state.rs 6.65 KB
Newer Older
1
2
3
// SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
// SPDX-License-Identifier: Apache-2.0

4
5
use anyhow::{Result, anyhow, bail};

6
7
8
9
use crate::common::protocols::DirectRequest;
use crate::common::protocols::MockEngineArgs;
use crate::replay::TraceCollector;
use crate::scheduler::{EngineCore, EnginePassResult};
10
11
12
13
14
15
16
17
18
19
use uuid::Uuid;

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub(crate) enum AggRequestPhase {
    QueuedAtRouter,
    Running,
}

pub(crate) struct AggRequestState {
    request: Option<DirectRequest>,
20
21
    pub(in crate::replay::offline) phase: AggRequestPhase,
    pub(in crate::replay::offline) prefill_completed: bool,
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
}

impl AggRequestState {
    pub(crate) fn new_queued(request: DirectRequest) -> Self {
        Self {
            request: Some(request),
            phase: AggRequestPhase::QueuedAtRouter,
            prefill_completed: false,
        }
    }

    pub(crate) fn new_running() -> Self {
        Self {
            request: None,
            phase: AggRequestPhase::Running,
            prefill_completed: false,
        }
    }

    pub(crate) fn take_queued_request(&mut self, uuid: Uuid) -> Result<DirectRequest> {
42
        if self.phase != AggRequestPhase::QueuedAtRouter {
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
            bail!("offline replay expected queued request state for {uuid}");
        }
        let request = self
            .request
            .take()
            .ok_or_else(|| anyhow!("offline replay missing queued request payload for {uuid}"))?;
        self.phase = AggRequestPhase::Running;
        Ok(request)
    }
}

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub(crate) enum DisaggPhase {
    QueuedPrefill,
    RunningPrefill,
    QueuedDecode,
    RunningDecode,
    Done,
}

pub(crate) struct DisaggRequestState {
    original: Option<DirectRequest>,
    #[cfg(test)]
    arrival_ms: f64,
67
    pub(in crate::replay::offline) phase: DisaggPhase,
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
    prefill_worker_idx: Option<usize>,
    decode_worker_idx: Option<usize>,
}

#[cfg(test)]
#[derive(Debug, Clone, PartialEq)]
pub(crate) struct DisaggRequestSnapshot {
    pub(crate) arrival_ms: f64,
    pub(crate) phase: DisaggPhase,
    pub(crate) prefill_worker_idx: Option<usize>,
    pub(crate) decode_worker_idx: Option<usize>,
}

impl DisaggRequestState {
    pub(crate) fn new(request: DirectRequest, arrival_ms: f64) -> Self {
        #[cfg(not(test))]
        let _ = arrival_ms;
        Self {
            original: Some(request),
            #[cfg(test)]
            arrival_ms,
            phase: DisaggPhase::QueuedPrefill,
            prefill_worker_idx: None,
            decode_worker_idx: None,
        }
    }

    pub(crate) fn original_request(&self) -> Result<&DirectRequest> {
        self.original
            .as_ref()
            .ok_or_else(|| anyhow!("offline disagg replay request payload was already released"))
    }

    pub(crate) fn build_prefill_request(&self) -> Result<DirectRequest> {
        let mut request = self.original_request()?.clone();
        request.max_output_tokens = 1;
        Ok(request)
    }

    pub(crate) fn start_prefill(&mut self, worker_idx: usize) {
        self.phase = DisaggPhase::RunningPrefill;
        self.prefill_worker_idx = Some(worker_idx);
    }

    pub(crate) fn queue_decode(&mut self) {
        self.phase = DisaggPhase::QueuedDecode;
    }

    pub(crate) fn start_decode(&mut self, worker_idx: usize) {
        self.phase = DisaggPhase::RunningDecode;
        self.decode_worker_idx = Some(worker_idx);
    }

    pub(crate) fn mark_done(&mut self) {
        self.phase = DisaggPhase::Done;
        self.original = None;
    }

    #[cfg(test)]
    pub(crate) fn debug_snapshot(&self) -> DisaggRequestSnapshot {
        DisaggRequestSnapshot {
            arrival_ms: self.arrival_ms,
            phase: self.phase,
            prefill_worker_idx: self.prefill_worker_idx,
            decode_worker_idx: self.decode_worker_idx,
        }
    }
}
136
137
138
139
140
141
142

pub(crate) struct OfflineWorkerState {
    core: EngineCore,
    busy: bool,
    in_flight: usize,
}

143
144
145
146
147
148
149
150
151
#[cfg(test)]
#[derive(Debug, Clone, PartialEq, Eq)]
pub(crate) struct OfflineWorkerSnapshot {
    pub(crate) busy: bool,
    pub(crate) in_flight: usize,
    pub(crate) ready: bool,
    pub(crate) drained: bool,
}

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
impl OfflineWorkerState {
    pub(crate) fn new(worker_idx: usize, args: MockEngineArgs, capture_kv_events: bool) -> Self {
        let core = match args.engine_type {
            crate::common::protocols::EngineType::Vllm => {
                if capture_kv_events {
                    EngineCore::Vllm(crate::scheduler::VllmCore::new_with_kv_capture(
                        args,
                        worker_idx as u64,
                    ))
                } else {
                    EngineCore::Vllm(crate::scheduler::VllmCore::new(args))
                }
            }
            crate::common::protocols::EngineType::Sglang => {
                if capture_kv_events {
                    EngineCore::Sglang(crate::scheduler::SglangCore::new_with_kv_capture(
                        args,
                        worker_idx as u64,
                    ))
                } else {
                    EngineCore::Sglang(crate::scheduler::SglangCore::new(args))
                }
            }
        };

        Self {
            core,
            busy: false,
            in_flight: 0,
        }
    }

    pub(crate) fn in_flight(&self) -> usize {
        debug_assert!(self.in_flight >= self.core.num_requests());
        self.in_flight
    }

    pub(crate) fn receive_request(&mut self, request: DirectRequest) {
        self.in_flight += 1;
        self.core.receive(request);
    }

    pub(crate) fn mark_completed(&mut self, completed_requests: usize) {
        self.in_flight = self.in_flight.saturating_sub(completed_requests);
    }

    pub(crate) fn mark_busy(&mut self) {
        self.busy = true;
    }

    pub(crate) fn mark_idle(&mut self) {
        self.busy = false;
    }

    pub(crate) fn is_ready(&self) -> bool {
        !self.busy && !self.core.is_empty()
    }

    pub(crate) fn is_drained(&self) -> bool {
        self.in_flight == 0 && !self.busy && self.core.is_empty()
    }

    pub(crate) fn execute_pass(
        &mut self,
        collector: &mut TraceCollector,
        now_ms: f64,
    ) -> EnginePassResult {
        self.core.execute_pass(collector, now_ms)
    }
221

222
223
224
225
    pub(crate) fn execute_hidden_pass(&mut self, now_ms: f64) -> EnginePassResult {
        self.core.execute_hidden_pass(now_ms)
    }

226
227
228
229
230
231
232
233
234
    #[cfg(test)]
    pub(crate) fn debug_snapshot(&self) -> OfflineWorkerSnapshot {
        OfflineWorkerSnapshot {
            busy: self.busy,
            in_flight: self.in_flight,
            ready: self.is_ready(),
            drained: self.is_drained(),
        }
    }
235
}