grpc.rs 8.26 KB
Newer Older
GuanLuo's avatar
GuanLuo committed
1
2
3
4
5
6
// SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
// SPDX-License-Identifier: Apache-2.0

use std::sync::Arc;

use crate::{
7
    discovery::{ModelManager, ModelWatcher},
GuanLuo's avatar
GuanLuo committed
8
9
10
11
    engines::StreamingEngineAdapter,
    entrypoint::{self, EngineConfig, input::common},
    grpc::service::kserve,
    kv_router::KvRouterConfig,
12
    model_card,
13
    namespace::is_global_namespace,
GuanLuo's avatar
GuanLuo committed
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
    types::openai::{
        chat_completions::{NvCreateChatCompletionRequest, NvCreateChatCompletionStreamResponse},
        completions::{NvCreateCompletionRequest, NvCreateCompletionResponse},
    },
};
use dynamo_runtime::transports::etcd;
use dynamo_runtime::{DistributedRuntime, Runtime};
use dynamo_runtime::{distributed::DistributedConfig, pipeline::RouterMode};

/// Build and run an KServe gRPC service
pub async fn run(runtime: Runtime, engine_config: EngineConfig) -> anyhow::Result<()> {
    let mut grpc_service_builder = kserve::KserveService::builder()
        .port(engine_config.local_model().http_port()) // [WIP] generalize port..
        .with_request_template(engine_config.local_model().request_template());

    let grpc_service = match engine_config {
        EngineConfig::Dynamic(_) => {
            let distributed_runtime = DistributedRuntime::from_settings(runtime.clone()).await?;
            let etcd_client = distributed_runtime.etcd_client();
            // This allows the /health endpoint to query etcd for active instances
            grpc_service_builder = grpc_service_builder.with_etcd_client(etcd_client.clone());
            let grpc_service = grpc_service_builder.build()?;
            match etcd_client {
                Some(ref etcd_client) => {
                    let router_config = engine_config.local_model().router_config();
                    // Listen for models registering themselves in etcd, add them to gRPC service
40
41
42
43
44
45
                    let namespace = engine_config.local_model().namespace().unwrap_or("");
                    let target_namespace = if is_global_namespace(namespace) {
                        None
                    } else {
                        Some(namespace.to_string())
                    };
GuanLuo's avatar
GuanLuo committed
46
47
48
49
                    run_watcher(
                        distributed_runtime,
                        grpc_service.state().manager_clone(),
                        etcd_client.clone(),
50
                        model_card::ROOT_PATH,
GuanLuo's avatar
GuanLuo committed
51
52
53
                        router_config.router_mode,
                        Some(router_config.kv_router_config),
                        router_config.busy_threshold,
54
                        target_namespace,
GuanLuo's avatar
GuanLuo committed
55
56
57
58
59
60
61
62
63
64
65
                    )
                    .await?;
                }
                None => {
                    // Static endpoints don't need discovery
                }
            }
            grpc_service
        }
        EngineConfig::StaticRemote(local_model) => {
            let card = local_model.card();
66
            let checksum = card.mdcsum();
GuanLuo's avatar
GuanLuo committed
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
            let router_mode = local_model.router_config().router_mode;

            let dst_config = DistributedConfig::from_settings(true); // true means static
            let distributed_runtime = DistributedRuntime::new(runtime.clone(), dst_config).await?;
            let grpc_service = grpc_service_builder.build()?;
            let manager = grpc_service.model_manager();

            let endpoint_id = local_model.endpoint_id();
            let component = distributed_runtime
                .namespace(&endpoint_id.namespace)?
                .component(&endpoint_id.component)?;
            let client = component.endpoint(&endpoint_id.name).client().await?;

            let kv_chooser = if router_mode == RouterMode::KV {
                Some(
                    manager
                        .kv_chooser_for(
                            &component,
                            card.kv_cache_block_size,
                            Some(local_model.router_config().kv_router_config),
                        )
                        .await?,
                )
            } else {
                None
            };

94
            let tokenizer_hf = card.tokenizer_hf()?;
GuanLuo's avatar
GuanLuo committed
95
96
97
            let chat_engine = entrypoint::build_routed_pipeline::<
                NvCreateChatCompletionRequest,
                NvCreateChatCompletionStreamResponse,
98
99
100
101
102
103
104
            >(
                card,
                &client,
                router_mode,
                None,
                kv_chooser.clone(),
                tokenizer_hf.clone(),
105
                None, // No prefill chooser in grpc static mode
106
            )
GuanLuo's avatar
GuanLuo committed
107
            .await?;
108
109
110
111
112
            manager.add_chat_completions_model(
                local_model.display_name(),
                checksum,
                chat_engine,
            )?;
GuanLuo's avatar
GuanLuo committed
113

