public_git
/
whisper.cpp
mirror of https://github.com/ggml-org/whisper.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Merge remote-tracking branch 'origin/master' into stream-pcm

This commit is contained in:
Ross Morsali 2026-03-16 14:27:44 +01:00
parent 02de44819b 79218f51d0
commit 200bcecbb1
621 changed files with 74698 additions and 22697 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
.devops/main-vulkan.Dockerfile Normal file
View File

@ -0,0 +1,20 @@
FROM ubuntu:24.04 AS build
WORKDIR /app
RUN apt-get update && \
apt-get install -y build-essential wget cmake git libvulkan-dev glslc \
&& rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
COPY .. .
RUN make base.en CMAKE_ARGS="-DGGML_VULKAN=1"
FROM ubuntu:24.04 AS runtime
WORKDIR /app
RUN apt-get update && \
apt-get install -y curl ffmpeg libsdl2-dev wget cmake git libvulkan1 mesa-vulkan-drivers \
&& rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
COPY --from=build /app /app
ENV PATH=/app/build/bin:$PATH
ENTRYPOINT [ "bash", "-c" ]

4
.github/workflows/bindings-go.yml vendored
View File

@ -13,10 +13,10 @@ jobs:
ubuntu-22:
runs-on: ubuntu-22.04
steps:
- uses: actions/setup-go@v5
- uses: actions/setup-go@v6
with:
go-version: '^1.23'
- uses: actions/checkout@v4
- uses: actions/checkout@v6
- run: |
cd bindings/go
make test

2
.github/workflows/bindings-ruby.yml vendored
View File

@ -17,5 +17,5 @@ jobs:
- uses: ruby/setup-ruby@v1
with:
ruby-version: '3.2'
- uses: actions/checkout@v4
- uses: actions/checkout@v6
- run: rake test

138
.github/workflows/build.yml vendored
View File

@ -67,7 +67,7 @@ jobs:
steps:
- name: Checkout with full history
uses: actions/checkout@v4
uses: actions/checkout@v6
with:
fetch-depth: 0
@ -127,7 +127,7 @@ jobs:
steps:
- name: Clone
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: Set up QEMU
uses: docker/setup-qemu-action@v3
@ -159,7 +159,7 @@ jobs:
steps:
- name: Clone
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: Set up QEMU
uses: docker/setup-qemu-action@v3
@ -174,10 +174,6 @@ jobs:
sed -i "s|archive.ubuntu.com|mirrors.kernel.org|g" /etc/apt/sources.list
sed -i "s|security.ubuntu.com|mirrors.kernel.org|g" /etc/apt/sources.list
apt-get update
apt-get install -y ca-certificates
sed -i "s|http://ports.ubuntu.com|https://mirror.kumi.systems|g" /etc/apt/sources.list
apt update
apt install -y build-essential libsdl2-dev cmake git
cmake -B build -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=armv8-a
@ -195,7 +191,7 @@ jobs:
steps:
- name: Clone
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: Set up QEMU
uses: docker/setup-qemu-action@v3
@ -210,10 +206,6 @@ jobs:
sed -i "s|archive.ubuntu.com|mirrors.kernel.org|g" /etc/apt/sources.list
sed -i "s|security.ubuntu.com|mirrors.kernel.org|g" /etc/apt/sources.list
apt-get update
apt-get install -y ca-certificates
sed -i "s|http://ports.ubuntu.com|https://mirror.kumi.systems|g" /etc/apt/sources.list
apt update
apt install -y build-essential libsdl2-dev cmake git
cmake -B build -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=armv7-a+fp
@ -231,7 +223,7 @@ jobs:
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: ccache
uses: hendrikmuhs/ccache-action@v1.2.16
@ -263,7 +255,7 @@ jobs:
#
# steps:
# - name: Clone
# uses: actions/checkout@v4
# uses: actions/checkout@v6
#
# - name: Build
# uses: cross-platform-actions/action@v0.27.0
@ -289,7 +281,7 @@ jobs:
steps:
- name: Clone
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: Set up QEMU
uses: docker/setup-qemu-action@v3
@ -323,7 +315,7 @@ jobs:
steps:
- name: Clone
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: Set up QEMU
uses: docker/setup-qemu-action@v3
@ -338,10 +330,6 @@ jobs:
sed -i "s|archive.ubuntu.com|mirrors.kernel.org|g" /etc/apt/sources.list
sed -i "s|security.ubuntu.com|mirrors.kernel.org|g" /etc/apt/sources.list
apt-get update
apt-get install -y ca-certificates
sed -i "s|http://ports.ubuntu.com|https://mirror.kumi.systems|g" /etc/apt/sources.list
apt update
apt install -y build-essential cmake libsdl2-dev git
cmake . -DWHISPER_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }} -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=armv8-a
@ -361,7 +349,7 @@ jobs:
steps:
- name: Clone
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: Set up QEMU
uses: docker/setup-qemu-action@v3
@ -376,10 +364,6 @@ jobs:
sed -i "s|archive.ubuntu.com|mirrors.kernel.org|g" /etc/apt/sources.list
sed -i "s|security.ubuntu.com|mirrors.kernel.org|g" /etc/apt/sources.list
apt-get update
apt-get install -y ca-certificates
sed -i "s|http://ports.ubuntu.com|https://mirror.kumi.systems|g" /etc/apt/sources.list
apt update
apt install -y build-essential cmake libsdl2-dev git
cmake . -DWHISPER_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }} -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=armv7-a+fp
@ -402,7 +386,7 @@ jobs:
steps:
- name: Clone
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: Set up QEMU
uses: docker/setup-qemu-action@v3
@ -417,10 +401,6 @@ jobs:
sed -i "s|archive.ubuntu.com|mirrors.kernel.org|g" /etc/apt/sources.list
sed -i "s|security.ubuntu.com|mirrors.kernel.org|g" /etc/apt/sources.list
apt-get update
apt-get install -y ca-certificates
sed -i "s|http://ports.ubuntu.com|https://mirror.kumi.systems|g" /etc/apt/sources.list
apt update
apt install -y clang build-essential cmake libsdl2-dev git
cmake . -DWHISPER_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }} -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_COMPILER=clang
@ -440,7 +420,7 @@ jobs:
steps:
- name: Clone
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: Set up QEMU
uses: docker/setup-qemu-action@v3
@ -480,7 +460,7 @@ jobs:
steps:
- name: Clone
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: add oneAPI to apt
shell: bash
@ -504,7 +484,7 @@ jobs:
- name: Clone
id: checkout
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: Build
id: cmake_build
@ -532,7 +512,7 @@ jobs:
steps:
- name: Clone
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: add oneAPI to apt
shell: bash
@ -556,7 +536,7 @@ jobs:
- name: Clone
id: checkout
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: Build
id: cmake_build
@ -581,7 +561,7 @@ jobs:
steps:
- name: Clone
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: Setup ${{ matrix.sys }}
uses: msys2/setup-msys2@v2
@ -636,7 +616,7 @@ jobs:
steps:
- name: Clone
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: Add msbuild to PATH
uses: microsoft/setup-msbuild@v2
@ -666,31 +646,31 @@ jobs:
- name: Upload SDL2.dll
if: matrix.sdl2 == 'ON'
uses: actions/upload-artifact@v4
uses: actions/upload-artifact@v6
with:
name: ${{ matrix.s2arc }}_SDL2.dll
path: build/bin/${{ matrix.build }}/SDL2.dll
- name: Upload whisper dll
uses: actions/upload-artifact@v4
uses: actions/upload-artifact@v6
with:
name: whisper_${{ matrix.arch }}.dll
path: build/bin/${{ matrix.build }}/whisper.dll
- name: Upload ggml dll
uses: actions/upload-artifact@v4
uses: actions/upload-artifact@v6
with:
name: ggml_${{ matrix.arch }}.dll
path: build/bin/${{ matrix.build }}/ggml.dll
- name: Upload ggml base dll
uses: actions/upload-artifact@v4
uses: actions/upload-artifact@v6
with:
name: ggml_base_${{ matrix.arch }}.dll
path: build/bin/${{ matrix.build }}/ggml-base.dll
- name: Upload ggml cpu dll
uses: actions/upload-artifact@v4
uses: actions/upload-artifact@v6
with:
name: ggml_cpu_${{ matrix.arch }}.dll
path: build/bin/${{ matrix.build }}/ggml-cpu.dll
@ -702,7 +682,7 @@ jobs:
- name: Upload binaries
if: matrix.sdl2 == 'ON' && ${{ needs.determine-tag.outputs.should_release }}
uses: actions/upload-artifact@v4
uses: actions/upload-artifact@v6
with:
name: whisper-bin-${{ matrix.arch }}.zip
path: whisper-bin-${{ matrix.arch }}.zip
@ -731,10 +711,10 @@ jobs:
steps:
- name: Clone
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: Export GitHub Actions cache environment variables
uses: actions/github-script@v7
uses: actions/github-script@v8
with:
script: |
core.exportVariable('ACTIONS_CACHE_URL', process.env.ACTIONS_CACHE_URL || '');
@ -788,7 +768,7 @@ jobs:
- name: Upload binaries
if: matrix.blas == 'ON' && matrix.sdl2 == 'ON' && ${{ needs.determine-tag.outputs.should_release }}
uses: actions/upload-artifact@v4
uses: actions/upload-artifact@v6
with:
name: whisper-blas-bin-${{ matrix.arch }}.zip
path: whisper-blas-bin-${{ matrix.arch }}.zip
@ -812,7 +792,7 @@ jobs:
sdl2_ver: 2.28.5
steps:
- name: Clone repository
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: Install Ninja
id: install_ninja
@ -997,7 +977,7 @@ jobs:
- name: Upload binaries
if: ${{ needs.determine-tag.outputs.should_release }}
uses: actions/upload-artifact@v4
uses: actions/upload-artifact@v6
with:
name: whisper-cublas-${{ matrix.cuda-toolkit }}-bin-${{ matrix.arch }}.zip
path: whisper-cublas-${{ matrix.cuda-toolkit }}-bin-${{ matrix.arch }}.zip
@ -1013,7 +993,7 @@ jobs:
steps:
- name: Clone
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: Setup emsdk
uses: mymindstorm/setup-emsdk@v14
@ -1036,7 +1016,7 @@ jobs:
steps:
- name: Checkout code
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: Configure
run: |
@ -1078,7 +1058,7 @@ jobs:
- name: Upload artifacts
if: ${{ needs.determine-tag.outputs.should_release }}
uses: actions/upload-artifact@v4
uses: actions/upload-artifact@v6
with:
path: whisper-${{ needs.determine-tag.outputs.tag_name }}-xcframework.zip
name: whisper-${{ needs.determine-tag.outputs.tag_name }}-xcframework.zip
@ -1090,12 +1070,12 @@ jobs:
steps:
- name: Clone
uses: actions/checkout@v4
uses: actions/checkout@v6
with:
path: whisper
- name: Install Java
uses: actions/setup-java@v4
uses: actions/setup-java@v5
with:
distribution: zulu
java-version: 21
@ -1119,10 +1099,10 @@ jobs:
steps:
- name: Clone
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: set up JDK 11
uses: actions/setup-java@v4
uses: actions/setup-java@v5
with:
java-version: '11'
distribution: 'temurin'
@ -1145,36 +1125,36 @@ jobs:
needs: ['windows']
runs-on: windows-latest
steps:
- uses: actions/checkout@v4
- uses: actions/checkout@v6
- name: Install Java
uses: actions/setup-java@v4
uses: actions/setup-java@v5
with:
distribution: zulu
java-version: 20
- name: Download Whisper Windows lib
uses: actions/download-artifact@v4
uses: actions/download-artifact@v7
with:
name: whisper_x64.dll
- name: Download GGML Windows lib
uses: actions/download-artifact@v4
uses: actions/download-artifact@v7
with:
name: ggml_x64.dll
- name: Download GGML Base Windows lib
uses: actions/download-artifact@v4
uses: actions/download-artifact@v7
with:
name: ggml_base_x64.dll
- name: Download GGML CPU Windows lib
uses: actions/download-artifact@v4
uses: actions/download-artifact@v7
with:
name: ggml_cpu_x64.dll
- name: Download SDL2.dll
uses: actions/download-artifact@v4
uses: actions/download-artifact@v7
with:
name: x64_SDL2.dll
@ -1221,7 +1201,7 @@ jobs:
Compress-Archive -Path "bindings/java/build/libs/whispercpp-*.jar" -DestinationPath "whispercpp.jar.zip"
- name: Upload jar
uses: actions/upload-artifact@v4
uses: actions/upload-artifact@v6
with:
name: whispercpp.jar.zip
path: whispercpp.jar.zip
@ -1245,7 +1225,7 @@ jobs:
steps:
- name: Clone
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: Test quantize
run: |
@ -1269,7 +1249,7 @@ jobs:
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
uses: actions/checkout@v6
with:
fetch-depth: 0
@ -1282,7 +1262,7 @@ jobs:
# Downloads all the artifacts from the previous jobs
- name: Download artifacts
id: download-artifact
uses: actions/download-artifact@v4
uses: actions/download-artifact@v7
with:
path: ./artifact
@ -1332,7 +1312,7 @@ jobs:
steps:
- name: Checkout code
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: Set environment variables
id: set_vars
@ -1358,7 +1338,7 @@ jobs:
steps:
- name: Checkout
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: Build
shell: bash
@ -1378,7 +1358,7 @@ jobs:
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
@ -1403,7 +1383,7 @@ jobs:
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
@ -1428,7 +1408,7 @@ jobs:
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
@ -1453,7 +1433,7 @@ jobs:
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
@ -1478,7 +1458,7 @@ jobs:
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
@ -1503,7 +1483,7 @@ jobs:
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: Test
id: ggml-ci
@ -1517,7 +1497,7 @@ jobs:
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: Test
id: ggml-ci
@ -1531,7 +1511,7 @@ jobs:
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: Test
id: ggml-ci
@ -1545,7 +1525,7 @@ jobs:
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: Test
id: ggml-ci
@ -1558,7 +1538,7 @@ jobs:
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: Test
id: ggml-ci
@ -1571,7 +1551,7 @@ jobs:
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: Test
id: ggml-ci

5
.github/workflows/docker.yml vendored
View File

@ -22,10 +22,11 @@ jobs:
- { tag: "main-musa", dockerfile: ".devops/main-musa.Dockerfile", platform: "linux/amd64" }
- { tag: "main-intel", dockerfile: ".devops/main-intel.Dockerfile", platform: "linux/amd64" }
- { tag: "main-cuda", dockerfile: ".devops/main-cuda.Dockerfile", platform: "linux/amd64" }
- { tag: "main-vulkan", dockerfile: ".devops/main-vulkan.Dockerfile", platform: "linux/amd64" }
steps:
- name: Check out the repo
uses: actions/checkout@v3
uses: actions/checkout@v6
- name: Set up QEMU
uses: docker/setup-qemu-action@v3
@ -67,7 +68,7 @@ jobs:
echo "tags=$TAGS" >> $GITHUB_OUTPUT
- name: Build and push Docker image (tagged)
uses: docker/build-push-action@v5
uses: docker/build-push-action@v6
with:
context: .
push: ${{ github.event_name == 'push' }}

6
.github/workflows/examples-wasm.yml vendored
View File

@ -22,10 +22,10 @@ jobs:
runs-on: ubuntu-latest
steps:
- name: Checkout
uses: actions/checkout@v4
uses: actions/checkout@v6
- name: Setup Pages
uses: actions/configure-pages@v4
uses: actions/configure-pages@v5
- name: Setup emsdk
uses: mymindstorm/setup-emsdk@v14
@ -88,7 +88,7 @@ jobs:
find staging -type f | sort
- name: Upload artifact
uses: actions/upload-pages-artifact@v3
uses: actions/upload-pages-artifact@v4
with:
path: ./staging

4
.github/workflows/examples.yml vendored
View File

@ -17,7 +17,7 @@ jobs:
node-version: [ 16.x, 18.x ]
steps:
- name: Clone
uses: actions/checkout@v1
uses: actions/checkout@v6
- name: Dependencies
run: |
@ -27,7 +27,7 @@ jobs:
sudo apt-get install libsdl2-dev
- name: Use Node.js ${{ matrix.node-version }}
uses: actions/setup-node@v1
uses: actions/setup-node@v6
with:
node-version: ${{ matrix.node-version }}
cache: 'npm'

2
LICENSE
View File

@ -1,6 +1,6 @@
MIT License
Copyright (c) 2023-2024 The ggml authors
Copyright (c) 2023-2026 The ggml authors
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal

17
README.md
View File

@ -443,11 +443,12 @@ ffmpeg -i samples/jfk.wav jfk.opus
### Images
We have two Docker images available for this project:
We have multiple Docker images available for this project:
1. `ghcr.io/ggml-org/whisper.cpp:main`: This image includes the main executable file as well as `curl` and `ffmpeg`. (platforms: `linux/amd64`, `linux/arm64`)
2. `ghcr.io/ggml-org/whisper.cpp:main-cuda`: Same as `main` but compiled with CUDA support. (platforms: `linux/amd64`)
3. `ghcr.io/ggml-org/whisper.cpp:main-musa`: Same as `main` but compiled with MUSA support. (platforms: `linux/amd64`)
4. `ghcr.io/ggml-org/whisper.cpp:main-vulkan`: Same as `main` but compiled with Vulkan support. (platforms: `linux/amd64`)
### Usage
@ -456,15 +457,27 @@ We have two Docker images available for this project:
docker run -it --rm \
-v path/to/models:/models \
whisper.cpp:main "./models/download-ggml-model.sh base /models"
# transcribe an audio file
docker run -it --rm \
-v path/to/models:/models \
-v path/to/audios:/audios \
whisper.cpp:main "whisper-cli -m /models/ggml-base.bin -f /audios/jfk.wav"
# transcribe an audio file in samples folder
docker run -it --rm \
-v path/to/models:/models \
whisper.cpp:main "whisper-cli -m /models/ggml-base.bin -f ./samples/jfk.wav"
# run the web server
docker run -it --rm -p "8080:8080" \
-v path/to/models:/models \
whisper.cpp:main "whisper-server --host 127.0.0.1 -m /models/ggml-base.bin"
# run the bench too on the small.en model using 4 threads
docker run -it --rm \
-v path/to/models:/models \
whisper.cpp:main "whisper-bench -m /models/ggml-small.en.bin -t 4"
```
## Installing with Conan
@ -750,7 +763,7 @@ argument to `whisper-cli`. In addition to this option a VAD model is also
required.
The way this works is that first the audio samples are passed through
the VAD model which will detect speech segments. Using this information the
the VAD model which will detect speech segments. Using this information,
only the speech segments that are detected are extracted from the original audio
input and passed to whisper for processing. This reduces the amount of audio
data that needs to be processed by whisper and can significantly speed up the

17
bindings/go/examples/go-model-download/main.go
View File

@ -282,13 +282,20 @@ func Download(ctx context.Context, p io.Writer, model, out string) (string, erro
default:
// Read body
n, err := resp.Body.Read(data)
if n > 0 {
if m, err := w.Write(data[:n]); err != nil {
return path, err
} else {
count += int64(m)
}
}
if err != nil {
DownloadReport(p, pct, count, resp.ContentLength)
if err == io.EOF {
DownloadReport(p, pct, count, resp.ContentLength)
return path, nil
}
return path, err
} else if m, err := w.Write(data[:n]); err != nil {
return path, err
} else {
count += int64(m)
}
}
}

101
bindings/ruby/README.md
View File

@ -247,6 +247,58 @@ whisper.transcribe("path/to/audio.wav", params)
```
### Tokens ###
Each segment has tokens.
To enable token timestamps, you need to set `Whisper::Params#token_timestamps = true`. Then, retrieve tokens from segments using `Whisper::Segment#each_token`.
```ruby
whisper = Whisper::Context.new("base.en")
params = Whisper::Params.new(token_timestamps: true)
whisper
.transcribe("path/to/audio.wav", params)
.each_segment do |segment|
segment.each_token do |token|
token => {start_time:, end_time:, text:, probability:}
st = "%05.2fs" % (start_time / 1000.0)
et = "%05.2fs" % (end_time / 1000.0)
prob = "%.1f%%" % (probability * 100)
puts "[#{st} --> #{et}] #{text} (#{prob})"
end
end
```
```
[00.00s --> 00.00s] [_BEG_] (84.2%)
[00.32s --> 00.37s] And (71.2%)
[00.37s --> 00.53s] so (98.5%)
[00.69s --> 00.85s] my (70.7%)
[00.85s --> 01.59s] fellow (99.5%)
[01.59s --> 02.10s] Americans (90.1%)
[02.85s --> 03.30s] , (28.4%)
[03.30s --> 04.14s] ask (79.8%)
[04.14s --> 04.28s] not (78.9%)
[05.03s --> 05.35s] what (93.3%)
[05.41s --> 05.74s] your (98.8%)
[05.74s --> 06.41s] country (99.6%)
[06.41s --> 06.74s] can (97.7%)
[06.74s --> 06.92s] do (99.0%)
[07.00s --> 07.00s] for (95.8%)
[07.01s --> 07.52s] you (98.5%)
[07.81s --> 08.05s] , (49.3%)
[08.19s --> 08.37s] ask (65.6%)
[08.37s --> 08.75s] what (98.8%)
[08.91s --> 09.04s] you (98.2%)
[09.04s --> 09.32s] can (96.9%)
[09.32s --> 09.38s] do (90.3%)
[09.44s --> 09.76s] for (91.8%)
[09.76s --> 09.99s] your (98.2%)
[10.02s --> 10.36s] country (99.6%)
[10.51s --> 10.99s] . (87.0%)
[11.00s --> 11.00s] [_TT_550] (7.6%)
```
### Models ###
You can see model information:
@ -323,7 +375,38 @@ whisper
end
```
The second argument `samples` may be an array, an object with `length` and `each` method, or a MemoryView. If you can prepare audio data as C array and export it as a MemoryView, whispercpp accepts and works with it with zero copy.
The second argument `samples` may be an array, an object with `length` and `each` method, or a MemoryView.
If you can prepare audio data as C array and export it as a MemoryView, whispercpp accepts and works with it with zero copy.
```ruby
require "torchaudio"
require "arrow-numo-narray"
require "whisper"
waveform, sample_rate = TorchAudio.load("test/fixtures/jfk.wav")
# Convert Torch::Tensor to Arrow::Array via Numo::NArray
samples = waveform.squeeze.numo.to_arrow.to_arrow_array
whisper = Whisper::Context.new("base")
whisper
# Arrow::Array exports MemoryView
.full(Whisper::Params.new, samples)
```
Custom context params
---------------------
You can use customize `Whisper::Context`'s behavior using `Whisper::Context::Params`.
```ruby
context_params = Whisper::Context::Params.new(
use_gpu: false,
flash_attn: false,
# etc
)
whisper = Whisper::Context.new("base", context_params)
```
Using VAD separately from ASR
-----------------------------
@ -334,13 +417,27 @@ VAD feature itself is useful. You can use it separately from ASR:
vad = Whisper::VAD::Context.new("silero-v6.2.0")
vad
.detect("path/to/audio.wav", Whisper::VAD::Params.new)
.each_with_index do |segment, index|
.each.with_index do |segment, index|
segment => {start_time: st, end_time: ed} # `Segment` responds to `#deconstruct_keys`
puts "[%{nth}: %{st} --> %{ed}]" % {nth: index + 1, st:, ed:}
end
```
You may also low level API `Whisper::VAD::Context#segments_from_samples` as such `Whisper::Context#full`:
```ruby
# Ruby Array
reader = WaveFile::Reader.new("path/to/audio.wav", WaveFile::Format.new(:mono, :float, 16000))
samples = reader.enum_for(:each_buffer).map(&:samples).flatten
# Or, object which exports MemoryView
waveform, sample_rate = TorchAudio.load("test/fixtures/jfk.wav")
samples = waveform.squeeze.numo.to_arrow.to_arrow_array
segments = vad.segments_from_samples(Whisper::VAD::Params.new, samples)
```
Development
-----------

