v1.2.1

Fixed

• Models with float16 (fp16) inputs and outputs now produce detections. Previously these models — including every TensorRT engine produced by EdgeFirst Studio with fp16 precision — completed without errors but returned no predictions, leaving predictions.parquet empty. Validated on Jetson platforms running TensorRT; other platforms and runtimes have not been verified yet.

v1.2.0

Added

• The F2 Studio explorer now shows a coloured deployability indicator next to every artifact in a training session. Green means this build of the profiler on this host can run the artifact; yellow means the artifact targets accelerator hardware that can't be confirmed without a runtime probe; red means the artifact targets a different SoC or the matching backend isn't built into this profiler. For yellow and red, a short inline reason explains the diagnosis at a glance — for example [Hailo-8L] (Hailo runtime and accessory not confirmed) or [NXP i.MX 95 Neutron] (TFLite backend not built into this profiler). A short legend appears below the list. The indicator is informational only and does not block selection — you can still attempt to profile any artifact.
• New pipelining guide explaining how the validation pipeline overlaps its six stages so the NPU stays busy while the CPU prepares the next frame, what guarantees the bounded-channel design gives you about latency and memory, and how to read stalls in the F4 Pipeline Stages panel. Linked from the README and the architecture overview.

Changed

• The training session list in the F2 Studio explorer now shows the most recently trained model at the top instead of the bottom. The session you most likely want to validate is the first one highlighted when you open an experiment.
• TensorRT engine-load failures now surface the underlying TensorRT runtime reason — plan-header size mismatch, version mismatch, compute-capability mismatch — instead of the previous generic "deserializeCudaEngine failed" message. Incompatible-artifact failures are now actionable from the profiler's error output alone.

Fixed

• Validation sessions in EdgeFirst Studio no longer show as "complete" while the profiler CLI is still running. Each stage of a newly-created validation session is now correctly displayed as not-yet-started until work for that stage begins, and the post-profiling analysis step stays visible as a separate stage that remains pending until the EdgeFirst Validator app picks it up and runs. If the validator fails to launch after the profiler finishes uploading results, the analysis stage is now marked with an error instead of being left in a misleading "queued" state.
• EdgeFirst Studio validation sessions launched from the F2 Studio tab for ONNX models now record the actual execution provider in the session name. A CUDA run on a Linux x86_64 host produces a session named like …-x86_64-onnx-cuda, a CoreML run on macOS produces …-macos-onnx-coreml, and a plain CPU run produces …-onnx-cpu. Previously every ONNX session ended in …-onnx-cpu regardless of which provider actually ran, which made CPU runs and accelerated runs indistinguishable in Studio's session listings. The execution-provider modal also now opens before the Studio session is created, so the name is correct on the first attempt rather than relying on a rename afterwards.
• TensorRT .engine artifacts produced by EdgeFirst Studio training sessions now load on Jetson targets. Previously every such artifact failed at deserialization, so TensorRT validation on Jetson required pointing the profiler at an .onnx and rebuilding the engine locally.

v1.1.0

Added

• Prebuilt libhailort_profiler.so now builds in CI for both linux-aarch64 (RPi5 / AI Kit) and linux-x86_64 (Hailo-8 PCIe). Each shim is built against HailoRT v4.23.0 compiled from source (no Hailo Developer Zone account required), uploaded as a workflow artifact, then bundled into the matching release archive and wheel via the same glob-based injection used for libtrt_shim.so. The wheel-injection and archive-packaging steps were generalized to iterate over shims/prebuilt/<arch>/*.so so future vendor shims drop in without further CI surgery.
• ProfilingStatus::Failed { error: String, frames_done, frames_total, total_time_ms } variant in src/tui/event.rs. Carries the human-readable failure message, the last-known progress at the moment of failure, and the wall-clock elapsed time captured at the transition (so the dashboard's "Elapsed" stops counting on a dead run). Future framework steps will promote the error field to a typed Diagnostic with structured code/stage/hint fields; the wire-level PROF: protocol is unchanged in this release.
• Strict pre-flight for --provider cuda (strict_cuda_preflight in src/inference/ort_backend.rs). Verifies each CUDA / cuDNN runtime SONAME the EP plugin needs is loadable; on a miss, the run hard-errors with the exact pip install nvidia-*-cu12 ... command instead of silently running on CPU. The check runs only when CUDA is the selected provider, so CPU users on partial setups never see CUDA-specific warnings. Paired with error_on_failure() on the CUDA EP dispatch so ort propagates registration failures instead of swallowing them.
• F4 Profiler completion popup now shows two persistent per-artifact progress bars (predictions.parquet and trace.pftrace) during Studio uploads, with byte/percent labels and pending/active/done/skipped state colouring. Replaces the previous ◌→✓ icon-only step indicators that flashed past too quickly for the user to see on fast uploads; each bar now stays at 100% green once its artifact finishes so the visual trail is preserved. A new artifact_bar_lines helper backed by an ArtifactState enum centralises the rendering; the Skipped variant covers the case where trace.pftrace was not produced but the session still completed (rendered dim with a "skipped (not produced)" label instead of a perpetually spinning ◌). UploadProgress now stores per-artifact byte counters (parquet_current/_total, trace_current/_total) that are fed by routing BackgroundEvent::DownloadProgress events on phase labels matching tui::event::UPLOAD_LABEL_PREDICTIONS / UPLOAD_LABEL_TRACE.
• F2 Studio validation screen now has an explicit "start new session" affordance. On the Complete / Failed page, Enter resets the phase state machine and returns to the Explorer with a freshly-loaded projects list; Esc resets and returns to the current breadcrumb level. The footer hint reads Press Enter to start a new session • Esc to return to Explorer. Previously the screen had no way out other than quitting and relaunching the binary.

