2. Add Apple Silicon (MPS) training support#8
Open
musicalplatypus wants to merge 8 commits into
Open
Conversation
MPS doesn't support float64 tensors. The previous .to(device).float() pattern moved float64 data to MPS first, which fails. Swapping to .float().to(device) casts to float32 on CPU then moves to device. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
torcheval's multiclass_confusion_matrix, multiclass_f1_score, and multiclass_auroc use sparse COO tensors internally, which are not supported on MPS. Move inputs to CPU for these metric computations. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add non_blocking=True to all training loop .to(device) calls to overlap CPU-GPU transfers with computation - Auto-enable native AMP (bfloat16) on MPS devices when not explicitly set; add --no-native-amp opt-out flag - Reduce per-batch GPU syncs: accumulate loss as detached tensor, defer .item() to SmoothedValue.update() at print time - Fix _get_device() to respect training_device config param instead of always auto-detecting - Add MPS memory reporting to MetricLogger (current_allocated_memory) - Tune default num_workers to 4 on macOS (spawn overhead vs 8 on Linux) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
MPS autocast defaults to float16 but GradScaler is unsupported, so small TinyML models suffer gradient underflow — all predictions collapse to a single class. Revert to opt-in only (--native-amp). Also adds macOS developer setup instructions to README. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
On MPS unified memory, non_blocking=True has negligible benefit (CPU and GPU share physical memory) and may cause subtle issues with certain DataLoader/model configurations. Revert to synchronous transfers which were proven working. The deferred .item() optimization in SmoothedValue is retained — that provides the main MPS performance win (7.8x faster metric logging path). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
run_tinyml_modelmaker.py: - Skip model catalog validation when nas_enabled=True - Guard params.update() against None model_description - Generate fallback model_description with generic_model=True, training_backend, and model_training_id for NAS train.py (timeseries_classification): - Fix nas_enabled check: was comparing bool True to string 'True' (str2bool converts arg to bool, so == 'True' always failed) train_cnn_search.py: - Fix MPS float64 crash: cast to float32 before .to(device) (MPS doesn't support float64, so .to(device).float() fails) tests/test_nas_support.py: - 9 tests: model validation bypass, fallback description, str2bool, argparse integration
MPS does not support float64 dtype. When DataLoader returns float64 tensors, .to(device) transfers them as-is to MPS, and the subsequent .long()/.float() conversion fails with TypeError. Fix: cast dtype BEFORE device transfer (.long().to(device) instead of .to(device).long()). Also add explicit dtype=torch.float32 to empty tensor creation (torch.tensor([]) defaults to float64). Fixed files (7): - timeseries_classification/test_onnx.py - image_classification/test_onnx.py - timeseries_anomalydetection/test_onnx.py (2 functions) - timeseries_anomalydetection/test_onnx_cls.py - timeseries_anomalydetection/train.py - timeseries_regression/test_onnx.py - timeseries_forecasting/test_onnx.py
Replace proprietary tinyml-mlbackend model references with built-in
open-source CNN implementations for arc fault, motor fault, and fan
imbalance classification. Previously these 10 models failed with
ValueError because the proprietary model files were not available.
New model classes: CNN_AF_3L_{200,300,700,1400}, CNN_MF_{1L,2L,3L}
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
musicalplatypus
changed the title
Adithya-Thonse
pushed a commit
that referenced
this pull request
Jun 12, 2026
Merge in TINYML-ALGO/tinyml-agent-skills from 2026/pranav_a to main * commit '8c3260bcdd549353b2cdbf3562bb3f32753fdddf': improving readme
Adithya-Thonse
added a commit
that referenced
this pull request
Jun 12, 2026
de8af16d Pull request #45: https://jira.itg.ti.com/browse/TINYML_ALGO-698 REVERT: e48ef1a Pull request #14: TINYML_ALGO-711: fixing readme REVERT: 16fc6a6 TINYML_ALGO-711: fixing readme REVERT: e3639d2 Pull request #13: removing pycache REVERT: f8bb3b7 removing pycache REVERT: dd38428 Pull request #12: restructuring agent skill REVERT: ff02a0e restructuring agent skill REVERT: d26c6a5 Pull request #11: fixing tiny ml name REVERT: 640ffd3 fixing tiny ml name REVERT: 4ee3a19 Pull request #10: 2026/pranav a REVERT: be83fc6 minor fixes REVERT: e3a5700 removed assets, included autoMP quant REVERT: 1af575a Pull request #9: correcting npu devices list REVERT: 31e9eb1 correcting npu devices list REVERT: 59b209b Pull request #8: improving readme REVERT: 8c3260b improving readme REVERT: 668916f Pull request #7: improving readme REVERT: 68686b3 improving readme REVERT: 814316e Pull request #6: fixes to readme and marketplace json REVERT: e4bc0b4 fixes to readme and marketplace json REVERT: 6a64208 Pull request #5: fixes to readme REVERT: 0f9c868 fixes to readme REVERT: 52f95ff Pull request #4: 2026/pranav a REVERT: 443295d fixes to readme REVERT: 1881112 fixes to readme and marketplace json REVERT: 229ab57 Pull request #3: 2026/pranav a REVERT: 6519104 minor readme fix REVERT: 38e9f9f minor readme fix REVERT: db81f81 Pull request #2: minor readme fix REVERT: 1c0737a minor readme fix REVERT: 0a0c02d Pull request #1: minor readme fix REVERT: b682335 minor readme fix REVERT: 062eb39 Initial Commit git-subtree-dir: tinyml-agent-skills git-subtree-split: de8af16d9e23de3e9bda3d811a0ebdece1178260
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
2 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
Adds full Apple Silicon (MPS/Metal) support for training across all task types. This enables GPU-accelerated training on Mac devices without CUDA.
Changes
MPS Device Compatibility
.float()before.to(device)to avoid MPS dtype errors (float64not supported on MPS)non_blocking=Truereverted — MPS doesn't support async transfers; reverted to synchronous.to(device)callstest_onnx.pyand training scripts to usefloat()instead ofdouble()MPS Performance Optimization
.item()in training loops —SmoothedValue.update()now stores detached tensors instead of calling.item()every batch, avoiding MPS command-buffer flush synchronization stalls--native-ampflag threading throughtimeseries_base.py(auto-enable on MPS was reverted due to gradient underflow — left as opt-in)torch.compileintegration — Added model compilation support for accelerated trainingNAS on MPS
test_nas_support.py)New Models
Files Changed (16 files)
Across
tinyml-modelmaker,tinyml-tinyverse,tinyml-modelzoo, andtinyml-modeloptimizationTesting
test_nas_support.py