Skills › Scientific Agent Skills › timesfm-forecasting › system_requirements.md

system_requirements.md

59 lines · skills/timesfm-forecasting/references/system_requirements.md at 5f2bc03

Runs code

Open on GitHub

Most marks come from matching words in the file, so treat them as pointers.

1# Resource and backend checks

3Run `python scripts/check_system.py --model v2.5` before initial model loading.

4Use `--json` for machine-readable output; exit 1 indicates a failed threshold.

5`--model v3.0` provides a separate resource estimate; it does not grant permission

6to use restricted weights or validate the 3.0 API.

8The checker measures total/available RAM, CUDA or MPS availability, free space

9on the configured Hub cache volume, Python and package importability. On macOS

10`vm_stat` uses the page size reported by the OS (often 16,384 bytes), not a fixed

114,096-byte multiplier. Unknown measurements warn instead of fabricating RAM.

13Thresholds are local heuristics, not published hardware minima or peak-RSS

14predictions. For 2.5 they block below 2 GiB available RAM or 2 GiB free cache

15space and warn below 4 GiB available RAM. Batch size recommendation is always 1

16until the user measures the actual workload. Temporary copies, compiler caches,

17activations, other processes and model/backend changes may require much more.

18No preflight can guarantee the process will avoid OOM.

Resource and backend checks

Run python scripts/check_system.py --model v2.5 before initial model loading. Use --json for machine-readable output; exit 1 indicates a failed threshold. --model v3.0 provides a separate resource estimate; it does not grant permission to use restricted weights or validate the 3.0 API.

The checker measures total/available RAM, CUDA or MPS availability, free space on the configured Hub cache volume, Python and package importability. On macOS vm_stat uses the page size reported by the OS (often 16,384 bytes), not a fixed 4,096-byte multiplier. Unknown measurements warn instead of fabricating RAM.

Thresholds are local heuristics, not published hardware minima or peak-RSS predictions. For 2.5 they block below 2 GiB available RAM or 2 GiB free cache space and warn below 4 GiB available RAM. Batch size recommendation is always 1 until the user measures the actual workload. Temporary copies, compiler caches, activations, other processes and model/backend changes may require much more. No preflight can guarantee the process will avoid OOM.

Raw lines 1 to 19

Runs code python scripts/check_system.py :3

20## Devices

22The 2.5 torch loader automatically uses CUDA:0 when CUDA is available, otherwise

23CPU. It does not auto-select Apple MPS even if PyTorch detects it. MPS detection

24is reported as informational/warning, and the effective 2.5 mode stays CPU.

25A tiny CUDA device may still OOM; its reported VRAM is total, not necessarily

26free. Multi-device count affects the upstream batch calculation; do not infer

27that this wrapper distributes weights across GPUs.

293.0 supports an explicit torch device and a separate MLX backend. See

30`timesfm3.md`. Neither GPU nor MLX performance was tested in this refresh.

Devices

The 2.5 torch loader automatically uses CUDA:0 when CUDA is available, otherwise CPU. It does not auto-select Apple MPS even if PyTorch detects it. MPS detection is reported as informational/warning, and the effective 2.5 mode stays CPU. A tiny CUDA device may still OOM; its reported VRAM is total, not necessarily free. Multi-device count affects the upstream batch calculation; do not infer that this wrapper distributes weights across GPUs.

3.0 supports an explicit torch device and a separate MLX backend. See timesfm3.md. Neither GPU nor MLX performance was tested in this refresh.

Raw lines 20 to 31

32## Install and cache

34Package 3.0.2 declares Python >=3.10, NumPy >=1.26.4,

35huggingface_hub >=0.28.0 and safetensors >=0.5.3. Its torch extra declares

36PyTorch >=2.0.0. The tested CPU stack uses Python 3.13, PyTorch 2.14.1,

37NumPy 2.5.3 and pandas 3.0.6. Platform wheel/interpreter support depends on

38those selected releases; do not equate the package's minimum with every backend's.

40XReg needs JAX and scikit-learn, not Flax. Current JAX 0.11.2 requires Python

41>=3.12. On macOS install CPU `jax`, not `jax[cuda]`. The upstream Flax extra

42also requests CUDA JAX; a CPU Flax setup needs individually compatible packages

43and was not runtime-tested here. Matplotlib/Pillow are for optional plots/GIFs.

45Cache precedence is `HF_HUB_CACHE`, otherwise `HF_HOME/hub`, otherwise

46`XDG_CACHE_HOME/huggingface/hub` (default `~/.cache/huggingface/hub`). The checker

47uses the nearest existing ancestor to check the correct filesystem before a

48new directory exists. The 2.5 float32 checkpoint is roughly 0.8 GB; this excludes

49framework packages, temporary files and compiler artifacts.

51```bash

52# Public model; network/download needed only when it is not already cached.

53hf download google/timesfm-2.5-200m-pytorch \

54 --revision 1d952420fba87f3c6dee4f240de0f1a0fbc790e3

55```

57Use `local_files_only=True` for offline loading of a populated cache. Do not

58force-download a checkpoint routinely. The [Hub CLI](https://huggingface.co/docs/huggingface_hub/en/guides/cli)

59is `hf`, replacing the old `huggingface-cli` spelling.

Install and cache

Package 3.0.2 declares Python >=3.10, NumPy >=1.26.4, huggingface_hub >=0.28.0 and safetensors >=0.5.3. Its torch extra declares PyTorch >=2.0.0. The tested CPU stack uses Python 3.13, PyTorch 2.14.1, NumPy 2.5.3 and pandas 3.0.6. Platform wheel/interpreter support depends on those selected releases; do not equate the package's minimum with every backend's.

XReg needs JAX and scikit-learn, not Flax. Current JAX 0.11.2 requires Python

=3.12. On macOS install CPU jax, not jax[cuda]. The upstream Flax extra

also requests CUDA JAX; a CPU Flax setup needs individually compatible packages and was not runtime-tested here. Matplotlib/Pillow are for optional plots/GIFs.

Cache precedence is HF_HUB_CACHE, otherwise HF_HOME/hub, otherwise XDG_CACHE_HOME/huggingface/hub (default ~/.cache/huggingface/hub). The checker uses the nearest existing ancestor to check the correct filesystem before a new directory exists. The 2.5 float32 checkpoint is roughly 0.8 GB; this excludes framework packages, temporary files and compiler artifacts.

# Public model; network/download needed only when it is not already cached.
hf download google/timesfm-2.5-200m-pytorch \
  --revision 1d952420fba87f3c6dee4f240de0f1a0fbc790e3

Use local_files_only=True for offline loading of a populated cache. Do not force-download a checkpoint routinely. The Hub CLI is hf, replacing the old huggingface-cli spelling.

Raw lines 32 to 59