client.rs 10 KB
Newer Older
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
1
/// Single shard Client
Nicolas Patry's avatar
Nicolas Patry committed
2
3
use crate::pb::generate::v2::text_generation_service_client::TextGenerationServiceClient;
use crate::pb::generate::v2::*;
4
5
6
use crate::{Chunk, Result};
use base64::engine::general_purpose::STANDARD;
use base64::Engine;
7
use grpc_metadata::InjectTelemetryContext;
8
use std::cmp::min;
9
use std::time::Duration;
Olivier Dehaene's avatar
Init  
Olivier Dehaene committed
10
use tonic::transport::{Channel, Uri};
11
use tracing::instrument;
Olivier Dehaene's avatar
Init  
Olivier Dehaene committed
12

13
14
static WARMUP_IMAGE_BASE64 :&str = "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";

Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
15
/// Text Generation Inference gRPC client
16
#[derive(Debug, Clone)]
Olivier Dehaene's avatar
Init  
Olivier Dehaene committed
17
pub struct Client {
Olivier Dehaene's avatar
Olivier Dehaene committed
18
    stub: TextGenerationServiceClient<Channel>,
Olivier Dehaene's avatar
Init  
Olivier Dehaene committed
19
20
21
}

impl Client {
Olivier Dehaene's avatar
Olivier Dehaene committed
22
23
24
    /// Returns a client connected to the given url
    pub async fn connect(uri: Uri) -> Result<Self> {
        let channel = Channel::builder(uri).connect().await?;
Olivier Dehaene's avatar
Init  
Olivier Dehaene committed
25

Olivier Dehaene's avatar
Olivier Dehaene committed
26
27
28
        Ok(Self {
            stub: TextGenerationServiceClient::new(channel),
        })
Olivier Dehaene's avatar
Init  
Olivier Dehaene committed
29
30
    }

Olivier Dehaene's avatar
Olivier Dehaene committed
31
32
    /// Returns a client connected to the given unix socket
    pub async fn connect_uds(path: String) -> Result<Self> {
Olivier Dehaene's avatar
Olivier Dehaene committed
33
        let channel = Channel::from_shared("http://[::]:50051".to_string())
Olivier Dehaene's avatar
Init  
Olivier Dehaene committed
34
35
36
37
            .unwrap()
            .connect_with_connector(tower::service_fn(move |_: Uri| {
                tokio::net::UnixStream::connect(path.clone())
            }))
Olivier Dehaene's avatar
Olivier Dehaene committed
38
            .await?;
Olivier Dehaene's avatar
Init  
Olivier Dehaene committed
39

Olivier Dehaene's avatar
Olivier Dehaene committed
40
41
42
        Ok(Self {
            stub: TextGenerationServiceClient::new(channel),
        })
Olivier Dehaene's avatar
Init  
Olivier Dehaene committed
43
44
    }

Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
45
    /// Returns a list of uris or unix sockets of all shards
Olivier Dehaene's avatar
Init  
Olivier Dehaene committed
46
47
    #[instrument(skip(self))]
    pub async fn service_discovery(&mut self) -> Result<Vec<String>> {
48
49
        let request = tonic::Request::new(ServiceDiscoveryRequest {}).inject_context();
        let response = self.stub.service_discovery(request).await?;
Olivier Dehaene's avatar
Init  
Olivier Dehaene committed
50
51
52
53
        let urls = response
            .into_inner()
            .urls
            .into_iter()
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
54
            // Remove unix socket prefix
Olivier Dehaene's avatar
Init  
Olivier Dehaene committed
55
56
57
58
59
60
61
62
            .map(|url| match url.strip_prefix("unix://") {
                None => url,
                Some(stripped_url) => stripped_url.to_string(),
            })
            .collect();
        Ok(urls)
    }

63
64
65
66
67
68
69
70
    /// Get model info
    #[instrument(skip(self))]
    pub async fn info(&mut self) -> Result<InfoResponse> {
        let request = tonic::Request::new(InfoRequest {}).inject_context();
        let response = self.stub.info(request).await?.into_inner();
        Ok(response)
    }

