fix(ci): install ti_mcu_nnc before tinyverse, bump to 2.1.2, add macOS step - #17
Merged
Adithya-Thonse merged 45 commits intoJul 28, 2026
Conversation
updates from the TI main branch. Most likely the 1.3.0 release
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
… utils, symlink 1. Add missing TinyMLQuantizationVersion import in timeseries/runner.py and vision/runner.py — NameError on every training run that reaches packaging. 2. Fix argv splice when native_amp=True + quantization: strip boolean flags before fixed-offset slicing, re-append after, preventing malformed argv. 3. Fix len(split_factor) called on float in dataset_utils.py — use the accumulated split_factors list instead of the original parameter. Fixed in both create_inter_file_split and create_intra_file_split. 4. Add else branch in dataset_load() for unknown annotation_format — raises ValueError instead of UnboundLocalError. 5. Remove duplicate os.symlink() call in make_symlink() — was creating the link twice, second call always failed with FileExistsError. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
… error handling 1. ConfigDict.__getstate__: add missing return — pickling now preserves state. 2. _parse_include_files: fix and→or — absolute/relative include paths now resolve correctly instead of always prepending base path. 3. TASK_CATEGORIES: use TASK_CATEGORY_TS_ANOMALYDETECTION instead of TASK_TYPE_GENERIC_TS_ANOMALYDETECTION — fixes wrong compilation parameters for anomaly detection tasks. 4. download_url: handle missing Content-Length header gracefully instead of crashing with TypeError on int(None). 5. download_files: track aggregate success across all URLs instead of reporting only the last download's status. 6. get_target_module: raise ValueError with available options instead of returning None — prevents confusing downstream AttributeError on NoneType. Applied to both timeseries and vision training modules. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
1. Set args.quit_event before compile_scr.run() so compilation can actually be cancelled (was set after run() returned — too late). 2. Fix typo 'anomlay_list.txt' → 'anomaly_list.txt' in dataset handling. 3. Move cleanup_special_chars() file write outside the read block so the file isn't truncated while the read handle is still open — prevents data loss if the write fails mid-way. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…, dead code 1. Persistent validation iterator: replace next(iter(valid_loader)) with a cycling iterator so architecture updates see different batches each step 2. Structural params: pass steps/multiplier/stem_multiplier from search to final model so the evaluated architecture matches what was searched 3. Best genotype tracking: select genotype with highest validation accuracy instead of always using the last epoch 4. Device abstraction: add get_device() (CUDA > MPS > CPU), replace all hardcoded .cuda() calls with .to(device) across search, model, architect 5. Differentiable resource penalty: replace inert scalar penalties with sum(softmax(alpha) * param_counts) normalized to [0,1], fully differentiable w.r.t. architecture parameters 6. Remove dead code: delete unused RNN genotypes, Genotype_RNN, save(), create_exp_dir(), and all RNN branches from architect Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The previous MPS optimization commit moved .item() from MetricLogger to SmoothedValue but it still fired on every batch update, causing a GPU command-buffer flush each time (no actual improvement). Fix by storing detached tensors in the deque and accumulating total on-device. Property accessors (median, avg, global_avg, max, value) call .item() lazily — only at print time (every print_freq batches). Benchmarked at 7.8x faster for the metric-logging path on MPS. Also reverts MPS memory reporting (torch.mps.current_allocated_memory) which introduced a new GPU sync that never existed before the previous commit. CUDA memory reporting is unchanged. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…kers persistent_workers=True keeps worker processes alive across epochs, but on macOS (spawn start method) the resource_tracker detects unreleased semaphores at exit. Add shutdown_data_loaders() to explicitly terminate workers after training completes. