Q: How can I run TimesFM 3.0 on Apple Silicon without PyTorch?

Use the MLX backend. It is provided in PR #476 under `src/timesfm3/mlx/` and keeps the existing PyTorch backend under `src/timesfm3/torch/`. The MLX version exposes the same `timesfm3_forecaster` interface, loads the same checkpoints, and matches PyTorch output to about `1e-6` max absolute error. A standalone implementation is available at `https://github.com/rachittshah/mlx-tsfm`. Quick setup: ```bash python -m pip install -U mlx git clone https://github.com/rachittshah/mlx-tsfm.git cd mlx-tsfm pytest tests/test_timesfm.py ``` When using the TimesFM repository with PR #476, import the forecaster from `timesfm3.mlx.timesfm3_forecaster` instead of the Torch module and use the same initialization and forecast calls. On an M4 Max, the 330M model reportedly takes about 20 ms per single forecast and reaches roughly 475 forecasts/second when batched.

Question 1

Why does torch_compile=True have no effect in TimesFM PyTorch, and how do I make torch.compile speed up inference?

Accepted Answer

The no-op is caused by `self = torch.compile(self)` inside `load_checkpoint`: `self` is a local variable, so rebinding it does not change the model object returned by `from_pretrained`.

Apply the fix from PR #459, or compile `forward` directly after loading:

```python
model = timesfm.TimesFM_2p5_200M_torch.from_pretrained(
    "google/timesfm-2.5-200m-pytorch"
)
model.model.forward = torch.compile(model.model.forward)
```

Equivalently, replace the ineffective line in `timesfm_2p5_torch.py` with:

```python
self.model.forward = torch.compile(self.model.forward)
```

The KV-cache slice assignment still causes graph breaks, but the attention and MLP matmuls are compiled, so inference is faster. On an A100 with batch=64, context=512, and horizon=128, this reduced latency from 108.8 ms to 73.7 ms (1.48x). Compiling the whole module after fixing the rebinding gave no measurable speedup in the reported benchmark.

Question 2

Why does TimesFM's `xreg_mode` behave opposite to the docstring for `xreg + timesfm` and `timesfm + xreg`?

Accepted Answer

The docstring is reversed in the affected implementation. Use the mode according to the code, not the documentation:

- To fit xreg on the targets first and then forecast residuals with TimesFM, use:
  ```python
  xreg_mode="xreg + timesfm"
  ```
- To generate the TimesFM forecast first and then fit xreg on its residuals, use:
  ```python
  xreg_mode="timesfm + xreg"
  ```

This is confirmed by the branch around L488–L494: `timesfm + xreg` uses `max(0, input_len - self.input_patch_len)` and is commented “For fitting residuals,” while `xreg + timesfm` uses `input_len`. Until the docstring is corrected, treat the documented descriptions as swapped.

Question 3

How to fix TimesFM 2.5 PyTorch from_pretrained error: unexpected keyword argument 'proxies', and load_checkpoint NotImplementedError?

Accepted Answer

This is fixed by timesfm PR #435. Upgrade to a release containing that fix, or install the latest source:

```bash
pip install -U git+https://github.com/google-research/timesfm.git
```

Then retry:

```python
import timesfm
tfm = timesfm.TimesFM_2p5_200M_torch.from_pretrained(
    "google/timesfm-2.5-200m-pytorch"
)
```

If you cannot upgrade, temporarily pin an older Hub client:

```bash
pip install "huggingface_hub<0.34"
# for example:
pip install huggingface_hub==0.33.5
```

This may conflict with newer `transformers`/`accelerate` releases. In affected timesfm versions, do not use `load_checkpoint(local_path)` as a workaround because the PyTorch subclass raises `NotImplementedError`; use a patched timesfm build instead. A source-level fix is to make the model constructor accept and ignore newer Mixin kwargs, e.g. `def __init__(self, config, **_ignored):`.

Question 4

How can I run TimesFM 3.0 on Apple Silicon without PyTorch?

Accepted Answer

Use the MLX backend. It is provided in PR #476 under `src/timesfm3/mlx/` and keeps the existing PyTorch backend under `src/timesfm3/torch/`. The MLX version exposes the same `timesfm3_forecaster` interface, loads the same checkpoints, and matches PyTorch output to about `1e-6` max absolute error. A standalone implementation is available at `https://github.com/rachittshah/mlx-tsfm`.

Quick setup:
```bash
python -m pip install -U mlx
git clone https://github.com/rachittshah/mlx-tsfm.git
cd mlx-tsfm
pytest tests/test_timesfm.py
```

When using the TimesFM repository with PR #476, import the forecaster from `timesfm3.mlx.timesfm3_forecaster` instead of the Torch module and use the same initialization and forecast calls. On an M4 Max, the 330M model reportedly takes about 20 ms per single forecast and reaches roughly 475 forecasts/second when batched.

timesfm

Overview

README Preview

FAQ (4)

同类型项目

browser-use

puppeteer

manim

PaddleOCR