Changed

• TUI provider modal now binds g (for GPU) instead of u for CUDA, with the hint text shortened to "Run with CUDA". m for CoreML and c for CPU are unchanged. ONNX.md updated to match.
• TUI-created Studio validation sessions now carry a platform-aware name and a multi-line description. The name is <model_stem>-<host>-<npu> (e.g. yolov5m-t-27e4_quant-u8-i8_smart-imx95-neutron) — the model artifact's filename has its runtime extension and any .imx95 / .hailo8l style hint stripped, then the runtime-resolved <host>-<npu> tag is appended. The host is detected from /sys/firmware/devicetree/base/compatible (covering imx8mp / imx95 / imx93 / orin-nano / rpi5) on Linux, short-circuits to macos / windows on those targets, and falls back to uname -m for DT-less Unix hosts; the NPU is derived from the model contents (TFLite NeutronGraph blob scan), the filename multi-suffix (Hailo .hailo8l.hef / .hailo8.hef / .hailo10.hef), and a host-side probe of libvx_delegate.so for generic TFLite on i.MX 8M Plus. Filename hints from Studio are treated as user-facing conventions only — the runtime host and the model contents are the source of truth, so a .imx95.tflite running on i.MX 8M Plus correctly resolves to imx8mp-vsi (or imx8mp-cpu), never imx95-neutron. The session description block is 4-7 lines (gracefully degrading on hosts where a probe is unavailable): hostname, kernel Version (sysname + release, no arch suffix), SoC / CPU model + core count, NPU name with a short detection note, total memory, and the profiler version. All Linux-specific probes (/sys/firmware/devicetree/*, /proc/cpuinfo, /proc/meminfo) sit behind cfg(target_os = "linux") and libc::uname sits behind cfg(unix), so the module compiles cleanly on macOS today and will accept a Windows backend without touching public API. New SoMs and accelerators extend two enums (HostTag, NpuTag) with one variant + one match arm each — see src/platform/.
• TUI validation workflow now downloads the model before creating the Studio session so the platform identification has the model bytes in hand. The session is created with the final name and description on the first call instead of being renamed afterwards. The download stage in set_session_stages covers both model + dataset; after the reorder, stages are initialized with the stage's progress label set to "Downloading dataset" to describe the remaining work (the model is already on disk). The stage itself is opened in running and is not explicitly transitioned to complete — same as on main — leaving the existing server-side stage machine to resolve it when a later stage moves to running. The "Downloading model" progress bar is unchanged from the user's perspective — only the Studio-side label and the moment of session creation have shifted. The blocking std::fs::read for the Neutron-blob scan is now tokio::fs::read so large models don't stall the tokio worker thread.
• tui/components/profiler.rs::detect_delegate is now a thin shim over the new platform::delegate_hint_for_host() helper; the inline /sys/firmware/devicetree/base/compatible probe lived in two places (also the configure-panel default) and is now centralized in src/platform/host.rs.