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…cripts The previous fix (f3cc7c0) added shutdown_data_loaders() to train.py files but missed all test_onnx scripts. These are the last step in the pipeline and create DataLoaders with num_workers=8 without cleanup, causing macOS resource_tracker to report leaked semaphores at process exit. Add shutdown_data_loaders() calls to all 6 test_onnx files and harden the function itself with try/except and gc.collect() to ensure worker processes fully release their semaphores before exit. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Enhanced shutdown_data_loaders to explicitly break the iterator's references to multiprocessing Queue/Lock/Event objects after calling _shutdown_workers(). Without this, the POSIX named semaphores inside those objects survive until Python's resource_tracker atexit handler, which reports them as leaked. By clearing the references eagerly, CPython's refcount-based deallocation calls sem_unlink immediately. Also added the missing shutdown_data_loaders call in image_classification/train.py, which was the only train.py that did not shut down its DataLoaders. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The previous approach (breaking Queue references for gc) didn't work because on Python <=3.11, _multiprocessing.SemLock's C dealloc only calls sem_close() — it never calls resource_tracker.unregister(). The resource_tracker therefore reports every semaphore as leaked regardless of cleanup efforts. (Fixed in Python 3.12+.) New approach: after _shutdown_workers() joins all worker processes, walk the iterator's Queue and Event objects to find their internal SemLock names, then explicitly call resource_tracker.unregister() for each one. This removes them from the resource_tracker's registry so its atexit handler has nothing to warn about. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…, eval efficiency - Wire up dead --compile-model arg to torch.compile (aot_eager for MPS, inductor for CUDA) - Add --native-amp flag for PyTorch native AMP (torch.amp.autocast) on CUDA and MPS - Add persistent_workers=True to DataLoaders (avoids respawning on macOS spawn) - Fix pin_memory logic to use torch.cuda.is_available() instead of broken gpu>0 check - Use optimizer.zero_grad(set_to_none=True) across all training loops - Fix O(n²) torch.cat accumulation in evaluate_classification with list-based collection - Move per-batch f1/confusion_matrix to epoch-end in classification eval Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…er config pipeline Wire torch.compile and native AMP opt-in flags from modelmaker params through timeseries_base.py argv construction to tinyml-tinyverse training scripts. Update ARCHITECTURE.md with Training Performance Optimizations section and PORTING_ASSESSMENT.md with post-porting development history. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…ve_paths() Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add TRAINING_BACKEND_TINYML_TINYVERSE, DATA_DIR_CLASSES, DATA_DIR_FILES, DATA_DIR_IMAGES constants. Use existing TRAINING_DEVICE_CUDA in params.py defaults instead of bare 'cuda' strings. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…nyml-modelmaker Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…er, etc.) Define typing.Protocol classes that formalize the implicit contracts already followed by ModelRunner, ModelTraining, ModelCompilation, and DatasetHandling. Uses structural subtyping so no existing classes need modification. Enables static type checking and documents the interface contracts for future implementations. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…ment analysis Documents the full tinyml-tensorlab architecture including all four sub-repos, pipeline flow, configuration system, model/quantization/NAS subsystems, a Mermaid architecture diagram, and 12 design/implementation improvement recommendations. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Restore matplotlib.use('Agg') in utils.py: removing it left
classification/anomaly/regression scripts without a headless backend
guard, causing potential display failures on CI/servers.