1
bindings/ruby/ext/extconf.rb
View File

@ -7,6 +7,7 @@ options = Options.new(cmake).to_s
have_library("gomp") rescue nil
libs = Dependencies.new(cmake, options).to_s
$CFLAGS << " -O3 -march=native"
$INCFLAGS << " -Isources/include -Isources/ggml/include -Isources/examples"
$LOCAL_LIBS << " #{libs}"
$cleanfiles << " build #{libs}"

24
bindings/ruby/ext/ruby_whisper.c
View File

@ -1,5 +1,3 @@
#include <ruby.h>
#include <ruby/memory_view.h>
#include "ruby_whisper.h"
VALUE mWhisper;
@ -35,7 +33,8 @@ static bool is_log_callback_finalized = false;
// High level API
extern VALUE ruby_whisper_segment_allocate(VALUE klass);
extern void init_ruby_whisper_context(VALUE *mWhisper);
extern VALUE init_ruby_whisper_context(VALUE *mWhisper);
extern void init_ruby_whisper_context_params(VALUE *cContext);
extern void init_ruby_whisper_params(VALUE *mWhisper);
extern void init_ruby_whisper_error(VALUE *mWhisper);
extern void init_ruby_whisper_segment(VALUE *mWhisper);
@ -164,6 +163,22 @@ void Init_whisper() {
rb_define_const(mWhisper, "LOG_LEVEL_DEBUG", INT2NUM(GGML_LOG_LEVEL_DEBUG));
rb_define_const(mWhisper, "LOG_LEVEL_CONT", INT2NUM(GGML_LOG_LEVEL_CONT));
rb_define_const(mWhisper, "AHEADS_NONE", INT2NUM(WHISPER_AHEADS_NONE));
rb_define_const(mWhisper, "AHEADS_N_TOP_MOST", INT2NUM(WHISPER_AHEADS_N_TOP_MOST));
rb_define_const(mWhisper, "AHEADS_CUSTOM", INT2NUM(WHISPER_AHEADS_CUSTOM));
rb_define_const(mWhisper, "AHEADS_TINY_EN", INT2NUM(WHISPER_AHEADS_TINY_EN));
rb_define_const(mWhisper, "AHEADS_TINY", INT2NUM(WHISPER_AHEADS_TINY));
rb_define_const(mWhisper, "AHEADS_BASE_EN", INT2NUM(WHISPER_AHEADS_BASE_EN));
rb_define_const(mWhisper, "AHEADS_BASE", INT2NUM(WHISPER_AHEADS_BASE));
rb_define_const(mWhisper, "AHEADS_SMALL_EN", INT2NUM(WHISPER_AHEADS_SMALL_EN));
rb_define_const(mWhisper, "AHEADS_SMALL", INT2NUM(WHISPER_AHEADS_SMALL));
rb_define_const(mWhisper, "AHEADS_MEDIUM_EN", INT2NUM(WHISPER_AHEADS_MEDIUM_EN));
rb_define_const(mWhisper, "AHEADS_MEDIUM", INT2NUM(WHISPER_AHEADS_MEDIUM));
rb_define_const(mWhisper, "AHEADS_LARGE_V1", INT2NUM(WHISPER_AHEADS_LARGE_V1));
rb_define_const(mWhisper, "AHEADS_LARGE_V2", INT2NUM(WHISPER_AHEADS_LARGE_V2));
rb_define_const(mWhisper, "AHEADS_LARGE_V3", INT2NUM(WHISPER_AHEADS_LARGE_V3));
rb_define_const(mWhisper, "AHEADS_LARGE_V3_TURBO", INT2NUM(WHISPER_AHEADS_LARGE_V3_TURBO));
rb_define_singleton_method(mWhisper, "lang_max_id", ruby_whisper_s_lang_max_id, 0);
rb_define_singleton_method(mWhisper, "lang_id", ruby_whisper_s_lang_id, 1);
rb_define_singleton_method(mWhisper, "lang_str", ruby_whisper_s_lang_str, 1);
@ -172,7 +187,8 @@ void Init_whisper() {
rb_define_singleton_method(mWhisper, "log_set", ruby_whisper_s_log_set, 2);
rb_define_private_method(rb_singleton_class(mWhisper), "finalize_log_callback", ruby_whisper_s_finalize_log_callback, 1);
init_ruby_whisper_context(&mWhisper);
cContext = init_ruby_whisper_context(&mWhisper);
init_ruby_whisper_context_params(&cContext);
init_ruby_whisper_params(&mWhisper);
init_ruby_whisper_error(&mWhisper);
init_ruby_whisper_segment(&mWhisper);

32
bindings/ruby/ext/ruby_whisper.h
View File

@ -1,6 +1,8 @@
#ifndef RUBY_WHISPER_H
#define RUBY_WHISPER_H
#include <ruby.h>
#include <ruby/memory_view.h>
#include "whisper.h"
typedef struct {
@ -14,6 +16,10 @@ typedef struct {
struct whisper_context *context;
} ruby_whisper;
typedef struct ruby_whisper_context_params {
struct whisper_context_params params;
} ruby_whisper_context_params;
typedef struct {
struct whisper_full_params params;
bool diarize;
@ -35,7 +41,7 @@ typedef struct {
typedef struct {
whisper_token_data *token_data;
const char *text;
VALUE text;
} ruby_whisper_token;
typedef struct {
@ -55,6 +61,13 @@ typedef struct {
struct whisper_vad_context *context;
} ruby_whisper_vad_context;
typedef struct parsed_samples_t {
float *samples;
int n_samples;
rb_memory_view_t memview;
bool memview_exported;
} parsed_samples_t;
#define GetContext(obj, rw) do { \
TypedData_Get_Struct((obj), ruby_whisper, &ruby_whisper_type, (rw)); \
if ((rw)->context == NULL) { \
@ -62,13 +75,28 @@ typedef struct {
} \
} while (0)
#define GetToken(obj, rwt) do { \
#define GetContextParams(obj, rwcp) do { \
TypedData_Get_Struct((obj), ruby_whisper_context_params, &ruby_whisper_context_params_type, (rwcp)); \
} while (0)
#define GetToken(obj, rwt) do { \
TypedData_Get_Struct((obj), ruby_whisper_token, &ruby_whisper_token_type, (rwt)); \
if ((rwt)->token_data == NULL) { \
rb_raise(rb_eRuntimeError, "Not initialized"); \
} \
} while (0)
#define GetVADContext(obj, rwvc) do { \
TypedData_Get_Struct((obj), ruby_whisper_vad_context, &ruby_whisper_vad_context_type, (rwvc)); \
if ((rwvc)->context == NULL) { \
rb_raise(rb_eRuntimeError, "Not initialized"); \
} \
} while (0)
#define GetVADParams(obj, rwvp) do { \
TypedData_Get_Struct((obj), ruby_whisper_vad_params, &ruby_whisper_vad_params_type, (rwvp)); \
} while (0)
#define GetVADSegments(obj, rwvss) do { \
TypedData_Get_Struct((obj), ruby_whisper_vad_segments, &ruby_whisper_vad_segments_type, (rwvss)); \
if ((rwvss)->segments == NULL) { \

340
bindings/ruby/ext/ruby_whisper_context.c
View File

@ -1,5 +1,3 @@
#include <ruby.h>
#include <ruby/memory_view.h>
#include "ruby_whisper.h"
extern ID id_to_s;
@ -20,6 +18,7 @@ extern VALUE eError;
extern VALUE cModel;
extern const rb_data_type_t ruby_whisper_params_type;
extern const rb_data_type_t ruby_whisper_context_params_type;
extern VALUE ruby_whisper_transcribe(int argc, VALUE *argv, VALUE self);
extern VALUE rb_whisper_model_s_new(VALUE context);
extern VALUE rb_whisper_segment_s_new(VALUE context, int index);
@ -27,6 +26,27 @@ extern void prepare_transcription(ruby_whisper_params *rwp, VALUE *context);
ID transcribe_option_names[1];
typedef struct fill_samples_args {
float *dest;
VALUE *src;
int n_samples;
} fill_samples_args;
typedef struct full_args {
VALUE *context;
VALUE *params;
float *samples;
int n_samples;
} full_args;
typedef struct full_parallel_args {
VALUE *context;
VALUE *params;
float *samples;
int n_samples;
int n_processors;
} full_parallel_args;
static void
ruby_whisper_free(ruby_whisper *rw)
{
@ -124,16 +144,25 @@ ruby_whisper_initialize(int argc, VALUE *argv, VALUE self)
{
ruby_whisper *rw;
VALUE whisper_model_file_path;
VALUE context_params;
struct whisper_context_params params;
// TODO: we can support init from buffer here too maybe another ruby object to expose
rb_scan_args(argc, argv, "01", &whisper_model_file_path);
rb_scan_args(argc, argv, "11", &whisper_model_file_path, &context_params);
TypedData_Get_Struct(self, ruby_whisper, &ruby_whisper_type, rw);
whisper_model_file_path = ruby_whisper_normalize_model_path(whisper_model_file_path);
if (!rb_respond_to(whisper_model_file_path, id_to_s)) {
rb_raise(rb_eRuntimeError, "Expected file path to model to initialize Whisper::Context");
}
rw->context = whisper_init_from_file_with_params(StringValueCStr(whisper_model_file_path), whisper_context_default_params());
if (NIL_P(context_params)) {
params = whisper_context_default_params();
} else {
ruby_whisper_context_params *rwcp;
GetContextParams(context_params, rwcp);
params = rwcp->params;
}
rw->context = whisper_init_from_file_with_params(StringValueCStr(whisper_model_file_path), params);
if (rw->context == NULL) {
rb_raise(rb_eRuntimeError, "error: failed to initialize whisper context");
}
@ -272,6 +301,147 @@ VALUE ruby_whisper_model_type(VALUE self)
return rb_str_new2(whisper_model_type_readable(rw->context));
}
static bool
check_memory_view(rb_memory_view_t *memview)
{
if (memview->format != NULL && strcmp(memview->format, "f") != 0) {
rb_warn("currently only format \"f\" is supported for MemoryView, but given: %s", memview->format);
return false;
}
if (memview->format != NULL && memview->ndim != 1) {
rb_warn("currently only 1 dimensional MemoryView is supported, but given: %zd", memview->ndim);
return false;
}
return true;
}
static VALUE
fill_samples(VALUE rb_args)
{
fill_samples_args *args = (fill_samples_args *)rb_args;
if (RB_TYPE_P(*args->src, T_ARRAY)) {
for (int i = 0; i < args->n_samples; i++) {
args->dest[i] = RFLOAT_VALUE(rb_ary_entry(*args->src, i));
}
} else {
// TODO: use rb_block_call
VALUE iter = rb_funcall(*args->src, id_to_enum, 1, rb_str_new2("each"));
for (int i = 0; i < args->n_samples; i++) {
// TODO: check if iter is exhausted and raise ArgumentError appropriately
VALUE sample = rb_funcall(iter, id_next, 0);
args->dest[i] = RFLOAT_VALUE(sample);
}
}
return Qnil;
}
struct parsed_samples_t
parse_samples(VALUE *samples, VALUE *n_samples)
{
bool memview_available = rb_memory_view_available_p(*samples);
struct parsed_samples_t parsed = {0};
parsed.memview_exported = false;
const bool is_array = RB_TYPE_P(*samples, T_ARRAY);
if (!NIL_P(*n_samples)) {
parsed.n_samples = NUM2INT(*n_samples);
if (is_array) {
if (RARRAY_LEN(*samples) < parsed.n_samples) {
rb_raise(rb_eArgError, "samples length %ld is less than n_samples %d", RARRAY_LEN(*samples), parsed.n_samples);
}
}
// Should check when samples.respond_to?(:length)?
} else {
if (is_array) {
if (RARRAY_LEN(*samples) > INT_MAX) {
rb_raise(rb_eArgError, "samples are too long");
}
parsed.n_samples = (int)RARRAY_LEN(*samples);
} else if (memview_available) {
bool memview_got = rb_memory_view_get(*samples, &parsed.memview, RUBY_MEMORY_VIEW_SIMPLE);
if (memview_got) {
parsed.memview_exported = check_memory_view(&parsed.memview);
if (!parsed.memview_exported) {
rb_memory_view_release(&parsed.memview);
parsed.memview = (rb_memory_view_t){0};
}
}
if (parsed.memview_exported) {
ssize_t n_samples_size = parsed.memview.byte_size / parsed.memview.item_size;
if (n_samples_size > INT_MAX) {
rb_memory_view_release(&parsed.memview);
rb_raise(rb_eArgError, "samples are too long: %zd", n_samples_size);
}
parsed.n_samples = (int)n_samples_size;
} else {
rb_warn("unable to get a memory view. falls back to Ruby object");
if (rb_respond_to(*samples, id_length)) {
parsed.n_samples = NUM2INT(rb_funcall(*samples, id_length, 0));
} else {
rb_raise(rb_eArgError, "samples must respond to :length");
}
}
} else if (rb_respond_to(*samples, id_length)) {
parsed.n_samples = NUM2INT(rb_funcall(*samples, id_length, 0));
} else {
rb_raise(rb_eArgError, "samples must respond to :length or be a MemoryView of an array of float when n_samples is not given");
}
}
if (parsed.memview_exported) {
parsed.samples = (float *)parsed.memview.data;
} else {
parsed.samples = ALLOC_N(float, parsed.n_samples);
fill_samples_args args = {
parsed.samples,
samples,
parsed.n_samples,
};
int state;
rb_protect(fill_samples, (VALUE)&args, &state);
if (state) {
xfree(parsed.samples);
rb_jump_tag(state);
}
}
return parsed;
}
VALUE
release_samples(VALUE rb_parsed_args)
{
parsed_samples_t *parsed_args = (parsed_samples_t *)rb_parsed_args;
if (parsed_args->memview_exported) {
rb_memory_view_release(&parsed_args->memview);
} else {
xfree(parsed_args->samples);
}
*parsed_args = (parsed_samples_t){0};
return Qnil;
}
static VALUE
full_body(VALUE rb_args)
{
full_args *args = (full_args *)rb_args;
ruby_whisper *rw;
ruby_whisper_params *rwp;
GetContext(*args->context, rw);
TypedData_Get_Struct(*args->params, ruby_whisper_params, &ruby_whisper_params_type, rwp);
prepare_transcription(rwp, args->context);
int result = whisper_full(rw->context, rwp->params, args->samples, args->n_samples);
return INT2NUM(result);
}
/*
* Run the entire model: PCM -> log mel spectrogram -> encoder -> decoder -> text
* Not thread safe for same context
@ -289,65 +459,17 @@ VALUE ruby_whisper_full(int argc, VALUE *argv, VALUE self)
rb_raise(rb_eArgError, "wrong number of arguments (given %d, expected 2..3)", argc);
}
ruby_whisper *rw;
ruby_whisper_params *rwp;
GetContext(self, rw);
VALUE params = argv[0];
TypedData_Get_Struct(params, ruby_whisper_params, &ruby_whisper_params_type, rwp);
VALUE samples = argv[1];
int n_samples;
rb_memory_view_t view;
const bool memory_view_available_p = rb_memory_view_available_p(samples);
if (argc == 3) {
n_samples = NUM2INT(argv[2]);
if (TYPE(samples) == T_ARRAY) {
if (RARRAY_LEN(samples) < n_samples) {
rb_raise(rb_eArgError, "samples length %ld is less than n_samples %d", RARRAY_LEN(samples), n_samples);
}
}
// Should check when samples.respond_to?(:length)?
} else {
if (TYPE(samples) == T_ARRAY) {
if (RARRAY_LEN(samples) > INT_MAX) {
rb_raise(rb_eArgError, "samples are too long");
}
n_samples = (int)RARRAY_LEN(samples);
} else if (memory_view_available_p) {
if (!rb_memory_view_get(samples, &view, RUBY_MEMORY_VIEW_SIMPLE)) {
view.obj = Qnil;
rb_raise(rb_eArgError, "unable to get a memory view");
}
ssize_t n_samples_size = view.byte_size / view.item_size;
if (n_samples_size > INT_MAX) {
rb_raise(rb_eArgError, "samples are too long");
}
n_samples = (int)n_samples_size;
} else if (rb_respond_to(samples, id_length)) {
n_samples = NUM2INT(rb_funcall(samples, id_length, 0));
} else {
rb_raise(rb_eArgError, "samples must respond to :length or be a MemoryView of an array of flaot when n_samples is not given");
}
}
float * c_samples = (float *)malloc(n_samples * sizeof(float));
if (memory_view_available_p) {
c_samples = (float *)view.data;
} else {
if (TYPE(samples) == T_ARRAY) {
for (int i = 0; i < n_samples; i++) {
c_samples[i] = RFLOAT_VALUE(rb_ary_entry(samples, i));
}
} else {
// TODO: use rb_block_call
VALUE iter = rb_funcall(samples, id_to_enum, 1, rb_str_new2("each"));
for (int i = 0; i < n_samples; i++) {
// TODO: check if iter is exhausted and raise ArgumentError appropriately
VALUE sample = rb_funcall(iter, id_next, 0);
c_samples[i] = RFLOAT_VALUE(sample);
}
}
}
prepare_transcription(rwp, &self);
const int result = whisper_full(rw->context, rwp->params, c_samples, n_samples);
VALUE n_samples = argc == 2 ? Qnil : argv[2];
struct parsed_samples_t parsed = parse_samples(&argv[1], &n_samples);
full_args args = {
&self,
&argv[0],
parsed.samples,
parsed.n_samples,
};
VALUE rb_result = rb_ensure(full_body, (VALUE)&args, release_samples, (VALUE)&parsed);
const int result = NUM2INT(rb_result);
if (0 == result) {
return self;
} else {
@ -355,6 +477,22 @@ VALUE ruby_whisper_full(int argc, VALUE *argv, VALUE self)
}
}
static VALUE
full_parallel_body(VALUE rb_args)
{
full_parallel_args *args = (full_parallel_args *)rb_args;
ruby_whisper *rw;
ruby_whisper_params *rwp;
GetContext(*args->context, rw);
TypedData_Get_Struct(*args->params, ruby_whisper_params, &ruby_whisper_params_type, rwp);
prepare_transcription(rwp, args->context);
int result = whisper_full_parallel(rw->context, rwp->params, args->samples, args->n_samples, args->n_processors);
return INT2NUM(result);
}
/*
* Split the input audio in chunks and process each chunk separately using whisper_full_with_state()
* Result is stored in the default state of the context
@ -372,19 +510,11 @@ static VALUE
ruby_whisper_full_parallel(int argc, VALUE *argv,VALUE self)
{
if (argc < 2 || argc > 4) {
rb_raise(rb_eArgError, "wrong number of arguments (given %d, expected 2..3)", argc);
rb_raise(rb_eArgError, "wrong number of arguments (given %d, expected 2..4)", argc);
}
ruby_whisper *rw;
ruby_whisper_params *rwp;
GetContext(self, rw);
VALUE params = argv[0];
TypedData_Get_Struct(params, ruby_whisper_params, &ruby_whisper_params_type, rwp);
VALUE samples = argv[1];
int n_samples;
VALUE n_samples = argc == 2 ? Qnil : argv[2];
int n_processors;
rb_memory_view_t view;
const bool memory_view_available_p = rb_memory_view_available_p(samples);
switch (argc) {
case 2:
n_processors = 1;
@ -396,56 +526,16 @@ ruby_whisper_full_parallel(int argc, VALUE *argv,VALUE self)
n_processors = NUM2INT(argv[3]);
break;
}
if (argc >= 3 && !NIL_P(argv[2])) {
n_samples = NUM2INT(argv[2]);
if (TYPE(samples) == T_ARRAY) {
if (RARRAY_LEN(samples) < n_samples) {
rb_raise(rb_eArgError, "samples length %ld is less than n_samples %d", RARRAY_LEN(samples), n_samples);
}
}
// Should check when samples.respond_to?(:length)?
} else if (memory_view_available_p) {
if (!rb_memory_view_get(samples, &view, RUBY_MEMORY_VIEW_SIMPLE)) {
view.obj = Qnil;
rb_raise(rb_eArgError, "unable to get a memory view");
}
ssize_t n_samples_size = view.byte_size / view.item_size;
if (n_samples_size > INT_MAX) {
rb_raise(rb_eArgError, "samples are too long");
}
n_samples = (int)n_samples_size;
} else {
if (TYPE(samples) == T_ARRAY) {
if (RARRAY_LEN(samples) > INT_MAX) {
rb_raise(rb_eArgError, "samples are too long");
}
n_samples = (int)RARRAY_LEN(samples);
} else if (rb_respond_to(samples, id_length)) {
n_samples = NUM2INT(rb_funcall(samples, id_length, 0));
} else {
rb_raise(rb_eArgError, "samples must respond to :length or be a MemoryView of an array of flaot when n_samples is not given");
}
}
float * c_samples = (float *)malloc(n_samples * sizeof(float));
if (memory_view_available_p) {
c_samples = (float *)view.data;
} else {
if (TYPE(samples) == T_ARRAY) {
for (int i = 0; i < n_samples; i++) {
c_samples[i] = RFLOAT_VALUE(rb_ary_entry(samples, i));
}
} else {
// FIXME: use rb_block_call
VALUE iter = rb_funcall(samples, id_to_enum, 1, rb_str_new2("each"));
for (int i = 0; i < n_samples; i++) {
// TODO: check if iter is exhausted and raise ArgumentError
VALUE sample = rb_funcall(iter, id_next, 0);
c_samples[i] = RFLOAT_VALUE(sample);
}
}
}
prepare_transcription(rwp, &self);
const int result = whisper_full_parallel(rw->context, rwp->params, c_samples, n_samples, n_processors);
struct parsed_samples_t parsed = parse_samples(&argv[1], &n_samples);
const full_parallel_args args = {
&self,
&argv[0],
parsed.samples,
parsed.n_samples,
n_processors,
};
const VALUE rb_result = rb_ensure(full_parallel_body, (VALUE)&args, release_samples, (VALUE)&parsed);
const int result = NUM2INT(rb_result);
if (0 == result) {
return self;
} else {
@ -631,7 +721,7 @@ ruby_whisper_get_model(VALUE self)
return rb_whisper_model_s_new(self);
}
void
VALUE
init_ruby_whisper_context(VALUE *mWhisper)
{
cContext = rb_define_class_under(*mWhisper, "Context", rb_cObject);
@ -669,4 +759,6 @@ init_ruby_whisper_context(VALUE *mWhisper)
rb_define_method(cContext, "each_segment", ruby_whisper_each_segment, 0);
rb_define_method(cContext, "model", ruby_whisper_get_model, 0);
return cContext;
}