Fixed

• 5000-image validation runs no longer drop the last pipeline_depth frames with PboDisconnected. The HAL ImageProcessor was moved into the preprocess thread closure, so when the preprocess loop exited at end-of-run the processor was dropped, tearing down its GL thread / PBO Sender while the inference thread still held depth-many PBO-backed staging tensors with only WeakSenders. Wrapping the processor in Arc<Mutex<>> and holding one Arc in the orchestrator scope past all thread joins keeps the GL Sender alive until inference fully drains, so every pre_tensor.map() succeeds. Mutex is uncontended in practice (only the preprocess thread ever calls convert()).
• Inference-thread skip log now names the failing file. The decode and preprocess skip sites already logged frame_id and path, but infer.rs only carried the error message — which sent the PboDisconnected regression hunt down the wrong rabbit hole. PreprocessedFrame now carries image_path through from DecodedFrame, and the infer-thread error WARN includes it.
• Pipeline summary splits the skip count by stage: decode=N preprocess=N infer=N instead of the previous lump-sum "decode or processing failure". OrchestratorResult gained decode_skipped / preprocess_skipped / infer_skipped fields (the existing skipped_frames is now their sum).
• yolo26 ONNX exports with model.end2end = false (pre-NMS head) now decode correctly and produce predictions.parquet. edgefirst-hal 0.23's DecoderVersion::is_end_to_end() treats the Yolo26 variant as unconditionally embedded-NMS, so the builder rejected the model with InvalidConfig("Yolo26 Segmentation Detection must have num_features be 6 + num_protos = ...") and the pipeline silently fell into timing-only mode. The profiler now rewrites decoder_version to "yolov8" whenever it sees decoder_version: "yolo26" paired with model.end2end: false — the head shape (4 box + N class + M mask coefs × anchors) is structurally identical to yolov8, so HAL's existing yolov8 DFL path handles it correctly. Compat shim only; reverts to passthrough once HAL gains an explicit end2end field.
• TUI validation sessions no longer stay stuck on the "Running" dashboard when the child pipeline reports a fatal error (e.g. libonnxruntime.so not found). BackgroundEvent::Error previously only set shared.pipeline_error and showed an error popup, leaving ProfilingStatus in Running forever — the dashboard kept counting elapsed time on zero frames and the F2 Studio tab stayed on RunningInference, even though the Studio backend session was correctly errored via report_session_error. A new ProfilingStatus::Failed variant gives the Error event a terminal target; the dispatcher now transitions Running → Failed, clears profiling_active, and bridges the failure to the Studio screen by enqueueing ValidationPhase::Failed on bg_tx. The error popup behaviour is unchanged.
• TUI now also surfaces silent child-exit failures (panic, SIGKILL, or any path that closes the child's stderr without emitting PROF:complete or PROF:error). ingest_prof_stream synthesises a BackgroundEvent::Error("Child process exited without producing a result...") on EOF when no terminal event was seen, so the same Running → Failed transition fires.
• TUI partial-failure runs (run_bench_internal invoke fails mid-loop at src/tui/bench.rs:166-173) now correctly stay in the Failed state. The child emits PROF:error for the failing iteration and then still emits PROF:complete with the partial stats from the iterations that did succeed; pre-fix, the parent's ...

v1.0.3

Fixed

• Validation sessions no longer hang at 100 % inference. The pipeline orchestrator's infer thread held a clone of the buffer-return channel sender for its error path and never dropped it before the post-loop drain — leaving itself as the last live sender so buffer_rx could never disconnect. The thread then blocked on for _buf in buffer_rx {} forever, which in turn blocked the engine hand-back, the thread joins, and the entire publish flow. Symptom was a progress bar pinned at 5000/5000 with no CPU, no memory growth, and no further output. The clone is now dropped immediately after drop(tx) so the drain terminates as soon as the decoder exits.
• ONNX Runtime per-op profiling no longer spams Maximum number of events reached, could not record profile event. partway through long runs. ORT's profiler uses a fixed-capacity 1 M-event ring buffer in upstream core/common/profiler.cc and the buffer is not exposed as a session config option in any released ORT version. A YOLO-class graph emits ~240–500 events per session.run, so a 5 000-frame run overflows the cap around frame 4 100 and loses every subsequent per-op event. OrtEngine now counts invocations and ends profiling early once ORT_PROFILE_INVOCATION_CAP (3 000 frames) is reached; the cached profile path is returned later by the trait end_profiling() call so the post-loop parser is unchanged. Aggregate stage timing (tensor_ref, session_run, output_copy) keeps running for every subsequent frame through the existing StagedTimer.

Changed

• Post-loop ORT profile emission now streams the JSON file event-by-event instead of serde_json::from_reader-ing the whole array into a Vec<OrtTraceEvent>. ORT writes one JSON object per line with comma separators (Chrome-tracing convention), so the new stream_ort_profile parser walks the file line-by-line, buffers nodes until the per-frame SequentialExecutor::Execute envelope event, then yields one OrtFrameProfile and frees the buffer. Combined with stream_emit_ort_trace_events, this drops peak post-loop memory from roughly 200 MB to ~48 KB for a 3 000-frame capped run and eliminates the multi-second pause that used to happen between the last inference and the start of the publish step.
• CLI now renders dataset, model, predictions, and trace transfers as indicatif progress bars matching the style used by edgefirst-cli. Downloads show decimal bytes (model artifact, trace) or item counts (dataset enumeration); uploads show decimal bytes. The TUI path is unchanged — it continues to consume the structured PROF: event stream and reports progress through its dashboard.
• Inference loop progress is now an indicatif bar instead of a per-100-frame [N/Total] X.X FPS | elapsed: Ys println!. The bar shows elapsed and ETA, the current/total frame counts with thousands separators, and a custom {fps} key formatted to one decimal place (avoids indicatif's default {per_sec} which renders the full f64). Skipped under --emit-prof (TUI mode) so the parent's PROF: event parser doesn't see ANSI escape sequences.
• Several "stuck at the end" silences now have status lines so neither the CLI nor the TUI parent process looks frozen: Inference complete. Finalizing results... between the last frame and stats computation, ORT profile: <path> before the streaming parse begins, Finalizing trace file... between the pipeline return and the pftrace BufWriter drop, and Publishing results to Studio... before the upload. Combined with the streaming ORT parse and the deadlock fix, the gap between "last frame" and "results published" is now visibly attributed instead of being one long silence.

