feat(training): untiled L3 baseline — per-step cycle counts for all 4 L3 models#21

Open

runwangdl wants to merge 19 commits into

develfrom

feat/untiling

Commits on May 10, 2026

feat(training): add CCT + CCT_LoRA to non-tiled siracusa training tests

runwangdl
and
claude
committed
feat(training): add L3 untiled baseline via fake-L1 shim + ResNet8 fixture

runwangdl
and
claude
committed
style: pre-commit fixes (yapf line wrap + clang-format 2-space indent)

runwangdl
and
claude
committed
fix(ci): drop ResNet8 untiled --l1 to 800 KB to fit ubuntu-latest 7 GB RAM

runwangdl
and
claude
committed
ci: --skipsim diagnostic for L3-untiled job + memory snapshot

runwangdl
and
claude
committed
feat(training): CCT/CCT_LoRA/MobileNetV1 untiled L3 + CI footprint summary

runwangdl
and
claude
committed
feat(ci): per-model sim gate + cycle counts in untiled-L3 summary

runwangdl
and
claude
committed
fix(untiled): drop CCT/CCT_LoRA + per-fixture needs_fake_l1 gate

runwangdl
and
claude
committed
fix(fake-l1): try real L1 first; restore CCT/CCT_LoRA + sim for all 4

runwangdl
and
claude
committed
feat(untiled): sed-based L1→L2 codegen rewrite + isolate to one CI job

runwangdl
and
claude
committed
fix(untiled): mchan_transfer_1d -> memcpy under DEEPLOY_L1_AS_L2

runwangdl
and
claude
committed
fix(untiled): skip MobileNetV1 sim in CI (known FC LSU crash)
runwangdl
committed
fix(untiled): cap MobileNet to 1-step/1-accum to fit FC L2 heap

runwangdl
and
claude
committed
fix(untiled): force MobileNet num_data_inputs=1
runwangdl
committed
Merge remote-tracking branch 'origin/devel' into feat/untiling
runwangdl
committed
fix(untiled): big CCT (img_size=32, emb_dim=128) needs --l1=800K
runwangdl
committed
ci(untiled): isolate big-CCT only — disable other 3 fixtures temporarily
runwangdl
committed
ci(untiled): re-enable L3-singlebuffer to capture big-CCT tiled cycle
runwangdl
committed

Commits on May 11, 2026

ci(untiled): drop n_steps=1 cap on CCT — use default 4-step schedule
runwangdl
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat(training): untiled L3 baseline — per-step cycle counts for all 4 L3 models#21

feat(training): untiled L3 baseline — per-step cycle counts for all 4 L3 models#21
runwangdl wants to merge 19 commits into
develfrom
feat/untiling

Commits on May 10, 2026

feat(training): add CCT + CCT_LoRA to non-tiled siracusa training tests

feat(training): add L3 untiled baseline via fake-L1 shim + ResNet8 fixture

style: pre-commit fixes (yapf line wrap + clang-format 2-space indent)

fix(ci): drop ResNet8 untiled --l1 to 800 KB to fit ubuntu-latest 7 GB RAM

ci: --skipsim diagnostic for L3-untiled job + memory snapshot

feat(training): CCT/CCT_LoRA/MobileNetV1 untiled L3 + CI footprint summary

feat(ci): per-model sim gate + cycle counts in untiled-L3 summary

fix(untiled): drop CCT/CCT_LoRA + per-fixture needs_fake_l1 gate

fix(fake-l1): try real L1 first; restore CCT/CCT_LoRA + sim for all 4

feat(untiled): sed-based L1→L2 codegen rewrite + isolate to one CI job

fix(untiled): mchan_transfer_1d -> memcpy under DEEPLOY_L1_AS_L2

fix(untiled): skip MobileNetV1 sim in CI (known FC LSU crash)

fix(untiled): cap MobileNet to 1-step/1-accum to fit FC L2 heap

fix(untiled): force MobileNet num_data_inputs=1

Merge remote-tracking branch 'origin/devel' into feat/untiling

fix(untiled): big CCT (img_size=32, emb_dim=128) needs --l1=800K

ci(untiled): isolate big-CCT only — disable other 3 fixtures temporarily

ci(untiled): re-enable L3-singlebuffer to capture big-CCT tiled cycle

Commits on May 11, 2026

ci(untiled): drop n_steps=1 cap on CCT — use default 4-step schedule