163
bindings/ruby/ext/ruby_whisper_context_params.c Normal file
View File

@ -0,0 +1,163 @@
#include "ruby_whisper.h"
#define NUM_PARAMS 6
#define DEF_BOOLEAN_ATTR_METHOD(name) \
static VALUE \
ruby_whisper_context_params_get_ ## name(VALUE self) { \
ruby_whisper_context_params *rwcp; \
GetContextParams(self, rwcp); \
return rwcp->params.name ? Qtrue : Qfalse; \
} \
static VALUE \
ruby_whisper_context_params_set_ ## name(VALUE self, VALUE value) { \
ruby_whisper_context_params *rwcp; \
GetContextParams(self, rwcp); \
rwcp->params.name = RTEST(value); \
return value; \
}
#define DEF_INT_ATTR_METHOD(name) \
static VALUE \
ruby_whisper_context_params_get_ ## name(VALUE self) { \
ruby_whisper_context_params *rwcp; \
GetContextParams(self, rwcp); \
return INT2NUM(rwcp->params.name); \
} \
static VALUE \
ruby_whisper_context_params_set_ ## name(VALUE self, VALUE value) { \
ruby_whisper_context_params *rwcp; \
GetContextParams(self, rwcp); \
rwcp->params.name = NUM2INT(value); \
return value; \
}
#define DEFINE_PARAM(param_name, nth) \
id_ ## param_name = rb_intern(#param_name); \
param_names[nth] = id_ ## param_name; \
rb_define_method(cContextParams, #param_name, ruby_whisper_context_params_get_ ## param_name, 0); \
rb_define_method(cContextParams, #param_name "=", ruby_whisper_context_params_set_ ## param_name, 1);
VALUE cContextParams;
static ID param_names[NUM_PARAMS];
static ID id_use_gpu;
static ID id_flash_attn;
static ID id_gpu_device;
static ID id_dtw_token_timestamps;
static ID id_dtw_aheads_preset;
static ID id_dtw_n_top;
static size_t
ruby_whisper_context_params_memsize(const void *p)
{
const ruby_whisper_context_params *rwcp = (ruby_whisper_context_params *)p;
if (!rwcp) {
return 0;
}
return sizeof(ruby_whisper_context_params);
}
const rb_data_type_t ruby_whisper_context_params_type = {
"ruby_whisper_context_params",
{0, RUBY_DEFAULT_FREE, ruby_whisper_context_params_memsize,},
0, 0,
0
};
static VALUE
ruby_whisper_context_params_s_allocate(VALUE klass)
{
ruby_whisper_context_params *rwcp;
return TypedData_Make_Struct(klass, ruby_whisper_context_params, &ruby_whisper_context_params_type, rwcp);
}
DEF_BOOLEAN_ATTR_METHOD(use_gpu);
DEF_BOOLEAN_ATTR_METHOD(flash_attn);
DEF_INT_ATTR_METHOD(gpu_device);
DEF_BOOLEAN_ATTR_METHOD(dtw_token_timestamps);
DEF_INT_ATTR_METHOD(dtw_aheads_preset);
static VALUE
ruby_whisper_context_params_get_dtw_n_top(VALUE self) {
ruby_whisper_context_params *rwcp;
GetContextParams(self, rwcp);
int dtw_n_top = rwcp->params.dtw_n_top;
return dtw_n_top == -1 ? Qnil : INT2NUM(dtw_n_top);
}
static VALUE
ruby_whisper_context_params_set_dtw_n_top(VALUE self, VALUE value) {
ruby_whisper_context_params *rwcp;
GetContextParams(self, rwcp);
rwcp->params.dtw_n_top = NIL_P(value) ? -1 : NUM2INT(value);
return value;
}
#define SET_PARAM_IF_SAME(param_name) \
if (id == id_ ## param_name) { \
ruby_whisper_context_params_set_ ## param_name(self, value); \
continue; \
}
static VALUE
ruby_whisper_context_params_initialize(int argc, VALUE *argv, VALUE self)
{
ruby_whisper_context_params *rwcp;
TypedData_Get_Struct(self, ruby_whisper_context_params, &ruby_whisper_context_params_type, rwcp);
rwcp->params = whisper_context_default_params();
VALUE kw_hash;
rb_scan_args_kw(RB_SCAN_ARGS_KEYWORDS, argc, argv, ":", &kw_hash);
if (NIL_P(kw_hash)) {
return Qnil;
}
VALUE values[NUM_PARAMS] = {Qundef};
rb_get_kwargs(kw_hash, param_names, 0, NUM_PARAMS, values);
ID id;
VALUE value;
for (int i = 0; i < NUM_PARAMS; i++) {
id = param_names[i];
value = values[i];
if (value == Qundef) {
continue;
}
SET_PARAM_IF_SAME(use_gpu)
SET_PARAM_IF_SAME(flash_attn)
SET_PARAM_IF_SAME(gpu_device)
SET_PARAM_IF_SAME(dtw_token_timestamps)
SET_PARAM_IF_SAME(dtw_aheads_preset)
SET_PARAM_IF_SAME(dtw_n_top)
}
return Qnil;
}
#undef SET_PARAM_IF_SAME
void
init_ruby_whisper_context_params(VALUE *cContext)
{
cContextParams = rb_define_class_under(*cContext, "Params", rb_cObject);
rb_define_alloc_func(cContextParams, ruby_whisper_context_params_s_allocate);
rb_define_method(cContextParams, "initialize", ruby_whisper_context_params_initialize, -1);
DEFINE_PARAM(use_gpu, 0)
DEFINE_PARAM(flash_attn, 1)
DEFINE_PARAM(gpu_device, 2)
DEFINE_PARAM(dtw_token_timestamps, 3)
DEFINE_PARAM(dtw_aheads_preset, 4)
DEFINE_PARAM(dtw_n_top, 5)
}
#undef DEFINE_PARAM
#undef DEF_INT_ATTR_METHOD
#undef DEF_BOOLEAN_ATTR_METHOD
#undef NUM_PARAMS

1
bindings/ruby/ext/ruby_whisper_model.c
View File

@ -1,4 +1,3 @@
#include <ruby.h>
#include "ruby_whisper.h"
extern const rb_data_type_t ruby_whisper_type;

1
bindings/ruby/ext/ruby_whisper_params.c
View File

@ -1,4 +1,3 @@
#include <ruby.h>
#include "ruby_whisper.h"
#define BOOL_PARAMS_SETTER(self, prop, value) \

1
bindings/ruby/ext/ruby_whisper_segment.c
View File

@ -1,4 +1,3 @@
#include <ruby.h>
#include "ruby_whisper.h"
#define N_KEY_NAMES 6

38
bindings/ruby/ext/ruby_whisper_token.c
View File

@ -1,4 +1,3 @@
#include <ruby.h>
#include "ruby_whisper.h"
#define N_KEY_NAMES 11
@ -25,12 +24,34 @@ ruby_whisper_token_memsize(const void *p)
if (!rwt) {
return 0;
}
return sizeof(rwt);
size_t size = sizeof(*rwt);
if (rwt->token_data) {
size += sizeof(*rwt->token_data);
}
return size;
}
static void
ruby_whisper_token_mark(void *p)
{
ruby_whisper_token *rwt = (ruby_whisper_token *)p;
rb_gc_mark(rwt->text);
}
static void
ruby_whisper_token_free(void *p)
{
ruby_whisper_token *rwt = (ruby_whisper_token *)p;
if (rwt->token_data) {
xfree(rwt->token_data);
rwt->token_data = NULL;
}
xfree(rwt);
}
static const rb_data_type_t ruby_whisper_token_type = {
"ruby_whisper_token",
{0, RUBY_DEFAULT_FREE, ruby_whisper_token_memsize,},
{ruby_whisper_token_mark, ruby_whisper_token_free, ruby_whisper_token_memsize,},
0, 0,
0
};
@ -41,19 +62,19 @@ ruby_whisper_token_allocate(VALUE klass)
ruby_whisper_token *rwt;
VALUE token = TypedData_Make_Struct(klass, ruby_whisper_token, &ruby_whisper_token_type, rwt);
rwt->token_data = NULL;
rwt->text = NULL;
rwt->text = Qnil;
return token;
}
VALUE
ruby_whisper_token_s_init(struct whisper_context *context, int i_segment, int i_token)
{
whisper_token_data token_data = whisper_full_get_token_data(context, i_segment, i_token);
const VALUE token = ruby_whisper_token_allocate(cToken);
ruby_whisper_token *rwt;
TypedData_Get_Struct(token, ruby_whisper_token, &ruby_whisper_token_type, rwt);
rwt->token_data = &token_data;
rwt->text = whisper_full_get_token_text(context, i_segment, i_token);
rwt->token_data = ALLOC(whisper_token_data);
*(rwt->token_data) = whisper_full_get_token_data(context, i_segment, i_token);
rwt->text = rb_str_new2(whisper_full_get_token_text(context, i_segment, i_token));
return token;
}
@ -183,10 +204,9 @@ ruby_whisper_token_get_text(VALUE self)
{
ruby_whisper_token *rwt;
GetToken(self, rwt);
return rb_str_new2(rwt->text);
return rwt->text;
}
/*
* Start time of the token.
*

5
bindings/ruby/ext/ruby_whisper_transcribe.cpp
View File

@ -1,4 +1,3 @@
#include <ruby.h>
#include "ruby_whisper.h"
#include "common-whisper.h"
#include <string>
@ -13,6 +12,7 @@ extern const rb_data_type_t ruby_whisper_params_type;
extern ID id_to_s;
extern ID id_call;
extern ID id_to_path;
extern ID transcribe_option_names[1];
extern void
@ -50,6 +50,9 @@ ruby_whisper_transcribe(int argc, VALUE *argv, VALUE self) {
rb_raise(rb_eRuntimeError, "Expected file path to wave file");
}
if (rb_respond_to(wave_file_path, id_to_path)) {
wave_file_path = rb_funcall(wave_file_path, id_to_path, 0);
}
std::string fname_inp = StringValueCStr(wave_file_path);
std::vector<float> pcmf32; // mono-channel F32 PCM

49
bindings/ruby/ext/ruby_whisper_vad_context.c
View File

@ -1,12 +1,23 @@
#include <ruby.h>
#include "ruby_whisper.h"
extern ID id_to_s;
extern VALUE cVADContext;
extern const rb_data_type_t ruby_whisper_vad_params_type;
extern VALUE ruby_whisper_vad_detect(VALUE self, VALUE file_path, VALUE params);
extern VALUE ruby_whisper_normalize_model_path(VALUE model_path);
extern parsed_samples_t parse_samples(VALUE *samples, VALUE *n_samples);
extern VALUE release_samples(VALUE parsed);
extern VALUE ruby_whisper_vad_segments_s_init(struct whisper_vad_segments *segments);
typedef struct segments_from_samples_args {
VALUE *context;
VALUE *params;
float *samples;
int n_samples;
} segments_from_samples_args;
static size_t
ruby_whisper_vad_context_memsize(const void *p)
@ -66,10 +77,46 @@ ruby_whisper_vad_context_initialize(VALUE self, VALUE model_path)
return Qnil;
}
static VALUE
segments_from_samples_body(VALUE rb_args)
{
segments_from_samples_args *args = (segments_from_samples_args *)rb_args;
ruby_whisper_vad_context *rwvc;
ruby_whisper_vad_params *rwvp;
GetVADContext(*args->context, rwvc);
GetVADParams(*args->params, rwvp);
struct whisper_vad_segments *segments = whisper_vad_segments_from_samples(rwvc->context, rwvp->params, args->samples, args->n_samples);
return ruby_whisper_vad_segments_s_init(segments);
}
static VALUE
ruby_whisper_vad_segments_from_samples(int argc, VALUE *argv, VALUE self)
{
if (argc < 2 || argc > 3) {
rb_raise(rb_eArgError, "wrong number of arguments (given %d, expected 2..3)", argc);
}
VALUE n_samples = argc == 2 ? Qnil : argv[2];
struct parsed_samples_t parsed = parse_samples(&argv[1], &n_samples);
segments_from_samples_args args = {
&self,
&argv[0],
parsed.samples,
parsed.n_samples,
};
VALUE segments = rb_ensure(segments_from_samples_body, (VALUE)&args, release_samples, (VALUE)&parsed);
return segments;
}
void init_ruby_whisper_vad_context(VALUE *mVAD)
{
cVADContext = rb_define_class_under(*mVAD, "Context", rb_cObject);
rb_define_alloc_func(cVADContext, ruby_whisper_vad_context_s_allocate);
rb_define_method(cVADContext, "initialize", ruby_whisper_vad_context_initialize, 1);
rb_define_method(cVADContext, "segments_from_samples", ruby_whisper_vad_segments_from_samples, -1);
rb_define_method(cVADContext, "detect", ruby_whisper_vad_detect, 2);
}

11
bindings/ruby/ext/ruby_whisper_vad_context_detect.cpp
View File

@ -1,4 +1,3 @@
#include <ruby.h>
#include "ruby_whisper.h"
#include "common-whisper.h"
#include <string>
@ -8,6 +7,8 @@
extern "C" {
#endif
extern ID id_to_path;
extern VALUE cVADSegments;
extern const rb_data_type_t ruby_whisper_vad_context_type;
@ -25,12 +26,12 @@ ruby_whisper_vad_detect(VALUE self, VALUE file_path, VALUE params) {
std::vector<std::vector<float>> pcmf32s;
whisper_vad_segments *segments;
TypedData_Get_Struct(self, ruby_whisper_vad_context, &ruby_whisper_vad_context_type, rwvc);
if (rwvc->context == NULL) {
rb_raise(rb_eRuntimeError, "Doesn't have referenxe to context internally");
}
GetVADContext(self, rwvc);
TypedData_Get_Struct(params, ruby_whisper_vad_params, &ruby_whisper_vad_params_type, rwvp);
if (rb_respond_to(file_path, id_to_path)) {
file_path = rb_funcall(file_path, id_to_path, 0);
}
cpp_file_path = StringValueCStr(file_path);
if (!read_audio_data(cpp_file_path, pcmf32, pcmf32s, false)) {

1
bindings/ruby/ext/ruby_whisper_vad_params.c
View File

@ -1,4 +1,3 @@
#include <ruby.h>
#include "ruby_whisper.h"
#define DEFINE_PARAM(param_name, nth) \

1
bindings/ruby/ext/ruby_whisper_vad_segment.c
View File

@ -1,4 +1,3 @@
#include <ruby.h>
#include "ruby_whisper.h"
#define N_KEY_NAMES 2

1
bindings/ruby/ext/ruby_whisper_vad_segments.c
View File

@ -1,4 +1,3 @@
#include <ruby.h>
#include "ruby_whisper.h"
extern ID id___method__;

45
bindings/ruby/sig/whisper.rbs
View File

@ -17,6 +17,21 @@ module Whisper
LOG_LEVEL_ERROR: Integer
LOG_LEVEL_DEBUG: Integer
LOG_LEVEL_CONT: Integer
AHEADS_NONE: Integer
AHEADS_N_TOP_MOST: Integer
AHEADS_CUSTOM: Integer
AHEADS_TINY_EN: Integer
AHEADS_TINY: Integer
AHEADS_BASE_EN: Integer
AHEADS_BASE: Integer
AHEADS_SMALL_EN: Integer
AHEADS_SMALL: Integer
AHEADS_MEDIUM_EN: Integer
AHEADS_MEDIUM: Integer
AHEADS_LARGE_V1: Integer
AHEADS_LARGE_V2: Integer
AHEADS_LARGE_V3: Integer
AHEADS_LARGE_V3_TURBO: Integer
def self.lang_max_id: () -> Integer
def self.lang_id: (string name) -> Integer
@ -37,8 +52,8 @@ module Whisper
# puts text
# end
#
def transcribe: (string, Params, ?n_processors: Integer) -> self
| (string, Params, ?n_processors: Integer) { (String) -> void } -> self
def transcribe: (path, Params, ?n_processors: Integer) -> self
| (path, Params, ?n_processors: Integer) { (String) -> void } -> self
def model_n_vocab: () -> Integer
def model_n_audio_ctx: () -> Integer
@ -120,6 +135,30 @@ module Whisper
def to_srt: () -> String
def to_webvtt: () -> String
class Params
def self.new: (
use_gpu: boolish,
flash_attn: boolish,
gpu_device: Integer,
dtw_token_timestamps: boolish,
dtw_aheads_preset: Integer,
dtw_n_top: Integer | nil,
) -> instance
def use_gpu=: (boolish) -> boolish
def use_gpu: () -> (true | false)
def flash_attn=: (boolish) -> boolish
def flash_attn: () -> (true | false)
def gpu_device=: (Integer) -> Integer
def gpu_device: () -> Integer
def dtw_token_timestamps=: (boolish) -> boolish
def dtw_token_timestamps: () -> (true | false)
def dtw_aheads_preset=: (Integer) -> Integer
def dtw_aheads_preset: () -> Integer
def dtw_n_top=: (Integer | nil) -> (Integer | nil)
def dtw_n_top: () -> (Integer | nil)
end
end
class Params
@ -603,6 +642,8 @@ module Whisper
class Context
def self.new: (String | path | ::URI::HTTP model_name_or_path) -> instance
def segments_from_samples: (Params, Array[Float] samples, ?Integer n_samples) -> Segments
| (Params, _Samples, ?Integer n_samples) -> Segments
def detect: (path wav_file_path, Params) -> Segments
end

82
bindings/ruby/test/test_context_params.rb Normal file
View File

@ -0,0 +1,82 @@
require_relative "helper"
class TestContextParams < TestBase
PARAM_NAMES = [
:use_gpu,
:flash_attn,
:gpu_device,
:dtw_token_timestamps,
:dtw_aheads_preset,
:dtw_n_top
]
def test_new
params = Whisper::Context::Params.new
assert_instance_of Whisper::Context::Params, params
end
def test_attributes
params = Whisper::Context::Params.new
assert_true params.use_gpu
params.use_gpu = false
assert_false params.use_gpu
assert_true params.flash_attn
params.flash_attn = false
assert_false params.flash_attn
assert_equal 0, params.gpu_device
params.gpu_device = 1
assert_equal 1, params.gpu_device
assert_false params.dtw_token_timestamps
params.dtw_token_timestamps = true
assert_true params.dtw_token_timestamps
assert_equal Whisper::AHEADS_NONE, params.dtw_aheads_preset
params.dtw_aheads_preset =Whisper::AHEADS_BASE
assert_equal Whisper::AHEADS_BASE, params.dtw_aheads_preset
assert_nil params.dtw_n_top
params.dtw_n_top = 6
assert_equal 6, params.dtw_n_top
params.dtw_n_top = nil
assert_nil params.dtw_n_top
end
def test_new_with_kw_args
params = Whisper::Context::Params.new(use_gpu: false)
assert_false params.use_gpu
end
def test_new_with_kw_wargs_non_existent
assert_raise ArgumentError do
Whisper::Context::Params.new(non_existent: "value")
end
end
data(PARAM_NAMES.collect {|param| [param, param]}.to_h)
def test_new_with_kw_args_default_values(param)
default_params = Whisper::Context::Params.new
default_value = default_params.send(param)
value = if param == :dtw_n_top
6
else
case default_value
in true | false
!default_value
in Integer
default_value + 1
end
end
params = Whisper::Context::Params.new(param => value)
assert_equal value, params.send(param)
PARAM_NAMES.reject {|name| name == param}.each do |name|
expected = default_params.send(name)
actual = params.send(name)
assert_equal expected, actual
end
end
end

11
bindings/ruby/test/test_token.rb
View File

@ -56,6 +56,17 @@ class TestToken < TestBase
@segment.each_token.collect(&:text)
end
def test_token_timestamps
params = Whisper::Params.new(token_timestamps: true)
whisper.transcribe(TestBase::AUDIO, params)
prev = -1
whisper.each_segment.first.each_token do |token|
assert token.start_time >= prev
assert token.end_time >= token.start_time
prev = token.end_time
end
end
def test_deconstruct_keys_with_nil
keys = %i[id tid probability log_probability pt ptsum t_dtw voice_length start_time end_time text]
expected = keys.collect {|key| [key, @token.send(key)] }.to_h

66
bindings/ruby/test/test_vad_context.rb
View File

@ -9,6 +9,25 @@ class TestVADContext < TestBase
def test_detect
context = Whisper::VAD::Context.new("silero-v6.2.0")
segments = context.detect(AUDIO, Whisper::VAD::Params.new)
assert_segments segments
end
def test_invalid_model_type
assert_raise TypeError do
Whisper::VAD::Context.new(Object.new)
end
end
def test_allocate
vad = Whisper::VAD::Context.allocate
assert_raise do
vad.detect(AUDIO, Whisper::VAD::Params.new)
end
end
private
def assert_segments(segments)
assert_instance_of Whisper::VAD::Segments, segments
i = 0
@ -35,16 +54,47 @@ class TestVADContext < TestBase
assert_equal 4, segments.length
end
def test_invalid_model_type
assert_raise TypeError do
Whisper::VAD::Context.new(Object.new)
sub_test_case "from samples" do
def setup
super
@vad = Whisper::VAD::Context.new("silero-v6.2.0")
@samples = File.read(AUDIO, nil, 78).unpack("s<*").collect {|i| i.to_f / 2**15}
end
end
def test_allocate
vad = Whisper::VAD::Context.allocate
assert_raise do
vad.detect(AUDIO, Whisper::VAD::Params.new)
def test_segments_from_samples
segments = @vad.segments_from_samples(Whisper::VAD::Params.new, @samples, @samples.length)
assert_segments segments
end
def test_segments_from_samples_without_length
segments = @vad.segments_from_samples(Whisper::VAD::Params.new, @samples)
assert_segments segments
end
def test_segments_from_samples_enumerator
samples = @samples.each
segments = @vad.segments_from_samples(Whisper::VAD::Params.new, samples, @samples.length)
assert_segments segments
end
def test_segments_from_samples_enumerator_without_length
samples = @samples.each
assert_raise ArgumentError do
@vad.segments_from_samples(Whisper::VAD::Params.new, samples)
end
end
def test_segments_from_samples_enumerator_with_too_large_length
samples = @samples.each.take(10).to_enum
assert_raise StopIteration do
@vad.segments_from_samples(Whisper::VAD::Params.new, samples, 11)
end
end
def test_segments_from_samples_with_memory_view
samples = JFKReader.new(AUDIO)
segments = @vad.segments_from_samples(Whisper::VAD::Params.new, samples)
assert_segments segments
end
end
end

20
bindings/ruby/test/test_whisper.rb
View File

@ -1,6 +1,7 @@
require_relative "helper"
require "stringio"
require "etc"
require "pathname"
# Exists to detect memory-related bug
Whisper.log_set ->(level, buffer, user_data) {}, nil
@ -20,6 +21,15 @@ class TestWhisper < TestBase
}
end
def test_whisper_pathname
@whisper = Whisper::Context.new("base.en")
params = Whisper::Params.new
@whisper.transcribe(Pathname(AUDIO), params) {|text|
assert_match(/ask not what your country can do for you, ask what you can do for your country/, text)
}
end
def test_transcribe_non_parallel
@whisper = Whisper::Context.new("base.en")
params = Whisper::Params.new
@ -207,6 +217,16 @@ class TestWhisper < TestBase
assert_match(/ask not what your country can do for you, ask what you can do for your country/, @whisper.each_segment.first.text)
end
def test_full_with_memroy_view_gc
samples = JFKReader.new(AUDIO)
@whisper.full(@params, samples)
GC.start
require "fiddle"
Fiddle::MemoryView.export samples do |view|
assert_equal 176000, view.to_s.unpack("#{view.format}*").length
end
end
def test_full_parallel
nprocessors = 2
@whisper.full_parallel(@params, @samples, @samples.length, nprocessors)

