From 18956baad9cde85bc55bd0113092e51786d9e1ee Mon Sep 17 00:00:00 2001
From: Ben Sima <ben@bsima.me>
Date: Mon, 2 Dec 2024 07:52:56 -0500
Subject: Delete Biz/Mynion.py and exllama

Mynion was a prototype and while it was cool and worked well, it is unused and
causing magma to build, which takes forever. I have settled on using ollama for
local inference and a hosted inference API for production.
---
 Biz/Bild/Deps/exllama.nix | 54 -----------------------------------------------
 Biz/Bild/Python.nix       |  2 --
 Biz/Bild/Sources.json     | 12 -----------
 3 files changed, 68 deletions(-)
 delete mode 100644 Biz/Bild/Deps/exllama.nix

(limited to 'Biz/Bild')

diff --git a/Biz/Bild/Deps/exllama.nix b/Biz/Bild/Deps/exllama.nix
deleted file mode 100644
index 434e9a9..0000000
--- a/Biz/Bild/Deps/exllama.nix
+++ /dev/null
@@ -1,54 +0,0 @@
-{ lib, sources, buildPythonPackage, pythonOlder
-, torch # tested on 2.0.1 and 2.1.0 (nightly) with cu118
-, safetensors, sentencepiece, ninja, cudaPackages, addOpenGLRunpath, which
-, libGL, gcc11 # cuda 11.7 requires g++ <12
-}:
-
-buildPythonPackage rec {
-  pname = "exllama";
-  version = sources.exllama.rev;
-  format = "setuptools";
-  disabled = pythonOlder "3.9";
-
-  src = sources.exllama;
-
-  # I only care about compiling for the Ampere architecture, which is what my
-  # RTX 3090 TI is, and for some reason (nix sandbox?) the torch extension
-  # builder
-  # cannot autodetect the arch
-  TORCH_CUDA_ARCH_LIST = "8.0;8.6+PTX";
-
-  CUDA_HOME = "${cudaPackages.cuda_nvcc}";
-
-  nativeBuildInputs = [
-    gcc11
-    which
-    libGL
-    addOpenGLRunpath
-    cudaPackages.cuda_nvcc
-    cudaPackages.cuda_cudart
-  ];
-
-  propagatedBuildInputs =
-    [ torch safetensors sentencepiece ninja cudaPackages.cudatoolkit ];
-
-  doCheck = false; # no tests currently
-  pythonImportsCheck = [
-    "exllama"
-    "exllama.cuda_ext"
-    "exllama.generator"
-    "exllama.lora"
-    "exllama.model"
-    "exllama.tokenizer"
-  ];
-
-  meta = with lib; {
-    description = ''
-      A more memory-efficient rewrite of the HF transformers implementation of
-      Llama for use with quantized weights.
-    '';
-    homepage = "https://github.com/jllllll/exllama";
-    license = licenses.mit;
-    maintainers = with maintainers; [ bsima ];
-  };
-}
diff --git a/Biz/Bild/Python.nix b/Biz/Bild/Python.nix
index 2385987..50a1779 100644
--- a/Biz/Bild/Python.nix
+++ b/Biz/Bild/Python.nix
@@ -4,8 +4,6 @@ _self: super: {
       with pysuper.pkgs.python3Packages;
       let dontCheck = p: p.overridePythonAttrs (_: { doCheck = false; });
       in {
-        exllama = callPackage ./Deps/exllama.nix { };
-        exllamav2 = callPackage ./Deps/exllamav2.nix { };
         interegular = callPackage ./Deps/interegular.nix { };
         llm-ollama = callPackage ./Deps/llm-ollama.nix { };
         mypy = dontCheck pysuper.mypy;
diff --git a/Biz/Bild/Sources.json b/Biz/Bild/Sources.json
index 3b1e4fd..c12b6ce 100644
--- a/Biz/Bild/Sources.json
+++ b/Biz/Bild/Sources.json
@@ -25,18 +25,6 @@
         "url_template": "https://github.com/<owner>/<repo>/archive/<rev>.tar.gz",
         "version": "0.7.0.8"
     },
-    "exllama": {
-        "branch": "master",
-        "description": "A more memory-efficient rewrite of the HF transformers implementation of Llama for use with quantized weights.",
-        "homepage": null,
-        "owner": "jllllll",
-        "repo": "exllama",
-        "rev": "3ddf3bd39bdff330623f3740cda4ae1537ef86d9",
-        "sha256": "0g87xm71jmw5bl4ya5dbk72fghhhwvrjqspaayq7zass16jixr1d",
-        "type": "tarball",
-        "url": "https://github.com/jllllll/exllama/archive/3ddf3bd39bdff330623f3740cda4ae1537ef86d9.tar.gz",
-        "url_template": "https://github.com/<owner>/<repo>/archive/<rev>.tar.gz"
-    },
     "ghc-exactprint": {
         "branch": "master",
         "description": "GHC version of haskell-src-exts exactPrint",
-- 
cgit v1.2.3