block_manager.rs 10.4 KB
Newer Older
1
// SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
2
3
4
// SPDX-License-Identifier: Apache-2.0

use super::*;
5
use anyhow::Result;
Ryan Olson's avatar
Ryan Olson committed
6
7
8
9
use dynamo_llm::block_manager::block::{
    data::logical::distributed_leader_worker::DistributedLeaderWorkerResources, locality::Logical,
};
use dynamo_llm::block_manager::{BasicMetadata, BlockParallelismStrategy};
10
use pyo3::PyResult;
Ryan Olson's avatar
Ryan Olson committed
11
12
13
14
use tokio_util::sync::CancellationToken;

mod controller;
mod distributed;
15

Ryan Olson's avatar
Ryan Olson committed
16
pub mod vllm;
17
18
19
20

/// Add bingings from this crate to the provided module
pub fn add_to_module(m: &Bound<'_, PyModule>) -> PyResult<()> {
    m.add_class::<BlockManager>()?;
Ryan Olson's avatar
Ryan Olson committed
21
22
23
24
25
26
27
28
    m.add_class::<distributed::KvbmWorker>()?;
    m.add_class::<distributed::KvbmLeader>()?;
    m.add_class::<controller::BlockManagerClient>()?;
    m.add_class::<controller::BlockPoolStatus>()?;
    m.add_class::<controller::ResetBlocksResponse>()?;

    vllm::add_to_module(m)?;

29
30
31
    Ok(())
}

Ryan Olson's avatar
Ryan Olson committed
32
33
34
35
36
37
38
39
40
41
42
43
type VllmBlockManager = dynamo_llm::block_manager::KvBlockManager<
    Logical<DistributedLeaderWorkerResources>,
    BasicMetadata,
>;

type VllmController = Arc<
    dynamo_llm::block_manager::controller::Controller<
        Logical<DistributedLeaderWorkerResources>,
        BasicMetadata,
    >,
>;

44
#[pyclass]
Ryan Olson's avatar
Ryan Olson committed
45
#[derive(Clone)]
46
pub struct BlockManager {
Ryan Olson's avatar
Ryan Olson committed
47
48
49
    inner: VllmBlockManager,
    drt: DistributedRuntime,
    _controller: Option<VllmController>,
50
51
}

Ryan Olson's avatar
Ryan Olson committed
52
// TODO: This is in desperate need of a massive refactor. We bind and instantiate this in Python, but we never actually use it.
53
#[pymethods]
Ryan Olson's avatar
Ryan Olson committed
54
#[allow(unused_variables)]
55
56
impl BlockManager {
    #[new]
Ryan Olson's avatar
Ryan Olson committed
57
    #[pyo3(signature = (worker_id, leader = None, page_size = 32, num_device_blocks = None, disable_device_pool = false))]
58
59
    fn new(
        worker_id: u64,
Ryan Olson's avatar
Ryan Olson committed
60
        leader: Option<distributed::KvbmLeader>,
61
        page_size: usize,
Ryan Olson's avatar
Ryan Olson committed
62
63
        num_device_blocks: Option<usize>,
        disable_device_pool: bool,
64
    ) -> PyResult<Self> {
Ryan Olson's avatar
Ryan Olson committed
65
        let cancel_token = CancellationToken::new();
66
67
68
        let mut config = dynamo_llm::block_manager::KvBlockManagerConfig::builder().runtime(
            dynamo_llm::block_manager::KvManagerRuntimeConfig::builder()
                .worker_id(worker_id)
Ryan Olson's avatar
Ryan Olson committed
69
                .cancellation_token(cancel_token.clone())
70
                .build()
71
                .map_err(to_pyerr)?,
72
        );
Ryan Olson's avatar
Ryan Olson committed
73
74
75
76

        let model_config = dynamo_llm::block_manager::KvManagerModelConfig::builder()
            .num_layers(1)
            .outer_dim(1)
77
            .page_size(page_size)
Ryan Olson's avatar
Ryan Olson committed
78
79
            .inner_dim(1);