- Restore logger and MSE logging in train_one_epoch_anomalydetection
and evaluate_anomalydetection: both lines were accidentally dropped
during the AMP refactor, silencing per-epoch reconstruction error.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…scripts Ensures data loaders are always shut down (releasing macOS POSIX semaphores) even when an exception is raised or an early return path is taken mid-function. Affects all 5 train.py scripts and all 6 test_onnx*.py scripts. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- nas/train_cnn_search.py: capture genotype after training step so best_genotype matches the architecture that achieved best_valid_acc; initialize to -inf so zero-accuracy epochs are correctly recorded - download_utils.py: guard content-length header against non-numeric values - config_dict.py: use os.path.isabs() for cross-platform include-path detection - nas/model.py: derive C_prev from cell.multiplier (actual genotype concat width) and validate it matches the network multiplier parameter - common/utils/utils.py: convert to float before .mean() on stacked tensors - nas/utils.py: validate gpu_index range before constructing cuda device Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The np.zeros((num_classes, num_classes)) at function entry was always overwritten by get_confusion_matrix at epoch end; it was also a latent TypeError when num_classes was None. Replaced with an explicit ValueError. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- dataset_utils.py: validate scalar split_factor in (0, 1) range; use split_factors list (not raw split_factor param) in post-normalization checks to avoid TypeError on float inputs - misc_utils.py: guard input_data_path before os.path.basename (prevents TypeError when both dataset_name and input_data_path are unset); use absolutized projects_path as base for model_packaged_path; fix archive name for flat run_names (filter(None,...) removes empty leading segment so 'myrun' → myrun.zip instead of _myrun.zip) - vision/runner.py: use v.get() for inference_time_us/sram/flash to avoid KeyError when a device entry is missing a metric Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Sections 1 (No Test Suite) and 6 (No Abstract Base Classes or Protocols) now accurately reflect what the PR stack addressed: the protocols module in ai_modules/protocols.py and the test_protocols.py test suite. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Resolved conflicts by preserving the try/finally structure from pr/macos-semaphore-fixes and integrating compile_model_if_enabled, get_amp_context, and get_grad_scaler additions from the performance PR. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Resolved __init__.py conflicts by keeping the ValueError-raising get_target_module() from pr/bug-fixes rather than the return-None pattern from pr/code-quality. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- NAS infer(): extend torch.no_grad() to cover forward pass + criterion (previously only covered device transfers; gradients computed needlessly) - constants.py get_default_data_dir_for_task(): raise ValueError on unknown task_category instead of silently returning DATA_DIR_CLASSES - utils.py evaluate_classification(): accumulate tensors on CPU (.cpu()) to avoid OOM from full-dataset tensors staying on GPU between batches - utils.py evaluate_classification(): use .squeeze() consistently for get_au_roc and get_confusion_matrix; remove now-redundant .cpu() calls - image_classification/train.py: fix AUC ROC log to use best['auc'] not best['f1'] (copy-paste bug) - misc_utils.py: apply filter(None, ...) to compilation archive name paths (same fix as training paths — leading underscore for bare run names) - ARCHITECTURE.md: fix CLI entry-point command and update summary table rows 1 and 6 to reflect addressed status Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…s refactor 0c1b23d moved path resolution out of ModelRunner and removed the tinyml_torchmodelopt.quantization import along with it, but two uses remain in run() at the packaging step. Any run that reaches model packaging raises NameError: name 'TinyMLQuantizationVersion' is not defined, which makes compilation fail on every platform. Restore the import, matching how params.py, misc_utils.py and timeseries_base.py already reference it. Verified: compiling an exported ONNX for F28P55 now completes and produces artifacts/mod.a, tvmgen_default.h and the deployment zip. All 20 Python files changed on this branch import cleanly.
Companion to 3607b15. The resolve_paths refactor in 0c1b23d dropped this import from both ai_modules/timeseries/runner.py and ai_modules/vision/runner.py; only the timeseries one was restored. vision/runner.py still raises NameError at lines 154 and 207 when a run reaches model packaging. Found by running pyflakes across every file this branch touches — the earlier import-only check could not catch it, since the name is referenced inside run() and never evaluated at import time.
image_classification/test_onnx.py and timeseries_anomalydetection/test_onnx_cls.py call .astype(np.float32) without importing numpy, so the nn_for_feature_extraction path raises NameError. Pre-existing rather than introduced here, but both files are already modified by this branch. Found with pyflakes; the name is only referenced inside a conditional branch, so it is invisible to an import-time check.
audio_classification/test_onnx.py calls .astype(np.float32) without importing numpy, raising NameError on the nn_for_feature_extraction path. The same fix exists on pr/mps-support, but that branch has never been merged into main, so it never propagated here. The sibling fixes for image_classification and timeseries_anomalydetection arrived via pr/macos-semaphore-fixes. Pre-existing upstream: TI's main carries the same bug. PR TexasInstruments#8 carries the fix back to them.
