[HiSparse] Add MHA hisparse support for MiniMax M3 (#31446)

Co-authored-by: Guangda Liu <bingps@users.noreply.github.com>

docs/cookbook/autoregressive/MiniMax/MiniMax-M3.mdx

@@ -129,6 +129,7 @@ The NVIDIA Blackwell recipes are validated single-node: **B200 at `--tp 8`** and
 
 - **Memory**: `--mem-fraction-static` reserves GPU memory for weights + KV pool; the rest is prefill **activation headroom**. The value scales with *free* memory per GPU (card capacity minus per-GPU weight), so it tracks the card more than the TP degree: **`0.65` on B200** (180 GB — less headroom once weights are resident) and **`0.75` on the larger-memory B300 / GB300** (`0.80` on AMD). Lower TP packs more weight per GPU, so a tighter config needs a *lower* value — B200 needs `0.65` even at `--tp 4`. Raising it past the validated value is fine only for low-concurrency single-stream serving; it OOMs under high concurrency or long context.
 - **Long context (32K+)**: keep `--mem-fraction-static` at the platform default and raise `--chunked-prefill-size` to `16384`. Decode TPOT stays roughly flat in context length thanks to sparse attention; 1K–128K prompts are validated.
+- **HiSparse for decode capacity**: on NVIDIA CUDA, HiSparse keeps the three dense layers on GPU, moves the 57 sparse-layer K/V caches to pinned host memory, and feeds selected block IDs directly to the swap-in kernel. For the released four-KV-head model, use `--tp 4` or greater, `--disable-radix-cache`, and `device_buffer_size >= 2048`. Enable it with `--enable-hisparse --hisparse-config='{"device_buffer_size":4096,"host_to_device_ratio":2}'` on the Triton launch command.
 - **Scaling TP**: B200 is documented at `--tp 8`; B300 / GB200 / GB300 at `--tp 4` (the single-node cross-family common denominator). On an 8-GPU B300 host you can also raise to `--tp 8` for more throughput / KV headroom.
 - **Expert parallelism**: to trade latency for throughput add `--ep` (see [Expert Parallelism Deployment](../../../docs/advanced_features/expert_parallelism)). On AMD, set `--ep` equal to `--tp`. Shared-experts fusion is automatically disabled when EP > 1; on AMD standard EP the server also disables `--enable-aiter-allreduce-fusion` automatically to preserve accuracy.
 - `--trust-remote-code` is required to load the MiniMax config / processor classes.

docs/docs/advanced_features/hisparse_guide.mdx

@@ -6,7 +6,7 @@ metatags:
 
 HiSparse reduces per-request GPU memory consumption during the decode phase by maintaining only a small "hot" KV buffer on GPU while keeping complete KV data in CPU pinned memory. Combined with PD disaggregation, it enables significantly higher decode concurrency.
 
-> **Prerequisites**: HiSparse works with models that use **DeepSeek Sparse Attention (DSA)** architectures (e.g., DeepSeek-V3.2, GLM-5.1) and **DeepSeek V4**. These models natively select a subset of tokens for attention, making it possible to keep only the top-k KV on GPU while storing the full KV in host memory — without accuracy loss. Additionally, HiSparse currently requires **PD disaggregation mode** and is enabled on the **decode instance** only.
+> **Prerequisites**: HiSparse works with models that use **DeepSeek Sparse Attention (DSA)** architectures (e.g., DeepSeek-V3.2, GLM-5.1), **DeepSeek V4**, and **MiniMax M3**. These models natively select a subset of tokens for attention, making it possible to keep only the top-k KV on GPU while storing the full KV in host memory — without accuracy loss. Additionally, HiSparse currently requires **PD disaggregation mode** and is enabled on the **decode instance** only.
 
 ## Why HiSparse?
 
@@ -165,6 +165,15 @@ python3 -m sglang.launch_server \
 
 > **Note**: For DSA models, `--kv-cache-dtype` defaults to `auto`, which resolves to `fp8_e4m3` on SM100+ (Blackwell) and `bfloat16` on older architectures. The DSA decode backend is automatically selected based on KV dtype (`bfloat16` → `flashmla_sparse`, `fp8_e4m3` → `flashmla_kv`), except for GLM DSA models on SM120/SM121 with `fp8_e4m3`, which use `flashinfer_sparse_mla`. DSA backend flags apply only to DSA models; DeepSeek V4 uses its own `dsv4` attention backend.
 
+### MiniMax M3
+
+Dense-layer K/V and index K stay on GPU; sparse-layer K/V use host memory plus a GPU working set.
+
+- Use TP 4 or greater, with the same TP size and PP 1 on both PD instances.
+- Use `--attention-backend triton`, `--mm-attention-backend triton_attn`, `--disable-prefill-cuda-graph`, and `--disable-radix-cache`.
+- Set `device_buffer_size` to at least 2048 in `--hisparse-config`; `top_k` does not override the model's selection width.
+- PD retraction backup is not supported. Use `--num-reserved-decode-tokens` to reserve capacity for the expected output length.
+
 ### Benchmark
 
 ```bash Command