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
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
// SPDX-FileCopyrightText: Copyright (c) 2024-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
// SPDX-License-Identifier: Apache-2.0
//! Block Manager for LLM KV Cache
//!
//! This module provides functionality for managing KV blocks in LLM attention
//! mechanisms. It handles storage allocation, block management, and safe access
//! patterns for both system memory and remote (NIXL) storage.
pub mod config;
mod state;
pub mod block;
pub mod connector;
pub mod distributed;
pub mod events;
pub mod kv_consolidator;
pub mod layout;
pub mod metrics_kvbm;
pub mod numa_allocator;
pub mod offload;
pub mod pool;
pub mod storage;
pub mod v2;
// dynamo rt integration
pub mod controller;
pub use crate::common::dtype::DType;
pub use block::{
BasicMetadata, BlockMetadata, Blocks, ImmutableBlock, MutableBlock,
locality::{self, LocalityProvider, LogicalResources},
nixl::{BlockDescriptorList, IsImmutable, IsMutable, MutabilityKind, RemoteBlock},
};
pub use config::*;
pub use layout::{LayoutConfig, LayoutConfigBuilder, LayoutError, LayoutType, nixl::NixlLayout};
pub use offload::{filter::OffloadFilter, request::BlockResult};
pub use pool::{BlockPool, ManagedBlockPool};
pub use storage::{
DeviceStorage, DiskStorage, PinnedStorage, Storage, StorageAllocator,
nixl::NixlRegisterableStorage,
};
pub use tokio_util::sync::CancellationToken;
use anyhow::{Context, Result};
use block::nixl::{BlockMutability, NixlBlockSet, RemoteBlocks, SerializedNixlBlockSet};
use derive_builder::Builder;
use nixl_sys::Agent as NixlAgent;
use serde::{Deserialize, Serialize};
use std::{
collections::HashMap,
sync::{Arc, RwLock},
};
use storage::nixl::MemType;
use tokio::sync::oneshot;
use validator::Validate;
pub type WorkerID = u64;
pub type ReferenceBlockManager = KvBlockManager<locality::Local, BasicMetadata>;
/// Represents the different cache levels for KV blocks
#[derive(Copy, Clone, Debug, Eq, Hash, PartialEq, Serialize, Deserialize)]
pub enum CacheLevel {
/// Represents KV blocks in GPU memory
G1,
/// Represents KV blocks in CPU memory
G2,
/// Represents KV blocks in Local NVMe storage
G3,
/// Represents KV blocks in Remote NVMe storage
G4,
}
/// Type of channel used to reset the block manager to a specific cache level
pub type BlockResetChannel = tokio::sync::broadcast::Receiver<CacheLevel>;
#[derive(Debug)]
struct CancelOnLastDrop {
cancellation_token: CancellationToken,
}
impl Drop for CancelOnLastDrop {
fn drop(&mut self) {
self.cancellation_token.cancel();
}
}
// When we construct the pool:
// 1. instantiate the runtime,
// 2. build layout::LayoutConfigs for each of the requested storage types
// 3. register the layouts with the NIXL agent if enabled
// 4. construct a Blocks object for each layout providing a unique block_set_idx
// for each layout type.
// 5. initialize the pools for each set of blocks
#[derive(Debug)]
pub struct KvBlockManager<Locality: LocalityProvider, Metadata: BlockMetadata> {
state: Arc<state::KvBlockManagerState<Locality, Metadata>>,
_cancellation_token: Arc<CancelOnLastDrop>,
block_size: usize,
}
impl<Locality: LocalityProvider, Metadata: BlockMetadata> Clone
for KvBlockManager<Locality, Metadata>
{
fn clone(&self) -> Self {
Self {
state: self.state.clone(),
_cancellation_token: self._cancellation_token.clone(),
block_size: self.block_size,
}
}
}
impl<Locality: LocalityProvider, Metadata: BlockMetadata> KvBlockManager<Locality, Metadata> {
/// Get the block size
pub fn block_size(&self) -> usize {
self.block_size
}
/// Get a reference to the disk block pool
pub fn disk(&self) -> Option<&dyn BlockPool<DiskStorage, Locality, Metadata>> {
self.state.disk()
}
/// Get a reference to the host block pool
pub fn host(&self) -> Option<&dyn BlockPool<PinnedStorage, Locality, Metadata>> {
self.state.host()
}
/// Get a reference to the device block pool
pub fn device(&self) -> Option<&dyn BlockPool<DeviceStorage, Locality, Metadata>> {
self.state.device()
}
/// Get the worker ID
pub fn worker_id(&self) -> WorkerID {
self.state.worker_id()
}
/// Onboard a set of blocks to the device pool
pub fn onboard_blocks<S: Storage>(
&self,
blocks: Vec<ImmutableBlock<S, Locality, Metadata>>,
targets: Option<Vec<MutableBlock<DeviceStorage, Locality, Metadata>>>,
) -> oneshot::Receiver<BlockResult<DeviceStorage, Locality, Metadata>> {
self.state.onboard_blocks(blocks, targets)
}
}
fn build_cancel_token(config: &mut KvBlockManagerConfig) -> Arc<CancelOnLastDrop> {
// The frontend of the KvBlockManager will take ownership of the cancellation token
// and will be responsible for cancelling the task when the KvBlockManager is dropped
let cancellation_token = config.runtime.cancellation_token.clone();
// The internal state will use a child token of the original token
config.runtime.cancellation_token = cancellation_token.child_token();
Arc::new(CancelOnLastDrop { cancellation_token })
}
impl<Metadata: BlockMetadata> KvBlockManager<locality::Local, Metadata> {
/// Create a new [KvBlockManager]
///
/// The returned object is a frontend to the [KvBlockManager] which owns the cancellation
/// tokens. When this object gets drop, the cancellation token will be cancelled and begin
/// the gracefully shutdown of the block managers internal state.
