TF32训练
==================

InternLM支持使用TF32训练模型。TensorFloat-32（TF32）是Nvidia在Ampere架构GPU上推出的专门运用于TensorCore的一种计算格式。其与其他常用数据格式的比较如下图：

InternLM supports training models using TF32. TensorFloat-32 (TF32) is a computational format introduced by Nvidia for TensorCores on Ampere architecture GPUs. Here's a comparison of TF32 with other data formats:

.. figure:: ../../imgs/tf32.png
   :scale: 50%
   :class: with-border

使用TF32的前置条件：

Prerequisites for using TF32:
 input data must be of type FP32 (single-precision floating-point) and the computations should be matrix multiplication, convolution and so on. 
1. 输入数据类型为FP32，且计算为矩阵乘法及卷积相关运算，才可以使用TF32作为TensorCore的中间计算类型。

Ampere GPU
2. Ampere架构的GPU。


值得注意的是，TF32仅仅是在使用TensorCore时的一种中间计算格式，并不是一个完全的数据类型。因此，为了区分不同的精度与计算格式（ ``BF16`` 、``FP16`` 、``FP32`` 、``TF32`` ），InternLM支持用户在 ``model config`` 中传入 ``torch.tf32`` 来表示想要使用TF32加速运算，本质上数据类型依旧为 ``FP32``。

It is noticed that TF32 is an intermediate calculation format when employing TensorCores. InternLM allows users to speficy ``torch.tf32`` in the model config to using TF32 acceleration while dtype is still ``torch.float32``.

.. code-block:: python

    model = dict(
        checkpoint=False,  # The proportion of layers for activation aheckpointing, the optional value are True/False/[0-1]
        num_attention_heads=NUM_ATTENTION_HEAD,
        embed_split_hidden=True,
        vocab_size=VOCAB_SIZE,
        embed_grad_scale=1,
        parallel_output=True,
        hidden_size=HIDDEN_SIZE,
        num_layers=NUM_LAYER,
        mlp_ratio=MLP_RATIO,
        apply_post_layer_norm=False,
        dtype="torch.tf32",  # Support: "torch.float16", "torch.half", "torch.bfloat16", "torch.float32", "torch.tf32"
        norm_type="rmsnorm",
        layer_norm_epsilon=1e-5,
        use_flash_attn=True,
        num_chunks=1,  # if num_chunks > 1, interleaved pipeline scheduler is used.
    )

InternLM会根据 ``model config`` 中的 ``dtype`` 字符串来判断真正的数据类型。InternLM通过设置以下变量来开启TF32训练。
InternLM enables TF32 training by setting the following variables.

.. code-block:: python

    torch.backends.cudnn.allow_tf32 = True
    torch.backends.cuda.matmul.allow_tf32 = True