GitHub.com/ggerganov/llama.cpp update (#3696)

* llama.cpp, github version instead of hardcoded version
* llama.cpp, check if model is specified, if yes, run it, if not, then download model
* Use entrypoint for custom llama.cpp invocation
* `llama.cpp` is just raw executable. This I think is our new pattern.
* To run chat use the entrypoint: `pkgx +brewkit -- run llama.cpp`

Co-authored-by: James Reynolds <magnsuviri@me.com>
Co-authored-by: Max Howell <mxcl@me.com>

projects/github.com/ggerganov/llama.cpp/README.md

@@ -1,14 +1,12 @@
 # getting started
 
 ```sh
-$ llama.cpp
+$ pkgx +brewkit -- run llama.cpp
-# ^^ default chat prompt with the OpenLLaMA model
+# ^^ default chat prompt with an appropriate hugging face model
 ```
 
-If you want to run `llama.cpp` with your own args specify them and chat mode
+If you want to run `llama.cpp` with your own args `pkgx llama.cpp $ARGS` is
-will be skipped.
+your friend.
-
-If you want to use a different model specify `--model`.
 
 # converting your own models
 
@@ -16,5 +14,7 @@ We provide a working `convert.py` from the llama.cpp project. To use it you
 need to launch it via a tea pkgenv:
 
 ```sh
-tea +github.com/ggerganov/llama.cpp convert.py path/to/your/model
+pkgx +llama.cpp -- convert.py path/to/your/model
+# ^^ the -- is necessary since `convert.py` is a not listed in the llama.cpp
+# provides list
 ```

projects/github.com/ggerganov/llama.cpp/entrypoint.sh

@@ -0,0 +1,32 @@
+#!/usr/bin/env -S pkgx +gum +aria2c bash
+
+set -eo pipefail
+test -n "$VERBOSE" && set -x
+
+MODEL_URL="https://huggingface.co/TheBloke/dolphin-2.1-mistral-7B-GGUF/resolve/main/dolphin-2.1-mistral-7b.Q4_0.gguf"
+MODEL_FILENAME=$(basename "$MODEL_URL")
+MODEL_DIR="${XDG_DATA_HOME:-$HOME/.local/share}"/llama.cpp
+
+if [ ! -f "$MODEL_DIR/$MODEL_FILENAME" ]; then
+  gum format <<EoMD
+  # downloading $MODEL_FILENAME
+  models will be placed: \`$PWD\`
+  > this may take a a few minutes…
+EoMD
+  echo #spacer
+  mkdir -p "$MODEL_DIR"
+  aria2c "$MODEL_URL" --dir="$MODEL_DIR"
+  gum format "# All done!"
+  echo #spacer
+fi
+
+D="$(cd "$(dirname "$0")" && pwd)"
+
+exec "$D"/bin/llama.cpp \
+    --model "$MODEL_DIR/$MODEL_FILENAME" \
+    -n 256 \
+    --repeat_penalty 1.0 \
+    --color \
+    --interactive \
+    --reverse-prompt "User:" \
+    --file "$D"/share/prompts/chat-with-bob.txt

projects/github.com/ggerganov/llama.cpp/llama-fetch

@@ -1,31 +0,0 @@
-#!/bin/sh
-
-set -e
-test -n "$VERBOSE" && set -x
-
-if test -f "$1"/VERSION && test $(cat "$1"/VERSION) = $2; then
-  exit
-fi
-
-mkdir -p "$1"
-cd "$1"
-
-tea gum format <<EoMD
-# downloading OpenLLaMA 3Bv2
-models will be placed: \`$PWD\`
-> this may take a a few minutes…
-EoMD
-
-tea wget \
-  --continue \
-  --output-document OpenLLaMA-3Bv2.ggml.f16.bin \
-  'https://huggingface.co/SlyEcho/open_llama_3b_v2_ggml/resolve/main/open-llama-3b-v2-f16.bin'
-
-tea wget \
-  --continue \
-  'https://huggingface.co/openlm-research/open_llama_3b_v2/resolve/main/tokenizer.model'
-
-echo $2 > VERSION
-
-tea gum format "# All done!"
-echo  #spacer

