hailo_model_zoo: Skill for Claude Code

Hailo Compile

Third stage of the DFC flow: take the quantized <model_name>_optimized.har from /hailo-optimize and produce a <model_name>.hef that runs on the Hailo accelerator.

This skill is shipped with the Hailo Model Zoo. It assumes the DFC wheel (hailo_sdk_client) is installed in the active virtualenv.

Inputs

Arg	Required	Notes
<optimized.har>	yes	Output of /hailo-optimize.
[hw_arch]	no	Read from the HAR if possible; otherwise default hailo10h.

Workflow

1. Resolve inputs

• Verify <optimized.har> exists.
• model_name = HAR filename stem (strip a trailing _optimized if present).
• hw_arch is read from the HAR; if unavailable, ask the user (or default to hailo10h).

2. Run compile

hailomz compile (and runner.compile() when driven through it) rewrites the HAR named after the network in the current directory — the same optimized HAR this stage reads from. Back up the optimized HAR under a distinct name first and checksum the file you're about to compile; losing the optimized HAR to a compile is easy to do and the loss is silent. Compile is also CPU-only, so set CUDA_VISIBLE_DEVICES="" before running it — otherwise it can hold a GPU that an optimization job elsewhere needs.

cp <optimized.har> <optimized.har>.bak
md5sum <optimized.har> > <optimized.har>.md5

from hailo_sdk_client import ClientRunner

runner = ClientRunner(har="<optimized.har>")
hef = runner.compile()
with open("<model_name>.hef", "wb") as f:
    f.write(hef)
runner.save_har("<model_name>_compiled_model.har")

The compiled HAR carries the metadata the profiler and hailo har extract consume. The _compiled_model.har suffix matches the DFC tutorial (DFC_3_Compilation_Tutorial.ipynb) and the compilation.rst documentation — downstream tooling expects this convention.

3. Verify

• Confirm <model_name>.hef exists and report its size and location.
• The compile log line Applying selected partition to N contexts gives the context count — report it. It's the number that explains a multi-context model's FPS.
• Inspect the HEF metadata with hailortcli parse-hef <model_name>.hef (requires HailoRT installed).
• Per-layer profiler (static) — hailo profiler <model_name>_compiled_model.har produces a static per-layer report. Note: for multi-context models (most large models), this report does not include performance / FPS numbers without runtime data.
• Profiler with runtime data (full FPS / utilization) — the documented two-step flow for accurate performance on multi-context models:

  hailortcli run2 -m raw measure-fw-actions --output-path runtime.json set-net <model_name>.hef
  hailo profiler <model_name>_compiled_model.har --runtime-data runtime.json --out-path runtime_profiler.html
  

• Smoke test on the device — hailortcli run2 set-net <model_name>.hef runs inference with random inputs as a quick sanity check (verify the exact form with hailortcli run --help — run2 always requires -m <mode>).