2
bindings/ruby/whispercpp.gemspec
View File

@ -3,7 +3,7 @@ require_relative "extsources"
Gem::Specification.new do |s|
s.name = "whispercpp"
s.authors = ["Georgi Gerganov", "Todd A. Fisher"]
s.version = '1.3.5'
s.version = '1.3.6'
s.description = %q{High-performance inference of OpenAI's Whisper automatic speech recognition (ASR) model via Ruby}
s.email = 'todd.fisher@gmail.com'
s.extra_rdoc_files = ['LICENSE', 'README.md']

2
build-xcframework.sh
View File

@ -559,7 +559,7 @@ xcodebuild -create-xcframework \
-framework $(pwd)/build-ios-device/framework/whisper.framework \
-debug-symbols $(pwd)/build-ios-device/dSYMs/whisper.dSYM \
-framework $(pwd)/build-macos/framework/whisper.framework \
-debug-symbols $(pwd)/build-macos/dSYMS/whisper.dSYM \
-debug-symbols $(pwd)/build-macos/dSYMs/whisper.dSYM \
-framework $(pwd)/build-visionos/framework/whisper.framework \
-debug-symbols $(pwd)/build-visionos/dSYMs/whisper.dSYM \
-framework $(pwd)/build-visionos-sim/framework/whisper.framework \

2
close-issue.yml
View File

@ -15,7 +15,7 @@ jobs:
issues: write
pull-requests: write
steps:
- uses: actions/stale@v5
- uses: actions/stale@v10
with:
exempt-issue-labels: "refactor,help wanted,good first issue,research,bug,roadmap"
days-before-issue-stale: 30

45
cmake/whisper-config.cmake.in
View File

@ -3,60 +3,25 @@ set(WHISPER_BUILD_COMMIT @WHISPER_BUILD_COMMIT@)
set(WHISPER_BUILD_NUMBER @WHISPER_BUILD_NUMBER@)
set(WHISPER_SHARED_LIB @BUILD_SHARED_LIBS@)
set(GGML_BLAS @GGML_BLAS@)
set(GGML_CUDA @GGML_CUDA@)
set(GGML_METAL @GGML_METAL@)
set(GGML_HIPBLAS @GGML_HIPBLAS@)
set(GGML_ACCELERATE @GGML_ACCELERATE@)
@PACKAGE_INIT@
set_and_check(WHISPER_INCLUDE_DIR "@PACKAGE_WHISPER_INCLUDE_INSTALL_DIR@")
set_and_check(WHISPER_LIB_DIR "@PACKAGE_WHISPER_LIB_INSTALL_DIR@")
set_and_check(WHISPER_BIN_DIR "@PACKAGE_WHISPER_BIN_INSTALL_DIR@")
# Ensure transient dependencies satisfied
find_package(Threads REQUIRED)
if (APPLE AND GGML_ACCELERATE)
find_library(ACCELERATE_FRAMEWORK Accelerate REQUIRED)
endif()
if (GGML_BLAS)
find_package(BLAS REQUIRED)
endif()
if (GGML_CUDA)
find_package(CUDAToolkit REQUIRED)
endif()
if (GGML_METAL)
find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
find_library(METAL_FRAMEWORK Metal REQUIRED)
find_library(METALKIT_FRAMEWORK MetalKit REQUIRED)
endif()
if (GGML_HIPBLAS)
find_package(hip REQUIRED)
find_package(hipblas REQUIRED)
find_package(rocblas REQUIRED)
endif()
find_package(ggml REQUIRED HINTS ${LLAMA_LIB_DIR}/cmake)
find_library(whisper_LIBRARY whisper
REQUIRED
HINTS ${WHISPER_LIB_DIR})
set(_whisper_link_deps "Threads::Threads" "@WHISPER_EXTRA_LIBS@")
set(_whisper_transient_defines "@WHISPER_TRANSIENT_DEFINES@")
HINTS ${WHISPER_LIB_DIR}
NO_CMAKE_FIND_ROOT_PATH
)
add_library(whisper UNKNOWN IMPORTED)
set_target_properties(whisper
PROPERTIES
INTERFACE_INCLUDE_DIRECTORIES "${WHISPER_INCLUDE_DIR}"
INTERFACE_LINK_LIBRARIES "${_whisper_link_deps}"
INTERFACE_COMPILE_DEFINITIONS "${_whisper_transient_defines}"
INTERFACE_LINK_LIBRARIES "ggml::ggml;ggml::ggml-base;"
IMPORTED_LINK_INTERFACE_LANGUAGES "CXX"
IMPORTED_LOCATION "${whisper_LIBRARY}"
INTERFACE_COMPILE_FEATURES cxx_std_11

8
examples/cli/cli.cpp
View File

@ -77,6 +77,7 @@ struct whisper_params {
bool log_score = false;
bool use_gpu = true;
bool flash_attn = true;
int32_t gpu_device = 0;
bool suppress_nst = false;
bool carry_initial_prompt = false;
@ -129,6 +130,10 @@ static char * requires_value_error(const std::string & arg) {
}
static bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
if (const char * env_device = std::getenv("WHISPER_ARG_DEVICE")) {
params.gpu_device = std::stoi(env_device);
}
for (int i = 1; i < argc; i++) {
std::string arg = argv[i];
@ -195,6 +200,7 @@ static bool whisper_params_parse(int argc, char ** argv, whisper_params & params
else if (arg == "-dtw" || arg == "--dtw") { params.dtw = ARGV_NEXT; }
else if (arg == "-ls" || arg == "--log-score") { params.log_score = true; }
else if (arg == "-ng" || arg == "--no-gpu") { params.use_gpu = false; }
else if (arg == "-dev" || arg == "--device") { params.gpu_device = std::stoi(ARGV_NEXT); }
else if (arg == "-fa" || arg == "--flash-attn") { params.flash_attn = true; }
else if (arg == "-nfa" || arg == "--no-flash-attn") { params.flash_attn = false; }
else if (arg == "-sns" || arg == "--suppress-nst") { params.suppress_nst = true; }
@ -276,6 +282,7 @@ static void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params
fprintf(stderr, " -dtw MODEL --dtw MODEL [%-7s] compute token-level timestamps\n", params.dtw.c_str());
fprintf(stderr, " -ls, --log-score [%-7s] log best decoder scores of tokens\n", params.log_score?"true":"false");
fprintf(stderr, " -ng, --no-gpu [%-7s] disable GPU\n", params.use_gpu ? "false" : "true");
fprintf(stderr, " -dev N, --device N [%-7d] GPU device ID (default: 0)\n", params.gpu_device);
fprintf(stderr, " -fa, --flash-attn [%-7s] enable flash attention\n", params.flash_attn ? "true" : "false");
fprintf(stderr, " -nfa, --no-flash-attn [%-7s] disable flash attention\n", params.flash_attn ? "false" : "true");
fprintf(stderr, " -sns, --suppress-nst [%-7s] suppress non-speech tokens\n", params.suppress_nst ? "true" : "false");
@ -1003,6 +1010,7 @@ int main(int argc, char ** argv) {
struct whisper_context_params cparams = whisper_context_default_params();
cparams.use_gpu = params.use_gpu;
cparams.gpu_device = params.gpu_device;
cparams.flash_attn = params.flash_attn;
if (!params.dtw.empty()) {

2
examples/common-ggml.cpp
View File

@ -73,6 +73,7 @@ bool ggml_common_quantize_0(
case GGML_FTYPE_MOSTLY_IQ1_M:
case GGML_FTYPE_MOSTLY_BF16:
case GGML_FTYPE_MOSTLY_MXFP4:
case GGML_FTYPE_MOSTLY_NVFP4:
{
fprintf(stderr, "%s: invalid model type %d\n", __func__, ftype);
return false;
@ -213,6 +214,7 @@ bool ggml_common_quantize_0(
case GGML_TYPE_TQ1_0:
case GGML_TYPE_TQ2_0:
case GGML_TYPE_MXFP4:
case GGML_TYPE_NVFP4:
case GGML_TYPE_COUNT:
{
fprintf(stderr, "%s: unsupported quantization type %d (%s)\n", __func__, ttype, ggml_type_name((ggml_type) ttype));

6910
examples/miniaudio.h
View File

File diff suppressed because it is too large Load Diff

22
examples/server/server.cpp
View File

@ -103,6 +103,7 @@ struct whisper_params {
bool no_timestamps = false;
bool use_gpu = true;
bool flash_attn = true;
int32_t gpu_device = 0;
bool suppress_nst = false;
bool no_context = true;
bool no_language_probabilities = false;
@ -179,6 +180,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
fprintf(stderr, " -sns, --suppress-nst [%-7s] suppress non-speech tokens\n", params.suppress_nst ? "true" : "false");
fprintf(stderr, " -nth N, --no-speech-thold N [%-7.2f] no speech threshold\n", params.no_speech_thold);
fprintf(stderr, " -ng, --no-gpu [%-7s] do not use gpu\n", params.use_gpu ? "false" : "true");
fprintf(stderr, " -dev N, --device N [%-7d] GPU device ID (default: 0)\n", params.gpu_device);
fprintf(stderr, " -fa, --flash-attn [%-7s] enable flash attention\n", params.flash_attn ? "true" : "false");
fprintf(stderr, " -nfa, --no-flash-attn [%-7s] disable flash attention\n", params.flash_attn ? "false" : "true");
fprintf(stderr, " -nlp, --no-language-probabilities [%-7s] exclude language probabilities from verbose_json output\n", params.no_language_probabilities ? "true" : "false");
@ -198,6 +200,10 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
}
bool whisper_params_parse(int argc, char ** argv, whisper_params & params, server_params & sparams) {
if (const char * env_device = std::getenv("WHISPER_ARG_DEVICE")) {
params.gpu_device = std::stoi(env_device);
}
for (int i = 1; i < argc; i++) {
std::string arg = argv[i];
@ -237,6 +243,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params, serve
else if (arg == "-oved" || arg == "--ov-e-device") { params.openvino_encode_device = argv[++i]; }
else if (arg == "-dtw" || arg == "--dtw") { params.dtw = argv[++i]; }
else if (arg == "-ng" || arg == "--no-gpu") { params.use_gpu = false; }
else if (arg == "-dev" || arg == "--device") { params.gpu_device = std::stoi(argv[++i]); }
else if (arg == "-fa" || arg == "--flash-attn") { params.flash_attn = true; }
else if (arg == "-nfa" || arg == "--no-flash-attn") { params.flash_attn = false; }
else if (arg == "-sns" || arg == "--suppress-nst") { params.suppress_nst = true; }
@ -643,6 +650,7 @@ int main(int argc, char ** argv) {
struct whisper_context_params cparams = whisper_context_default_params();
cparams.use_gpu = params.use_gpu;
cparams.gpu_device = params.gpu_device;
cparams.flash_attn = params.flash_attn;
if (!params.dtw.empty()) {
@ -740,9 +748,9 @@ int main(int argc, char ** argv) {
<body>
<h1>Whisper.cpp Server</h1>
<h2>/inference</h2>
<h2>)" + sparams.request_path + sparams.inference_path + R"(</h2>
<pre>
curl 127.0.0.1:)" + std::to_string(sparams.port) + R"(/inference \
curl 127.0.0.1:)" + std::to_string(sparams.port) + sparams.request_path + sparams.inference_path + R"( \
-H "Content-Type: multipart/form-data" \
-F file="@&lt;file-path&gt;" \
-F temperature="0.0" \
@ -759,7 +767,7 @@ int main(int argc, char ** argv) {
<div>
<h2>Try it out</h2>
<form action="/inference" method="POST" enctype="multipart/form-data">
<form action=")" + sparams.request_path + sparams.inference_path + R"(" method="POST" enctype="multipart/form-data">
<label for="file">Choose an audio file:</label>
<input type="file" id="file" name="file" accept="audio/*" required><br>
@ -803,6 +811,7 @@ int main(int argc, char ** argv) {
{
fprintf(stderr, "error: no 'file' field in the request\n");
const std::string error_resp = "{\"error\":\"no 'file' field in the request\"}";
res.status = 400;
res.set_content(error_resp, "application/json");
return;
}
@ -829,6 +838,7 @@ int main(int argc, char ** argv) {
std::string error_resp = "{\"error\":\"Failed to execute ffmpeg command.\"}";
const bool is_converted = convert_to_wav(temp_filename, error_resp);
if (!is_converted) {
res.status = 500;
res.set_content(error_resp, "application/json");
return;
}
@ -838,6 +848,7 @@ int main(int argc, char ** argv) {
{
fprintf(stderr, "error: failed to read WAV file '%s'\n", temp_filename.c_str());
const std::string error_resp = "{\"error\":\"failed to read WAV file\"}";
res.status = 400;
res.set_content(error_resp, "application/json");
std::remove(temp_filename.c_str());
return;
@ -849,6 +860,7 @@ int main(int argc, char ** argv) {
{
fprintf(stderr, "error: failed to read audio data\n");
const std::string error_resp = "{\"error\":\"failed to read audio data\"}";
res.status = 400;
res.set_content(error_resp, "application/json");
return;
}
@ -927,7 +939,7 @@ int main(int argc, char ** argv) {
wparams.logprob_thold = params.logprob_thold;
wparams.no_timestamps = params.no_timestamps;
wparams.token_timestamps = !params.no_timestamps && params.response_format == vjson_format;
wparams.token_timestamps = !params.no_timestamps;
wparams.no_context = params.no_context;
wparams.suppress_nst = params.suppress_nst;
@ -1119,6 +1131,7 @@ int main(int argc, char ** argv) {
{
fprintf(stderr, "error: no 'model' field in the request\n");
const std::string error_resp = "{\"error\":\"no 'model' field in the request\"}";
res.status = 400;
res.set_content(error_resp, "application/json");
return;
}
@ -1127,6 +1140,7 @@ int main(int argc, char ** argv) {
{
fprintf(stderr, "error: 'model': %s not found!\n", model.c_str());
const std::string error_resp = "{\"error\":\"model not found!\"}";
res.status = 400;
res.set_content(error_resp, "application/json");
return;
}

5
examples/talk-llama/CMakeLists.txt
View File

@ -22,18 +22,19 @@ if (WHISPER_SDL2)
llama-kv-cache-iswa.cpp
llama-memory-recurrent.cpp
llama-memory-hybrid.cpp
llama-memory-hybrid-iswa.cpp
llama-memory.cpp
llama-mmap.cpp
llama-model-loader.cpp
llama-model-saver.cpp
llama-model.cpp
llama-quant.cpp
llama-sampling.cpp
llama-sampler.cpp
llama-vocab.cpp
unicode.cpp
unicode-data.cpp
${SRC_MODELS})
target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS})
target_include_directories(${TARGET} PRIVATE . ${SDL2_INCLUDE_DIRS})
target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
install(TARGETS ${TARGET} RUNTIME)

20
examples/talk-llama/llama-adapter.cpp
View File

@ -146,11 +146,9 @@ llama_adapter_lora_weight * llama_adapter_lora::get_weight(ggml_tensor * w) {
return nullptr;
}
static void llama_adapter_lora_init_impl(const char * path_lora, llama_adapter_lora & adapter) {
static void llama_adapter_lora_init_impl(llama_model & model, const char * path_lora, llama_adapter_lora & adapter) {
LLAMA_LOG_INFO("%s: loading lora adapter from '%s' ...\n", __func__, path_lora);
llama_model & model = adapter.model;
ggml_context * ctx_init;
gguf_init_params meta_gguf_params = {
/* .no_alloc = */ true,
@ -413,17 +411,17 @@ static void llama_adapter_lora_init_impl(const char * path_lora, llama_adapter_l
}
}
// update number of nodes used
model.n_lora_nodes += adapter.get_n_nodes();
// register adapter with model
model.loras.insert(&adapter);
LLAMA_LOG_INFO("%s: loaded %zu tensors from lora file\n", __func__, adapter.ab_map.size()*2);
}
llama_adapter_lora * llama_adapter_lora_init(llama_model * model, const char * path_lora) {
llama_adapter_lora * adapter = new llama_adapter_lora(*model);
llama_adapter_lora * adapter = new llama_adapter_lora();
try {
llama_adapter_lora_init_impl(path_lora, *adapter);
llama_adapter_lora_init_impl(*model, path_lora, *adapter);
return adapter;
} catch (const std::exception & err) {
LLAMA_LOG_ERROR("%s: failed to apply lora adapter: %s\n", __func__, err.what());
@ -473,12 +471,8 @@ int32_t llama_adapter_meta_val_str_by_index(const llama_adapter_lora * adapter,
return snprintf(buf, buf_size, "%s", it->second.c_str());
}
void llama_adapter_lora_free(llama_adapter_lora * adapter) {
// update number of nodes used
GGML_ASSERT(adapter->model.n_lora_nodes >= adapter->get_n_nodes());
adapter->model.n_lora_nodes -= adapter->get_n_nodes();
delete adapter;
void llama_adapter_lora_free(llama_adapter_lora *) {
// deprecated: adapters are freed by llama_model's destructor
}
uint64_t llama_adapter_get_alora_n_invocation_tokens(const struct llama_adapter_lora * adapter) {

7
examples/talk-llama/llama-adapter.h
View File

@ -39,6 +39,8 @@ private:
std::vector<ggml_tensor *> tensors; // per layer
};
using llama_adapter_cvec_ptr = std::shared_ptr<llama_adapter_cvec>;
//
// llama_adapter_lora
//
@ -59,8 +61,6 @@ struct llama_adapter_lora_weight {
};
struct llama_adapter_lora {
llama_model & model;
// map tensor name to lora_a_b
std::unordered_map<std::string, llama_adapter_lora_weight> ab_map;
@ -75,7 +75,7 @@ struct llama_adapter_lora {
// activated lora (aLoRA)
std::vector<llama_token> alora_invocation_tokens;
llama_adapter_lora(llama_model & model) : model(model) {}
llama_adapter_lora() = default;
~llama_adapter_lora() = default;
llama_adapter_lora_weight * get_weight(ggml_tensor * w);
@ -86,3 +86,4 @@ struct llama_adapter_lora {
};
using llama_adapter_loras = std::unordered_map<llama_adapter_lora *, float>;
using llama_adapter_loras_ptr = std::unique_ptr<llama_adapter_loras>;

352
examples/talk-llama/llama-arch.cpp
View File

@ -4,6 +4,7 @@
#include <map>
#include <set>
#include <vector>
static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_CLIP, "clip" }, // dummy, only used by llama-quantize
@ -26,6 +27,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_NEO_BERT, "neo-bert" },
{ LLM_ARCH_JINA_BERT_V2, "jina-bert-v2" },
{ LLM_ARCH_JINA_BERT_V3, "jina-bert-v3" },
{ LLM_ARCH_EUROBERT, "eurobert" },
{ LLM_ARCH_BLOOM, "bloom" },
{ LLM_ARCH_STABLELM, "stablelm" },
{ LLM_ARCH_QWEN, "qwen" },
@ -37,6 +39,8 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_QWEN3NEXT, "qwen3next" },
{ LLM_ARCH_QWEN3VL, "qwen3vl" },
{ LLM_ARCH_QWEN3VLMOE, "qwen3vlmoe" },
{ LLM_ARCH_QWEN35, "qwen35" },
{ LLM_ARCH_QWEN35MOE, "qwen35moe" },
{ LLM_ARCH_PHI2, "phi2" },
{ LLM_ARCH_PHI3, "phi3" },
{ LLM_ARCH_PHIMOE, "phimoe" },
@ -72,15 +76,18 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_CHATGLM, "chatglm" },
{ LLM_ARCH_GLM4, "glm4" },
{ LLM_ARCH_GLM4_MOE, "glm4moe" },
{ LLM_ARCH_GLM_DSA, "glm-dsa" },
{ LLM_ARCH_BITNET, "bitnet" },
{ LLM_ARCH_T5, "t5" },
{ LLM_ARCH_T5ENCODER, "t5encoder" },
{ LLM_ARCH_JAIS, "jais" },
{ LLM_ARCH_JAIS2, "jais2" },
{ LLM_ARCH_NEMOTRON, "nemotron" },
{ LLM_ARCH_NEMOTRON_H, "nemotron_h" },
{ LLM_ARCH_NEMOTRON_H_MOE, "nemotron_h_moe" },
{ LLM_ARCH_EXAONE, "exaone" },
{ LLM_ARCH_EXAONE4, "exaone4" },
{ LLM_ARCH_EXAONE_MOE, "exaone-moe" },
{ LLM_ARCH_RWKV6, "rwkv6" },
{ LLM_ARCH_RWKV6QWEN2, "rwkv6qwen2" },
{ LLM_ARCH_RWKV7, "rwkv7" },
@ -116,9 +123,12 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_RND1, "rnd1" },
{ LLM_ARCH_PANGU_EMBED, "pangu-embedded" },
{ LLM_ARCH_MISTRAL3, "mistral3" },
{ LLM_ARCH_MIMO2, "mimo2" },
{ LLM_ARCH_PADDLEOCR, "paddleocr" },
{ LLM_ARCH_MIMO2, "mimo2" },
{ LLM_ARCH_STEP35, "step35" },
{ LLM_ARCH_LLAMA_EMBED, "llama-embed" },
{ LLM_ARCH_MAINCODER, "maincoder" },
{ LLM_ARCH_KIMI_LINEAR, "kimi-linear" },
{ LLM_ARCH_UNKNOWN, "(unknown)" },
};
@ -160,6 +170,8 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_EXPERT_FEED_FORWARD_LENGTH, "%s.expert_feed_forward_length" },
{ LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, "%s.expert_shared_feed_forward_length" },
{ LLM_KV_EXPERT_CHUNK_FEED_FORWARD_LENGTH, "%s.expert_chunk_feed_forward_length" },
{ LLM_KV_SWIGLU_CLAMP_EXP, "%s.swiglu_clamp_exp" },
{ LLM_KV_SWIGLU_CLAMP_SHEXP, "%s.swiglu_clamp_shexp" },
{ LLM_KV_USE_PARALLEL_RESIDUAL, "%s.use_parallel_residual" },
{ LLM_KV_TENSOR_DATA_LAYOUT, "%s.tensor_data_layout" },
{ LLM_KV_EXPERT_COUNT, "%s.expert_count" },
@ -173,6 +185,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_EXPERT_GROUP_SCALE, "%s.expert_group_scale" },
{ LLM_KV_EXPERTS_PER_GROUP, "%s.experts_per_group" },
{ LLM_KV_MOE_EVERY_N_LAYERS, "%s.moe_every_n_layers" },
{ LLM_KV_MOE_LATENT_SIZE, "%s.moe_latent_size" },
{ LLM_KV_NEXTN_PREDICT_LAYERS, "%s.nextn_predict_layers" },
{ LLM_KV_NUM_DEEPSTACK_LAYERS, "%s.n_deepstack_layers" },
{ LLM_KV_POOLING_TYPE, "%s.pooling_type" },
@ -190,6 +203,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_EMBEDDING_SCALE, "%s.embedding_scale" },
{ LLM_KV_TOKEN_SHIFT_COUNT, "%s.token_shift_count" },
{ LLM_KV_INTERLEAVE_MOE_LAYER_STEP, "%s.interleave_moe_layer_step" },
{ LLM_KV_FULL_ATTENTION_INTERVAL, "%s.full_attention_interval" },
{ LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" },
{ LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" },
@ -217,22 +231,28 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_ATTENTION_TEMPERATURE_SCALE, "%s.attention.temperature_scale" },
{ LLM_KV_ATTENTION_KEY_LENGTH_MLA, "%s.attention.key_length_mla" },
{ LLM_KV_ATTENTION_VALUE_LENGTH_MLA, "%s.attention.value_length_mla" },
{ LLM_KV_ATTENTION_KEY_LENGTH_SWA, "%s.attention.key_length_swa" },
{ LLM_KV_ATTENTION_VALUE_LENGTH_SWA, "%s.attention.value_length_swa" },
{ LLM_KV_ATTENTION_INDEXER_HEAD_COUNT, "%s.attention.indexer.head_count" },
{ LLM_KV_ATTENTION_INDEXER_KEY_LENGTH, "%s.attention.indexer.key_length" },
{ LLM_KV_ATTENTION_INDEXER_TOP_K, "%s.attention.indexer.top_k" },
{ LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" },
{ LLM_KV_ROPE_DIMENSION_SECTIONS, "%s.rope.dimension_sections" },
{ LLM_KV_ROPE_FREQ_BASE, "%s.rope.freq_base" },
{ LLM_KV_ROPE_FREQ_BASE_SWA, "%s.rope.freq_base_swa" },
{ LLM_KV_ROPE_SCALE_LINEAR, "%s.rope.scale_linear" },
{ LLM_KV_ROPE_SCALING_TYPE, "%s.rope.scaling.type" },
{ LLM_KV_ROPE_SCALING_FACTOR, "%s.rope.scaling.factor" },
{ LLM_KV_ROPE_SCALING_ATTN_FACTOR, "%s.rope.scaling.attn_factor" },
{ LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, "%s.rope.scaling.original_context_length" },
{ LLM_KV_ROPE_SCALING_FINETUNED, "%s.rope.scaling.finetuned" },
{ LLM_KV_ROPE_SCALING_YARN_LOG_MUL, "%s.rope.scaling.yarn_log_multiplier" },
{ LLM_KV_ROPE_SCALING_YARN_EXT_FACTOR, "%s.rope.scaling.yarn_ext_factor" },
{ LLM_KV_ROPE_SCALING_YARN_ATTN_FACTOR, "%s.rope.scaling.yarn_attn_factor" },
{ LLM_KV_ROPE_SCALING_YARN_BETA_FAST, "%s.rope.scaling.yarn_beta_fast" },
{ LLM_KV_ROPE_SCALING_YARN_BETA_SLOW, "%s.rope.scaling.yarn_beta_slow" },
{ LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" },
{ LLM_KV_ROPE_DIMENSION_COUNT_SWA, "%s.rope.dimension_count_swa" },
{ LLM_KV_ROPE_DIMENSION_SECTIONS, "%s.rope.dimension_sections" },
{ LLM_KV_ROPE_FREQ_BASE, "%s.rope.freq_base" },
{ LLM_KV_ROPE_FREQ_BASE_SWA, "%s.rope.freq_base_swa" },
{ LLM_KV_ROPE_SCALE_LINEAR, "%s.rope.scale_linear" },
{ LLM_KV_ROPE_SCALING_TYPE, "%s.rope.scaling.type" },
{ LLM_KV_ROPE_SCALING_FACTOR, "%s.rope.scaling.factor" },
{ LLM_KV_ROPE_SCALING_ATTN_FACTOR, "%s.rope.scaling.attn_factor" },
{ LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, "%s.rope.scaling.original_context_length" },
{ LLM_KV_ROPE_SCALING_FINETUNED, "%s.rope.scaling.finetuned" },
{ LLM_KV_ROPE_SCALING_YARN_LOG_MUL, "%s.rope.scaling.yarn_log_multiplier" },
{ LLM_KV_ROPE_SCALING_YARN_EXT_FACTOR, "%s.rope.scaling.yarn_ext_factor" },
{ LLM_KV_ROPE_SCALING_YARN_ATTN_FACTOR, "%s.rope.scaling.yarn_attn_factor" },
{ LLM_KV_ROPE_SCALING_YARN_BETA_FAST, "%s.rope.scaling.yarn_beta_fast" },
{ LLM_KV_ROPE_SCALING_YARN_BETA_SLOW, "%s.rope.scaling.yarn_beta_slow" },
{ LLM_KV_SPLIT_NO, "split.no" },
{ LLM_KV_SPLIT_COUNT, "split.count" },
@ -245,6 +265,8 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_SSM_GROUP_COUNT, "%s.ssm.group_count" },
{ LLM_KV_SSM_DT_B_C_RMS, "%s.ssm.dt_b_c_rms" },
{ LLM_KV_KDA_HEAD_DIM, "%s.kda.head_dim" },
{ LLM_KV_WKV_HEAD_SIZE, "%s.wkv.head_size" },
{ LLM_KV_POSNET_EMBEDDING_LENGTH, "%s.posnet.embedding_length" },
@ -332,6 +354,7 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
{ LLM_TENSOR_FFN_DOWN_EXP, "blk.%d.ffn_down.%d" },
{ LLM_TENSOR_FFN_UP_EXP, "blk.%d.ffn_up.%d" },
{ LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" },
{ LLM_TENSOR_FFN_GATE_UP_EXPS, "blk.%d.ffn_gate_up_exps" },
{ LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" },
{ LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" },
{ LLM_TENSOR_ATTN_POST_NORM, "blk.%d.post_attention_norm" },
@ -343,6 +366,8 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
{ LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" },
{ LLM_TENSOR_FFN_DOWN_SHEXP, "blk.%d.ffn_down_shexp" },
{ LLM_TENSOR_FFN_EXP_PROBS_B, "blk.%d.exp_probs_b" },
{ LLM_TENSOR_FFN_LATENT_DOWN, "blk.%d.ffn_latent_down" },
{ LLM_TENSOR_FFN_LATENT_UP, "blk.%d.ffn_latent_up" },
{ LLM_TENSOR_ATTN_NORM_2, "blk.%d.attn_norm_2" },
{ LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" },
{ LLM_TENSOR_LAYER_OUT_NORM, "blk.%d.layer_output_norm" },
@ -353,12 +378,14 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
{ LLM_TENSOR_TOKEN_TYPES, "token_types" },
{ LLM_TENSOR_CLS, "cls" },
{ LLM_TENSOR_CLS_OUT, "cls.output" },
{ LLM_TENSOR_CLS_NORM, "cls.norm" },
{ LLM_TENSOR_ENC_OUTPUT_NORM, "enc.output_norm" },
{ LLM_TENSOR_FFN_GATE_INP_SHEXP, "blk.%d.ffn_gate_inp_shexp" },
{ LLM_TENSOR_SSM_A_NOSCAN, "blk.%d.ssm_a" },
{ LLM_TENSOR_SSM_CONV1D, "blk.%d.ssm_conv1d" },
{ LLM_TENSOR_SSM_DT, "blk.%d.ssm_dt" },
{ LLM_TENSOR_SSM_BETA_ALPHA, "blk.%d.ssm_ba" },
{ LLM_TENSOR_SSM_ALPHA, "blk.%d.ssm_alpha" },
{ LLM_TENSOR_SSM_IN, "blk.%d.ssm_in" },
{ LLM_TENSOR_SSM_NORM, "blk.%d.ssm_norm" },
{ LLM_TENSOR_SSM_OUT, "blk.%d.ssm_out" },
@ -370,6 +397,15 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
{ LLM_TENSOR_SSM_DT_NORM, "blk.%d.ssm_dt_norm" },
{ LLM_TENSOR_SSM_B_NORM, "blk.%d.ssm_b_norm" },
{ LLM_TENSOR_SSM_C_NORM, "blk.%d.ssm_c_norm" },
{ LLM_TENSOR_SSM_CONV1D_Q, "blk.%d.ssm_conv1d_q" },
{ LLM_TENSOR_SSM_CONV1D_K, "blk.%d.ssm_conv1d_k" },
{ LLM_TENSOR_SSM_CONV1D_V, "blk.%d.ssm_conv1d_v" },
{ LLM_TENSOR_SSM_F_A, "blk.%d.ssm_f_a" },
{ LLM_TENSOR_SSM_F_B, "blk.%d.ssm_f_b" },
{ LLM_TENSOR_SSM_BETA, "blk.%d.ssm_beta" },
{ LLM_TENSOR_SSM_G_A, "blk.%d.ssm_g_a" },
{ LLM_TENSOR_SSM_G_B, "blk.%d.ssm_g_b" },
{ LLM_TENSOR_SSM_NORM, "blk.%d.ssm_norm" },
{ LLM_TENSOR_ATTN_Q_A_NORM, "blk.%d.attn_q_a_norm" },
{ LLM_TENSOR_ATTN_KV_A_NORM, "blk.%d.attn_kv_a_norm" },
{ LLM_TENSOR_ATTN_Q_A, "blk.%d.attn_q_a" },
@ -496,6 +532,10 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
{ LLM_TENSOR_VISEXP_FFN_GATE, "blk.%d.vis_gate" },
{ LLM_TENSOR_VISEXP_FFN_DOWN, "blk.%d.vis_down" },
{ LLM_TENSOR_VISEXP_FFN_UP, "blk.%d.vis_up" },
{ LLM_TENSOR_INDEXER_K_NORM, "blk.%d.indexer.k_norm" },
{ LLM_TENSOR_INDEXER_PROJ, "blk.%d.indexer.proj" },
{ LLM_TENSOR_INDEXER_ATTN_K, "blk.%d.indexer.attn_k" },
{ LLM_TENSOR_INDEXER_ATTN_Q_B, "blk.%d.indexer.attn_q_b" },
};
static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
@ -709,6 +749,7 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
case LLM_ARCH_INTERNLM2:
case LLM_ARCH_GRANITE:
case LLM_ARCH_ERNIE4_5:
case LLM_ARCH_PADDLEOCR:
case LLM_ARCH_SMOLLM3:
case LLM_ARCH_DREAM:
case LLM_ARCH_LLADA:
@ -787,6 +828,20 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
LLM_TENSOR_CLS,
LLM_TENSOR_CLS_OUT,
};
case LLM_ARCH_EUROBERT:
return {
LLM_TENSOR_TOKEN_EMBD,
LLM_TENSOR_OUTPUT_NORM,
LLM_TENSOR_ATTN_NORM,
LLM_TENSOR_ATTN_Q,
LLM_TENSOR_ATTN_K,
LLM_TENSOR_ATTN_V,
LLM_TENSOR_ATTN_OUT,
LLM_TENSOR_FFN_NORM,
LLM_TENSOR_FFN_GATE,
LLM_TENSOR_FFN_UP,
LLM_TENSOR_FFN_DOWN,
};
case LLM_ARCH_MODERN_BERT:
return {
LLM_TENSOR_TOKEN_EMBD,
@ -800,6 +855,7 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
LLM_TENSOR_FFN_NORM,
LLM_TENSOR_CLS,
LLM_TENSOR_CLS_OUT,
LLM_TENSOR_CLS_NORM,
};
case LLM_ARCH_JINA_BERT_V2:
return {
@ -952,11 +1008,11 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
LLM_TENSOR_ATTN_OUT,
LLM_TENSOR_ATTN_QKV,
LLM_TENSOR_ATTN_GATE,
LLM_TENSOR_FFN_NORM,
LLM_TENSOR_FFN_GATE_INP,
LLM_TENSOR_FFN_GATE_EXPS,
LLM_TENSOR_FFN_DOWN_EXPS,
LLM_TENSOR_FFN_UP_EXPS,
LLM_TENSOR_FFN_GATE_UP_EXPS,
LLM_TENSOR_FFN_GATE_INP_SHEXP,
LLM_TENSOR_FFN_GATE_SHEXP,
LLM_TENSOR_FFN_DOWN_SHEXP,
@ -969,6 +1025,64 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
LLM_TENSOR_SSM_NORM,
LLM_TENSOR_SSM_OUT,
};
case LLM_ARCH_QWEN35:
return {
LLM_TENSOR_TOKEN_EMBD,
LLM_TENSOR_OUTPUT_NORM,
LLM_TENSOR_OUTPUT,
LLM_TENSOR_ATTN_NORM,
LLM_TENSOR_ATTN_POST_NORM,
LLM_TENSOR_ATTN_Q,
LLM_TENSOR_ATTN_Q_NORM,
LLM_TENSOR_ATTN_K,
LLM_TENSOR_ATTN_K_NORM,
LLM_TENSOR_ATTN_V,
LLM_TENSOR_ATTN_OUT,
LLM_TENSOR_ATTN_QKV,
LLM_TENSOR_ATTN_GATE,
LLM_TENSOR_FFN_GATE,
LLM_TENSOR_FFN_DOWN,
LLM_TENSOR_FFN_UP,
LLM_TENSOR_SSM_A_NOSCAN,
LLM_TENSOR_SSM_CONV1D,
LLM_TENSOR_SSM_DT,
LLM_TENSOR_SSM_BETA,
LLM_TENSOR_SSM_ALPHA,
LLM_TENSOR_SSM_NORM,
LLM_TENSOR_SSM_OUT,
};
case LLM_ARCH_QWEN35MOE:
return {
LLM_TENSOR_TOKEN_EMBD,
LLM_TENSOR_OUTPUT_NORM,
LLM_TENSOR_OUTPUT,
LLM_TENSOR_ATTN_NORM,
LLM_TENSOR_ATTN_POST_NORM,
LLM_TENSOR_ATTN_Q,
LLM_TENSOR_ATTN_Q_NORM,
LLM_TENSOR_ATTN_K,
LLM_TENSOR_ATTN_K_NORM,
LLM_TENSOR_ATTN_V,
LLM_TENSOR_ATTN_OUT,
LLM_TENSOR_ATTN_QKV,
LLM_TENSOR_ATTN_GATE,
LLM_TENSOR_FFN_GATE_INP,
LLM_TENSOR_FFN_GATE_EXPS,
LLM_TENSOR_FFN_DOWN_EXPS,
LLM_TENSOR_FFN_UP_EXPS,
LLM_TENSOR_FFN_GATE_UP_EXPS,
LLM_TENSOR_FFN_GATE_INP_SHEXP,
LLM_TENSOR_FFN_GATE_SHEXP,
LLM_TENSOR_FFN_DOWN_SHEXP,
LLM_TENSOR_FFN_UP_SHEXP,
LLM_TENSOR_SSM_A_NOSCAN,
LLM_TENSOR_SSM_CONV1D,
LLM_TENSOR_SSM_DT,
LLM_TENSOR_SSM_BETA,
LLM_TENSOR_SSM_ALPHA,
LLM_TENSOR_SSM_NORM,
LLM_TENSOR_SSM_OUT,
};
case LLM_ARCH_QWEN3VL:
case LLM_ARCH_CHAMELEON:
case LLM_ARCH_HUNYUAN_DENSE:
@ -976,6 +1090,7 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
LLM_TENSOR_TOKEN_EMBD,
LLM_TENSOR_OUTPUT_NORM,
LLM_TENSOR_OUTPUT,
LLM_TENSOR_CLS_OUT,
LLM_TENSOR_ATTN_NORM,
LLM_TENSOR_ATTN_Q,
LLM_TENSOR_ATTN_Q_NORM,
@ -1497,6 +1612,7 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
LLM_TENSOR_FFN_GATE_EXPS,
LLM_TENSOR_FFN_DOWN_EXPS,
LLM_TENSOR_FFN_UP_EXPS,
LLM_TENSOR_FFN_GATE_UP_EXPS,
LLM_TENSOR_FFN_GATE_INP_SHEXP,
LLM_TENSOR_FFN_GATE_SHEXP,
LLM_TENSOR_FFN_DOWN_SHEXP,
@ -1549,6 +1665,12 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
LLM_TENSOR_FFN_DOWN,
LLM_TENSOR_ATTN_POST_NORM,
LLM_TENSOR_FFN_POST_NORM,
LLM_TENSOR_NEXTN_EH_PROJ,
LLM_TENSOR_NEXTN_EMBED_TOKENS,
LLM_TENSOR_NEXTN_ENORM,
LLM_TENSOR_NEXTN_HNORM,
LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD,
LLM_TENSOR_NEXTN_SHARED_HEAD_NORM,
};
case LLM_ARCH_GLM4_MOE:
return {
@ -1581,6 +1703,46 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD,
LLM_TENSOR_NEXTN_SHARED_HEAD_NORM,
};
case LLM_ARCH_GLM_DSA:
return {
LLM_TENSOR_TOKEN_EMBD,
LLM_TENSOR_OUTPUT_NORM,
LLM_TENSOR_OUTPUT,
LLM_TENSOR_ATTN_NORM,
LLM_TENSOR_ATTN_Q_A_NORM,
LLM_TENSOR_ATTN_KV_A_NORM,
LLM_TENSOR_ATTN_Q,
LLM_TENSOR_ATTN_Q_A,
LLM_TENSOR_ATTN_Q_B,
LLM_TENSOR_ATTN_KV_A_MQA,
LLM_TENSOR_ATTN_KV_B,
LLM_TENSOR_ATTN_K_B,
LLM_TENSOR_ATTN_V_B,
LLM_TENSOR_ATTN_OUT,
LLM_TENSOR_FFN_NORM,
LLM_TENSOR_FFN_GATE,
LLM_TENSOR_FFN_UP,
LLM_TENSOR_FFN_DOWN,
LLM_TENSOR_FFN_GATE_INP,
LLM_TENSOR_FFN_GATE_EXPS,
LLM_TENSOR_FFN_DOWN_EXPS,
LLM_TENSOR_FFN_UP_EXPS,
LLM_TENSOR_FFN_GATE_INP_SHEXP,
LLM_TENSOR_FFN_GATE_SHEXP,
LLM_TENSOR_FFN_DOWN_SHEXP,
LLM_TENSOR_FFN_UP_SHEXP,
LLM_TENSOR_FFN_EXP_PROBS_B,
LLM_TENSOR_INDEXER_K_NORM,
LLM_TENSOR_INDEXER_PROJ,
LLM_TENSOR_INDEXER_ATTN_K,
LLM_TENSOR_INDEXER_ATTN_Q_B,
LLM_TENSOR_NEXTN_EH_PROJ,
LLM_TENSOR_NEXTN_EMBED_TOKENS,
LLM_TENSOR_NEXTN_ENORM,
LLM_TENSOR_NEXTN_HNORM,
LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD,
LLM_TENSOR_NEXTN_SHARED_HEAD_NORM,
};
case LLM_ARCH_BITNET:
return {
LLM_TENSOR_TOKEN_EMBD,
@ -1659,6 +1821,20 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
LLM_TENSOR_FFN_GATE,
LLM_TENSOR_FFN_DOWN,
};
case LLM_ARCH_JAIS2:
return {
LLM_TENSOR_TOKEN_EMBD,
LLM_TENSOR_OUTPUT_NORM,
LLM_TENSOR_OUTPUT,
LLM_TENSOR_ATTN_NORM,
LLM_TENSOR_ATTN_Q,
LLM_TENSOR_ATTN_K,
LLM_TENSOR_ATTN_V,
LLM_TENSOR_ATTN_OUT,
LLM_TENSOR_FFN_NORM,
LLM_TENSOR_FFN_UP,
LLM_TENSOR_FFN_DOWN,
};
case LLM_ARCH_NEMOTRON_H:
return {
LLM_TENSOR_TOKEN_EMBD,
@ -1706,6 +1882,8 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
LLM_TENSOR_FFN_UP_EXPS,
LLM_TENSOR_FFN_DOWN_EXPS,
LLM_TENSOR_FFN_EXP_PROBS_B,
LLM_TENSOR_FFN_LATENT_DOWN,
LLM_TENSOR_FFN_LATENT_UP,
// MoE shared expert layer
LLM_TENSOR_FFN_DOWN_SHEXP,
LLM_TENSOR_FFN_UP_SHEXP,
@ -1728,6 +1906,38 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
LLM_TENSOR_FFN_UP,
LLM_TENSOR_FFN_POST_NORM,
};
case LLM_ARCH_EXAONE_MOE:
return {
LLM_TENSOR_TOKEN_EMBD,
LLM_TENSOR_OUTPUT_NORM,
LLM_TENSOR_OUTPUT,
LLM_TENSOR_ROPE_FREQS,
LLM_TENSOR_ATTN_NORM,
LLM_TENSOR_ATTN_Q,
LLM_TENSOR_ATTN_Q_NORM,
LLM_TENSOR_ATTN_K,
LLM_TENSOR_ATTN_K_NORM,
LLM_TENSOR_ATTN_V,
LLM_TENSOR_ATTN_OUT,
LLM_TENSOR_FFN_NORM,
LLM_TENSOR_FFN_GATE,
LLM_TENSOR_FFN_DOWN,
LLM_TENSOR_FFN_UP,
LLM_TENSOR_FFN_GATE_INP,
LLM_TENSOR_FFN_GATE_EXPS,
LLM_TENSOR_FFN_DOWN_EXPS,
LLM_TENSOR_FFN_UP_EXPS,
LLM_TENSOR_FFN_GATE_SHEXP,
LLM_TENSOR_FFN_UP_SHEXP,
LLM_TENSOR_FFN_DOWN_SHEXP,
LLM_TENSOR_FFN_EXP_PROBS_B,
LLM_TENSOR_NEXTN_EH_PROJ,
LLM_TENSOR_NEXTN_EMBED_TOKENS,
LLM_TENSOR_NEXTN_ENORM,
LLM_TENSOR_NEXTN_HNORM,
LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD,
LLM_TENSOR_NEXTN_SHARED_HEAD_NORM,
};
case LLM_ARCH_RWKV6:
return {
LLM_TENSOR_TOKEN_EMBD,
@ -2234,6 +2444,35 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
LLM_TENSOR_FFN_UP_EXPS,
LLM_TENSOR_FFN_EXP_PROBS_B,
};
case LLM_ARCH_STEP35:
return {
LLM_TENSOR_TOKEN_EMBD,
LLM_TENSOR_OUTPUT_NORM,
LLM_TENSOR_OUTPUT,
LLM_TENSOR_ROPE_FREQS,
LLM_TENSOR_ROPE_FACTORS_LONG,
LLM_TENSOR_ROPE_FACTORS_SHORT,
LLM_TENSOR_ATTN_NORM,
LLM_TENSOR_ATTN_Q,
LLM_TENSOR_ATTN_Q_NORM,
LLM_TENSOR_ATTN_K,
LLM_TENSOR_ATTN_K_NORM,
LLM_TENSOR_ATTN_V,
LLM_TENSOR_ATTN_GATE,
LLM_TENSOR_ATTN_OUT,
LLM_TENSOR_FFN_NORM,
LLM_TENSOR_FFN_GATE,
LLM_TENSOR_FFN_DOWN,
LLM_TENSOR_FFN_UP,
LLM_TENSOR_FFN_GATE_INP,
LLM_TENSOR_FFN_GATE_EXPS,
LLM_TENSOR_FFN_DOWN_EXPS,
LLM_TENSOR_FFN_UP_EXPS,
LLM_TENSOR_FFN_GATE_SHEXP,
LLM_TENSOR_FFN_UP_SHEXP,
LLM_TENSOR_FFN_DOWN_SHEXP,
LLM_TENSOR_FFN_EXP_PROBS_B,
};
case LLM_ARCH_GPTJ:
case LLM_ARCH_UNKNOWN:
return {
@ -2256,6 +2495,54 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
LLM_TENSOR_FFN_DOWN,
LLM_TENSOR_FFN_UP,
};
case LLM_ARCH_KIMI_LINEAR:
return {
LLM_TENSOR_TOKEN_EMBD,
LLM_TENSOR_OUTPUT_NORM,
LLM_TENSOR_OUTPUT,
LLM_TENSOR_ROPE_FREQS,
LLM_TENSOR_ATTN_NORM,
LLM_TENSOR_ATTN_Q,
LLM_TENSOR_ATTN_K,
LLM_TENSOR_ATTN_V,
LLM_TENSOR_ATTN_OUT,
LLM_TENSOR_FFN_NORM,
// Dense FFN (layer 0 only)
LLM_TENSOR_FFN_GATE,
LLM_TENSOR_FFN_DOWN,
LLM_TENSOR_FFN_UP,
// MoE FFN (layers 1+)
LLM_TENSOR_FFN_GATE_INP,
LLM_TENSOR_FFN_GATE_EXPS,
LLM_TENSOR_FFN_DOWN_EXPS,
LLM_TENSOR_FFN_UP_EXPS,
LLM_TENSOR_FFN_EXP_PROBS_B,
// Shared experts
LLM_TENSOR_FFN_GATE_SHEXP,
LLM_TENSOR_FFN_DOWN_SHEXP,
LLM_TENSOR_FFN_UP_SHEXP,
// KDA (using SSM_ enum prefix, keeping GGUF names for backward compat)
LLM_TENSOR_SSM_CONV1D_Q,
LLM_TENSOR_SSM_CONV1D_K,
LLM_TENSOR_SSM_CONV1D_V,
LLM_TENSOR_SSM_F_A,
LLM_TENSOR_SSM_F_B,
LLM_TENSOR_SSM_BETA,
LLM_TENSOR_SSM_A,
LLM_TENSOR_SSM_G_A,
LLM_TENSOR_SSM_G_B,
LLM_TENSOR_SSM_DT,
LLM_TENSOR_SSM_NORM,
// MLA
LLM_TENSOR_ATTN_Q_A,
LLM_TENSOR_ATTN_Q_B,
LLM_TENSOR_ATTN_Q_A_NORM,
LLM_TENSOR_ATTN_KV_A_MQA,
LLM_TENSOR_ATTN_KV_B,
LLM_TENSOR_ATTN_K_B,
LLM_TENSOR_ATTN_V_B,
LLM_TENSOR_ATTN_KV_A_NORM,
};
default:
GGML_ABORT("unknown architecture for tensor mapping");
}
@ -2279,6 +2566,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
{LLM_TENSOR_OUTPUT, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
{LLM_TENSOR_CLS, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
{LLM_TENSOR_CLS_OUT, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
{LLM_TENSOR_CLS_NORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
{LLM_TENSOR_DENSE_2_OUT, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}}, // Dense layer output
{LLM_TENSOR_DENSE_3_OUT, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}}, // Dense layer output
{LLM_TENSOR_OUTPUT_NORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
@ -2331,6 +2619,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
{LLM_TENSOR_SSM_X, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_SSM_DT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_SSM_OUT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_SSM_ALPHA, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_SSM_BETA_ALPHA, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_TIME_MIX_W1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_TIME_MIX_W2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
@ -2359,6 +2648,15 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
{LLM_TENSOR_SSM_C_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_SSM_D, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_SSM_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
// Kimi KDA - Conv tensors are 4D [d_conv, 1, d_inner, 1], reshaped to 2D at runtime
{LLM_TENSOR_SSM_CONV1D_Q, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_SSM_CONV1D_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_SSM_CONV1D_V, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_SSM_F_A, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_SSM_F_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_SSM_BETA, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_SSM_G_A, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_SSM_G_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_TIME_MIX_LERP_X, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_TIME_MIX_LN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_CHANNEL_MIX_LERP_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
@ -2401,6 +2699,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
{LLM_TENSOR_FFN_DOWN_EXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
{LLM_TENSOR_FFN_GATE_EXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
{LLM_TENSOR_FFN_UP_EXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
{LLM_TENSOR_FFN_GATE_UP_EXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
{LLM_TENSOR_FFN_DOWN_CHEXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
{LLM_TENSOR_FFN_GATE_CHEXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
{LLM_TENSOR_FFN_UP_CHEXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
@ -2448,6 +2747,10 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
{LLM_TENSOR_VISEXP_FFN_GATE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_VISEXP_FFN_DOWN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_VISEXP_FFN_UP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_INDEXER_K_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_INDEXER_PROJ, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_INDEXER_ATTN_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_INDEXER_ATTN_Q_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
// NextN/MTP tensors are currently ignored (reserved for future MTP support)
// These tensors only exist in the last layer(s) and are treated as output tensors
{LLM_TENSOR_NEXTN_EH_PROJ, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
@ -2456,6 +2759,9 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
{LLM_TENSOR_NEXTN_HNORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
{LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
{LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
// Nemotron 3 Super
{LLM_TENSOR_FFN_LATENT_DOWN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_FFN_LATENT_UP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
};
LLM_KV::LLM_KV(llm_arch arch, const char * suffix) : arch(arch), suffix(suffix) {}
@ -2493,6 +2799,15 @@ std::string LLM_TN_IMPL::str() const {
return name;
}
std::vector<llm_arch> llm_arch_all() {
std::vector<llm_arch> ret;
ret.reserve(LLM_ARCH_NAMES.size());
for (const auto & [arch, _] : LLM_ARCH_NAMES) {
ret.push_back(arch);
}
return ret;
}
const char * llm_arch_name(llm_arch arch) {
auto it = LLM_ARCH_NAMES.find(arch);
if (it == LLM_ARCH_NAMES.end()) {
@ -2540,6 +2855,9 @@ bool llm_arch_is_hybrid(const llm_arch & arch) {
case LLM_ARCH_NEMOTRON_H:
case LLM_ARCH_NEMOTRON_H_MOE:
case LLM_ARCH_QWEN3NEXT:
case LLM_ARCH_KIMI_LINEAR:
case LLM_ARCH_QWEN35:
case LLM_ARCH_QWEN35MOE:
return true;
default:
return false;

