Registry / ai-ml / torchtitan

torchtitan

JSON →
library0.2.2pypypiunverified

A PyTorch native platform for training generative AI models. Current version 0.2.2, release cadence is approximately monthly.

pip install torchtitan
INSTALL
IMPORT
SIG · TORCHTITAN
T
torchtitan
ai-mlpythonv0.2.2
harness data pending
Install & Compatibility
Where this runs

No compatibility data collected yet for this library.

Code
Verified usage

Verified import paths — ran on the pinned version, not inferred.

ParallelDims
from torchtitan.parallelisms import ParallelDims
from torchtitan.parallel import ParallelDims
Module renamed from 'parallel' to 'parallelisms' in v0.2.0
TrainSpec
from torchtitan.config_manager import TrainSpec
from torchtitan import TrainSpec
TrainSpec is not top-level; must import from config_manager
JobConfig
from torchtitan.config_manager import JobConfig
ModelConverter
from torchtitan.models import ModelConverter

Basic setup: configure, define parallelism, and convert model.

import os from torchtitan import JobConfig, ParallelDims, TrainSpec from torchtitan.models import ModelConverter # Initialize config config = JobConfig() config.model = 'llama3' # or 'diffusion' etc. config.hidden_size = 4096 config.num_layers = 32 # Setup parallelism dims = ParallelDims(dp=1, tp=1, pp=1) # Create train spec spec = TrainSpec(model_config=config, parallel_dims=dims) # Convert model for training model = ModelConverter(spec).convert() print(f'Model architecture: {type(model).__name__}')
Debug
Known issues
breakingIn v0.2.0, the import path for parallelism was changed from `torchtitan.parallel` to `torchtitan.parallelisms`. Old code using `from torchtitan.parallel import ParallelDims` will break.
fix
Update import to `from torchtitan.parallelisms import ParallelDims`
affects: <0.2.0
breakingIn v0.2.0, `TrainSpec` was moved from top-level `torchtitan` to `torchtitan.config_manager`. Old import `from torchtitan import TrainSpec` will fail.
fix
Use `from torchtitan.config_manager import TrainSpec`
affects: <0.2.0
gotchaTorchTitan models require specific PyTorch nightly versions. Using stable PyTorch may cause undefined behavior or errors.
fix
Use `pip install torch --pre` or install PyTorch nightly build matching the TorchTitan version.
affects: all
deprecatedThe `ModelConverter` API is being phased out in favor of direct model construction via `torchtitan.models.build_model`. Check release notes.
fix
Prefer `from torchtitan.models import build_model` and use `model = build_model(config)`
affects: >=0.2.0
Upgrade
Version history
0.2.2latest on PyPI · released Feb 20, 2026
Audit
Dependencies
torchrequiredRequired for tensor computation and model definition
torchvisionoptionalRequired for vision-related models (e.g., diffusion)
torchaooptionalRequired for some quantization/optimization features
Agent activity
22 hits · last 30 days
node
20
OpenAI (training)
1
Resources
torchtitan — pip install torchtitan · libregistry