amp.rst

Automatic Mixed Precision package - torch.amp

torch.amp

torch.amp

torch.amp provides convenience methods for mixed precision, where some operations use the torch.float32 (float) datatype and other operations use lower precision floating point datatype (lower_precision_fp): torch.float16 (half) or torch.bfloat16. Some ops, like linear layers and convolutions, are much faster in lower_precision_fp. Other ops, like reductions, often require the dynamic range of float32. Mixed precision tries to match each op to its appropriate datatype.

Ordinarily, "automatic mixed precision training" with datatype of torch.float16 uses torch.autocast and torch.cpu.amp.GradScaler or torch.cuda.amp.GradScaler together, as shown in the CUDA Automatic Mixed Precision examples<amp-examples> and CUDA Automatic Mixed Precision recipe. However, torch.autocast and torch.GradScaler are modular, and may be used separately if desired. As shown in the CPU example section of torch.autocast, "automatic mixed precision training/inference" on CPU with datatype of torch.bfloat16 only uses torch.autocast.

For CUDA and CPU, APIs are also provided separately:

• torch.autocast("cuda", args...) is equivalent to torch.cuda.amp.autocast(args...).
• torch.autocast("cpu", args...) is equivalent to torch.cpu.amp.autocast(args...). For CPU, only lower precision floating point datatype of torch.bfloat16 is supported for now.
• torch.GradScaler("cuda", args...) is equivalent to torch.cuda.amp.GradScaler(args...).
• torch.GradScaler("cpu", args...) is equivalent to torch.cpu.amp.GradScaler(args...).

torch.autocast and torch.cpu.amp.autocast are new in version 1.10.

local

Autocasting

torch.amp.autocast_mode

is_autocast_available

torch

autocast

torch.cuda.amp

autocast

custom_fwd

custom_bwd

torch.cpu.amp

autocast

Gradient Scaling

If the forward pass for a particular op has float16 inputs, the backward pass for that op will produce float16 gradients. Gradient values with small magnitudes may not be representable in float16. These values will flush to zero ("underflow"), so the update for the corresponding parameters will be lost.

To prevent underflow, "gradient scaling" multiplies the network's loss(es) by a scale factor and invokes a backward pass on the scaled loss(es). Gradients flowing backward through the network are then scaled by the same factor. In other words, gradient values have a larger magnitude, so they don't flush to zero.

Each parameter's gradient (.grad attribute) should be unscaled before the optimizer updates the parameters, so the scale factor does not interfere with the learning rate.

Note

AMP/fp16 may not work for every model! For example, most bf16-pretrained models cannot operate in the fp16 numerical range of max 65504 and will cause gradients to overflow instead of underflow. In this case, the scale factor may decrease under 1 as an attempt to bring gradients to a number representable in the fp16 dynamic range. While one may expect the scale to always be above 1, our GradScaler does NOT make this guarantee to maintain performance. If you encounter NaNs in your loss or gradients when running with AMP/fp16, verify your model is compatible.

torch.cuda.amp

GradScaler

Autocast Op Reference

Op Eligibility

Ops that run in float64 or non-floating-point dtypes are not eligible, and will run in these types whether or not autocast is enabled.

Only out-of-place ops and Tensor methods are eligible. In-place variants and calls that explicitly supply an out=... Tensor are allowed in autocast-enabled regions, but won't go through autocasting. For example, in an autocast-enabled region a.addmm(b, c) can autocast, but a.addmm_(b, c) and a.addmm(b, c, out=d) cannot. For best performance and stability, prefer out-of-place ops in autocast-enabled regions.

Ops called with an explicit dtype=... argument are not eligible, and will produce output that respects the dtype argument.

CUDA Op-Specific Behavior

The following lists describe the behavior of eligible ops in autocast-enabled regions. These ops always go through autocasting whether they are invoked as part of a torch.nn.Module, as a function, or as a torch.Tensor method. If functions are exposed in multiple namespaces, they go through autocasting regardless of the namespace.

Ops not listed below do not go through autocasting. They run in the type defined by their inputs. However, autocasting may still change the type in which unlisted ops run if they're downstream from autocasted ops.

If an op is unlisted, we assume it's numerically stable in float16. If you believe an unlisted op is numerically unstable in float16, please file an issue.

