> ## Documentation Index > Fetch the complete documentation index at: https://lmsysorg-dsv4-1.mintlify.site/llms.txt > Use this file to discover all available pages before exploring further. # Attention Backend SGLang supports a large variety of attention backends. Each of them has different pros and cons. You can test them according to your needs. Selecting an optimal attention backend is crucial for maximizing your performance. Different backends excel in various scenarios, so choose based on your model, hardware, and use case. Not all backends are supported on all platforms and model architectures. If you don't specify `--attention-backend`, SGLang makes a best effort to automatically select the most performant backend based on your hardware and model architecture. ## Support Matrix The support matrix is split into two parts: MHA (standard attention) and MLA (multi-head latent attention). For an explanation of the key differences between MHA and MLA, please see the [SGLang documentation on DeepSeek MLA](/cookbook/autoregressive/DeepSeek/DeepSeek-V3#4-2-4-mla-optimizations) and the original [DeepSeek MLA paper](https://arxiv.org/pdf/2405.04434). ### MHA Backends
**Backend** **Page Size > 1 (native)** **FP8 KV Cache** **FP4 KV Cache** **Spec topk=1** **Spec topk>1** **Sliding Window** **MultiModal**
**FlashInfer** ✅ ✅ ❌ ✅ ✅ ✅ ❌
**FA3 (FlashAttention 3)** ✅ ✅ ❌ ✅ ✅ ✅ ✅
**FA4 (FlashAttention 4)** 128 ❌ ✅ ✅ ✅ ✅ ✅
**Triton** ❌ ✅ ✅ ✅ ✅ ✅ ✅
**Torch Native (SDPA)** ❌ ✅ ✅ ❌ ❌ ❌ ✅
**FlexAttention (PyTorch)** ❌ ❌ ✅ ❌ ❌ ❌ ❌
**TRTLLM MHA** 16, 32 or 64 ✅ ✅ ✅ ❌ ✅ ❌
**Dual Chunk FlashAttention** ✅ ❌ ❌ ❌ ❌ ❌ ❌
**HPC-Ops** 64 ✅ ❌ ❌ ❌ ❌ ❌
**AITER (ROCm)** ✅ ✅ ❌ ✅ ✅ ✅ ✅
**Wave (ROCm)** ✅ ❌ ❌ ❌ ❌ ❌ ❌
**Ascend (NPU)** ✅ ❌ ❌ ✅ ❌ ✅ ✅
**Intel XPU** ✅ ❌ ❌ ❌ ❌ ✅ ❌
**Intel AMX (CPU)** ❌ ❌ ❌ ❌ ❌ ❌ ❌
### MLA Backends
**Backend** **Native Page Sizes** **FP8 KV Cache** **FP4 KV Cache** **Chunked Prefix Cache** **Spec topk=1** **Spec topk>1**
**FlashInfer MLA** 1 ❌ ✅ ✅ ✅ ❌
**FlashMLA** 64 ✅ ✅ ✅ ✅ ❌
**Cutlass MLA** 128 ✅ ✅ ✅ ✅ ❌
**TRTLLM MLA (Blackwell)** 32 or 64 ✅ ✅ ✅ ✅ ❌
**CuteDSL MLA (Blackwell)** 32 or 64 ✅ ❌ ✅ ✅ ❌
**TokenSpeed MLA (Blackwell)** 32 or 64 ✅ (required) ❌ ✅ ✅ ❌
**FA3 (FlashAttention 3)** n/a ❌ ❌ ✅ ✅ ⚠️ (page\_size=1 only)
**Triton** n/a ❌ ❌ ❌ ✅ ⚠️ (page\_size=1 only)
**FA4** 1 ❌ ✅ ✅ ❌ ❌
**Ascend MLA (NPU)** 128 ❌ ❌ ❌ ❌ ❌
Multimodal attention is selected by `--mm-attention-backend`. The "MultiModal" column indicates whether a corresponding multimodal implementation exists for that backend family. * DSA is specifically designed for [DeepSeek V3.2](https://lmsys.org/blog/2025-09-29-deepseek-V32/). See the [DSA Attention Backend](#dsa-attention-backend) section and [DeepSeek V3.2 deployment guide](/cookbook/autoregressive/DeepSeek/DeepSeek-V3_2) for details. For the KV4 FA4 scenario, FA4 requires using a different --decode-attention-backend to run. Except for trtllm\_mha