[diffusion] feat: support request-scoped skip-softmax attention (#37959)

docs/docs/sglang-diffusion/api/cli.mdx

@@ -118,6 +118,7 @@ Use `sglang generate --help` and `sglang serve --help` for the full argument lis
 - `--num-inference-steps {STEPS}` and `--seed {SEED}`
 - `--num-outputs-per-prompt {N}` / `--num-outputs {N}`: generate multiple outputs for each prompt. A scalar seed expands as `seed + output_index`.
 - `--quality {lossless,extra-high,high}`: cumulative request-level optimization tier. `lossless` (default) keeps the selected deployment's reference path and all unconditional bit-exact replacements. `extra-high` adds only request-gated DiT/VAE kernel fusions; the tier does not itself enable sparse, caching, or other approximate paths. `high` includes the complete `extra-high` set and may also enable model-owned approximate optimizations. Separately configured quantization, attention, or caching options still apply. Support and validation constraints are model-specific.
+- `skip_softmax_params` (online request only): explicitly enables lossy BLASST/Skip-Softmax attention for compatible FA self-attention layers. See [Attention Backends](../attention_backends#request-scoped-skip-softmax).
 - `--height {HEIGHT}`, `--width {WIDTH}`, `--num-frames {N}`, `--fps {FPS}`
 - `--output-path {PATH}`, `--output-file-name {NAME}`, `--save-output`, `--return-frames`
 

docs/docs/sglang-diffusion/attention_backends.mdx

@@ -789,6 +789,53 @@ dense switching; under ring parallelism the target must be ring-capable. Note
 `sage_attn` / `sage_attn_3` are lossy (quantized attention) — validate quality
 on your workload.
 
+### Request-scoped Skip Softmax
+
+Skip Softmax (BLASST) keeps the QK matmul, but skips the exponential,
+softmax-state update, V load, and PV matmul for attention tiles whose estimated
+softmax mass is below a threshold. It is an explicit lossy optimization backed
+by the FlashInfer kernels shipped with SGLang's pinned dependencies. See the
+[BLASST paper](https://arxiv.org/abs/2512.12087) and NVIDIA's
+[video-generation study](https://github.com/NVIDIA/TensorRT-LLM/blob/main/docs/source/blogs/tech_blog/blog28_Accelerating_Video_Generation_with_GEMM_Quantization_Attention_Quantization_and_Skip_Softmax_Attention_in_TensorRT-LLM.md)
+for the algorithm and measured quality/performance trade-offs.
+
+Pass `skip_softmax_params` in one image or video request. The runtime routes
+compatible self-attention layers through the FA/FlashInfer path for that batch;
+cross-attention keeps its normal backend, and the next request restores the
+server default.
+
+```python
+client.videos.create(
+    model="<MODEL_PATH_OR_ID>",
+    prompt="...",
+    extra_body={
+        "skip_softmax_params": {
+            "threshold_scale_factor": 500.0,
+            "start_step": 14,
+        }
+    },
+)
+```
+
+- `threshold_scale_factor` is required and must be positive. The kernel uses
+  `threshold_scale_factor / context_length`; larger values skip more work and
+  usually increase quality loss.
+- `start_step` is the zero-based denoising step at which sparse execution
+  starts. It defaults to `0`; keeping early high-noise steps dense is generally
+  safer.
+
+There is intentionally no default threshold. Calibrate the threshold and start
+step against output-quality metrics for each model, resolution, step count, and
+deployment backend; values tuned for one workload are not portable.
+
+Current support is FP16/BF16, head dimension 128 or 256, and unmasked
+self-attention on Hopper SM90 and Blackwell SM100/SM103/SM107. Ulysses sequence
+parallelism is supported because the kernel runs after its all-to-all. Ring
+Attention, attention masks, `torch.compile`, and breakable CUDA graphs reject
+the request instead of silently running dense attention. Models that merely
+offer an FA backend do not automatically qualify: the runtime still checks the
+GPU, dtype, head dimension, attention role, and execution mode.
+
 ### Using SpargeAttention
 
 Install the optional CUDA extension, then select the backend explicitly: