TimesFM (Time Series Foundation Model) is a pretrained time-series foundation model developed by Google Research for time-series forecasting.
Python
feat(timesfm3/mlx): cap context at global_context before decode
rachittshahf40faab9c7cd
diff --git a/src/timesfm3/mlx/timesfm3_forecaster.py b/src/timesfm3/mlx/timesfm3_forecaster.pyindex bbfb5b7..f88038f 100644--- a/src/timesfm3/mlx/timesfm3_forecaster.py+++ b/src/timesfm3/mlx/timesfm3_forecaster.py@@ -42,6 +42,10 @@ class _ModelConfig:median_quantile_index: int = 4# mx.compile the forward pass (fuses kernels, cuts dispatch overhead).compile: bool = True+ # Longest context fed to the model. Contexts beyond this are truncated to+ # their most recent `max_context_length` points before decode, matching the+ # torch backend's `global_context` cap (`_MAX_CONTEXT_LENGTH`).+ max_context_length: int = 15360# Hugging Face download options.cache_dir: str | None = Nonerevision: str | None = None@@ -101,6 +105,11 @@ class TimesFM3Forecaster:def context_length(self) -> int:return self.model.config.input_patch_len+ @property+ def global_context(self) -> int:+ """Longest context the model runs on; longer inputs are truncated to it."""+ return self.config.max_context_length+def predict(self,context: np.ndarray,@@ -176,7 +185,8 @@ class TimesFM3Forecaster:# Group same-length contexts so each group runs through a single batched forward pass.# decode() is per-series independent (running stats, RevIN, detrending are per-batch-row), so# batching is numerically identical to looping but scales throughput near-linearly.- arrs = [np.asarray(c, dtype=np.float32).reshape(-1) for c in contexts]+ cap = self.config.max_context_length+ arrs = [np.asarray(c, dtype=np.float32).reshape(-1)[-cap:] for c in contexts]groups: dict[int, list[int]] = {}for i, a in enumerate(arrs):groups.setdefault(a.shape[0], []).append(i)diff --git a/src/timesfm3/mlx/timesfm3_forecaster_test.py b/src/timesfm3/mlx/timesfm3_forecaster_test.pyindex 193ebee..b5c1fe1 100644--- a/src/timesfm3/mlx/timesfm3_forecaster_test.py+++ b/src/timesfm3/mlx/timesfm3_forecaster_test.py@@ -102,6 +102,27 @@ class TimesFM3MlxRealWeightsTest(unittest.TestCase):self.assertEqual(np.asarray(out.forecast).shape, (64,))self.assertEqual(np.asarray(out.quantiles).shape, (64, 9))+ def test_global_context_default_matches_torch(self):+ # The MLX backend must expose the same 15,360-step context cap as the torch+ # backend so the two are backend-equivalent for very long contexts.+ forecaster = timesfm3_forecaster.TimesFM3Forecaster.from_pretrained(_CHECKPOINT)+ self.assertEqual(forecaster.global_context, 15360)++ def test_context_truncated_to_global_context(self):+ # A context longer than the cap must be truncated to its last+ # `global_context` points before decode, so forecasting the full series+ # equals forecasting only that tail.+ cap = 256+ forecaster = timesfm3_forecaster.TimesFM3Forecaster.from_pretrained(+ _CHECKPOINT, max_context_length=cap+ )+ self.assertEqual(forecaster.global_context, cap)+ ctx = np.sin(np.linspace(0, 60, 512)).astype(np.float32)+ out_full = forecaster.predict(ctx, horizon=32, return_quantiles=True)+ out_tail = forecaster.predict(ctx[-cap:], horizon=32, return_quantiles=True)+ np.testing.assert_allclose(out_full.forecast, out_tail.forecast, atol=1e-5)+ np.testing.assert_allclose(out_full.quantiles, out_tail.quantiles, atol=1e-5)+@unittest.skipUnless(_torch_available(), "requires torch as the parity oracle")def test_parity_with_torch_backend(self):from ..torch import timesfm3_forecaster as torch_forecaster