projects/github.com/ggerganov/llama.cpp/llama.cpp

@@ -1,37 +0,0 @@
-#!/bin/sh
-
-set -e
-test -n "$VERBOSE" && set -x
-
-D="$(cd "$(dirname "$0")"/.. && pwd)"
-VERSION="$(basename "$D")"
-MODEL_DIR="${XDG_DATA_HOME:-$HOME/.local/share}/models/OpenLLaMA"
-
-"$D"/libexec/llama-fetch "$MODEL_DIR" "$VERSION"
-
-if [ "$1" = '--fetch' ]; then
-  exit
-fi
-
-if [ $# -eq 0 ]; then
-  exec "$D"/libexec/llama.cpp \
-    --model "$MODEL_DIR/OpenLLaMA-3Bv2.ggml.f16.bin" \
-    --ctx-size 512 \
-    --batch-size 1024 \
-    --n-predict 256 \
-    --keep 48 \
-    --repeat_penalty 1.0 \
-    --color \
-    --interactive \
-    --reverse-prompt "User:" \
-    --file "$D"/share/prompts/chat-with-bob.txt
-fi
-
-# detect --model arg or not
-for arg in "$@"; do
-  if [ "$arg" = "--model" -o  "$arg" = "-m" ]; then
-    exec "$D"/libexec/llama.cpp "$@"
-  fi
-done
-
-exec "$D"/libexec/llama.cpp "$@" --model "$MODEL_DIR/OpenLLaMA-3Bv2.ggml.f16.bin"

projects/github.com/ggerganov/llama.cpp/package.yml

@@ -1,9 +1,10 @@
 distributable:
-  url: https://github.com/ggerganov/llama.cpp/archive/refs/tags/master-fff0e0e.tar.gz
+  url: https://github.com/ggerganov/llama.cpp/archive/refs/tags/b{{version.raw}}.tar.gz
   strip-components: 1
 
 versions:
-  - 2023.07.20
+  github: ggerganov/llama.cpp/tags
+  strip: /^b/
 
 display-name:
   LLaMA.cpp
@@ -20,7 +21,7 @@ platforms:
 
 dependencies:
   python.org: ~3.11
-  tea.xyz: ^0  # the scripts use tea/cli
+  pkgx.sh: ^1
 
 build:
   dependencies:
@@ -28,28 +29,29 @@ build:
   env:
     VIRTUAL_ENV: ${{prefix}}/venv
   script:
+    - make --jobs {{hw.concurrency}}
+
     - |
-      make --jobs {{hw.concurrency}}
+      install -D main {{prefix}}/bin/llama.cpp
-      install -D main {{prefix}}/libexec/llama.cpp
+      install -D props/entrypoint.sh {{prefix}}/entrypoint.sh
-      install -D props/llama.cpp {{prefix}}/bin/llama.cpp
+      install -D ggml-metal.metal {{prefix}}/bin/ggml-metal.metal
 
     - |
       mkdir -p {{prefix}}/share
       mv prompts {{prefix}}/share
-      mv props/llama-fetch {{prefix}}/libexec
 
     - |
       install -D convert.py $VIRTUAL_ENV/bin/convert.py
       python-venv-stubber.sh convert.py
 
-    - python -m venv $VIRTUAL_ENV
     - |
+      python -m venv $VIRTUAL_ENV
       source $VIRTUAL_ENV/bin/activate
       pip install -r requirements.txt
       deactivate
 
-test: |
+test:
-  {{prefix}}/libexec/llama.cpp --help
+  '{{prefix}}/bin/llama.cpp --help'
-  # testing more than this requires downloading the models 😬
+  # ^^ testing more than this requires downloading the models 😬
 
-entrypoint: llama.cpp
+entrypoint: ./entrypoint.sh