Use TimesFM 2.5 from GitHub repo (not PyPI)

handler.py

@@ -1,6 +1,6 @@
 """
 Custom handler for HuggingFace Inference Endpoints.
-Uses timesfm pip package (v1.3.0) with the 2.0 PyTorch model.
+Uses TimesFM 2.5 (200M) installed from GitHub repo.
 """
 
 import numpy as np
@@ -14,18 +14,19 @@ class EndpointHandler:
 
         torch.set_float32_matmul_precision("high")
 
-        self.model = timesfm.TimesFm(
-            hparams=timesfm.TimesFmHparams(
-                backend="gpu" if torch.cuda.is_available() else "cpu",
-                per_core_batch_size=32,
-                horizon_len=128,
-                num_layers=20,
-                context_len=1024,
-                use_positional_embedding=False,
-            ),
-            checkpoint=timesfm.TimesFmCheckpoint(
-                huggingface_repo_id="google/timesfm-2.0-500m-pytorch",
-            ),
+        self.model = timesfm.TimesFM_2p5_200M_torch.from_pretrained(
+            "google/timesfm-2.5-200m-pytorch"
+        )
+        self.model.compile(
+            timesfm.ForecastConfig(
+                max_context=1024,
+                max_horizon=128,
+                normalize_inputs=True,
+                use_continuous_quantile_head=True,
+                force_flip_invariance=True,
+                infer_is_positive=False,
+                fix_quantile_crossing=True,
+            )
         )
 
     def __call__(self, data: dict[str, Any]) -> dict[str, Any]:
@@ -37,14 +38,9 @@ class EndpointHandler:
             return {"error": "inputs must be a non-empty list of numbers"}
 
         input_array = [np.array(inputs, dtype=np.float64)]
-        frequency_input = [0]  # 0 = no specific frequency
-
-        point, quantiles = self.model.forecast(
-            input_array,
-            freq=frequency_input,
-        )
+        point, quantiles = self.model.forecast(horizon=horizon, inputs=input_array)
 
         return {
-            "point_forecast": point[0][:horizon].tolist(),
-            "quantile_forecast": quantiles[0][:horizon].tolist(),
+            "point_forecast": point[0].tolist(),
+            "quantile_forecast": quantiles[0].tolist(),
         }

requirements.txt

@@ -1,3 +1,2 @@
-timesfm==1.3.0
-torch
+git+https://github.com/google-research/timesfm.git@master#egg=timesfm[torch]
 numpy