42
examples/talk-llama/llama-arch.h
View File

@ -4,6 +4,7 @@
#include <string>
#include <set>
#include <vector>
//
// gguf constants (sync with gguf.py)
@ -30,6 +31,7 @@ enum llm_arch {
LLM_ARCH_NEO_BERT,
LLM_ARCH_JINA_BERT_V2,
LLM_ARCH_JINA_BERT_V3,
LLM_ARCH_EUROBERT,
LLM_ARCH_BLOOM,
LLM_ARCH_STABLELM,
LLM_ARCH_QWEN,
@ -41,6 +43,8 @@ enum llm_arch {
LLM_ARCH_QWEN3NEXT,
LLM_ARCH_QWEN3VL,
LLM_ARCH_QWEN3VLMOE,
LLM_ARCH_QWEN35,
LLM_ARCH_QWEN35MOE,
LLM_ARCH_PHI2,
LLM_ARCH_PHI3,
LLM_ARCH_PHIMOE,
@ -76,15 +80,18 @@ enum llm_arch {
LLM_ARCH_CHATGLM,
LLM_ARCH_GLM4,
LLM_ARCH_GLM4_MOE,
LLM_ARCH_GLM_DSA,
LLM_ARCH_BITNET,
LLM_ARCH_T5,
LLM_ARCH_T5ENCODER,
LLM_ARCH_JAIS,
LLM_ARCH_JAIS2,
LLM_ARCH_NEMOTRON,
LLM_ARCH_NEMOTRON_H,
LLM_ARCH_NEMOTRON_H_MOE,
LLM_ARCH_EXAONE,
LLM_ARCH_EXAONE4,
LLM_ARCH_EXAONE_MOE,
LLM_ARCH_RWKV6,
LLM_ARCH_RWKV6QWEN2,
LLM_ARCH_RWKV7,
@ -120,9 +127,12 @@ enum llm_arch {
LLM_ARCH_RND1,
LLM_ARCH_PANGU_EMBED,
LLM_ARCH_MISTRAL3,
LLM_ARCH_PADDLEOCR,
LLM_ARCH_MIMO2,
LLM_ARCH_STEP35,
LLM_ARCH_LLAMA_EMBED,
LLM_ARCH_MAINCODER,
LLM_ARCH_KIMI_LINEAR,
LLM_ARCH_UNKNOWN,
};
@ -164,6 +174,8 @@ enum llm_kv {
LLM_KV_EXPERT_FEED_FORWARD_LENGTH,
LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH,
LLM_KV_EXPERT_CHUNK_FEED_FORWARD_LENGTH,
LLM_KV_SWIGLU_CLAMP_EXP,
LLM_KV_SWIGLU_CLAMP_SHEXP,
LLM_KV_USE_PARALLEL_RESIDUAL,
LLM_KV_TENSOR_DATA_LAYOUT,
LLM_KV_EXPERT_COUNT,
@ -177,6 +189,7 @@ enum llm_kv {
LLM_KV_EXPERT_GROUP_SCALE,
LLM_KV_EXPERTS_PER_GROUP,
LLM_KV_MOE_EVERY_N_LAYERS,
LLM_KV_MOE_LATENT_SIZE,
LLM_KV_NEXTN_PREDICT_LAYERS,
LLM_KV_NUM_DEEPSTACK_LAYERS,
LLM_KV_POOLING_TYPE,
@ -194,6 +207,7 @@ enum llm_kv {
LLM_KV_EMBEDDING_SCALE,
LLM_KV_TOKEN_SHIFT_COUNT,
LLM_KV_INTERLEAVE_MOE_LAYER_STEP,
LLM_KV_FULL_ATTENTION_INTERVAL,
LLM_KV_ATTENTION_HEAD_COUNT,
LLM_KV_ATTENTION_HEAD_COUNT_KV,
@ -221,8 +235,14 @@ enum llm_kv {
LLM_KV_ATTENTION_TEMPERATURE_SCALE,
LLM_KV_ATTENTION_KEY_LENGTH_MLA,
LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
LLM_KV_ATTENTION_KEY_LENGTH_SWA,
LLM_KV_ATTENTION_VALUE_LENGTH_SWA,
LLM_KV_ATTENTION_INDEXER_HEAD_COUNT,
LLM_KV_ATTENTION_INDEXER_KEY_LENGTH,
LLM_KV_ATTENTION_INDEXER_TOP_K,
LLM_KV_ROPE_DIMENSION_COUNT,
LLM_KV_ROPE_DIMENSION_COUNT_SWA,
LLM_KV_ROPE_DIMENSION_SECTIONS,
LLM_KV_ROPE_FREQ_BASE,
LLM_KV_ROPE_FREQ_BASE_SWA,
@ -249,6 +269,8 @@ enum llm_kv {
LLM_KV_SSM_GROUP_COUNT,
LLM_KV_SSM_DT_B_C_RMS,
LLM_KV_KDA_HEAD_DIM,
LLM_KV_WKV_HEAD_SIZE,
LLM_KV_TOKENIZER_MODEL,
@ -356,6 +378,7 @@ enum llm_tensor {
LLM_TENSOR_FFN_DOWN_EXPS, // merged experts
LLM_TENSOR_FFN_GATE_EXPS,
LLM_TENSOR_FFN_UP_EXPS,
LLM_TENSOR_FFN_GATE_UP_EXPS,
LLM_TENSOR_FFN_DOWN_SHEXP,
LLM_TENSOR_FFN_GATE_SHEXP,
LLM_TENSOR_FFN_UP_SHEXP,
@ -363,6 +386,8 @@ enum llm_tensor {
LLM_TENSOR_FFN_GATE_CHEXPS,
LLM_TENSOR_FFN_UP_CHEXPS,
LLM_TENSOR_FFN_EXP_PROBS_B,
LLM_TENSOR_FFN_LATENT_DOWN,
LLM_TENSOR_FFN_LATENT_UP,
LLM_TENSOR_ATTN_Q_NORM,
LLM_TENSOR_ATTN_K_NORM,
LLM_TENSOR_LAYER_OUT_NORM,
@ -397,6 +422,16 @@ enum llm_tensor {
LLM_TENSOR_SSM_NORM,
LLM_TENSOR_SSM_OUT,
LLM_TENSOR_SSM_BETA_ALPHA, // qwen3next
LLM_TENSOR_SSM_ALPHA, // qwen3.5
// Kimi Linear KDA (using SSM_ prefix for consistency)
LLM_TENSOR_SSM_CONV1D_Q, // kimi: Q conv1d weight
LLM_TENSOR_SSM_CONV1D_K, // kimi: K conv1d weight
LLM_TENSOR_SSM_CONV1D_V, // kimi: V conv1d weight
LLM_TENSOR_SSM_F_A, // kimi: forget gate projection A
LLM_TENSOR_SSM_F_B, // kimi: forget gate projection B
LLM_TENSOR_SSM_BETA, // kimi: beta mixing coefficient and qwen3.5
LLM_TENSOR_SSM_G_A, // kimi: output gate projection A
LLM_TENSOR_SSM_G_B, // kimi: output gate projection B
LLM_TENSOR_TIME_MIX_W0,
LLM_TENSOR_TIME_MIX_W1,
LLM_TENSOR_TIME_MIX_W2,
@ -473,6 +508,7 @@ enum llm_tensor {
LLM_TENSOR_ENC_OUTPUT_NORM,
LLM_TENSOR_CLS,
LLM_TENSOR_CLS_OUT,
LLM_TENSOR_CLS_NORM,
LLM_TENSOR_CONV1D,
LLM_TENSOR_CONVNEXT_DW,
LLM_TENSOR_CONVNEXT_NORM,
@ -497,6 +533,10 @@ enum llm_tensor {
LLM_TENSOR_VISEXP_FFN_GATE,
LLM_TENSOR_VISEXP_FFN_DOWN,
LLM_TENSOR_VISEXP_FFN_UP,
LLM_TENSOR_INDEXER_K_NORM,
LLM_TENSOR_INDEXER_PROJ,
LLM_TENSOR_INDEXER_ATTN_K,
LLM_TENSOR_INDEXER_ATTN_Q_B,
LLM_TENSOR_NEXTN_EH_PROJ,
LLM_TENSOR_NEXTN_EMBED_TOKENS,
LLM_TENSOR_NEXTN_ENORM,
@ -575,6 +615,8 @@ struct llm_tensor_info {
ggml_op op;
};
std::vector<llm_arch> llm_arch_all();
const char * llm_arch_name(llm_arch arch);
llm_arch llm_arch_from_string(const std::string & name);

4
examples/talk-llama/llama-batch.cpp
View File

@ -394,11 +394,13 @@ llama_ubatch llama_batch_allocr::ubatch_reserve(uint32_t n_seq_tokens, uint32_t
clear();
split_reset();
const int64_t n_pos_all = (int64_t) n_tokens*n_pos_per_embd;
auto udata = std::make_shared<llama_ubatch::data_t>();
udata->token .resize(n_tokens);
udata->embd .clear();
udata->pos .resize(n_tokens);
udata->pos .resize(n_pos_all);
udata->n_seq_id .resize(n_tokens);
udata->seq_id .resize(n_tokens);
udata->seq_id_unq.resize(0);

22
examples/talk-llama/llama-chat.cpp
View File

@ -57,6 +57,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
{ "minicpm", LLM_CHAT_TEMPLATE_MINICPM },
{ "exaone3", LLM_CHAT_TEMPLATE_EXAONE_3 },
{ "exaone4", LLM_CHAT_TEMPLATE_EXAONE_4 },
{ "exaone-moe", LLM_CHAT_TEMPLATE_EXAONE_MOE },
{ "rwkv-world", LLM_CHAT_TEMPLATE_RWKV_WORLD },
{ "granite", LLM_CHAT_TEMPLATE_GRANITE },
{ "gigachat", LLM_CHAT_TEMPLATE_GIGACHAT },
@ -137,6 +138,9 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
} else if (tmpl_contains("[gMASK]<sop>")) {
return LLM_CHAT_TEMPLATE_CHATGLM_4;
} else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|user|>")) {
if (tmpl_contains("<|tool_declare|>")) {
return LLM_CHAT_TEMPLATE_EXAONE_MOE;
}
return tmpl_contains("</s>") ? LLM_CHAT_TEMPLATE_FALCON_3 : LLM_CHAT_TEMPLATE_GLMEDGE;
} else if (tmpl_contains("<|{{ item['role'] }}|>") && tmpl_contains("<|begin_of_image|>")) {
return LLM_CHAT_TEMPLATE_GLMEDGE;
@ -229,7 +233,7 @@ int32_t llm_chat_apply_template(
llm_chat_template tmpl,
const std::vector<const llama_chat_message *> & chat,
std::string & dest, bool add_ass) {
// Taken from the research: https://github.com/ggerganov/llama.cpp/issues/5527
// Taken from the research: https://github.com/ggml-org/llama.cpp/issues/5527
std::stringstream ss;
if (tmpl == LLM_CHAT_TEMPLATE_CHATML) {
// chatml template
@ -576,6 +580,22 @@ int32_t llm_chat_apply_template(
if (add_ass) {
ss << "[|assistant|]";
}
} else if (tmpl == LLM_CHAT_TEMPLATE_EXAONE_MOE) {
for (auto message : chat) {
std::string role(message->role);
if (role == "system") {
ss << "<|system|>\n" << trim(message->content) << "<|endofturn|>\n";
} else if (role == "user") {
ss << "<|user|>\n" << trim(message->content) << "<|endofturn|>\n";
} else if (role == "assistant") {
ss << "<|assistant|>\n" << trim(message->content) << "<|endofturn|>\n";
} else if (role == "tool") {
ss << "<|tool|>\n" << trim(message->content) << "<|endofturn|>\n";
}
}
if (add_ass) {
ss << "<|assistant|>\n";
}
} else if (tmpl == LLM_CHAT_TEMPLATE_RWKV_WORLD) {
// this template requires the model to have "\n\n" as EOT token
for (size_t i = 0; i < chat.size(); i++) {

1
examples/talk-llama/llama-chat.h
View File

@ -36,6 +36,7 @@ enum llm_chat_template {
LLM_CHAT_TEMPLATE_MINICPM,
LLM_CHAT_TEMPLATE_EXAONE_3,
LLM_CHAT_TEMPLATE_EXAONE_4,
LLM_CHAT_TEMPLATE_EXAONE_MOE,
LLM_CHAT_TEMPLATE_RWKV_WORLD,
LLM_CHAT_TEMPLATE_GRANITE,
LLM_CHAT_TEMPLATE_GIGACHAT,

1060
examples/talk-llama/llama-context.cpp
View File

File diff suppressed because it is too large Load Diff

55
examples/talk-llama/llama-context.h
View File

@ -4,6 +4,7 @@
#include "llama-cparams.h"
#include "llama-graph.h"
#include "llama-adapter.h"
#include "llama-impl.h"
#include "ggml-cpp.h"
#include "ggml-opt.h"
@ -40,6 +41,14 @@ struct llama_context {
~llama_context();
// reserve a new backend scheduler (if needed)
// for example, when:
// - changing loras
// - changing samplers
// - changing attention type
// - etc.
void sched_reserve();
void synchronize();
const llama_model & get_model() const;
@ -96,16 +105,11 @@ struct llama_context {
void set_causal_attn(bool value);
void set_warmup(bool value);
void set_adapter_lora(
llama_adapter_lora * adapter,
float scale);
void set_adapters_lora(llama_adapter_lora ** adapters, size_t n_adapters, float * scales);
bool rm_adapter_lora(
llama_adapter_lora * adapter);
bool adapters_lora_are_same(llama_adapter_lora ** adapters, size_t n_adapters, float * scales);
void clear_adapter_lora();
bool apply_adapter_cvec(
bool set_adapter_cvec(
const float * data,
size_t len,
int32_t n_embd,
@ -204,7 +208,7 @@ private:
// Make sure enough space is available for outputs.
// Returns max number of outputs for which space was reserved.
uint32_t output_reserve(int32_t n_outputs, const llama_batch & batch);
uint32_t output_reserve(int32_t n_outputs);
void output_reorder();
@ -252,43 +256,36 @@ private:
const llama_model & model;
llama_cparams cparams;
llama_adapter_cvec cvec;
llama_adapter_loras loras;
llama_cparams cparams;
llama_adapter_cvec_ptr cvec;
llama_adapter_loras_ptr loras;
llama_cross cross; // TODO: tmp for handling cross-attention - need something better probably
std::unique_ptr<llama_memory_i> memory;
// decode output (2-dimensional array: [n_outputs][n_vocab])
size_t logits_size = 0; // capacity (of floats) for logits
float * logits = nullptr;
buffer_view<float> logits = {nullptr, 0};
// embeddings output (2-dimensional array: [n_outputs][n_embd])
// populated only when pooling_type == LLAMA_POOLING_TYPE_NONE
size_t embd_size = 0; // capacity (of floats) for embeddings
float * embd = nullptr;
buffer_view<float> embd = {nullptr, 0};
// TODO: simplify
struct sampling_info {
// !samplers.empty() to check if any samplers are active
std::map<llama_seq_id, llama_sampler *> samplers;
float * logits = nullptr;
size_t logits_size = 0;
llama_token * sampled = nullptr;
size_t sampled_size = 0;
float * probs = nullptr;
size_t probs_size = 0;
llama_token * candidates = nullptr;
size_t candidates_size = 0;
buffer_view<float> logits = {nullptr, 0};
buffer_view<llama_token> sampled = {nullptr, 0};
buffer_view<float> probs = {nullptr, 0};
buffer_view<llama_token> candidates = {nullptr, 0};
std::vector<uint32_t> logits_count;
std::vector<uint32_t> probs_count;
std::vector<uint32_t> candidates_count;
// optimization
std::vector<llama_token> token_ids_full_vocab;
};
@ -314,6 +311,8 @@ private:
ggml_backend_sched_ptr sched;
bool sched_need_reserve = true;
ggml_backend_t backend_cpu = nullptr;
std::vector<ggml_backend_ptr> backends;

5
examples/talk-llama/llama-cparams.h
View File

@ -30,10 +30,15 @@ struct llama_cparams {
bool causal_attn;
bool offload_kqv;
bool flash_attn;
bool auto_fa;
bool fused_gdn_ar; // use fused gated delta net (autoregressive)
bool fused_gdn_ch; // use fused gated delta net (chunked)
bool auto_fgdn;
bool no_perf;
bool warmup;
bool op_offload;
bool kv_unified;
bool pipeline_parallel;
enum llama_pooling_type pooling_type;