pub async fn new(mut config: KvBlockManagerConfig) -> Result<Self> {
let _cancellation_token = build_cancel_token(&mut config);
let block_size = config.model.page_size;
// Create the internal state
let state = state::KvBlockManagerState::<locality::Local, Metadata>::new(config).await?;
Ok(Self {
state,
_cancellation_token,
block_size,
})
}
/// Exports the local blockset configuration as a serialized object.
pub fn export_local_blockset(&self) -> Result<SerializedNixlBlockSet> {
self.state.export_local_blockset()
}
/// Imports a remote blockset configuration from a serialized object.
pub fn import_remote_blockset(
&self,
serialized_blockset: SerializedNixlBlockSet,
) -> Result<()> {
self.state.import_remote_blockset(serialized_blockset)
}
/// Get a [`Vec<RemoteBlock<IsImmutable>>`] from a [`BlockDescriptorList`]
pub fn get_remote_blocks_immutable(
&self,
bds: &BlockDescriptorList,
) -> Result<Vec<RemoteBlock<IsImmutable>>> {
self.state.get_remote_blocks_immutable(bds)
}
/// Get a [`Vec<RemoteBlock<IsMutable>>`] from a [`BlockDescriptorList`]
pub fn get_remote_blocks_mutable(
&self,
bds: &BlockDescriptorList,
) -> Result<Vec<RemoteBlock<IsMutable>>> {
self.state.get_remote_blocks_mutable(bds)
}
}
impl<R: LogicalResources, Metadata: BlockMetadata> KvBlockManager<locality::Logical<R>, Metadata> {
pub async fn new(mut config: KvBlockManagerConfig, logical_resources: R) -> Result<Self> {
let block_size = config.model.page_size;
let _cancellation_token = build_cancel_token(&mut config);
let state = state::KvBlockManagerState::<locality::Logical<R>, Metadata>::new(
config,
logical_resources,
)
.await?;
Ok(Self {
state,
_cancellation_token,
block_size,
})
}
}
#[cfg(all(test, feature = "testing-full"))]
mod tests {
use super::*;
use crate::tokens::Tokens;
use std::sync::atomic::{AtomicU64, Ordering};
// Atomic Counter for Worker ID
static WORKER_ID: AtomicU64 = AtomicU64::new(1337);
pub fn create_reference_block_manager_config_with_counts(
device: usize,
host: usize,
disk: usize,
) -> KvBlockManagerConfig {
let worker_id = WORKER_ID.fetch_add(1, Ordering::SeqCst);
// Check if we're already in a Tokio runtime context
let async_runtime = if tokio::runtime::Handle::try_current().is_ok() {
None // If we're already in a runtime, don't create a new one
} else {
// Only create a new runtime if not already in one
Some(Arc::new(tokio::runtime::Runtime::new().unwrap()))
};
let builder = KvBlockManagerConfig::builder()
.runtime(
KvManagerRuntimeConfig::builder()
.worker_id(worker_id)
.enable_nixl()
.async_runtime(async_runtime)
.build()
.unwrap(),
)
.model(
KvManagerModelConfig::builder()
.num_layers(3)
.outer_dim(2)
.page_size(4)
.inner_dim(16)
.build()
.unwrap(),
);
let builder = if disk > 0 {
builder.disk_layout(
KvManagerLayoutConfig::builder()
.num_blocks(disk)
.allocator(storage::DiskAllocator)
.build()
.unwrap(),
)
} else {
builder
};
let builder = if host > 0 {
builder.host_layout(
KvManagerLayoutConfig::builder()
.num_blocks(host)
.allocator(storage::PinnedAllocator::default())
.build()
.unwrap(),
)
} else {
builder
};
let builder = if device > 0 {
builder.device_layout(
KvManagerLayoutConfig::builder()
.num_blocks(device)
.allocator(storage::DeviceAllocator::new(0).unwrap())
.build()
.unwrap(),
)
} else {
builder
};
builder.build().unwrap()
}
pub fn create_reference_block_manager_config() -> KvBlockManagerConfig {
create_reference_block_manager_config_with_counts(8, 16, 16)
}
pub async fn create_reference_block_manager() -> ReferenceBlockManager {
ReferenceBlockManager::new(create_reference_block_manager_config())
.await
.unwrap()
}
#[tokio::test]
async fn test_reference_block_manager_inherited_async_runtime() {
dynamo_runtime::logging::init();
let _block_manager = create_reference_block_manager().await;
}
// This tests mimics the behavior of two unique kvbm workers exchanging blocksets
// Each KvBlockManager is a unique worker in this test, each has its resources including
// it's own worker_ids, nixl_agent, and block pools.