114
115
116
117
118
119
120
121
122
123
124
125
126
            let completions_engine = entrypoint::build_routed_pipeline::<
                NvCreateCompletionRequest,
                NvCreateCompletionResponse,
            >(
                card,
                &client,
                router_mode,
                None,
                kv_chooser,
                tokenizer_hf,
                None, // No prefill chooser in grpc static mode
            )
            .await?;
127
128
129
130
131
            manager.add_completions_model(
                local_model.display_name(),
                checksum,
                completions_engine,
            )?;
GuanLuo's avatar
GuanLuo committed
132
133
134
135
136
137
138

            grpc_service
        }
        EngineConfig::StaticFull { engine, model, .. } => {
            let grpc_service = grpc_service_builder.build()?;
            let engine = Arc::new(StreamingEngineAdapter::new(engine));
            let manager = grpc_service.model_manager();
139
140
141
            let checksum = model.card().mdcsum();
            manager.add_completions_model(model.service_name(), checksum, engine.clone())?;
            manager.add_chat_completions_model(model.service_name(), checksum, engine)?;
GuanLuo's avatar
GuanLuo committed
142
143
144
145
146
147
148
149
150
            grpc_service
        }
        EngineConfig::StaticCore {
            engine: inner_engine,
            model,
            ..
        } => {
            let grpc_service = grpc_service_builder.build()?;
            let manager = grpc_service.model_manager();
151
            let checksum = model.card().mdcsum();
GuanLuo's avatar
GuanLuo committed
152

153
154
155
156
157
158
159
            let tokenizer_hf = model.card().tokenizer_hf()?;
            let chat_pipeline =
                common::build_pipeline::<
                    NvCreateChatCompletionRequest,
                    NvCreateChatCompletionStreamResponse,
                >(model.card(), inner_engine.clone(), tokenizer_hf.clone())
                .await?;
160
            manager.add_chat_completions_model(model.service_name(), checksum, chat_pipeline)?;
GuanLuo's avatar
GuanLuo committed
161
162
163
164

            let cmpl_pipeline = common::build_pipeline::<
                NvCreateCompletionRequest,
                NvCreateCompletionResponse,
165
            >(model.card(), inner_engine, tokenizer_hf)
GuanLuo's avatar
GuanLuo committed
166
            .await?;
167
            manager.add_completions_model(model.service_name(), checksum, cmpl_pipeline)?;
GuanLuo's avatar
GuanLuo committed
168
169
170
171
172
173
174
175
176
177
            grpc_service
        }
    };
    grpc_service.run(runtime.primary_token()).await?;
    runtime.shutdown(); // Cancel primary token
    Ok(())
}

/// Spawns a task that watches for new models in etcd at network_prefix,
/// and registers them with the ModelManager so that the HTTP service can use them.
178
#[allow(clippy::too_many_arguments)]
GuanLuo's avatar
GuanLuo committed
179
180
181
182
183
184
185
186
async fn run_watcher(
    runtime: DistributedRuntime,
    model_manager: Arc<ModelManager>,
    etcd_client: etcd::Client,
    network_prefix: &str,
    router_mode: RouterMode,
    kv_router_config: Option<KvRouterConfig>,
    busy_threshold: Option<f64>,
187
    target_namespace: Option<String>,
GuanLuo's avatar
GuanLuo committed
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
) -> anyhow::Result<()> {
    let watch_obj = ModelWatcher::new(
        runtime,
        model_manager,
        router_mode,
        kv_router_config,
        busy_threshold,
    );
    tracing::info!("Watching for remote model at {network_prefix}");
    let models_watcher = etcd_client.kv_get_and_watch_prefix(network_prefix).await?;
    let (_prefix, _watcher, receiver) = models_watcher.dissolve();

    // [gluo NOTE] This is different from http::run_watcher where it alters the HTTP service
    // endpoint being exposed, gRPC doesn't have the same concept as the KServe service
    // only has one kind of inference endpoint.

    // Pass the sender to the watcher
    let _watcher_task = tokio::spawn(async move {
206
        watch_obj.watch(receiver, target_namespace.as_deref()).await;
GuanLuo's avatar
GuanLuo committed
207
208
209
210
    });

    Ok(())
}