v1.0.2

Added

• Prebuilt libtrt_shim.so now ships in linux-aarch64 release archives and wheels. Jetson Orin users on JetPack 6.x can pip install edgefirst-profiler or curl install.sh | bash and run TensorRT models out of the box — no more "TensorRT shim not found" with a "build it on the Jetson" follow-up. The shim is built in CI from the NVIDIA L4T TensorRT 10.3 container. Users on other JetPack versions can still build from source via shims/trt-shim/README.md.

Changed

• TensorRT shim search path now checks binary-relative locations first: dirname(profiler_binary)/libtrt_shim.so for flat archives and manual drop-ins, and dirname(profiler_binary)/../lib/libtrt_shim.so for pip-installed wheels. $TRT_SHIM_PATH, LD_LIBRARY_PATH, and the /usr/local/lib / /usr/lib fallbacks remain.

v1.0.1

Fixed

• Release archives on EdgeFirstAI/profiler-cli GitHub Releases now ship a .sha256 sidecar alongside each .tar.gz / .zip. The v1.0.0 release was missing these sidecars, causing install.sh and install.ps1 to fail their checksum-verification step with a 404 on the first install attempt (worked around manually for v1.0.0 by uploading sidecars after the fact).
• install.sh no longer prints tmpdir: unbound variable during cleanup on success. The script's EXIT trap referenced a local variable that had gone out of scope by the time the trap fired; the variable is now process-scoped and the trap defensively defaults if unset.

Changed

• CI builders for the wheel and binary matrix switched from ubuntu-latest and ubuntu-22.04-arm to ubuntu-22.04-large and ubuntu-22.04-arm-large respectively. The larger SKUs cut wall-clock build times on the long-pole entries.
• CHANGELOG body reformatted with one paragraph per line so GitHub Release pages, PyPI project descriptions, and other markdown consumers can wrap the text to fit their layout instead of preserving the source-file hard wraps.

v1.0.0

Initial public release of the EdgeFirst Profiler CLI — an on-target
profiling agent for AI vision pipelines, designed to characterize
real-world inference performance on the hardware your model will
actually run on.

Platforms

• Linux x86_64 (glibc 2.17+ / manylinux2014)
• Linux aarch64 (glibc 2.17+ / manylinux2014)
• macOS arm64 (macOS 11+)

Windows support is planned and not yet shipped; the installer prints
a clear message for Windows users in the meantime.

Edge hardware support

• NVIDIA Jetson Orin / Orin Nano — CUDA via ONNX Runtime, optional
  TensorRT execution provider.
• NXP i.MX 95 — Neutron NPU via the TFLite delegate.
• NXP i.MX 8M Plus — VSI NPU via the TFLite delegate.
• Raspberry Pi 5 — CPU baseline plus optional accelerators.
• Kinara Ara-2 — DVM models via the ara2-proxy daemon.
• Hailo-8 / Hailo-8L — via HailoRT.
• Apple Silicon — CoreML execution provider via ONNX Runtime.

Execution providers

ONNX Runtime is the default, lowest-common-denominator backend. Vendor
backends opt in per build: TFLite (with NXP Neutron and VSI delegates),
Hailo, TensorRT, Apple CoreML.

Features

• Per-operator timing dashboard with a live terminal UI (TUI).
• Pipeline orchestrator with 4-thread pipelined validation for
  realistic-throughput measurements.
• EdgeFirst Studio integration: launch profiling sessions from Studio,
  report status and progress, publish traces and validation runs to
  your Studio account.
• Offline-first: runs without a Studio account; Studio sign-in is
  only required to publish or pull data.

Distribution

• Standalone binary archives via GitHub Releases on
  EdgeFirstAI/profiler-cli
  (install.sh / install.ps1 resolve the latest version).
• Python wheels on PyPI as edgefirst-profiler — the wheel ships the
  same native binary as the standalone archive, installed onto your
  PATH via pip install edgefirst-profiler.