Code

ggml : add graph tensor allocator#2411

Code

Merged

slaren merged 13 commits intomasterggerganov/llama.cpp:masterfrom ggml-allocatorggerganov/llama.cpp:ggml-allocatorCopy head branch name to clipboardJul 30, 2023

+812-88

Commits

Commits on Jul 27, 2023

• ggml : add graph tensor allocator

  slaren

  committed
  768ecfc

  Copy full SHA for 768ecfc

• adjust buffer size to account for alignment

  slaren

  committed
  598a9ad

  Copy full SHA for 598a9ad

• fix mpi build

  slaren

  committed
  8afe392

  Copy full SHA for 8afe392

• allow using the allocator with opencl

  slaren

  committed
  8fa5483

  Copy full SHA for 8fa5483

• add list of ops that support in-place

  slaren

  committed
  f67179a

  Copy full SHA for f67179a

• ggml : don't calculate data pointer of unallocated tensors when creating a view with an offset

  slaren

  committed
  64584d5

  Copy full SHA for 64584d5

• llama.cpp : free allocator when deleting context, cleanup

  slaren

  committed
  af7bd42

  Copy full SHA for af7bd42

• replace n_views and n_children in ggml_tensor with a hash table in the allocator

  slaren

  committed
  e39e62b

  Copy full SHA for e39e62b

• ggml : refactor ggml_view_Nd into ggml_view_tensor_offset

  slaren

  committed
  e592a17

  Copy full SHA for e592a17

• llama.cpp : fix embeddings output

  slaren

  committed
  ba0ab56

  Copy full SHA for ba0ab56

• llama.cpp : fix embeddings input

  slaren

  committed
  966c069

  Copy full SHA for 966c069

• llama.cpp : better memory usage prints with allocator

  slaren

  committed
  cd4a8cd

  Copy full SHA for cd4a8cd

Commits on Jul 30, 2023

• rename ggml_allocator to ggml_allocr

  Show description for 570aa7c
  slaren

  committed
  570aa7c

  Copy full SHA for 570aa7c