Skip to main content

qualia_core_db/inference/lab/
timeline.rs

1//! Decode timeline: host phase counters + optional GPU profile + tok/s.
2
3use std::path::Path;
4use std::time::Instant;
5
6use crate::hardware_passport::measure_decode_proxy_tok_s;
7use crate::llm_bench::{
8    phase_snapshot, reset_phase_metrics, reset_resident_path_counts, resident_path_counts,
9};
10#[cfg(feature = "gpu-runtime")]
11use crate::llm_gpu_profiler::{self, Phase};
12
13pub use etalon_inference_lab::timeline::DecodeTimeline;
14/// Run a short decode with phase counters; enable GPU timestamps if supported.
15pub fn run_decode_timeline(model: &Path, tokens: u32) -> DecodeTimeline {
16    let _state = super::state::TrialState::capture();
17    let tokens = tokens.max(1).min(64);
18    reset_phase_metrics();
19    reset_resident_path_counts();
20    #[cfg(feature = "gpu-runtime")]
21    {
22        llm_gpu_profiler::set_enabled(true);
23        llm_gpu_profiler::reset();
24    }
25
26    let t0 = Instant::now();
27    let tok_s = measure_decode_proxy_tok_s(model, tokens);
28    let wall_ms = t0.elapsed().as_secs_f64() * 1e3;
29    let (hits, falls) = resident_path_counts();
30    let snap = phase_snapshot();
31    let host_phase_ns = format!(
32        "{{\"load_ns\":{},\"prefill_ns\":{},\"prefill_tokens\":{},\"decode_ns\":{},\"decode_tokens\":{},\"decode_forward_ns\":{},\"decode_output_ns\":{}}}",
33        snap.load_ns,
34        snap.prefill_ns,
35        snap.prefill_tokens,
36        snap.decode_ns,
37        snap.decode_tokens,
38        snap.decode_forward_ns,
39        snap.decode_output_ns
40    );
41
42    let gpu_phase_ns = {
43        #[cfg(feature = "gpu-runtime")]
44        {
45            let mut gpu_parts = Vec::new();
46            for pt in llm_gpu_profiler::snapshot() {
47                if pt.calls > 0 || pt.total_ns > 0 {
48                    gpu_parts.push(format!(
49                        "\"{}\":{{\"ns\":{},\"calls\":{}}}",
50                        pt.phase.label(),
51                        pt.total_ns,
52                        pt.calls
53                    ));
54                }
55            }
56            format!("{{{}}}", gpu_parts.join(","))
57        }
58        #[cfg(not(feature = "gpu-runtime"))]
59        {
60            "{}".to_string()
61        }
62    };
63
64    let mut notes = Vec::new();
65    #[cfg(feature = "gpu-runtime")]
66    if !llm_gpu_profiler::enabled() {
67        notes.push("GPU timestamps not active (device or flag)".into());
68    }
69    #[cfg(not(feature = "gpu-runtime"))]
70    {
71        notes.push("GPU timestamps not compiled (gpu-runtime feature off)".into());
72    }
73    if falls > 0 {
74        notes.push(format!("resident fallbacks={falls}"));
75    }
76    if hits == 0 && tok_s.is_some() {
77        notes.push("no resident hits counted - check path".into());
78    }
79    #[cfg(feature = "gpu-runtime")]
80    let _ = Phase::COUNT;
81
82    #[cfg(feature = "gpu-runtime")]
83    llm_gpu_profiler::set_enabled(false);
84
85    DecodeTimeline {
86        model: model.display().to_string(),
87        tokens,
88        tok_s,
89        wall_ms,
90        resident_hits: hits,
91        resident_fallbacks: falls,
92        host_phase_ns,
93        gpu_phase_ns,
94        notes: notes.join("; "),
95    }
96}