71
72
73
74
75
76
77
78
    /// Get model health
    #[instrument(skip(self))]
    pub async fn health(&mut self) -> Result<HealthResponse> {
        let request = tonic::Request::new(HealthRequest {}).inject_context();
        let response = self.stub.health(request).await?.into_inner();
        Ok(response)
    }

Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
79
    /// Clear the past generations cache
Olivier Dehaene's avatar
Init  
Olivier Dehaene committed
80
    #[instrument(skip(self))]
81
82
    pub async fn clear_cache(&mut self, batch_id: Option<u64>) -> Result<()> {
        let request = tonic::Request::new(ClearCacheRequest { id: batch_id }).inject_context();
83
        self.stub.clear_cache(request).await?;
Olivier Dehaene's avatar
Init  
Olivier Dehaene committed
84
85
86
        Ok(())
    }

87
88
89
90
91
    /// Filter a cached batch
    #[instrument(skip(self))]
    pub async fn filter_batch(
        &mut self,
        batch_id: u64,
92
93
        request_ids: Vec<u64>,
    ) -> Result<Option<CachedBatch>> {
94
95
        let request = tonic::Request::new(FilterBatchRequest {
            batch_id,
96
            request_ids,
97
98
99
100
101
102
        })
        .inject_context();
        let filtered_batch = self.stub.filter_batch(request).await?.into_inner();
        Ok(filtered_batch.batch)
    }

103
104
105
    /// Warmup on a max size batch
    ///
    /// Returns the maximum amount of tokens supported by the hardware
106
    #[instrument(skip_all)]
107
108
109
110
    pub async fn warmup(
        &mut self,
        max_input_length: u32,
        max_prefill_tokens: u32,
OlivierDehaene's avatar
OlivierDehaene committed
111
        max_total_tokens: u32,
112
        max_batch_size: Option<usize>,
113
    ) -> Result<Option<u32>> {
114
115
116
117
        let mut n_tokens = 0;
        let mut requests = Vec::new();
        // Create requests
        while n_tokens < max_prefill_tokens {
OlivierDehaene's avatar
OlivierDehaene committed
118
            let truncate = min(max_input_length, max_prefill_tokens - n_tokens);
119

120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
            let mut input_chunks = Vec::new();
            input_chunks
                .push(Chunk::Text("_test ".to_string().repeat(max_input_length as usize)).into());
            if n_tokens == 0 {
                input_chunks.push(
                    Chunk::Image(Image {
                        // Safe unwrap, because we control the data.
                        data: STANDARD.decode(WARMUP_IMAGE_BASE64).unwrap(),
                        mimetype: "image/jpeg;base64".to_string(),
                    })
                    .into(),
                );
            }

            // Send stringly-typed inputs for compatibility for backends that haven't
            // been updated to support chunks.
136
137
            let mut inputs = String::new();
            inputs.push_str(&"_test ".to_string().repeat(max_input_length as usize));
Nicolas Patry's avatar
Nicolas Patry committed
138
139
140
            if n_tokens == 0 {
                // 1 request is enough to test vision heads.
                // Sending images on other queries messes up easily with truncation.
141
142
143
                inputs.push_str(&format!(
                    "![](data:image/jpeg;base64,{WARMUP_IMAGE_BASE64})",
                ));
Nicolas Patry's avatar
Nicolas Patry committed
144
            }
145

146
147
            requests.push(Request {
                id: 0,
148
149
150
                input_chunks: Some(Input {
                    chunks: input_chunks,
                }),
151
                inputs,
152
                // We truncate the input on the server side to be sure that it has the correct size
OlivierDehaene's avatar
OlivierDehaene committed
153
                truncate,
154
155
156
157
158
159
160
161
162
                // Set sampling parameters to also take these ops into account in the max memory
                parameters: Some(NextTokenChooserParameters {
                    temperature: 0.9,
                    top_k: 10,
                    top_p: 0.9,
                    typical_p: 0.9,
                    do_sample: false,
                    seed: 0,
                    repetition_penalty: 1.2,
163
                    frequency_penalty: 0.1,
164
                    watermark: true,
drbh's avatar
drbh committed
165
166
                    grammar: String::new(),
                    grammar_type: GrammarType::None as i32,
167
168
                }),
                stopping_parameters: Some(StoppingCriteriaParameters {
OlivierDehaene's avatar
OlivierDehaene committed
169
                    max_new_tokens: max_total_tokens - truncate,
170
                    stop_sequences: vec![],
OlivierDehaene's avatar
OlivierDehaene committed
171
                    ignore_eos_token: true,
172
173
                }),
                prefill_logprobs: true,
Nicolas Patry's avatar
Nicolas Patry committed
174
                top_n_tokens: 20,
175
176
            });
            n_tokens += max_input_length;
177
178
179
180
181