//
// This test is meant to mimic the behavior of the basic nixl integration test found here:
// https://github.com/ai-dynamo/nixl/blob/main/src/bindings/rust/src/tests.rs
// TODO: This test doesn't work because NIXL doesn't support partial metadata in the rust bindings.
#[ignore]
#[tokio::test]
async fn test_reference_block_managers() {
dynamo_runtime::logging::init();
// create two block managers - mimics two unique dynamo workers
let kvbm_0 = create_reference_block_manager().await;
let kvbm_1 = create_reference_block_manager().await;
assert_ne!(kvbm_0.worker_id(), kvbm_1.worker_id());
// in dynamo, we would exchange the blocksets via the discovery plane
let blockset_0 = kvbm_0.export_local_blockset().unwrap();
let blockset_1 = kvbm_1.export_local_blockset().unwrap();
// in dynamo, we would be watching the discovery plane for remote blocksets
kvbm_0.import_remote_blockset(blockset_1).unwrap();
kvbm_1.import_remote_blockset(blockset_0).unwrap();
// Worker 0
// Allocate 4 mutable blocks on the host
let _blocks_0 = kvbm_0.host().unwrap().allocate_blocks(4).await.unwrap();
// // Create a BlockDescriptorList for the mutable blocks
// // let blockset_0 = BlockDescriptorList::from_mutable_blocks(&blocks_0).unwrap();
// let blockset_0 = blocks_0.as_block_descriptor_set().unwrap();
// // Worker 1
// // Create a RemoteBlock list from blockset_0
// let _blocks_1 = kvbm_1.host().unwrap().allocate_blocks(4).await.unwrap();
// let mut _remote_blocks_0 = kvbm_1.get_remote_blocks_mutable(&blockset_0).unwrap();
// TODO(#967) - Enable with TransferEngine
// // Create a TransferRequestPut for the mutable blocks
// let transfer_request = TransferRequestPut::new(&blocks_0, &mut remote_blocks_0).unwrap();
// // Validate blocks - this could be an expensive operation
// // TODO: Create an ENV trigger debug flag which will call this on every transfer request
// // In this case, we expect an error because we have overlapping blocks as we are sending to/from the same blocks
// // because we are using the wrong target (artifact of the test setup allowing variable to cross what woudl be
// // worker boundaries)
// assert!(transfer_request.validate_blocks().is_err());
// // This is proper request - PUT from worker 1 (local) to worker 0 (remote)
// let transfer_request = TransferRequestPut::new(&blocks_1, &mut remote_blocks_0).unwrap();
// assert!(transfer_request.validate_blocks().is_ok());
// // Execute the transfer request
// transfer_request.execute().unwrap();
// let mut put_request = PutRequestBuilder::<_, _>::builder();
// put_request.from(&blocks_1).to(&mut remote_blocks_0);
// // Create a Put request direct between two local blocks
// // split the blocks into two vecs each with 2 blocks
// let mut blocks_1 = blocks_1;
// let slice_0 = blocks_1.split_off(2);
// let mut slice_1 = blocks_1;
// let transfer_request = TransferRequestPut::new(&slice_0, &mut slice_1).unwrap();
// assert!(transfer_request.validate_blocks().is_ok());
// // Execute the transfer request
// transfer_request.execute().unwrap();
}
#[tokio::test]
async fn test_offload() -> Result<()> {
dynamo_runtime::logging::init();
let block_manager = create_reference_block_manager().await;
let device = block_manager.device().unwrap();
let tokens = Tokens::from(vec![1, 2, 3, 4]);
let token_sequence = tokens.into_sequence(4, Some(0));
let token_block = token_sequence.blocks().first().unwrap();
let mut device_block = device.allocate_blocks(1).await?.into_iter().next().unwrap();
device_block.apply_token_block(token_block.clone())?;
let immutable_device_blocks = device.register_blocks(vec![device_block]).await.unwrap();
assert_eq!(immutable_device_blocks.len(), 1);
tokio::time::sleep(std::time::Duration::from_millis(100)).await;
// It should now be on host and disk.
let host_blocks = block_manager
.host()
.unwrap()
.match_sequence_hashes(vec![immutable_device_blocks[0].sequence_hash()].as_slice())
.await
.unwrap();
assert_eq!(host_blocks.len(), 1);
let disk_blocks = block_manager
.disk()
.unwrap()
.match_sequence_hashes(vec![immutable_device_blocks[0].sequence_hash()].as_slice())
.await
.unwrap();
assert_eq!(disk_blocks.len(), 1);
Ok(())
}
}