Skip to content

TimesFM (Time Series Foundation Model) is a pretrained time-series foundation model developed by Google Research for time-series forecasting.

Python

fix: improve PEFT device consistency and XReg output slicing

darkpowerxo18d5eb2d441e
diff --git a/peft/adapters.py b/peft/adapters.py
index 951cfe1..acc6eee 100644
--- a/peft/adapters.py
+++ b/peft/adapters.py
@@ -110,9 +110,10 @@ class DoRALinear(nn.Module):
in_f = base_linear.in_features
out_f = base_linear.out_features
+ dev = base_linear.weight.device
- self.lora_A = nn.Parameter(torch.empty(in_f, rank))
- self.lora_B = nn.Parameter(torch.zeros(rank, out_f))
+ self.lora_A = nn.Parameter(torch.empty(in_f, rank, device=dev))
+ self.lora_B = nn.Parameter(torch.zeros(rank, out_f, device=dev))
nn.init.kaiming_uniform_(self.lora_A, a=math.sqrt(5))
# Magnitude vector — initialised from pretrained column norms.
@@ -272,7 +273,8 @@ def load_adapter_weights(model: nn.Module, path: str) -> None:
The model must already have adapters injected (via ``inject_adapters``)
before calling this function.
"""
- tensors = load_file(path, device="cpu")
+ device = str(next(model.parameters()).device)
+ tensors = load_file(path, device=device)
trainable = {n for n, p in model.named_parameters() if p.requires_grad}
missing = trainable - set(tensors.keys())
if missing:
diff --git a/src/timesfm/utils/xreg_lib.py b/src/timesfm/utils/xreg_lib.py
index 4355261..2759b67 100644
--- a/src/timesfm/utils/xreg_lib.py
+++ b/src/timesfm/utils/xreg_lib.py
@@ -509,9 +509,9 @@ class BatchedInContextXRegLinear(BatchedInContextXRegBase):
@ x_tr_j.T
@ y_tr_j
)
- outputs.append(np.array(x_te_j @ beta_hat))
+ outputs.append(np.array(x_te_j @ beta_hat)[:tel])
if debug_info:
- outputs_context.append(np.array(x_tr_raw_j @ beta_hat))
+ outputs_context.append(np.array(x_tr_raw_j @ beta_hat)[:trl])
train_idx += trl
test_idx += tel