            // Check max_batch_size
            if Some(requests.len()) == max_batch_size {
                break;
            }
182
183
184
185
186
187
188
189
190
        }

        let batch = Batch {
            id: 0,
            size: requests.len() as u32,
            requests,
            max_tokens: 0,
        };

191
192
193
194
195
196
197
        let request = tonic::Request::new(WarmupRequest {
            batch: Some(batch),
            max_input_length,
            max_prefill_tokens,
            max_total_tokens,
        })
        .inject_context();
198
199
        let response = self.stub.warmup(request).await?.into_inner();
        Ok(response.max_supported_total_tokens)
200
201
    }

Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
202
203
    /// Generate one token for each request in the given batch
    ///
204
    /// Returns Generation for each request in batch
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
205
    /// and the next cached batch
206
    #[instrument(skip_all, fields(id = &batch.id, size = &batch.size))]
207
208
209
    pub async fn prefill(
        &mut self,
        batch: Batch,
210
    ) -> Result<(Vec<Generation>, Option<CachedBatch>, PrefillTimings)> {
211
212
        let request = tonic::Request::new(PrefillRequest { batch: Some(batch) }).inject_context();
        let response = self.stub.prefill(request).await?.into_inner();
213
214
215
216
217
        Ok((
            response.generations,
            response.batch,
            PrefillTimings::new(response.forward_ns, response.decode_ns, response.total_ns),
        ))
Olivier Dehaene's avatar
Init  
Olivier Dehaene committed
218
219
    }

220
    /// Generate one token for each request in the given cached batches
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
221
    ///
222
    /// Returns Generation for each request in batches
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
223
    /// and the next cached batch
224
    #[instrument(skip_all, fields(size = batches.iter().map(|batch|{batch.size}).sum::<u32>()))]
225
    pub async fn decode(
Olivier Dehaene's avatar
Init  
Olivier Dehaene committed
226
        &mut self,
227
        batches: Vec<CachedBatch>,
228
    ) -> Result<(Vec<Generation>, Option<CachedBatch>, DecodeTimings)> {
229
230
        let request = tonic::Request::new(DecodeRequest { batches }).inject_context();
        let response = self.stub.decode(request).await?.into_inner();
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
        Ok((
            response.generations,
            response.batch,
            DecodeTimings::new(
                response.concat_ns,
                response.forward_ns,
                response.decode_ns,
                response.total_ns,
            ),
        ))
    }
}

pub struct PrefillTimings {
    pub forward: Duration,
    pub decode: Duration,
    pub total: Duration,
}

impl PrefillTimings {
    fn new(forward_ns: u64, decode_ns: u64, total_ns: u64) -> Self {
        Self {
            forward: Duration::from_nanos(forward_ns),
            decode: Duration::from_nanos(decode_ns),
            total: Duration::from_nanos(total_ns),
        }
    }
}

pub struct DecodeTimings {
    pub concat: Option<Duration>,
    pub forward: Duration,
    pub decode: Duration,
    pub total: Duration,
}

impl DecodeTimings {
    fn new(concat_ns: Option<u64>, forward_ns: u64, decode_ns: u64, total_ns: u64) -> Self {
        Self {
270
            concat: concat_ns.map(Duration::from_nanos),
271
272
273
274
            forward: Duration::from_nanos(forward_ns),
            decode: Duration::from_nanos(decode_ns),
            total: Duration::from_nanos(total_ns),
        }
Olivier Dehaene's avatar
Olivier Dehaene committed
275
    }
Olivier Dehaene's avatar
Init  
Olivier Dehaene committed
276
}