ggml_flash_attn_back

Backward pass for Flash Attention.
Used during training to compute gradients through attention.

Provides 'R' bindings to the 'GGML' tensor library for machine
learning, designed primarily for 'Vulkan' GPU acceleration with full CPU
fallback. 'Vulkan' support is auto-detected at build time on Linux (when
'libvulkan-dev' and 'glslc' are installed) and on Windows (when 'Vulkan'
'SDK' is installed and 'VULKAN_SDK' environment variable is set); all
operations fall back to CPU transparently when no GPU is available.
Implements tensor operations, neural network layers, quantization, and a
'Keras'-like sequential model API for building and training networks.
Includes 'AdamW' (Adam with Weight decay) and 'SGD' (Stochastic Gradient
Descent) optimizers with 'MSE' (Mean Squared Error) and cross-entropy
losses. Also provides a dynamic 'autograd' engine ('PyTorch'-style) with
data-parallel training via 'dp_train()', broadcast arithmetic, 'f16'
(half-precision) support on 'Vulkan' GPU, and a multi-head attention layer
for building Transformer architectures. Serves as backend for 'LLM' (Large
Language Model) inference via 'llamaR' and Stable Diffusion image
generation via 'sdR'. See <https://github.com/ggml-org/ggml> for more
information about the underlying library.

Yuri Baramykov

ggmlR

'GGML' Tensor Operations for Machine Learning

Georgi Gerganov

Jeffrey Quesnelle

Bowen Peng

Mozilla Foundation 

ggml_flash_attn_back function

<dl><dt>ctx</dt>
<dd>GGML context</dd>
<dt>q</dt>
<dd>Query tensor (same as forward pass)</dd>
<dt>k</dt>
<dd>Key tensor (same as forward pass)</dd>
<dt>v</dt>
<dd>Value tensor (same as forward pass)</dd>
<dt>d</dt>
<dd>Gradient tensor from upstream (same shape as forward output)</dd>
<dt>masked</dt>
<dd>Logical: whether causal masking was used in forward pass</dd></dl>

Arguments

Flash Attention Backward (Graph) — ggml_flash_attn_back

<dl>

<dt>ctx</dt>
<dd>GGML context</dd>


<dt>q</dt>
<dd>Query tensor (same as forward pass)</dd>


<dt>k</dt>
<dd>Key tensor (same as forward pass)</dd>


<dt>v</dt>
<dd>Value tensor (same as forward pass)</dd>


<dt>d</dt>
<dd>Gradient tensor from upstream (same shape as forward output)</dd>


<dt>masked</dt>
<dd>Logical: whether causal masking was used in forward pass</dd>

</dl>

ggml_flash_attn_back: Flash Attention Backward (Graph)

Description

Usage

Value

Arguments