12
examples/talk-llama/llama-ext.h Normal file
View File

@ -0,0 +1,12 @@
#pragma once
#include "llama-context.h"
#include "ggml.h"
#include "stdint.h"
// Reserve a new compute graph. It is valid until the next call to llama_graph_reserve.
LLAMA_API struct ggml_cgraph * llama_graph_reserve(
struct llama_context * ctx,
uint32_t n_tokens,
uint32_t n_seqs,
uint32_t n_outputs);

16
examples/talk-llama/llama-grammar.cpp
View File

@ -2,7 +2,7 @@
#include "llama-impl.h"
#include "llama-vocab.h"
#include "llama-sampling.h"
#include "llama-sampler.h"
#include <cmath>
#include <algorithm>
@ -601,7 +601,7 @@ const char * llama_grammar_parser::parse_sequence(
throw std::runtime_error(std::string("expecting an int at ") + pos);
}
const char * int_end = parse_int(pos);
uint64_t min_times = std::stoul(std::string(pos, int_end - pos));
uint64_t min_times = std::stoull(std::string(pos, int_end - pos));
pos = parse_space(int_end, is_nested);
uint64_t max_times = UINT64_MAX; // default: no max limit
@ -614,7 +614,7 @@ const char * llama_grammar_parser::parse_sequence(
if (is_digit_char(*pos)) {
const char * int_end = parse_int(pos);
max_times = std::stoul(std::string(pos, int_end - pos));
max_times = std::stoull(std::string(pos, int_end - pos));
pos = parse_space(int_end, is_nested);
}
@ -1160,13 +1160,13 @@ struct llama_grammar * llama_grammar_init_impl(
// if there is a grammar, parse it
// rules will be empty (default) if there are parse errors
if (!parser.parse(grammar_str) || parser.rules.empty()) {
fprintf(stderr, "%s: failed to parse grammar\n", __func__);
LLAMA_LOG_ERROR("failed to parse grammar\n");
return nullptr;
}
// Ensure that there is a "root" node.
if (parser.symbol_ids.find("root") == parser.symbol_ids.end()) {
fprintf(stderr, "%s: grammar does not contain a 'root' symbol\n", __func__);
// Ensure that the grammar contains the start symbol
if (parser.symbol_ids.find(grammar_root) == parser.symbol_ids.end()) {
LLAMA_LOG_ERROR("grammar does not contain a '%s' symbol\n", grammar_root);
return nullptr;
}
@ -1195,7 +1195,7 @@ struct llama_grammar * llama_grammar_init_impl(
continue;
}
if (llama_grammar_detect_left_recursion(vec_rules, i, &rules_visited, &rules_in_progress, &rules_may_be_empty)) {
LLAMA_LOG_ERROR("unsupported grammar, left recursion detected for nonterminal at index %zu", i);
LLAMA_LOG_ERROR("unsupported grammar, left recursion detected for nonterminal at index %zu\n", i);
return nullptr;
}
}

719
examples/talk-llama/llama-graph.cpp
View File

File diff suppressed because it is too large Load Diff

141
examples/talk-llama/llama-graph.h
View File

@ -24,6 +24,7 @@ class llama_kv_cache_context;
class llama_kv_cache_iswa_context;
class llama_memory_recurrent_context;
class llama_memory_hybrid_context;
class llama_memory_hybrid_iswa_context;
// certain models (typically multi-modal) can produce different types of graphs
enum llm_graph_type {
@ -105,7 +106,7 @@ using llm_graph_input_ptr = std::unique_ptr<llm_graph_input_i>;
class llm_graph_input_embd : public llm_graph_input_i {
public:
llm_graph_input_embd() = default;
llm_graph_input_embd(int64_t n_embd) : n_embd(n_embd) {}
virtual ~llm_graph_input_embd() = default;
void set_input(const llama_ubatch * ubatch) override;
@ -114,6 +115,8 @@ public:
ggml_tensor * tokens = nullptr; // I32 [n_batch]
ggml_tensor * embd = nullptr; // F32 [n_embd, n_batch]
const int64_t n_embd = 0;
};
class llm_graph_input_pos : public llm_graph_input_i {
@ -314,6 +317,39 @@ public:
const llama_kv_cache_context * mctx;
};
// V-less input for the KV cache
// ref: https://github.com/ggml-org/llama.cpp/pull/19067
class llm_graph_input_attn_k : public llm_graph_input_i {
public:
llm_graph_input_attn_k(
const llama_hparams & hparams,
const llama_cparams & cparams,
const llama_kv_cache_context * mctx) :
hparams(hparams),
cparams(cparams),
mctx(mctx) {
}
~llm_graph_input_attn_k() = default;
void set_input(const llama_ubatch * ubatch) override;
bool can_reuse(const llm_graph_params & params) override;
ggml_tensor * get_k_idxs() const { return self_k_idxs; }
ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; }
ggml_tensor * self_k_idxs = nullptr; // I64 [n_batch]
ggml_tensor * self_kq_mask = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
ggml_tensor * self_kq_mask_cnv = nullptr; // [n_kv, n_batch/n_stream, 1, n_stream]
const llama_hparams hparams;
const llama_cparams cparams;
const llama_kv_cache_context * mctx;
};
class llm_graph_input_attn_kv_iswa : public llm_graph_input_i {
public:
llm_graph_input_attn_kv_iswa(
@ -397,6 +433,62 @@ public:
const llama_memory_hybrid_context * mctx;
};
class llm_graph_input_mem_hybrid_k : public llm_graph_input_i {
public:
llm_graph_input_mem_hybrid_k(
const llama_cparams & cparams,
std::unique_ptr<llm_graph_input_attn_k> inp_attn,
std::unique_ptr<llm_graph_input_rs> inp_rs,
const llama_memory_hybrid_context * mctx) :
inp_attn(std::move(inp_attn)),
inp_rs(std::move(inp_rs)),
cparams(cparams),
mctx(mctx) { }
virtual ~llm_graph_input_mem_hybrid_k() = default;
void set_input(const llama_ubatch * ubatch) override;
bool can_reuse(const llm_graph_params & params) override;
std::unique_ptr<llm_graph_input_attn_k> inp_attn;
std::unique_ptr<llm_graph_input_rs> inp_rs;
llm_graph_input_attn_k * get_attn() const { return inp_attn.get(); }
llm_graph_input_rs * get_recr() const { return inp_rs.get(); }
const llama_cparams cparams;
const llama_memory_hybrid_context * mctx;
};
class llm_graph_input_mem_hybrid_iswa : public llm_graph_input_i {
public:
llm_graph_input_mem_hybrid_iswa(
const llama_cparams & cparams,
std::unique_ptr<llm_graph_input_attn_kv_iswa> inp_attn,
std::unique_ptr<llm_graph_input_rs> inp_rs,
const llama_memory_hybrid_iswa_context * mctx) :
inp_attn(std::move(inp_attn)),
inp_rs(std::move(inp_rs)),
cparams(cparams),
mctx(mctx) { }
virtual ~llm_graph_input_mem_hybrid_iswa() = default;
void set_input(const llama_ubatch * ubatch) override;
bool can_reuse(const llm_graph_params & params) override;
std::unique_ptr<llm_graph_input_attn_kv_iswa> inp_attn;
std::unique_ptr<llm_graph_input_rs> inp_rs;
llm_graph_input_attn_kv_iswa * get_attn() const { return inp_attn.get(); }
llm_graph_input_rs * get_recr() const { return inp_rs.get(); }
const llama_cparams cparams;
const llama_memory_hybrid_iswa_context * mctx;
};
class llm_graph_input_sampling : public llm_graph_input_i {
public:
llm_graph_input_sampling(std::map<llama_seq_id, llama_sampler *> samplers) :
@ -537,7 +629,7 @@ public:
virtual ~llm_graph_result() = default;
ggml_tensor * get_tokens() const { return t_tokens; }
ggml_tensor * get_inp_tokens() const { return t_inp_tokens; }
ggml_tensor * get_logits() const { return t_logits; }
ggml_tensor * get_embd() const { return t_embd; }
ggml_tensor * get_embd_pooled() const { return t_embd_pooled; }
@ -564,7 +656,8 @@ public:
void set_params(const llm_graph_params & params);
// important graph nodes
ggml_tensor * t_tokens = nullptr;
ggml_tensor * t_inp_tokens = nullptr;
ggml_tensor * t_inp_embd = nullptr; // [n_embd_inp, n_tokens]
ggml_tensor * t_logits = nullptr;
ggml_tensor * t_embd = nullptr;
ggml_tensor * t_embd_pooled = nullptr;
@ -671,10 +764,11 @@ struct llm_graph_context {
ggml_tensor * cur,
int il) const;
// do mat_mul, while optionally apply lora
// do mat_mul, while optionally apply lora and per-tensor scale
ggml_tensor * build_lora_mm(
ggml_tensor * w,
ggml_tensor * cur) const;
ggml_tensor * cur,
ggml_tensor * w_s = nullptr) const;
// do mat_mul_id, while optionally apply lora
ggml_tensor * build_lora_mm_id(
@ -717,11 +811,14 @@ struct llm_graph_context {
int64_t n_expert_used,
llm_ffn_op_type type_op,
bool norm_w,
bool scale_w,
float w_scale,
llama_expert_gating_func_type gating_op,
int il,
ggml_tensor * probs_in = nullptr) const;
ggml_tensor * probs_in = nullptr,
ggml_tensor * gate_up_exps = nullptr,
ggml_tensor * up_exps_s = nullptr,
ggml_tensor * gate_exps_s = nullptr,
ggml_tensor * down_exps_s = nullptr) const;
ggml_tensor * build_moe_ffn(
ggml_tensor * cur,
@ -738,11 +835,15 @@ struct llm_graph_context {
int64_t n_expert_used,
llm_ffn_op_type type_op,
bool norm_w,
bool scale_w,
float w_scale,
llama_expert_gating_func_type gating_op,
int il,
ggml_tensor * probs_in = nullptr) const;
ggml_tensor * probs_in = nullptr,
ggml_tensor * gate_up_exps = nullptr,
ggml_tensor * gate_up_exps_b = nullptr,
ggml_tensor * up_exps_s = nullptr,
ggml_tensor * gate_exps_s = nullptr,
ggml_tensor * down_exps_s = nullptr) const;
//
// inputs
@ -801,6 +902,21 @@ struct llm_graph_context {
ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
ggml_tensor * kq_b,
ggml_tensor * sinks, // [n_head_q]
ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v] // TODO: remove
float kq_scale,
int il) const;
llm_graph_input_attn_k * build_attn_inp_k() const;
ggml_tensor * build_attn(
llm_graph_input_attn_k * inp,
ggml_tensor * wo,
ggml_tensor * wo_b,
ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
ggml_tensor * kq_b,
ggml_tensor * sinks, // [n_head_q]
ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
float kq_scale,
int il) const;
@ -880,6 +996,9 @@ struct llm_graph_context {
//
llm_graph_input_mem_hybrid * build_inp_mem_hybrid() const;
llm_graph_input_mem_hybrid_k * build_inp_mem_hybrid_k() const;
llm_graph_input_mem_hybrid_iswa * build_inp_mem_hybrid_iswa() const;
//
// pooling
@ -889,7 +1008,8 @@ struct llm_graph_context {
ggml_tensor * cls,
ggml_tensor * cls_b,
ggml_tensor * cls_out,
ggml_tensor * cls_out_b) const;
ggml_tensor * cls_out_b,
ggml_tensor * cls_norm) const;
//
// sampling (backend sampling)
@ -903,6 +1023,7 @@ struct llm_graph_context {
void build_dense_out(
ggml_tensor * dense_2,
ggml_tensor * dense_2_b,
ggml_tensor * dense_3) const;
};

97
examples/talk-llama/llama-hparams.cpp
View File

@ -62,6 +62,14 @@ uint32_t llama_hparams::n_gqa(uint32_t il) const {
return n_head/n_head_kv;
}
uint32_t llama_hparams::n_rot(uint32_t il) const {
if (il < n_layer) {
return is_swa(il) ? n_rot_swa : n_rot_full;
}
GGML_ABORT("fatal error");
}
uint32_t llama_hparams::n_embd_inp() const {
uint32_t n_embd_inp = n_embd;
@ -72,20 +80,36 @@ uint32_t llama_hparams::n_embd_inp() const {
return n_embd_inp;
}
uint32_t llama_hparams::get_n_embd_out() const {
return n_embd_out > 0 ? n_embd_out : n_embd;
uint32_t llama_hparams::n_embd_out() const {
return n_embd_out_impl > 0 ? n_embd_out_impl : n_embd;
}
uint32_t llama_hparams::n_embd_head_k(uint32_t il) const {
if (il < n_layer) {
return is_swa(il) ? n_embd_head_k_swa : n_embd_head_k_full;
}
GGML_ABORT("fatal error");
}
uint32_t llama_hparams::n_embd_head_v(uint32_t il) const {
if (il < n_layer) {
return is_swa(il) ? n_embd_head_v_swa : n_embd_head_v_full;
}
GGML_ABORT("fatal error");
}
uint32_t llama_hparams::n_embd_k_gqa(uint32_t il) const {
const uint32_t n_head_kv = this->n_head_kv(il);
return n_embd_head_k * n_head_kv;
return n_embd_head_k(il) * n_head_kv;
}
uint32_t llama_hparams::n_embd_v_gqa(uint32_t il) const {
const uint32_t n_head_kv = this->n_head_kv(il);
return n_embd_head_v * n_head_kv;
return n_embd_head_v(il) * n_head_kv;
}
bool llama_hparams::is_n_embd_k_gqa_variable() const {
@ -139,6 +163,13 @@ uint32_t llama_hparams::n_embd_r() const {
return n_embd * (n_shortconv_l_cache - 1);
}
if (n_embd_head_kda != 0) {
// for Kimi KDA layers
// Conv state for Q, K, V: 3 * (d_conv - 1) * n_head * head_dim
const uint32_t d_inner = n_head() * n_embd_head_kda; // 32 * 128 = 4096
return 3 * (ssm_d_conv > 0 ? ssm_d_conv - 1 : 3) * d_inner;
}
// TODO: maybe support other convolution strides than 1
// NOTE: since the first column of the conv_state is shifted out each time, it's not actually needed
// Corresponds to Mamba's conv_states size
@ -151,6 +182,13 @@ uint32_t llama_hparams::n_embd_s() const {
return n_embd * wkv_head_size;
}
if (n_embd_head_kda != 0) {
// for Kimi KDA layers
// Full recurrent state: head_dim * head_dim * n_head
// h tensor shape for delta attention: [head_dim, head_dim, n_head]
return n_embd_head_kda * n_embd_head_kda * n_head(); // 128 * 128 * 32 = 524288
}
// corresponds to Mamba's ssm_states size
return ssm_d_state * ssm_d_inner;
}
@ -175,6 +213,21 @@ bool llama_hparams::is_swa(uint32_t il) const {
GGML_ABORT("fatal error");
}
bool llama_hparams::is_mla() const {
assert((n_embd_head_k_mla_impl == 0 && n_embd_head_v_mla_impl == 0) ||
(n_embd_head_k_mla_impl != 0 && n_embd_head_v_mla_impl != 0));
return n_embd_head_k_mla_impl != 0 && n_embd_head_v_mla_impl != 0;
}
uint32_t llama_hparams::n_embd_head_k_mla() const {
return is_mla() ? n_embd_head_k_mla_impl : n_embd_head_k();
}
uint32_t llama_hparams::n_embd_head_v_mla() const {
return is_mla() ? n_embd_head_v_mla_impl : n_embd_head_v();
}
bool llama_hparams::has_kv(uint32_t il) const {
if (n_layer_kv_from_start >= 0) {
if (il < (uint32_t) n_layer_kv_from_start) {
@ -200,42 +253,6 @@ uint32_t llama_hparams::n_layer_kv() const {
return res;
}
bool llama_hparams::is_masked_swa(uint32_t n_swa, llama_swa_type swa_type, llama_pos p0, llama_pos p1) {
assert(p0 >= 0 && p1 >= 0);
switch (swa_type) {
case LLAMA_SWA_TYPE_NONE:
{
} break;
case LLAMA_SWA_TYPE_STANDARD:
{
if (p1 - p0 >= (int32_t) n_swa) {
return true;
}
} break;
case LLAMA_SWA_TYPE_CHUNKED:
{
const llama_pos pos_chunk_start = (p1 / n_swa) * n_swa;
if (p0 < pos_chunk_start) {
return true;
}
} break;
case LLAMA_SWA_TYPE_SYMMETRIC:
{
const int32_t half_n_swa = (int32_t) n_swa / 2;
const int32_t pos_diff = p1 - p0;
// Mask if outside the symmetric window
if (pos_diff < -half_n_swa || pos_diff > half_n_swa) {
return true;
}
} break;
}
return false;
}
bool llama_hparams::use_mrope() const {
return rope_sections[0] > 0 && rope_sections[1] > 0;
}

89
examples/talk-llama/llama-hparams.h
View File

@ -3,6 +3,7 @@
#include "llama.h"
#include <array>
#include <cassert>
// bump if necessary
#define LLAMA_MAX_LAYERS 512
@ -41,19 +42,25 @@ struct llama_hparams {
uint32_t n_ctx_train; // context size the model was trained on
uint32_t n_embd;
uint32_t n_embd_features = 0;
uint32_t n_layer;
int32_t n_layer_kv_from_start = -1; // if non-negative, the first n_layer_kv_from_start layers have KV cache
uint32_t n_rot;
uint32_t n_embd_head_k; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads
uint32_t n_embd_head_v; // dimension of values (d_v) aka n_embd_head
uint32_t n_expert = 0;
uint32_t n_expert_used = 0;
uint32_t n_rel_attn_bkts = 0;
// different head size for full_attention and SWA layers
uint32_t n_embd_head_k_full; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads
uint32_t n_embd_head_v_full; // dimension of values (d_v) aka n_embd_head
uint32_t n_embd_head_k_swa;
uint32_t n_embd_head_v_swa;
// different RoPE dimensions for full_attention and SWA layers
uint32_t n_rot_full;
uint32_t n_rot_swa;
// note: deepseek2 using MLA converts into MQA with larger heads, then decompresses to MHA
uint32_t n_embd_head_k_mla = 0;
uint32_t n_embd_head_v_mla = 0;
uint32_t n_embd_head_k_mla_impl = 0;
uint32_t n_embd_head_v_mla_impl = 0;
// for WavTokenizer
struct llama_hparams_posnet posnet;
@ -82,6 +89,7 @@ struct llama_hparams {
bool expert_weights_norm = false;
uint32_t expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_NONE;
uint32_t moe_every_n_layers = 0;
uint32_t moe_latent_size = 0;
uint32_t nextn_predict_layers = 0;
float f_norm_eps;
@ -136,6 +144,9 @@ struct llama_hparams {
uint32_t ssm_dt_rank = 0;
uint32_t ssm_n_group = 0;
// for Kimi Linear KDA
uint32_t n_embd_head_kda = 0;
// for hybrid state space models
std::array<bool, LLAMA_MAX_LAYERS> recurrent_layer_arr;
@ -163,7 +174,7 @@ struct llama_hparams {
uint32_t n_cls_out = 1;
// output embedding dimension (0 = use n_embd)
uint32_t n_embd_out = 0;
uint32_t n_embd_out_impl = 0;
// llama4 smallthinker
uint32_t n_moe_layer_step = 0;
@ -190,11 +201,16 @@ struct llama_hparams {
std::array<float, LLAMA_MAX_LAYERS> xielu_beta;
std::array<float, LLAMA_MAX_LAYERS> xielu_eps;
// DSA (deepseek sparse attention)
uint32_t indexer_n_head = 0;
uint32_t indexer_head_size = 0;
uint32_t indexer_top_k = 0;
// qwen3vl deepstack
uint32_t n_deepstack_layers = 0;
// needed by encoder-decoder models (e.g. T5, FLAN-T5)
// ref: https://github.com/ggerganov/llama.cpp/pull/8141
// ref: https://github.com/ggml-org/llama.cpp/pull/8141
llama_token dec_start_token_id = LLAMA_TOKEN_NULL;
uint32_t dec_n_layer = 0;
@ -202,6 +218,11 @@ struct llama_hparams {
enum llama_rope_type rope_type = LLAMA_ROPE_TYPE_NONE;
enum llama_rope_scaling_type rope_scaling_type_train = LLAMA_ROPE_SCALING_TYPE_NONE;
// Step35: optional per-layer clamps for (Swi)GLU
std::array<float, LLAMA_MAX_LAYERS> swiglu_clamp_exp; // clamping for expert FFN
std::array<float, LLAMA_MAX_LAYERS> swiglu_clamp_shexp; // shared expert
// this value n_pattern means that every nth layer is dense (i.e. non-SWA)
// dense_first means whether the pattern is start with a dense layer
// note that if n_pattern == 0, all layers are SWA
@ -234,11 +255,17 @@ struct llama_hparams {
uint32_t n_gqa(uint32_t il = 0) const;
uint32_t n_rot(uint32_t il = 0) const;
// dimension of main + auxiliary input embeddings
uint32_t n_embd_inp() const;
// dimension of output embeddings
uint32_t get_n_embd_out() const;
uint32_t n_embd_out() const;
// dimension of key/value embeddings for each head (per layer)
uint32_t n_embd_head_k(uint32_t il = 0) const;
uint32_t n_embd_head_v(uint32_t il = 0) const;
// dimension of key embeddings across all k-v heads
uint32_t n_embd_k_gqa(uint32_t il = 0) const;
@ -268,15 +295,57 @@ struct llama_hparams {
bool is_swa(uint32_t il) const;
// note: currently only support if either all or none of the layers are MLA
bool is_mla() const;
uint32_t n_embd_head_k_mla() const;
uint32_t n_embd_head_v_mla() const;
bool has_kv(uint32_t il) const;
// number of layers for which has_kv() returns true
uint32_t n_layer_kv() const;
// note that this function uses different SWA parameters from those in the hparams
// note: inlined on purpose for performance reasons
// TODO: think of a better place for this function
// TODO: pack the SWA params in a struct?
static bool is_masked_swa(uint32_t n_swa, llama_swa_type swa_type, llama_pos p0, llama_pos p1);
static bool is_masked_swa(uint32_t n_swa, llama_swa_type swa_type, llama_pos p0, llama_pos p1) {
assert(p0 >= 0 && p1 >= 0);
switch (swa_type) {
case LLAMA_SWA_TYPE_NONE:
{
} break;
case LLAMA_SWA_TYPE_STANDARD:
{
if (p1 - p0 >= (int32_t) n_swa) {
return true;
}
} break;
case LLAMA_SWA_TYPE_CHUNKED:
{
const llama_pos pos_chunk_start = (p1 / n_swa) * n_swa;
if (p0 < pos_chunk_start) {
return true;
}
} break;
case LLAMA_SWA_TYPE_SYMMETRIC:
{
const int32_t half_n_swa = (int32_t) n_swa / 2;
const int32_t pos_diff = p1 - p0;
// Mask if outside the symmetric window
if (pos_diff < -half_n_swa || pos_diff > half_n_swa) {
return true;
}
} break;
}
return false;
}
bool use_mrope() const;
};