being incompatible with FA4, all other decode backends behave as shown in the table. Speculative decoding topk: `topk` is the number of draft tokens sampled per step from the draft model. `topk = 1` follows classic EAGLE; `topk > 1` explores multiple branches and requires backend support in both draft and verification paths. Page size controls how many tokens are grouped into a KV cache block. For the prefix cache to take effect, the number of tokens must fill at least one complete page. For example, if your prompt is only 32 tokens and `page_size = 64`, it won't fill a complete page and cannot be matched in the prefix cache (pages cannot be padded). With 65 tokens and `page_size = 64`, only the first page of 64 tokens will be cached and matched; the remaining 1 token is discarded. Use `page_size = 1` for maximum prefix reuse (token-level matching). Note that higher page sizes generally improve attention kernel performance, so prefer `page_size > 1` when prefix cache reuse is not critical. Many backends that do not natively operate on pages can emulate `page_size > 1` at the wrapper layer by expanding page tables to per-token indices. The "Page Size > 1 (native)" column indicates true in-kernel paging. Some backends require fixed native page sizes and cannot be reduced/emulated differently: TRTLLM MHA (16/32/64), TRTLLM MLA (32/64), CuteDSL MLA (32/64), FlashMLA (64), Cutlass MLA (128), Ascend (128), HPC-Ops (64). MLA page-size constraints: * FlashInfer MLA: page\_size = 1. * FlashMLA: page\_size = 64. * Cutlass MLA: page\_size = 128. * TRTLLM MLA: page\_size ∈ \{32, 64}. * CuteDSL MLA: page\_size ∈ \{32, 64} (decode-only; prefill falls back to `trtllm_mla` when unset). * TokenSpeed MLA: page\_size ∈ \{32, 64} (Blackwell SM100/SM12x only; requires `--kv-cache-dtype fp8_e4m3`). ### GDN Attention Backends GDN (Gated Delta Network) is a linear attention mechanism with O(n) complexity, used in hybrid models that alternate GDN linear attention layers with standard full attention layers. GDN is **not** selected via `--attention-backend`; it is automatically activated when the model architecture requires it (e.g., Qwen 3.5, Qwen 3 Next, Jet Nemotron, Jet VLM). The GDN linear attention layers have their own kernel backends, selected via `--linear-attn-backend` (default: `triton`). You can override the kernel per phase with `--linear-attn-decode-backend` and `--linear-attn-prefill-backend`. On SM100/SM103 with CUDA 13+, SGLang automatically selects FlashInfer for GDN prefill when the per-phase override is unset, the base linear-attention backend is Triton, recurrent state is BF16, key/value head dimensions are 128, dynamic chunking and page-major KV layout are disabled, and `--chunked-prefill-size` is between 1 and 8192. Radix caching may be disabled or use `no_buffer`, `extra_buffer`, or `extra_buffer_lazy`; the extra-buffer paths use state checkpoints.
