LLM Laboratory

Date: 05.01.2026

PyTorch dtype compatibility — quick cheat sheet

The table below shows, dtype in API that we can use in torch_dtype=... for PyTorch versions

PyTorch version	Supported dtype
≥ 1.1	`torch.float16` (`torch.half`) `torch.float32` (`torch.float`) `torch.float64` (`torch.double`)
≥ 1.6	`torch.float16` (`torch.half`) `torch.float32` (`torch.float`) `torch.float64` (`torch.double`) `torch.bfloat16`
≥ 2.2	`torch.float16` (`torch.half`) `torch.float32` (`torch.float`) `torch.float64` (`torch.double`) `torch.bfloat16` `torch.float8_e4m3fn` `torch.float8_e5m2`
≥ 2.9	`torch.float16` (`torch.half`) `torch.float32` (`torch.float`) `torch.float64` (`torch.double`) `torch.bfloat16` `torch.float8_e4m3fn` `torch.float8_e5m2` `torch.float8_e4m3fnuz` `torch.float8_e5m2fnuz` `torch.float8_e8m0fnu` `torch.float4_e2m1fn_x2`