80
        config = config.model(model_config.build().map_err(to_pyerr)?);
Ryan Olson's avatar
Ryan Olson committed
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

        let (leader, drt) = if let Some(leader) = leader {
            let (leader, rt) = leader.dissolve();

            if !disable_device_pool {
                config = config.device_layout(
                    dynamo_llm::block_manager::KvManagerLayoutConfig::builder()
                        .num_blocks(leader.num_device_blocks())
                        .logical(Some(BlockParallelismStrategy::LeaderWorkerSharded))
                        .build()
                        .map_err(to_pyerr)?,
                );
            }

            if leader.num_host_blocks() > 0 {
                tracing::info!("Using {} host blocks", leader.num_host_blocks());
                config = config.host_layout(
                    dynamo_llm::block_manager::KvManagerLayoutConfig::builder()
                        .num_blocks(leader.num_host_blocks())
                        .logical(Some(BlockParallelismStrategy::LeaderWorkerSharded))
                        .build()
                        .map_err(to_pyerr)?,
                );
            }

            if leader.num_disk_blocks() > 0 {
                tracing::info!("Using {} disk blocks", leader.num_disk_blocks());
                config = config.disk_layout(
                    dynamo_llm::block_manager::KvManagerLayoutConfig::builder()
                        .num_blocks(leader.num_disk_blocks())
                        .logical(Some(BlockParallelismStrategy::LeaderWorkerSharded))
                        .build()
                        .map_err(to_pyerr)?,
                );
            }
            (Some(leader), rt)
        } else {
            tracing::info!("Leader not provided. Block transfer functionality will be disabled.");

            // let num_device_blocks = num_device_blocks
            //     .expect("num_device_blocks must be provided if leader is not provided");

            // config = config.device_layout(
            //     dynamo_llm::block_manager::KvManagerLayoutConfig::builder()
            //         .num_blocks(num_device_blocks)
            //         .logical(Some(BlockParallelismStrategy::LeaderWorkerSharded))
            //         .build()
            //         .map_err(to_pyerr)?,
            // );

            unimplemented!("Leader not provided");
            // (
            //     None,
            //     Arc::new(
            //         tokio::runtime::Builder::new_multi_thread()
            //             .enable_all()
            //             .build()
            //             .map_err(to_pyerr)?,
            //     ),
            // )
        };

        let rt = drt.inner().runtime().primary();

145
        let config = config.build().map_err(to_pyerr)?;
146
        Ok(BlockManager {
Ryan Olson's avatar
Ryan Olson committed
147
148
149
150
151
152
153
154
155
156
157
158
159
160
            inner: rt
                .block_on(async {
                    let resources =
                        DistributedLeaderWorkerResources::new(leader, cancel_token.child_token())?;

                    dynamo_llm::block_manager::KvBlockManager::<
                        Logical<DistributedLeaderWorkerResources>,
                        BasicMetadata,
                    >::new(config, resources)
                    .await
                })
                .map_err(to_pyerr)?,
            drt,
            _controller: None,
161
162
163
        })
    }

Ryan Olson's avatar
Ryan Olson committed
164
165
    fn block_size(&self) -> usize {
        self.inner.block_size()
166
167
    }

Ryan Olson's avatar
Ryan Olson committed
168
169
170
171
172
    fn init_controller(&mut self, component: Component) -> PyResult<()> {
        if self._controller.is_some() {
            tracing::warn!("Controller already initialized. Ignoring init_controller call.");
            return Ok(());
        }
173

Ryan Olson's avatar
Ryan Olson committed
174
175
176
177
178
179
180
181
182
183
        let block_manager = self.inner.clone();
        let controller = self
            .drt
            .inner()
            .runtime()
            .primary()
            .block_on(controller::Controller::new(
                block_manager,
                component.inner.clone(),
            ))
184
            .map_err(to_pyerr)?;
Ryan Olson's avatar
Ryan Olson committed
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202

        self._controller = Some(Arc::new(controller));

        let instance_id = component
            .inner
            .drt()
            .primary_lease()
            .map(|lease| lease.id())
            .ok_or_else(|| to_pyerr(anyhow::anyhow!("no instance id")))?;

        tracing::info!(
            "Dynamo KVBM Controller: {}.{}:{}",
            component.inner.namespace().name(),
            component.inner.name(),
            instance_id
        );

