Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
attention_backends.md 8.93 KB
Newer Older
raojy's avatar
raw_vllm  
raojy committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216 
# Attention Backend Feature Support

This document is auto-generated by `tools/pre_commit/generate_attention_backend_docs.py`.
It shows the feature support for each registered attention backend
based on the checks in `AttentionBackend.validate_configuration()`.

**Do not edit this file manually.** Run the following command to
regenerate it:

```bash
python tools/pre_commit/generate_attention_backend_docs.py
```

## Setting the Attention Backend

### Command Line

There are two ways to specify the backend from the command line:

**Option 1: Using `--attention-backend` (simple)**

```bash
vllm serve <model> --attention-backend FLASH_ATTN
```

**Option 2: Using `--attention-config.backend` / `-ac.backend` (structured config)**

```bash
# Dot notation
vllm serve <model> --attention-config.backend FLASH_ATTN
vllm serve <model> -ac.backend FLASH_ATTN

# JSON format
vllm serve <model> --attention-config '{"backend": "FLASH_ATTN"}'
vllm serve <model> -ac '{"backend": "FLASH_ATTN"}'
```

> **Note:** `--attention-backend` and `--attention-config.backend` are mutually
> exclusive. Use one or the other, not both.

### Python API

Use `AttentionConfig` with the `LLM` class:

```python
from vllm import LLM
from vllm.config import AttentionConfig
from vllm.v1.attention.backends.registry import AttentionBackendEnum

# Method 1: Using AttentionConfig with enum
llm = LLM(
    model="Qwen/Qwen3-0.6B",
    attention_config=AttentionConfig(backend=AttentionBackendEnum.FLASH_ATTN),
)

# Method 2: Using attention_backend parameter with string
llm = LLM(
    model="Qwen/Qwen3-0.6B",
    attention_backend="FLASH_ATTN",
)
```

## Backend Selection Behavior

### Manual Selection

When you explicitly set a backend via `--attention-backend` or `AttentionConfig`:

1. The backend is **validated** against your configuration (model dtype, head
   size, compute capability, etc.)
2. If the backend **doesn't support** your configuration, an error is raised
   with the specific reason
3. If valid, the backend is used

Example error when selecting an incompatible backend:

```text
ValueError: Selected backend FLASHMLA is not valid for this configuration.
Reason: ['compute capability not supported']
```

### Automatic Selection

When no backend is specified (the default):

1. vLLM iterates through backends in **priority order** (see tables below)
2. Each backend is validated against your configuration
3. The **first compatible backend** is selected
4. If no backend is compatible, an error is raised listing all backends and
   their incompatibility reasons

## Backend Priority (CUDA)

When no backend is explicitly selected, vLLM chooses the first
compatible backend from these priority-ordered lists.

Priority is **1 = highest** (tried first).

### Standard Attention (MHA, MQA, GQA)

**Blackwell (SM 10.x):**

| Priority | Backend |
|----------|---------|
| 1 | `FLASHINFER` |
| 2 | `FLASH_ATTN` |
| 3 | `TRITON_ATTN` |
| 4 | `FLEX_ATTENTION` |

**Ampere/Hopper (SM 8.x-9.x):**

| Priority | Backend |
|----------|---------|
| 1 | `FLASH_ATTN` |
| 2 | `FLASHINFER` |
| 3 | `TRITON_ATTN` |
| 4 | `FLEX_ATTENTION` |

### MLA Attention (DeepSeek-style)

**Blackwell (SM 10.x):**

| Priority | Backend |
|----------|---------|
| 1 | `FLASHINFER_MLA` |
| 2 | `CUTLASS_MLA` |
| 3 | `FLASH_ATTN_MLA` |
| 4 | `FLASHMLA` |
| 5 | `TRITON_MLA` |
| 6 | `FLASHMLA_SPARSE` |
| 7 | `FLASHINFER_MLA_SPARSE` |

**Ampere/Hopper (SM 8.x-9.x):**

| Priority | Backend |
|----------|---------|
| 1 | `FLASH_ATTN_MLA` |
| 2 | `FLASHMLA` |
| 3 | `FLASHINFER_MLA` |
| 4 | `TRITON_MLA` |
| 5 | `FLASHMLA_SPARSE` |

> **Note:** ROCm and CPU platforms have their own selection logic. See the platform-specific documentation for details.

## Legend

| Column | Description |
|--------|-------------|
| **Dtypes** | Supported model data types (fp16, bf16, fp32) |
| **KV Dtypes** | Supported KV cache data types (`auto`, `fp8`, `fp8_e4m3`, etc.) |
| **Block Sizes** | Supported KV cache block sizes (%N means multiples of N) |
| **Head Sizes** | Supported attention head sizes |
| **Sink** | Attention sink support (for StreamingLLM) |
| **Sparse** | Sparse attention support (MLA only) |
| **MM Prefix** | Multimodal prefix full attention support |
| **DCP** | Decode Context Parallelism support (`--decode-context-parallel-size`) |
| **Attention Types** | Supported attention patterns (Decoder, Encoder, Enc-Dec) |
| **Compute Cap.** | Required CUDA compute capability (N/A for non-CUDA backends) |