CUDA Ops that can autocast to float16

__matmul__, addbmm, addmm, addmv, addr, baddbmm, bmm, chain_matmul, multi_dot, conv1d, conv2d, conv3d, conv_transpose1d, conv_transpose2d, conv_transpose3d, GRUCell, linear, LSTMCell, matmul, mm, mv, prelu, RNNCell

CUDA Ops that can autocast to float32

__pow__, __rdiv__, __rpow__, __rtruediv__, acos, asin, binary_cross_entropy_with_logits, cosh, cosine_embedding_loss, cdist, cosine_similarity, cross_entropy, cumprod, cumsum, dist, erfinv, exp, expm1, group_norm, hinge_embedding_loss, kl_div, l1_loss, layer_norm, log, log_softmax, log10, log1p, log2, margin_ranking_loss, mse_loss, multilabel_margin_loss, multi_margin_loss, nll_loss, norm, normalize, pdist, poisson_nll_loss, pow, prod, reciprocal, rsqrt, sinh, smooth_l1_loss, soft_margin_loss, softmax, softmin, softplus, sum, renorm, tan, triplet_margin_loss

CUDA Ops that promote to the widest input type

These ops don't require a particular dtype for stability, but take multiple inputs and require that the inputs' dtypes match. If all of the inputs are float16, the op runs in float16. If any of the inputs is float32, autocast casts all inputs to float32 and runs the op in float32.

addcdiv, addcmul, atan2, bilinear, cross, dot, grid_sample, index_put, scatter_add, tensordot

Some ops not listed here (e.g., binary ops like add) natively promote inputs without autocasting's intervention. If inputs are a mixture of float16 and float32, these ops run in float32 and produce float32 output, regardless of whether autocast is enabled.

Prefer binary_cross_entropy_with_logits over binary_cross_entropy

The backward passes of torch.nn.functional.binary_cross_entropy (and torch.nn.BCELoss, which wraps it) can produce gradients that aren't representable in float16. In autocast-enabled regions, the forward input may be float16, which means the backward gradient must be representable in float16 (autocasting float16 forward inputs to float32 doesn't help, because that cast must be reversed in backward). Therefore, binary_cross_entropy and BCELoss raise an error in autocast-enabled regions.

Many models use a sigmoid layer right before the binary cross entropy layer. In this case, combine the two layers using torch.nn.functional.binary_cross_entropy_with_logits or torch.nn.BCEWithLogitsLoss. binary_cross_entropy_with_logits and BCEWithLogits are safe to autocast.

CPU Op-Specific Behavior

The following lists describe the behavior of eligible ops in autocast-enabled regions. These ops always go through autocasting whether they are invoked as part of a torch.nn.Module, as a function, or as a torch.Tensor method. If functions are exposed in multiple namespaces, they go through autocasting regardless of the namespace.

Ops not listed below do not go through autocasting. They run in the type defined by their inputs. However, autocasting may still change the type in which unlisted ops run if they're downstream from autocasted ops.

If an op is unlisted, we assume it's numerically stable in bfloat16. If you believe an unlisted op is numerically unstable in bfloat16, please file an issue.

CPU Ops that can autocast to bfloat16

conv1d, conv2d, conv3d, bmm, mm, baddbmm, addmm, addbmm, linear, matmul, _convolution

CPU Ops that can autocast to float32

conv_transpose1d, conv_transpose2d, conv_transpose3d, avg_pool3d, binary_cross_entropy, grid_sampler, grid_sampler_2d, _grid_sampler_2d_cpu_fallback, grid_sampler_3d, polar, prod, quantile, nanquantile, stft, cdist, trace, view_as_complex, cholesky, cholesky_inverse, cholesky_solve, inverse, lu_solve, orgqr, inverse, ormqr, pinverse, max_pool3d, max_unpool2d, max_unpool3d, adaptive_avg_pool3d, reflection_pad1d, reflection_pad2d, replication_pad1d, replication_pad2d, replication_pad3d, mse_loss, ctc_loss, kl_div, multilabel_margin_loss, fft_fft, fft_ifft, fft_fft2, fft_ifft2, fft_fftn, fft_ifftn, fft_rfft, fft_irfft, fft_rfft2, fft_irfft2, fft_rfftn, fft_irfftn, fft_hfft, fft_ihfft, linalg_matrix_norm, linalg_cond, linalg_matrix_rank, linalg_solve, linalg_cholesky, linalg_svdvals, linalg_eigvals, linalg_eigvalsh, linalg_inv, linalg_householder_product, linalg_tensorinv, linalg_tensorsolve, fake_quantize_per_tensor_affine, eig, geqrf, lstsq, _lu_with_info, qr, solve, svd, symeig, triangular_solve, fractional_max_pool2d, fractional_max_pool3d, adaptive_max_pool3d, multilabel_margin_loss_forward, linalg_qr, linalg_cholesky_ex, linalg_svd, linalg_eig, linalg_eigh, linalg_lstsq, linalg_inv_ex

CPU Ops that promote to the widest input type

These ops don't require a particular dtype for stability, but take multiple inputs and require that the inputs' dtypes match. If all of the inputs are bfloat16, the op runs in bfloat16. If any of the inputs is float32, autocast casts all inputs to float32 and runs the op in float32.

cat, stack, index_copy

Some ops not listed here (e.g., binary ops like add) natively promote inputs without autocasting's intervention. If inputs are a mixture of bfloat16 and float32, these ops run in float32 and produce float32 output, regardless of whether autocast is enabled.