, num_warps=8),
-    ],
-    key=["inner_dim"],
-)
 @triton.jit
 def _fused_scale_shift_4d_kernel(
     output_ptr,
@@ -416,9 +406,12 @@ def fuse_scale_shift_kernel(
         x_2d = x.view(rows, C)
         output_2d = output.view(rows, C)
 
-        def grid(meta):
-            return (rows, triton.cdiv(C, meta["BLOCK_N"]))
-
+        # Autotuning this bandwidth-bound kernel is much more expensive than
+        # the launch itself on causal video models.  A capped power-of-two
+        # tile is fastest or within noise across the production hidden sizes.
+        block_n = max(64, min(512, triton.next_power_of_2(C)))
+        num_warps = 2 if block_n == 64 else 4
+        grid = (rows, triton.cdiv(C, block_n))
         num_frames = scale.shape[1]
         assert (
             L % num_frames == 0
@@ -454,6 +447,8 @@ def fuse_scale_shift_kernel(
             L,
             num_frames,
             frame_seqlen,
+            BLOCK_N=block_n,
+            num_warps=num_warps,
         )
     else:
         # 2D: [B, C] or [1, C]  -> treat as [B, 1, C] and broadcast over L

python/sglang/multimodal_gen/.claude/skills/sglang-diffusion-benchmark-profile/existing-fast-paths.md

@@ -77,6 +77,10 @@ framework-specific optimization workflow.
 - Locations: `elementwise.py`, `layernorm.py`, `fused_scale_shift_gate.py`, `qwen_image.py`, `triton/scale_shift.py`
 - Use cases: `x * (1 + scale) + shift`, `a * (k + b) + c`, and Qwen-style `(layernorm/residual layernorm) + scale/shift + gate select`.
 - Constraints: `x` must be CUDA and contiguous. `scale/shift` support 0D/1D/2D/3D/4D broadcast. 4D `[B, F, 1, C]` requires `L % F == 0`.
+- Causal-video cold start: the 4D path uses a static capped power-of-two
+  column tile rather than Triton autotuning. Do not reintroduce request-time
+  autotuning here: LingBot-World calls this path once per transformer block,
+  and tuning overhead can dominate its first denoise step.
 - NPU fallback: `scale_shift.py` swaps to `npu_fallback` native path.
 - Validation: `test/registered/kernels/ops/diffusion/test_qwen_image_modulation.py`.