Expand description
TensorFlow Profiler-style execution graphs for Candle runs.
Record one representative JSONL trace (instrument::TraceSession), build an graph::ExecutionGraph,
and inspect it via CLI or HTML (viewer::render_evidence_html).
There is no static Rust analysis — the graph comes only from what executed.
Re-exports§
pub use evidence::build_evidence;pub use evidence::Comparison;pub use evidence::EvidencePacket;pub use evidence::SpanComparison;pub use graph::build_from_trace;pub use graph::ExecutionGraph;pub use instrument::candle;pub use instrument::MemoryRecord;pub use instrument::OpRecord;pub use instrument::ProfileRun;pub use instrument::SpanGuard;pub use instrument::SpanId;pub use instrument::SpanKind;pub use instrument::TensorRecord;pub use instrument::TraceSession;pub use nsight::GpuEvidenceStatus;pub use nsight::NsightEvidence;pub use phase::ExecutionPhase;pub use phase::ExecutionStep;pub use trace::TimingMode;pub use trace::parse_trace;pub use trace::write_jsonl;pub use trace::TraceDocument;pub use trace::TraceEvent;pub use trace::SCHEMA as TRACE_SCHEMA;
Modules§
- cli
- Trace-file CLI helpers.
- evidence
- One bounded evidence packet for agents, reports, comparisons, and the HTML viewer.
- graph
- Execution graph derived purely from trace documents.
- instrument
- Post-run trace instrumentation — TensorFlow Profiler-style span API.
- nsight
- Stable normalization seam for official
nsys stats --format csvreports. - phase
- Train vs inference execution phase for trace metadata.
- trace
- TensorFlow-Profiler-style trace protocol (
candle-graph/trace/6). - viewer
- Dependency-free interactive HTML visualizer for
candle-graph/viewer/4evidence payloads.