fix: harden torch.compile support against real-hardware failure modes - #22

Merged

Adithya-Thonse merged 17 commits into

TexasInstruments:mainfrom

musicalplatypus:pr/compile-hardening

Aug 6, 2026

Commits on Jul 29, 2026

fix(ci): trigger on tinyml-tinyverse changes too

t5fkg8d44d-beep
and
claude
committed
c94f6dc
fix: validate torch.compile with a warmup pass so failures actually fall back to eager mode

t5fkg8d44d-beep
and
claude
committed
6b8d3c5
fix: restore training mode unconditionally after torch.compile warmup

t5fkg8d44d-beep
and
claude
committed
6f892fd
fix: pass input_shape to compile_model_if_enabled at all 4 call sites

t5fkg8d44d-beep
and
claude
committed
3fd2bed
docs: clarify compile warmup only validates one graph variant, not the training-mode graph
t5fkg8d44d-beep
committed
57ff586
fix: unwrap torch.compile wrapper before ONNX/TorchScript export
t5fkg8d44d-beep
committed
d0f420a
fix: recursively unwrap compiled submodules before export, not just top-level

t5fkg8d44d-beep
and
claude
committed
4ff3e62
fix: skip torch.compile when quantization is enabled — FX tracing can't handle a compiled model

t5fkg8d44d-beep
and
claude
committed
7f9c0f7
test: strengthen skip-compile tests with torch.compile mock and INFO-log assertion
t5fkg8d44d-beep
committed
106e2f9
fix: strip torch.compile wrapper prefix from saved checkpoints

t5fkg8d44d-beep
and
claude
committed
e0d03c3
fix: strip _orig_mod prefix from EMA checkpoint keys too, symmetric save+load

t5fkg8d44d-beep
and
claude
committed
1ed7eff
test: add asymmetric compiled-save/uncompiled-load EMA case — the real production shape
t5fkg8d44d-beep
committed
2268e22
fix(ci): run tinyml-tinyverse tests

t5fkg8d44d-beep
and
claude
committed
aff21fd
fix: strip _orig_mod prefix in load_weights() as defense-in-depth

t5fkg8d44d-beep
and
claude
committed
0326b50
fix: resume_from_checkpoint unsafe deserialization + old-checkpoint compat

t5fkg8d44d-beep
and
claude
committed
898bc4a
fix(tests): use real argparse.Namespace for checkpoint round-trip tests

t5fkg8d44d-beep
and
claude
committed
0179c2d

Commits on Aug 6, 2026

Merge remote-tracking branch 'upstream/main' into pr/compile-hardening
t5fkg8d44d-beep
committed
58b5089

Navigation Menu

Uh oh!

fix: harden torch.compile support against real-hardware failure modes - #22

fix: harden torch.compile support against real-hardware failure modes#22
Adithya-Thonse merged 17 commits into
TexasInstruments:mainfrom
musicalplatypus:pr/compile-hardening

Commits on Jul 29, 2026

fix(ci): trigger on tinyml-tinyverse changes too

fix: validate torch.compile with a warmup pass so failures actually fall back to eager mode

fix: restore training mode unconditionally after torch.compile warmup

fix: pass input_shape to compile_model_if_enabled at all 4 call sites

docs: clarify compile warmup only validates one graph variant, not the training-mode graph

fix: unwrap torch.compile wrapper before ONNX/TorchScript export

fix: recursively unwrap compiled submodules before export, not just top-level

fix: skip torch.compile when quantization is enabled — FX tracing can't handle a compiled model

test: strengthen skip-compile tests with torch.compile mock and INFO-log assertion

fix: strip torch.compile wrapper prefix from saved checkpoints

fix: strip _orig_mod prefix from EMA checkpoint keys too, symmetric save+load

test: add asymmetric compiled-save/uncompiled-load EMA case — the real production shape

fix(ci): run tinyml-tinyverse tests

fix: strip _orig_mod prefix in load_weights() as defense-in-depth

fix: resume_from_checkpoint unsafe deserialization + old-checkpoint compat

fix(tests): use real argparse.Namespace for checkpoint round-trip tests

Commits on Aug 6, 2026

Merge remote-tracking branch 'upstream/main' into pr/compile-hardening