Backend Decode Prefill / Extend Spec Decoding (Target Verify)
Triton (CUDA) ✅ ✅ ✅
Triton (AMD/ROCm) ✅ ✅ ✅
Triton (NPU) ✅ ✅ ❌
Triton (CPU) ✅ ✅ ❌
CuTe DSL (CUDA only) ✅ ❌ ❌
FlashInfer (CUDA, SM90/SM100/SM103) ✅ ✅ ✅ linear chain; tree falls back to Triton
GDN models are hybrid: the full-attention layers still require a standard `--attention-backend`. Platform constraints for the full-attention backend on hybrid GDN models: * **Blackwell SM120 (e.g., RTX PRO 6000 Blackwell)**: `triton` or `flashinfer` for prefill/full attention; `trtllm_mha` is supported for `--decode-attention-backend` only. * **Other Blackwell variants (including SM100 B200/GB200)**: `triton`, `trtllm_mha`, or `fa4` only. * **NPU (Ascend)**: `ascend` only. * **AMD (ROCm)**: `triton` recommended. * **Other CUDA (Hopper, Ampere, etc.)**: auto-selection works; no special constraints. ### DSA Attention Backend DSA (DeepSeek Sparse Attention) is a native sparse attention mechanism used by [DeepSeek V3.2](https://lmsys.org/blog/2025-09-29-deepseek-V32/). It is activated automatically when the model architecture requires it and is selected via `--attention-backend dsa` (deprecated alias: `nsa`). Internally, the DSA backend dispatches to different sub-backends for prefill and decode phases. You can override these with `--dsa-prefill-backend` and `--dsa-decode-backend`:
Sub-backend Prefill Decode Notes
flashmla\_sparse ✅ ✅ Default prefill on Hopper and Blackwell (BF16)
flashmla\_sparse\_q8 ✅ ❌ Native FP8 (q8×kv8) sparse prefill on Hopper (SM90); requires --kv-cache-dtype fp8\_e4m3
flashmla\_kv ✅ ✅ Default for FP8 on Hopper (prefill + decode)
flashmla\_auto ✅ ❌ Picks flashmla\_sparse or flashmla\_kv by KV cache dtype
fa3 ✅ ✅ Default decode on Hopper (BF16)
trtllm ✅ ✅ Default decode on Blackwell (BF16); default for FP8 on Blackwell (prefill + decode)
tilelang ✅ ✅ Default on AMD (ROCm)
aiter ✅ ✅ AMD-specific kernel library (requires aiter package)
#### Gemma 4 on `aiter` (ROCm) Gemma 4's hybrid attention mixes sliding-window layers (`head_dim` 256) with full-attention layers (`head_dim` 512) and shares KV across layers. The `aiter` backend supports this and is allow-listed for Gemma 4, so you can serve it with `--attention-backend aiter` on MI300X/MI350X. Under the hood `aiter` routes the `head_dim`-512 prefill through the Triton `unified_attention` kernel (the CK batch-prefill path caps `head_dim` at 256), honors each layer's softmax scale, and resolves the sliding-window KV pool for speculative-decode draft workers. Non–sliding-window models (`head_dim` ≤ 256) are unaffected and keep using the CK path. For Gemma 4 on `aiter`, use bf16 KV cache (`--kv-cache-dtype auto`). Do **not** use `--kv-cache-dtype fp8_e4m3`: batched decode currently returns incorrect output under fp8 KV on this path (a pre-existing aiter fp8-decode issue, tracked separately). bf16 KV is correct at every batch size. For deployment examples, see the [DeepSeek V3.2 deployment guide](/cookbook/autoregressive/DeepSeek/DeepSeek-V3_2). ### Hybrid attention (different backends for prefill vs decode) (Experimental) Hybrid attention is an experimental feature. You can mix-and-match attention backends for prefill and decode. This is useful when one backend excels at prefill and another excels at decode. For the implementation details, please see `python/sglang/srt/layers/attention/hybrid_attn_backend.py`. ```bash Command theme={null} # Example: Prefill with FA4, Decode with TRTLLM MLA (Blackwell) python3 -m sglang.launch_server \ --model-path nvidia/DeepSeek-R1-FP4 \ --tp 8 \ --attention-backend trtllm_mla \ --moe-runner-backend flashinfer_trtllm \ --quantization modelopt_fp4 \ --prefill-attention-backend fa4 ``` #### Speculative decoding with hybrid attention Hybrid attention also works with speculative decoding. The backend used for draft decoding and target verification depends on `--speculative-attention-mode`: * `--speculative-attention-mode decode` (recommended): draft/verify use the decode backend. * `--speculative-attention-mode prefill` (default): draft/verify use the prefill backend. Constraints when combining hybrid attention with speculative decoding: * If any attention backend is `trtllm_mha`, speculative decoding supports only `--speculative-eagle-topk 1`. * For paged MHA backends with `--page-size > 1` and `--speculative-eagle-topk > 1`, only `flashinfer` is supported. * CUDA Graph: the decode backend is always captured; the prefill backend is captured only when `--speculative-attention-mode prefill`. If you set only one of `--prefill-attention-backend` or `--decode-attention-backend`, the unspecified phase inherits `--attention-backend`. If both are specified and differ, SGLang automatically enables a hybrid wrapper to dispatch to the chosen backend per phase. ## Attention Backend Selection Guide (CUDA) If the `--attention-backend` argument is not specified, SGLang automatically selects the best backend based on the hardware (CUDA) and model architecture. ### Automatic Selection Logic **1. MHA Models (e.g., Llama, Qwen)** * **Hopper (e.g., H100, H200)**: Defaults to `fa3` if using CUDA 12.3+ and the model configuration is supported. * **Blackwell (e.g., B200)**: Defaults to `trtllm_mha`, unless using speculative decoding with `topk > 1`. * **Other Architectures (Ampere, Ada, etc.)**: Defaults to `flashinfer` if available; otherwise falls back to `triton`. **2. MLA Models (e.g., DeepSeek V3)** * **Hopper**: Defaults to `fa3` (requires CUDA 12.3+). * **Blackwell**: Defaults to `flashinfer`; `trtllm_mla` is auto-selected for DeepSeek V3 models specifically. * **Other Architectures**: Defaults to `triton`. ## User Guide ### Launch Command for Different Attention Backends * FlashInfer (Default for Non-Hopper Machines, e.g., A100, A40) ```bash Command theme={null} python3 -m sglang.launch_server \ --model meta-llama/Meta-Llama-3.1-8B-Instruct \ --attention-backend flashinfer python3 -m sglang.launch_server \ --tp 8 \ --model deepseek-ai/DeepSeek-V3 \ --attention-backend flashinfer \ --trust-remote-code ``` * FlashAttention 3 (Default for Hopper Machines, e.g., H100, H200, H20) ```bash Command theme={null} python3 -m sglang.launch_server \ --model meta-llama/Meta-Llama-3.1-8B-Instruct \ --attention-backend fa3 python3 -m sglang.launch_server \ --tp 8 \ --model deepseek-ai/DeepSeek-V3 \ --trust-remote-code \ --attention-backend fa3 ``` * Triton ```bash Command theme={null} python3 -m sglang.launch_server \ --model meta-llama/Meta-Llama-3.1-8B-Instruct \ --attention-backend triton python3 -m sglang.launch_server \ --tp 8 \ --model deepseek-ai/DeepSeek-V3 \ --attention-backend triton \ --trust-remote-code ``` * FlashMLA ```bash Command theme={null} python3 -m sglang.launch_server \ --tp 8 \ --model deepseek-ai/DeepSeek-R1 \ --attention-backend flashmla \ --trust-remote-code python3 -m sglang.launch_server \ --tp 8 \ --model deepseek-ai/DeepSeek-R1 \ --attention-backend flashmla \ --kv-cache-dtype fp8_e4m3 \ --trust-remote-code ``` * TRTLLM MLA (Optimized for Blackwell Architecture, e.g., B200) ```bash Command theme={null} python3 -m sglang.launch_server \ --tp 8 \ --model deepseek-ai/DeepSeek-R1 \ --attention-backend trtllm_mla \ --trust-remote-code ``` * TRTLLM MLA with FP8 KV Cache (Higher concurrency, lower memory footprint) ```bash Command theme={null} python3 -m sglang.launch_server \ --tp 8 \ --model deepseek-ai/DeepSeek-R1 \ --attention-backend trtllm_mla \ --kv-cache-dtype fp8_e4m3 \ --trust-remote-code ``` * TRTLLM MHA (Optimized for Blackwell Architecture, e.g., B200) ```bash Command theme={null} python3 -m sglang.launch_server \ --tp 4 \ --model Qwen/Qwen3.5-35B-A3B-FP8 \ --attention-backend trtllm_mha \ --trust-remote-code ``` * TRTLLM MHA (XQA backend) (Optimized for SM90 and SM120, e.g., H20, H200, 5090) Note