**Symbols:** ✅ = Supported, ❌ = Not supported

## Standard Attention (MHA, MQA, GQA) Backends

| Backend | Version | Dtypes | KV Dtypes | Block Sizes | Head Sizes | Sink | MM Prefix | DCP | Attention Types | Compute Cap. |
|---------|---------|--------|-----------|-------------|------------|------|-----------|-----|-----------------|--------------|
| `CPU_ATTN` |  | fp16, bf16, fp32 | `auto` | Any | 32, 64, 80, 96, 112, 128, 160, 192, 224, 256 | ❌ | ❌ | ❌ | All | N/A |
| `FLASHINFER` | Native† | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32, 64 | 64, 128, 256 | ❌ | ❌ | ✅ | Decoder | 7.x-9.x |
| `FLASHINFER` | TRTLLM† | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32, 64 | 64, 128, 256 | ✅ | ❌ | ✅ | Decoder | 10.x |
| `FLASH_ATTN` | FA2* | fp16, bf16 | `auto`, `bfloat16` | %16 | Any | ❌ | ❌ | ✅ | All | ≥8.0 |
| `FLASH_ATTN` | FA3* | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | Any | ✅ | ❌ | ✅ | All | 9.x |
| `FLASH_ATTN` | FA4* | fp16, bf16 | `auto`, `bfloat16` | %16 | Any | ❌ | ❌ | ✅ | All | ≥10.0 |
| `FLASH_ATTN_DIFFKV` |  | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ✅ | Decoder | Any |
| `FLEX_ATTENTION` |  | fp16, bf16, fp32 | `auto`, `bfloat16` | Any | Any | ❌ | ✅ | ❌ | Decoder, Encoder Only | Any |
| `ROCM_AITER_FA` |  | fp16, bf16 | `auto` | 16, 32 | 64, 128, 256 | ❌ | ❌ | ❌ | Decoder | N/A |
| `ROCM_AITER_UNIFIED_ATTN` |  | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ❌ | All | N/A |
| `ROCM_ATTN` |  | fp16, bf16, fp32 | `auto` | 16, 32, 544 | 32, 64, 80, 96, 128, 160, 192, 224, 256 | ❌ | ❌ | ❌ | All | N/A |
| `TREE_ATTN` |  | fp16, bf16 | `auto` | %16 | 32, 64, 96, 128, 160, 192, 224, 256 | ❌ | ❌ | ❌ | Decoder | Any |
| `TRITON_ATTN` |  | fp16, bf16, fp32 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | Any | ✅ | ✅ | ❌ | All | Any |

> **†** FlashInfer uses TRTLLM attention on Blackwell (SM100), which supports sinks. Disable via `--attention-config.use_trtllm_attention=0`.
>
> **\*** Specify the FlashAttention version via `--attention-config.flash_attn_version=2`, `3`, or `4`. Default is FA4 on SM100+ (Blackwell), FA3 on SM90 (Hopper), FA2 otherwise.

## MLA (Multi-head Latent Attention) Backends

MLA uses separate backends for prefill and decode phases.

### Prefill Backends

The prefill backend is selected at runtime based on hardware and
configuration.

| Backend | Description | Compute Cap. | Enable | Disable | Notes |
|---------|-------------|--------------|--------|---------|-------|
| TRT-LLM Ragged‡ | TensorRT-LLM ragged attention | 10.x | Default on SM100 | `-ac.use_trtllm_ragged_deepseek_prefill=0` | DeepSeek R1 dims only |
| FlashInfer | FlashInfer CUTLASS backend | 10.x | `-ac.disable_flashinfer_prefill=0` | `-ac.disable_flashinfer_prefill=1` | DeepSeek R1 dims only |
| cuDNN | cuDNN-based attention | 10.x | `-ac.use_cudnn_prefill=1` | `-ac.use_cudnn_prefill=0` |  |
| FlashAttention | FlashAttention varlen (FA2/FA3) | Any | Default fallback | Use other backends | FA3 on SM90, FA2 otherwise |

> **‡** TRT-LLM Ragged is the default on Blackwell (SM100).
> On other GPUs, FlashAttention is used as the default.

### Decode Backends

| Backend | Dtypes | KV Dtypes | Block Sizes | Head Sizes | Sink | Sparse | MM Prefix | DCP | Attention Types | Compute Cap. |
|---------|--------|-----------|-------------|------------|------|--------|-----------|-----|-----------------|--------------|
| `CUTLASS_MLA` | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3` | 128 | Any | ❌ | ❌ | ❌ | ✅ | Decoder | 10.x |
| `FLASHINFER_MLA` | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3` | 32, 64 | Any | ❌ | ❌ | ❌ | ❌ | Decoder | 10.x |
| `FLASHINFER_MLA_SPARSE` | fp16, bf16 | `auto`, `bfloat16` | 32, 64 | 576 | ❌ | ✅ | ❌ | ❌ | Decoder | 10.x |
| `FLASHMLA` | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3` | 64 | Any | ❌ | ❌ | ❌ | ✅ | Decoder | 9.x-10.x |
| `FLASHMLA_SPARSE` | bf16 | `auto`, `bfloat16`, `fp8_ds_mla` | 64 | 576 | ❌ | ✅ | ❌ | ❌ | Decoder | 9.x-10.x |
| `FLASH_ATTN_MLA` | fp16, bf16 | `auto`, `bfloat16` | %16 | Any | ❌ | ❌ | ❌ | ✅ | Decoder | 9.x |
| `ROCM_AITER_MLA` | fp16, bf16 | `auto` | 1 | Any | ❌ | ❌ | ❌ | ❌ | Decoder | N/A |
| `ROCM_AITER_MLA_SPARSE` | fp16, bf16 | `auto` | Any | 576 | ❌ | ❌ | ❌ | ❌ | Decoder | N/A |
| `ROCM_AITER_TRITON_MLA` | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ❌ | ❌ | Decoder | N/A |
| `TRITON_MLA` | fp16, bf16 | `auto`, `bfloat16` | Any | Any | ❌ | ❌ | ❌ | ✅ | Decoder | Any |