test(gap9): run the fine-tuning experiments in CI

[runwangdl]

What

19 training cases in a job of their own, gap9-training-experiments, covering four families of parameter-efficient fine-tuning.

family	model	Mcyc / step
channel-wise	ResNet8SplitBlock	42.66
channel-wise	MobileNetV1SplitAlt	38.53
qlora	CCT	53.89
qlora	MobileNetV1	42.28
lora	ResNet8 (PEFT)	56.52
lora	MobileNetV1	51.13
layerwise	ResNet8 last1 / 2 / 4 / 6	16.23 / 19.71 / 29.58 / 38.64
layerwise	MobileNetV1 last1 … last24	18.73 / 22.14 / 24.83 / 26.60 / 28.72 / 32.92 / 36.60 / 39.70 / 44.28

Every one was run end-to-end on gvsoc against this commit and reported Errors: 0. The cycle counts above are from those runs and are recorded next to each entry in the config.

What is deliberately absent

Two candidates were tried on this same tree and fail, so they are not registered. The reason sits beside the table so a later fix has something to check against:

model	failure
ResNet8_QLORA	faults on device — cluster PE7 LSU Invalid access
ResNet8_PEFT_QLORA	tiler assert in minimizeRectangle: per-channel transfers with the channel index written into the batch offset

The second is the same defect that a channel split whose halves are read twice runs into, which is why the channel-wise graphs here are built to read each half once. It is worth fixing on its own.

Why a separate job

gap9-training-tiled-l3-singlebuffer already runs close to an hour. Adding 19 cases to it would roughly double that, and an experiment going red would gate the mainline configurations. The new job runs in parallel and gates nothing.

Overrides

The per-model overrides carry what the verification runs actually used — CC_STACK_SIZE=4096, SLAVESTACKSIZE=512 — rather than the tuned values of the base models these graphs derive from (MobileNetV1 is tuned to cc_stack 8192). Registering settings nobody exercised would put a configuration into CI that was never measured.

Assets

Existing convention only: network.onnx, inputs.npz, outputs.npz per training directory plus the optimizer's network.onnx. Intermediate build output is not included. 22 MB total; check-added-large-files passes.

🤖 Generated with Claude Code