that TRTLLM XQA backend only works well for pagesize 64. ```bash Command theme={null} python3 -m sglang.launch_server \ --tp 4 \ --model Qwen/Qwen3.5-35B-A3B-FP8 \ --decode-attention-backend trtllm_mha \ --trust-remote-code ``` * HPC-Ops (MHA kernels from [HPC-Ops](https://github.com/Tencent/hpc-ops) by the Tencent Hunyuan AI Infra team; Hopper (SM90) only, requires installing the `hpc` package from source, page size 64, bf16 or fp8\_e4m3 KV cache, head\_dim 128, q/kv head group 4 or 8) ```bash Command theme={null} python3 -m sglang.launch_server \ --model Qwen/Qwen3-30B-A3B-Instruct-2507-FP8 \ --attention-backend hpc_ops \ --page-size 64 \ --trust-remote-code # FP8 models should also set --kv-cache-dtype fp8_e4m3 to run the FP8 attention # kernels. This enables the fused QKNorm+RoPE+FP8-quant+StoreKV op, which is # currently wired for Hunyuan V3 and requires per-rank (q_heads, kv_heads) of # (64, 8) or (8, 1), e.g. --tp 1 or --tp 8 for Hy3. python3 -m sglang.launch_server \ --tp 8 \ --model tencent/Hy3-FP8 \ --attention-backend hpc_ops \ --kv-cache-dtype fp8_e4m3 \ --page-size 64 \ --trust-remote-code ``` * FlashAttention 4 (MHA & MLA) ```bash Command theme={null} # FA4 for both prefill and decode on SM90/SM100 python3 -m sglang.launch_server \ --model-path Qwen/Qwen3-30B-A3B-Instruct-2507-FP8 \ --attention-backend fa4 \ --page-size 128 \ --trust-remote-code python3 -m sglang.launch_server \ --tp 8 \ --model deepseek-ai/DeepSeek-R1 \ --prefill-attention-backend fa4 \ --trust-remote-code ``` * Ascend ```bash Command theme={null} python3 -m sglang.launch_server \ --model meta-llama/Meta-Llama-3.1-8B-Instruct \ --attention-backend ascend ``` * Intel XPU ```bash Command theme={null} python3 -m sglang.launch_server \ --model meta-llama/Meta-Llama-3.1-8B-Instruct \ --attention-backend intel_xpu ``` * Wave ```bash Command theme={null} python3 -m sglang.launch_server \ --model meta-llama/Meta-Llama-3.1-8B-Instruct \ --attention-backend wave ``` * FlexAttention ```bash Command theme={null} python3 -m sglang.launch_server \ --model meta-llama/Meta-Llama-3.1-8B-Instruct \ --attention-backend flex_attention ``` * Torch Native ```bash Command theme={null} python3 -m sglang.launch_server \ --model meta-llama/Meta-Llama-3.1-8B-Instruct \ --attention-backend torch_native ``` ## Steps to add a new attention backend To add a new attention backend, you can learn from the existing backends (`python/sglang/srt/layers/attention/triton_backend.py`, `python/sglang/srt/layers/attention/flashattention_backend.py`) and follow the steps below. Linear attention kernel backends (GDN, KDA) follow a different pattern. They implement `LinearAttnKernelBase` in `python/sglang/srt/layers/attention/linear/kernels/` and are dispatched by `GDNKernelDispatcher` / `KDAKernelDispatcher` rather than registered via `@register_attention_backend`. 1. Run without cuda graph. Support the two forward functions * forward\_extend * Will be used for prefill, prefill with KV cache, and target verification * It will be called once per layer * forward\_decode * Will be used for normal decode, and draft decode * It will be called once per layer * init\_forward\_metadata * Initialize the class and common metadata shared by all layers * Call the plan function for optimizations like split\_kv * It will be called once per forward 2. Run with cuda graph. It has two phases (capture and replay) and you need to implement three functions * init\_cuda\_graph\_state * It will be called once during life time * Create all common shared buffers * init\_forward\_metadata\_capture\_cuda\_graph * It will be called before capturing a cuda graph * It is similar to init\_forward\_metadata but write the medatada to some pre-defined buffers * init\_forward\_metadata\_replay\_cuda\_graph * It will be called before replaying a cuda graph * This function is in the critical path and needs to be fast