mistralrs_core/paged_attention/
mod.rs

1/// The higher-level manager of the blocks allocated. Operations performed by the block engine do
2/// not directly change memory.
3mod block_engine;
4mod block_engine_sequence;
5/// This is the lower-level manager of the cache. It manages swapping and copying the blocks and
6/// actually allocates the KV cache for the CPU and GPU. It is used by the LLMEngine to execute
7/// operations issued by the scheduler.
8mod cache_engine;
9mod config;
10mod layers;
11mod scheduler;
12pub const _PAD_SLOT_ID: i64 = -1;
13
14pub use block_engine::{BlockEngine, BlockTables, LogicalTokenBlock, PhysicalTokenBlock};
15pub use block_engine_sequence::BlockEngineSequence;
16pub use cache_engine::{CacheConfig, CacheEngine, PagedCacheType};
17use candle_core::{DType, Device};
18pub use config::{ModelConfigLike, ModelConfigMetadata};
19pub use layers::PagedAttention;
20pub use scheduler::{
21    PagedAttentionScheduler, PagedAttentionSchedulerConfig, PagedAttentionSchedulerOutput,
22};
23
24use crate::MemoryUsage;
25use tracing::{info, warn};
26
27pub const DEFAULT_PAGED_ATTENTION_BLOCK_SIZE: usize = 32;
28
29/// All memory counts in MB. Default for block size is 32.
30#[derive(Clone, Copy)]
31pub struct PagedAttentionConfig {
32    pub(crate) block_size: Option<usize>,
33    pub(crate) mem_gpu: MemoryGpuConfig,
34    pub(crate) cache_type: PagedCacheType,
35}
36
37impl PagedAttentionConfig {
38    pub fn new(
39        block_size: Option<usize>,
40        mem_gpu: MemoryGpuConfig,
41        cache_type: PagedCacheType,
42    ) -> anyhow::Result<Self> {
43        Ok(Self {
44            block_size,
45            mem_gpu,
46            cache_type,
47        })
48    }
49}
50
51#[derive(Debug, Clone, Copy, PartialEq)]
52pub enum AttentionImplementation {
53    Eager,
54    PagedAttention,
55}
56
57#[derive(Clone, Copy)]
58#[cfg_attr(feature = "pyo3_macros", pyo3::pyclass)]
59pub enum MemoryGpuConfig {
60    MbAmount(usize),
61    Utilization(f32),
62    ContextSize(usize),
63}
64
65// See `pagedattention.cu` CALL_V1_LAUNCHER_BLOCK_SIZE
66const SUPPORTED_BLOCK_SIZE: &[usize] = &[8, 16, 32];
67
68const SIZE_IN_MB: usize = 1024 * 1024;
69
70macro_rules! mb_to_blocks {
71    ($mb_size:expr, $dtype_size:expr, $block_size:expr, $config:expr) => {
72        $mb_size
73            / $dtype_size
74            / $block_size
75            / $config.num_kv_heads()
76            / ($config.k_head_dim().max($config.v_head_dim()))
77            / $config.num_layers()
78            / 2
79    };
80}
81
82macro_rules! ctxt_to_blocks {
83    ($context_len:expr, $dtype_size:expr, $block_size:expr, $config:expr) => {
84        $context_len
85            * $dtype_size
86            * $config.num_kv_heads()
87            * ($config.k_head_dim().max($config.v_head_dim()))
88            * $config.num_layers()
89            * 2
90    };
91}
92
93/// Memory values are in MBs or a percentage in [0,1]. Specify block size or the default is 32.
94#[allow(clippy::too_many_arguments)]
95pub fn calculate_cache_config(
96    mem_gpu: MemoryGpuConfig,
97    block_size: Option<usize>,
98    dtype: DType,
99    cache_type: PagedCacheType,
100    config: &dyn ModelConfigLike,
101    device: &Device,
102    layer_devices: &[Option<Device>],
103    silent: bool,
104) -> anyhow::Result<CacheConfig> {
105    let block_size = block_size.unwrap_or(DEFAULT_PAGED_ATTENTION_BLOCK_SIZE);
106    if !SUPPORTED_BLOCK_SIZE.contains(&block_size) {
107        anyhow::bail!("Block size must be in {SUPPORTED_BLOCK_SIZE:?}, got {block_size}");
108    }
109    let dtype = cache_type.to_dtype(dtype);
110    let dtype_size = dtype.size_in_bytes();
111
112    let mut min_mem_gpu = usize::MAX;
113    for dev in layer_devices {
114        let device = dev.as_ref().unwrap_or(device);
115
116        #[allow(clippy::cast_possible_truncation, clippy::cast_precision_loss)]
117        let mem_gpu = match mem_gpu {
118            MemoryGpuConfig::MbAmount(v) => v,
119            MemoryGpuConfig::Utilization(f) => {
120                let free = MemoryUsage.get_memory_available(device)? as f32 / SIZE_IN_MB as f32;
121                let total = MemoryUsage.get_total_memory(device)? as f32 / SIZE_IN_MB as f32;
122                let used = total - free;
123                (total * f - used) as usize
124            }
125            MemoryGpuConfig::ContextSize(toks) => {
126                ctxt_to_blocks!(toks, dtype_size, block_size, config) / SIZE_IN_MB
127            }
128        };
129        min_mem_gpu = min_mem_gpu.min(mem_gpu);
130    }
131
132    // // Cap at kv cache for max seq len
133    // let mem_for_toks =
134    //     ctxt_to_blocks!(config.max_seq_len(), dtype_size, block_size, config) / SIZE_IN_MB;
135    // let mem_gpu = min_mem_gpu.min(mem_for_toks);
136
137    // Cap Metal GPU memory to the wired (non‑paged) allocation limit reported by the kernel (`iogpu.wired_limit_mb`).
138    // Users can raise this limit with `sudo sysctl -w iogpu.wired_limit_mb=<desired_mb>`.
139    let mem_gpu = if matches!(device, Device::Metal(_)) {
140        let metal_cap_mb = MemoryUsage.get_total_memory(device)? / SIZE_IN_MB;
141
142        info!("Metal GPU wired limit is {metal_cap_mb} MB.");
143
144        if min_mem_gpu > metal_cap_mb {
145            if !silent {
146                warn!(
147                    "Capping Metal GPU memory allocation from {} MB to {} MB (limited by iogpu.wired_limit_mb). \
148To raise this cap run: `sudo sysctl -w iogpu.wired_limit_mb=<desired_mb>`.",
149                    min_mem_gpu,
150                    metal_cap_mb
151                );
152            }
153            metal_cap_mb
154        } else {
155            min_mem_gpu
156        }
157    } else {
158        min_mem_gpu
159    };
160
161    let num_gpu_blocks = mb_to_blocks!(mem_gpu * SIZE_IN_MB, dtype_size, block_size, config);
162    if num_gpu_blocks == 0 {
163        anyhow::bail!("Num GPU blocks is 0. This means there is not enough memory. Either reduce the memory amount/utilization/context size or disable PagedAttention.");
164    }
165
166    if !silent {
167        info!("Allocating {mem_gpu} MB for PagedAttention KV cache per GPU");
168        info!("PagedAttention KV cache type is {dtype:?}");
169        info!("Using PagedAttention with block size {block_size} and {num_gpu_blocks} GPU blocks: available context length is {} tokens", num_gpu_blocks*block_size);
170    }
171    Ok(CacheConfig {
172        block_size,
173        num_gpu_blocks,
174        cache_type,
175    })
176}
mistralrs_core/paged_attention/mod.rs

mistralrs_core/paged_attention/
mod.rs