Three bug fixes for MPS/Apple Silicon users, verified by 7/7 passing TestEndToEnd on macOS ARM64: - Isolate onnxsim.simplify() in a subprocess on darwin/arm64 so the SIGSEGV during Python interpreter shutdown cannot kill the main process before packaging and compilation run (quant_base.py) - Set PYTORCH_ENABLE_MPS_FALLBACK=1 inside _get_device() when MPS is selected, so ops unsupported on MPS fall back to CPU transparently; previously this was only set in the e2e test environment (timeseries_base.py, audio_base.py, image_base.py) - Fix MPS float64 crash in all reference test/train scripts: move dtype cast (.long()/.float()) before .to(device) so float64 tensors are converted on CPU before being moved to MPS, which rejects float64 (8 test_onnx.py and train.py files across all task types) Also remove spurious f-prefix from 63 f-strings with no placeholders across modelmaker, tinyverse, modelzoo, and agent-skills. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…ization The auto-quantization calibration and eval loops moved float64 feature tensors to the device before casting, so MPS raised "Cannot convert a MPS Tensor to float64 dtype" and killed the run right after training finished — the model trained and exported, then quantization aborted and nothing compiled. main already carries the equivalent fixes for tinyverse utils.py and the reference scripts; this was the one remaining site.
Backfills main with the two pieces that lived only on the MPS branch: the nas_enabled model-description fallback (so a NAS id such as NAS_m is not rejected as an unknown model) and tests/test_nas_support.py. Both now ship in the dedicated NAS PR.
…S step Three bugs caused tinyml_tinyverse to silently not be installed: 1. Wrong order: ti_mcu_nnc was installed after pip install -e tinyml-tinyverse, so pip tried to fetch ti_mcu_nnc 2.1.2 (from pyproject.toml) during the tinyverse install and failed. 2. Version mismatch: CI installed ti_mcu_nnc 2.1.1 but tinyml-tinyverse/ pyproject.toml requires 2.1.2. 3. Silent failure: the Install dependencies step is a multi-command run block with no set -e. When pip install -e tinyml-tinyverse exits 1, bash continues and the step appears green because pip install pytest (the last command) succeeds. The module is never installed. 4. macOS had no ti_mcu_nnc install step at all. Fix: move all ti_mcu_nnc installs before Install dependencies, bump to 2.1.2, add a macOS ARM64 step, and add set -e so install failures are visible. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This was referenced Jul 28, 2026
Merged
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
The Tier 1 Component Tests have been failing on all platforms with
ModuleNotFoundError: No module named 'tinyml_tinyverse'. This is a CI configuration bug, not a code bug —tinyml_tinyversenever gets installed due to three compounding issues.Root causes
1. Wrong install order
pip install -e tinyml-tinyverseruns beforeti_mcu_nncis installed. Sincetinyml-tinyverse/pyproject.tomlliststi_mcu_nncas a required dependency, pip tries to fetch it during the tinyverse install — and fails.2. Version mismatch
The CI workflow installs
ti_mcu_nnc 2.1.1, buttinyml-tinyverse/pyproject.tomlrequires2.1.2. The URL for2.1.1cannot satisfy the declared dependency.3. Silent failure (the sneaky one)
The
Install dependenciesstep is a multi-commandrun:block withoutset -e. Whenpip install -e tinyml-tinyverseexits with code 1, bash continues. The finalpip install pytestsucceeds, so the step is marked green.tinyml_tinyverseis never installed, and the failure only surfaces at test collection time — two steps later.4. macOS had no
ti_mcu_nncinstall stepLinux and Windows had explicit steps; macOS relied entirely on tinyverse's own dependency resolution, which was broken by the above.
Fix
ti_mcu_nncinstall steps beforeInstall dependencies2.1.1 → 2.1.2to matchtinyml-tinyverse/pyproject.tomlti_mcu_nncinstall stepset -etoInstall dependenciesso future failures surface immediatelyVerification
Confirmed that checking out
upstream/mainand running the tests locally produces the identicalModuleNotFoundError— the failure predates all open PRs and is purely a CI configuration issue.🤖 Generated with Claude Code