ibm-research-ttm-r3-NPU
#NPU Ascend910 TTM TinyTimeMixer
Model: ibm-research/ttm-r3 (revision: aca5d4956c59726b320c1562a6eaebb7fd7ec9b9, local sha: see config)
Task: time-series forecasting (context 512 -> prediction 30, multivariate)
Architecture: TinyTimeMixerForDecomposedPrediction (decomposed trend-residual, d_model=128, num_layers=10, patch 9x9, student_t distribution, multi-quantile head)
Weights URL: https://huggingface.co/ibm-research/ttm-r3
Data Contract
Source: Synthetic deterministic series + ETTh1-style sinusoidal (seed 42, same as notebook)
Schema: past_values [B, T, C] = [2, 512, 1], past_observed_mask [2,512,1] all ones
Channel: 1 (univariate), frequency token default, no exogenous
Preprocessing: StandardScaler (std scaling) internal to model, no external categorical encoding
Split: Single inference window per series, no train/test leakage, deterministic generation: sin(0.02t)+noise, cos(0.015 t)+noise
Dtype: FP32, no quantization
Environment
NPU: Ascend910_9362, 2 devices, used npu:0 (Phy 4, 0000:0A:00.0)
Toolkit: CANN 8.5.1, torch 2.9.0+cpu, torch_npu 2.9.0.post1, transformers 4.57.6, huggingface_hub 0.34.0, tokenizers 0.22.0, tsfm_public 0.3.9.dev4, Python 3.11.14
Command: python inference.py (default, no args, uses npu:0)
Dependencies
torch==2.9.0
torch_npu==2.9.0.post1
transformers==4.57.6
huggingface_hub==0.34.0
tokenizers==0.22.0
datasets==5.0.1
numpy==1.26.4
pandas==3.0.2
scikit-learn==1.7.0
einops>=0.7
granite-tsfm @ git+https://github.com/ibm-granite/granite-tsfm.git@v0.2.22
deprecated
NPU Inference
pip install -r requirements.txt
python inference.py
loads TinyTimeMixerForDecomposedPrediction from ibm-research/ttm-r3, moves to npu:0, runs synchronize-wrapped forward, prints device/dtype/input/output and benchmark.
Real Results (npu:0, seed 42)
Input: past [2,512,1] mean 0.1481 std 0.6927, first 5 series0 [0.1927,0.1687,0.1301,-0.1506,0.1478]
Output: [2,30,1] mean -0.3287 min -0.8937 max 0.3000, first window series0 [-0.6887,-0.7043,-0.7207,-0.7327,-0.7471], last window [-0.8841,-0.8872,-0.8896,-0.8937,-0.8922]
Device verified: model param npu:0, output npu:0, torch.npu.is_available True, npu-smi shows Ascend910_9362
Consistency
CPU vs NPU: same weights, same synthetic input, same FP32, seed 42
Max_abs 0.00025, mean_abs 0.00013, shape [2,30,1]
Threshold: atol 5e-4 rtol 1e-3 (FP32, NPU fused conv/attention causes 1e-4-scale drift, documented)
Result: PASS with python scripts/compare_outputs.py --atol 5e-4 --rtol 1e-3
First (compile): 0.2887s (includes NPU kernel compile)
Stable: avg 0.0173s min 0.0169s max 0.0181s p50 0.0172s p90 0.0180s p95 0.0181s over 10 runs after 3 warmup, torch.npu.synchronize each
Throughput: 115.43 windows/s (batch 2, ctx 512, horizon 30) ; 230.86 rows? per second equivalent
Batch/context/horizon: 2 / 512 / 30 / 1ch / FP32
Evidence
assets/agent_workflow.png – full workflow: inspect, download, cpu baseline, npu verify, compare, benchmark, validate
assets/npu_device_call.png – npu-smi, check_npu.py, param device, input/output device
assets/model_result.png – default python inference.py complete stdout
Generated via scripts/render_xterm_evidence.mjs --style raw from real logs (no Pillow simulation).
Limitations
Only past_values + observed_mask path tested; frequency token, exogenous, static categorical not exercised
Synthetic data only; real ETTh1/ETT datasets would require additional preprocessing but same model path
FP32 only; FP16/BF16 not validated
Single NPU (910B) tested; multi-NPU not benchmarked