        Ok(())
203
    }
Ryan Olson's avatar
Ryan Olson committed
204
}
205

Ryan Olson's avatar
Ryan Olson committed
206
207
208
209
impl BlockManager {
    #[inline(always)]
    pub fn get_block_manager(&self) -> &VllmBlockManager {
        &self.inner
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

#[derive(Default)]
pub struct BlockManagerBuilder {
    worker_id: u64,
    leader: Option<distributed::KvbmLeader>,
    page_size: usize,
    disable_device_pool: bool,
}

impl BlockManagerBuilder {
    pub fn new() -> Self {
        Self {
            page_size: 32, // default consistent with BlockManager::new
            ..Default::default()
        }
    }

    pub fn worker_id(mut self, id: u64) -> Self {
        self.worker_id = id;
        self
    }
    pub fn page_size(mut self, ps: usize) -> Self {
        self.page_size = ps;
        self
    }
    pub fn leader(mut self, l: distributed::KvbmLeader) -> Self {
        self.leader = Some(l);
        self
    }
    pub fn disable_device_pool(mut self, yes: bool) -> Self {
        self.disable_device_pool = yes;
        self
    }

    /// Async build (call from an async context).
    pub async fn build(self) -> Result<BlockManager> {
        let worker_id = self.worker_id;
        let leader = self.leader.ok_or_else(|| {
            anyhow::anyhow!("leader is required (runtime is always taken from leader)")
        })?;

        // Get (inner leader handle, runtime) from the provided leader.
        let (leader_inner, drt) = leader.dissolve();

        let cancel_token = CancellationToken::new();

        // Runtime & model config
        let runtime_config = dynamo_llm::block_manager::KvManagerRuntimeConfig::builder()
            .worker_id(worker_id)
            .cancellation_token(cancel_token.clone())
            .build()?;

        let mut config =
            dynamo_llm::block_manager::KvBlockManagerConfig::builder().runtime(runtime_config);

        let model_config = dynamo_llm::block_manager::KvManagerModelConfig::builder()
            .num_layers(1)
            .outer_dim(1)
            .page_size(self.page_size)
            .inner_dim(1)
            .build()?;

        config = config.model(model_config);

        // Layouts derived from leader’s counts
        if !self.disable_device_pool {
            config = config.device_layout(
                dynamo_llm::block_manager::KvManagerLayoutConfig::builder()
                    .num_blocks(leader_inner.num_device_blocks())
                    .logical(Some(BlockParallelismStrategy::LeaderWorkerSharded))
                    .build()?,
            );
        }

        if leader_inner.num_host_blocks() > 0 {
            config = config.host_layout(
                dynamo_llm::block_manager::KvManagerLayoutConfig::builder()
                    .num_blocks(leader_inner.num_host_blocks())
                    .logical(Some(BlockParallelismStrategy::LeaderWorkerSharded))
                    .build()?,
            );
        }

        if leader_inner.num_disk_blocks() > 0 {
            config = config.disk_layout(
                dynamo_llm::block_manager::KvManagerLayoutConfig::builder()
                    .num_blocks(leader_inner.num_disk_blocks())
                    .logical(Some(BlockParallelismStrategy::LeaderWorkerSharded))
                    .build()?,
            );
        }

        let config = config.build()?;

        let resources =
            DistributedLeaderWorkerResources::new(Some(leader_inner), cancel_token.child_token())?;

        let inner = dynamo_llm::block_manager::KvBlockManager::<
            Logical<DistributedLeaderWorkerResources>,
            BasicMetadata,
        >::new(config, resources)
        .await?;

        Ok(BlockManager {
            inner,
            drt,
            _controller: None,
        })
    }
}