8
examples/talk-llama/llama-impl.cpp
View File

@ -100,18 +100,18 @@ std::string format(const char * fmt, ...) {
std::string llama_format_tensor_shape(const std::vector<int64_t> & ne) {
char buf[256];
snprintf(buf, sizeof(buf), "%5" PRId64, ne.at(0));
snprintf(buf, sizeof(buf), "%6" PRId64, ne.at(0));
for (size_t i = 1; i < ne.size(); i++) {
snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %5" PRId64, ne.at(i));
snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %6" PRId64, ne.at(i));
}
return buf;
}
std::string llama_format_tensor_shape(const struct ggml_tensor * t) {
char buf[256];
snprintf(buf, sizeof(buf), "%5" PRId64, t->ne[0]);
snprintf(buf, sizeof(buf), "%6" PRId64, t->ne[0]);
for (int i = 1; i < GGML_MAX_DIMS; i++) {
snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %5" PRId64, t->ne[i]);
snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %6" PRId64, t->ne[i]);
}
return buf;
}

14
examples/talk-llama/llama-impl.h
View File

@ -49,6 +49,16 @@ struct time_meas {
int64_t & t_acc;
};
template <typename T>
struct buffer_view {
T * data;
size_t size = 0;
bool has_data() const {
return data && size > 0;
}
};
void replace_all(std::string & s, const std::string & search, const std::string & replace);
// TODO: rename to llama_format ?
@ -60,4 +70,6 @@ std::string llama_format_tensor_shape(const struct ggml_tensor * t);
std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i);
#define LLAMA_TENSOR_NAME_FATTN "__fattn__"
#define LLAMA_TENSOR_NAME_FATTN "__fattn__"
#define LLAMA_TENSOR_NAME_FGDN_AR "__fgdn_ar__"
#define LLAMA_TENSOR_NAME_FGDN_CH "__fgdn_ch__"

4
examples/talk-llama/llama-kv-cache-iswa.cpp
View File

@ -218,7 +218,9 @@ llama_memory_context_ptr llama_kv_cache_iswa::init_update(llama_context * lctx,
}
bool llama_kv_cache_iswa::get_can_shift() const {
return kv_base->get_size() == kv_swa->get_size();
return kv_base->get_can_shift() &&
kv_swa->get_can_shift() &&
kv_base->get_size() == kv_swa->get_size();
}
void llama_kv_cache_iswa::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {

385
examples/talk-llama/llama-kv-cache.cpp
View File

@ -97,6 +97,8 @@ llama_kv_cache::llama_kv_cache(
__func__, hparams.n_embd_v_gqa_max());
}
const bool is_mla = hparams.is_mla();
for (uint32_t il = 0; il < hparams.n_layer; il++) {
if (!hparams.has_kv(il)) {
LLAMA_LOG_DEBUG("%s: layer %3d: does not have KV cache\n", __func__, il);
@ -130,18 +132,21 @@ llama_kv_cache::llama_kv_cache(
throw std::runtime_error("failed to create ggml context for kv cache");
}
ggml_tensor * k = ggml_new_tensor_3d(ctx, type_k, n_embd_k_gqa, kv_size, n_stream);
ggml_tensor * v = ggml_new_tensor_3d(ctx, type_v, n_embd_v_gqa, kv_size, n_stream);
const bool has_k = true;
const bool has_v = !is_mla;
ggml_format_name(k, "cache_k_l%d", il);
ggml_format_name(v, "cache_v_l%d", il);
ggml_tensor * k = has_k ? ggml_new_tensor_3d(ctx, type_k, n_embd_k_gqa, kv_size, n_stream) : nullptr;
ggml_tensor * v = has_v ? ggml_new_tensor_3d(ctx, type_v, n_embd_v_gqa, kv_size, n_stream) : nullptr;
has_k && ggml_format_name(k, "cache_k_l%d", il);
has_v && ggml_format_name(v, "cache_v_l%d", il);
std::vector<ggml_tensor *> k_stream;
std::vector<ggml_tensor *> v_stream;
for (uint32_t s = 0; s < n_stream; ++s) {
k_stream.push_back(ggml_view_2d(ctx, k, n_embd_k_gqa, kv_size, k->nb[1], s*k->nb[2]));
v_stream.push_back(ggml_view_2d(ctx, v, n_embd_v_gqa, kv_size, v->nb[1], s*v->nb[2]));
k_stream.push_back(has_k ? ggml_view_2d(ctx, k, n_embd_k_gqa, kv_size, k->nb[1], s*k->nb[2]) : nullptr);
v_stream.push_back(has_v ? ggml_view_2d(ctx, v, n_embd_v_gqa, kv_size, v->nb[1], s*v->nb[2]) : nullptr);
}
map_layer_ids[il] = layers.size();
@ -578,7 +583,7 @@ llama_kv_cache::slot_info_vec_t llama_kv_cache::prepare(const std::vector<llama_
break;
}
// remeber the position that we found
// remember the position that we found
res.push_back(sinfo_new);
// store the old state of the cells in the recovery stack
@ -647,7 +652,10 @@ bool llama_kv_cache::update(llama_context * lctx, bool do_shift, const stream_co
const auto & layer = layers[il];
ggml_backend_tensor_copy(layer.k_stream[ssrc], layer.k_stream[sdst]);
ggml_backend_tensor_copy(layer.v_stream[ssrc], layer.v_stream[sdst]);
if (layer.v_stream[ssrc]) {
ggml_backend_tensor_copy(layer.v_stream[ssrc], layer.v_stream[sdst]);
}
}
}
}
@ -852,7 +860,7 @@ llama_kv_cache::slot_info llama_kv_cache::find_slot(const llama_ubatch & ubatch,
const llama_seq_id seq_id_cell = cells.seq_get(idx);
// SWA mask
if (is_masked_swa(pos_cell, cells.seq_pos_max(seq_id_cell) + 1)) {
if (llama_hparams::is_masked_swa(n_swa, swa_type, pos_cell, cells.seq_pos_max(seq_id_cell) + 1)) {
can_use = true;
}
}
@ -966,6 +974,13 @@ void llama_kv_cache::apply_ubatch(const slot_info & sinfo, const llama_ubatch &
}
bool llama_kv_cache::get_can_shift() const {
// Step35 uses per-layer RoPE dims; K-shift assumes a single global n_rot.
if (model.arch == LLM_ARCH_STEP35) {
return false;
}
if (hparams.n_pos_per_embd() > 1) {
return false;
}
return true;
}
@ -1018,8 +1033,8 @@ ggml_tensor * llama_kv_cache::get_k(ggml_context * ctx, int32_t il, uint32_t n_k
const uint32_t ns = sinfo.s1 - sinfo.s0 + 1;
return ggml_view_4d(ctx, k,
hparams.n_embd_head_k, hparams.n_head_kv(il), n_kv, ns,
ggml_row_size(k->type, hparams.n_embd_head_k),
hparams.n_embd_head_k(il), hparams.n_head_kv(il), n_kv, ns,
ggml_row_size(k->type, hparams.n_embd_head_k(il)),
ggml_row_size(k->type, n_embd_k_gqa),
ggml_row_size(k->type, n_embd_k_gqa*kv_size),
ggml_row_size(k->type, n_embd_k_gqa*kv_size)*sinfo.s0);
@ -1041,8 +1056,8 @@ ggml_tensor * llama_kv_cache::get_v(ggml_context * ctx, int32_t il, uint32_t n_k
if (!v_trans) {
// note: v->nb[1] <= v->nb[2]
return ggml_view_4d(ctx, v,
hparams.n_embd_head_v, hparams.n_head_kv(il), n_kv, ns,
ggml_row_size(v->type, hparams.n_embd_head_v), // v->nb[1]
hparams.n_embd_head_v(il), hparams.n_head_kv(il), n_kv, ns,
ggml_row_size(v->type, hparams.n_embd_head_v(il)), // v->nb[1]
ggml_row_size(v->type, n_embd_v_gqa), // v->nb[2]
ggml_row_size(v->type, n_embd_v_gqa*kv_size), // v->nb[3]
ggml_row_size(v->type, n_embd_v_gqa*kv_size)*sinfo.s0);
@ -1050,8 +1065,8 @@ ggml_tensor * llama_kv_cache::get_v(ggml_context * ctx, int32_t il, uint32_t n_k
// note: v->nb[1] > v->nb[2]
return ggml_view_4d(ctx, v,
n_kv, hparams.n_head_kv(il), hparams.n_embd_head_v, ns,
ggml_row_size(v->type, kv_size*hparams.n_embd_head_v), // v->nb[1]
n_kv, hparams.n_head_kv(il), hparams.n_embd_head_v(il), ns,
ggml_row_size(v->type, kv_size*hparams.n_embd_head_v(il)), // v->nb[1]
ggml_row_size(v->type, kv_size), // v->nb[2]
ggml_row_size(v->type, kv_size*n_embd_v_gqa), // v->nb[3]
ggml_row_size(v->type, kv_size*n_embd_v_gqa)*sinfo.s0);
@ -1237,6 +1252,197 @@ void llama_kv_cache::set_input_k_shift(ggml_tensor * dst) const {
}
}
struct args_set_input_kq_mask {
const llama_hparams & hparams;
const llama_ubatch * ubatch;
const std::vector<llama_kv_cells> & v_cells;
const std::vector<uint32_t> & seq_to_stream;
uint32_t n_swa;
llama_swa_type swa_type;
int64_t n_kv;
int64_t n_stream;
int64_t n_tps;
};
template<bool causal, bool swa, bool is_2d, bool alibi>
static void set_input_kq_mask_impl(const args_set_input_kq_mask & args, float * data) {
//const auto & hparams = args.hparams;
const auto & ubatch = args.ubatch;
const auto & v_cells = args.v_cells;
const auto & seq_to_stream = args.seq_to_stream;
const uint32_t n_swa = args.n_swa;
const llama_swa_type swa_type = args.swa_type;
const int64_t n_kv = args.n_kv;
const int64_t n_stream = args.n_stream;
const int64_t n_tps = args.n_tps;
// the min position in the batch for each sequence
llama_pos seq_pos_min[LLAMA_MAX_SEQ];
std::fill(seq_pos_min, seq_pos_min + LLAMA_MAX_SEQ, INT32_MAX);
for (uint32_t i = 0; i < ubatch->n_tokens; ++i) {
const llama_seq_id seq_id = ubatch->seq_id[i][0];
seq_pos_min[seq_id] = std::min(seq_pos_min[seq_id], ubatch->pos[i]);
}
for (uint32_t s = 0; s < n_stream; ++s) {
// bookkeeping of the KQ mask cells that could change for other tokens of the same sequence
std::unordered_map<llama_seq_id, uint32_t> seq_srct;
std::unordered_map<llama_seq_id, std::vector<uint32_t>> seq_idxs;
for (uint32_t ii = 0; ii < n_tps; ++ii) {
const uint32_t i = s*n_tps + ii;
const llama_seq_id seq_id = ubatch->seq_id[i][0];
const auto & cells = v_cells.at(seq_to_stream[seq_id]);
llama_pos p0 = -1;
const llama_pos p1 = ubatch->pos[i];
// for M-RoPE
const llama_pos p1_x = is_2d ? ubatch->pos[i + ubatch->n_tokens*2] : 0;
const llama_pos p1_y = is_2d ? ubatch->pos[i + ubatch->n_tokens] : 0;
const uint64_t idst = n_kv*i;
// for tokens of the same sequence, the mask is mostly the same, so we can reuse it
// the only cells that could change are the ones that are with similar positions as the
// ones in the batch (i.e. due to causal masking, SWA, etc.)
// keep track of those cells and shortcut the loop to save time
// note: this optimization is not compatible with Alibi position encoding
// ref: https://github.com/ggml-org/llama.cpp/pull/18842
bool prev = false;
auto & idxs = seq_idxs[seq_id];
if (!alibi) {
if (seq_srct.find(seq_id) != seq_srct.end()) {
const uint32_t srct = seq_srct[seq_id];
const uint64_t idst_prev = n_kv*srct;
std::copy(data + idst_prev, data + idst_prev + n_kv, data + idst);
prev = true;
} else {
idxs.clear();
idxs.reserve(ubatch->n_tokens + n_swa + 32);
seq_srct[seq_id] = i;
}
}
for (uint32_t jj = 0; jj < n_kv; ++jj) {
uint32_t j = jj;
// we have an exiting mask for this sequence -> update just seq_idxs
if (!alibi) {
if (prev) {
if (jj >= idxs.size()) {
break;
}
j = idxs[jj];
}
}
if (cells.is_empty(j)) {
goto skip;
}
// mask the token if not the same sequence
if (!cells.seq_has(j, seq_id)) {
goto skip;
}
p0 = cells.pos_get(j);
if (!alibi) {
if (!prev) {
// record all cells for which: p0 >= seq_pos_min[seq_id] - n_swa - 32
if (p0 + (int32_t) (n_swa + 32) >= seq_pos_min[seq_id]) {
idxs.push_back(j);
}
}
}
if (causal) {
// mask future tokens
if (p0 > p1) {
goto skip;
}
// M-RoPE causal mask
if (is_2d) {
if (p0 == p1) {
const auto & p0_ext = cells.ext_get(j);
if (p0_ext.is_2d_gt(p1_x, p1_y)) {
goto skip;
}
}
}
}
// apply SWA if any
if (swa) {
if (llama_hparams::is_masked_swa(n_swa, swa_type, p0, p1)) {
goto skip;
}
}
if (alibi) {
data[idst + j] = -std::abs(p0 - p1);
} else {
data[idst + j] = 0.0f;
}
continue;
skip:
data[idst + j] = -INFINITY;
}
}
}
}
template<bool causal, bool swa, bool is_2d>
static void set_input_kq_mask_impl(const args_set_input_kq_mask & args, float * data) {
const bool alibi = args.hparams.use_alibi;
if (alibi) {
set_input_kq_mask_impl<causal, swa, is_2d, true> (args, data);
} else {
set_input_kq_mask_impl<causal, swa, is_2d, false>(args, data);
}
}
template<bool causal, bool swa>
static void set_input_kq_mask_impl(const args_set_input_kq_mask & args, float * data) {
const bool is_2d = args.ubatch->is_pos_2d();
if (is_2d) {
set_input_kq_mask_impl<causal, swa, true> (args, data);
} else {
set_input_kq_mask_impl<causal, swa, false>(args, data);
}
}
template<bool causal>
static void set_input_kq_mask_impl(const args_set_input_kq_mask & args, float * data) {
const bool swa = args.swa_type != LLAMA_SWA_TYPE_NONE;
if (swa) {
set_input_kq_mask_impl<causal, true> (args, data);
} else {
set_input_kq_mask_impl<causal, false>(args, data);
}
}
void llama_kv_cache::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const {
const uint32_t n_tokens = ubatch->n_tokens;
@ -1251,74 +1457,29 @@ void llama_kv_cache::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * u
// n_tps == n_tokens_per_stream
const int64_t n_tps = n_tokens/n_stream;
std::fill(data, data + ggml_nelements(dst), -INFINITY);
//const int64_t t_start = ggml_time_us();
// Use only the previous KV cells of the correct sequence for each token of the ubatch.
// It's assumed that if a token in the batch has multiple sequences, they are equivalent.
// Example with a cache of 10 tokens, 2 tokens populated in cache and 3 tokens in batch:
// Causal mask:
// xxx-------
// xxxx------
// xxxxx-----
// Non-causal mask:
// xxxxx-----
// xxxxx-----
// xxxxx-----
// To visualize the mask, see https://github.com/ggml-org/llama.cpp/pull/12615
// TODO: optimize this section
for (uint32_t h = 0; h < 1; ++h) {
for (uint32_t s = 0; s < n_stream; ++s) {
for (uint32_t ii = 0; ii < n_tps; ++ii) {
const uint32_t i = s*n_tps + ii;
const args_set_input_kq_mask args = {
/*.hparams =*/ hparams,
/*.ubatch =*/ ubatch,
/*.v_cells =*/ v_cells,
/*.seq_to_stream =*/ seq_to_stream,
/*.n_swa =*/ n_swa,
/*.swa_type =*/ swa_type,
/*.n_kv =*/ n_kv,
/*.n_stream =*/ n_stream,
/*.n_tps =*/ n_tps,
};
const llama_seq_id seq_id = ubatch->seq_id[i][0];
const auto & cells = v_cells[seq_to_stream[seq_id]];
const llama_pos p1 = ubatch->pos[i];
// for M-RoPE
const bool is_2d = ubatch->is_pos_2d();
const llama_pos p1_x = is_2d ? ubatch->pos[i + ubatch->n_tokens*2] : 0;
const llama_pos p1_y = is_2d ? ubatch->pos[i + ubatch->n_tokens] : 0;
const uint64_t idst = n_kv*(h*n_stream*n_tps + s*n_tps + ii);
for (uint32_t j = 0; j < n_kv; ++j) {
if (cells.is_empty(j)) {
continue;
}
// mask the token if not the same sequence
if (!cells.seq_has(j, seq_id)) {
continue;
}
const llama_pos p0 = cells.pos_get(j);
// mask future tokens
if (causal_attn && p0 > p1) {
continue;
}
// M-RoPE causal mask
if (causal_attn && is_2d && p0 == p1) {
const auto & p0_ext = cells.ext_get(j);
if (p0_ext.is_2d_gt(p1_x, p1_y)) {
continue;
}
}
// apply SWA if any
if (is_masked_swa(p0, p1)) {
continue;
}
data[idst + j] = hparams.use_alibi ? -std::abs(p0 - p1) : 0.0f;
}
}
}
if (causal_attn) {
set_input_kq_mask_impl<true> (args, data);
} else {
set_input_kq_mask_impl<false>(args, data);
}
//const int64_t t_end = ggml_time_us();
//LLAMA_LOG_ERROR("%s: kq mask time: %0.3f ms\n", __func__, (t_end - t_start)/1000.0);
}
void llama_kv_cache::set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const {
@ -1370,7 +1531,7 @@ size_t llama_kv_cache::size_v_bytes() const {
size_t size_v_bytes = 0;
for (const auto & layer : layers) {
size_v_bytes += ggml_nbytes(layer.v);
size_v_bytes += layer.v ? ggml_nbytes(layer.v) : 0;
}
return size_v_bytes;
@ -1383,7 +1544,8 @@ ggml_tensor * llama_kv_cache::build_rope_shift(
ggml_tensor * shift,
ggml_tensor * factors,
float freq_base,
float freq_scale) const {
float freq_scale,
uint32_t il) const {
const auto & n_ctx_orig = cparams.n_ctx_orig_yarn;
const auto & yarn_ext_factor = cparams.yarn_ext_factor;
@ -1391,7 +1553,7 @@ ggml_tensor * llama_kv_cache::build_rope_shift(
const auto & yarn_beta_slow = cparams.yarn_beta_slow;
const auto & yarn_attn_factor = cparams.yarn_attn_factor;
const auto & n_rot = hparams.n_rot;
const auto & n_rot = hparams.n_rot(il);
const auto & rope_type = hparams.rope_type == LLAMA_ROPE_TYPE_MROPE || hparams.rope_type == LLAMA_ROPE_TYPE_IMROPE
// @ngxson : this is a workaround
// for M-RoPE, we want to rotate the whole vector when doing KV shift
@ -1445,9 +1607,6 @@ ggml_cgraph * llama_kv_cache::build_graph_shift(llm_graph_result * res, llama_co
auto * ctx = res->get_ctx();
auto * gf = res->get_gf();
const auto & n_embd_head_k = hparams.n_embd_head_k;
//const auto & n_embd_head_v = hparams.n_embd_head_v;
auto inp = std::make_unique<llm_graph_input_k_shift>(this);
inp->k_shift = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, (int64_t) get_size()*n_stream);
@ -1461,6 +1620,10 @@ ggml_cgraph * llama_kv_cache::build_graph_shift(llm_graph_result * res, llama_co
const int64_t n_head_kv = hparams.n_head_kv(il);
const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
const auto n_rot = hparams.n_rot(il);
const auto n_embd_head_k = hparams.n_embd_head_k(il);
const auto n_embd_nope = hparams.n_lora_kv > 0 ? n_embd_head_k - n_rot : 0;
const float freq_base_l = model.get_rope_freq_base (cparams, il);
const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
@ -1468,12 +1631,12 @@ ggml_cgraph * llama_kv_cache::build_graph_shift(llm_graph_result * res, llama_co
ggml_tensor * k =
ggml_view_3d(ctx, layer.k,
n_embd_head_k, n_head_kv, get_size()*n_stream,
n_rot, n_head_kv, get_size()*n_stream,
ggml_row_size(layer.k->type, n_embd_head_k),
ggml_row_size(layer.k->type, n_embd_k_gqa),
0);
ggml_row_size(layer.k->type, n_embd_nope));
ggml_tensor * cur = build_rope_shift(cparams, ctx, k, inp->k_shift, rope_factors, freq_base_l, freq_scale_l);
ggml_tensor * cur = build_rope_shift(cparams, ctx, k, inp->k_shift, rope_factors, freq_base_l, freq_scale_l, il);
ggml_build_forward_expand(gf, cur);
}
@ -1483,10 +1646,6 @@ ggml_cgraph * llama_kv_cache::build_graph_shift(llm_graph_result * res, llama_co
return gf;
}
bool llama_kv_cache::is_masked_swa(llama_pos p0, llama_pos p1) const {
return llama_hparams::is_masked_swa(n_swa, swa_type, p0, p1);
}
void llama_kv_cache::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
GGML_UNUSED(flags);
@ -1599,8 +1758,10 @@ void llama_kv_cache::state_write_meta(llama_io_write_i & io, const cell_ranges_t
io.write(&pos, sizeof(pos));
io.write(&n_seq_id, sizeof(n_seq_id));
// TODO: we also need to save llama_kv_cell_ext when apply_ubatch() support loading it
// see: https://github.com/ggml-org/llama.cpp/pull/16825#issuecomment-3460868350
if (hparams.n_pos_per_embd() > 1) {
const llama_kv_cell_ext ext = cells.ext_get(i);
io.write(&ext, sizeof(ext));
}
for (const auto & seq_id : seq_ids) {
io.write(&seq_id, sizeof(seq_id));
@ -1618,8 +1779,6 @@ void llama_kv_cache::state_write_data(llama_io_write_i & io, const cell_ranges_t
io.write(&v_trans, sizeof(v_trans));
io.write(&n_layer, sizeof(n_layer));
std::vector<uint8_t> tmp_buf;
// Iterate and write all the keys first, each row is a cell
// Get whole range at a time
for (const auto & layer : layers) {
@ -1637,7 +1796,7 @@ void llama_kv_cache::state_write_data(llama_io_write_i & io, const cell_ranges_t
const uint64_t k_size_row = ggml_row_size(k->type, n_embd_k_gqa);
io.write(&k_size_row, sizeof(k_size_row));
// Read each range of cells of k_size length each into tmp_buf and write out
// Read each range of cells of k_size length and write out
for (const auto & range : cr.data) {
const size_t range_size = range.second - range.first;
const size_t buf_size = range_size * k_size_row;
@ -1652,6 +1811,9 @@ void llama_kv_cache::state_write_data(llama_io_write_i & io, const cell_ranges_t
const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
auto * v = layer.v_stream[cr.strm];
if (!v) {
continue;
}
// Write value type
const int32_t v_type_i = (int32_t) v->type;
@ -1661,7 +1823,7 @@ void llama_kv_cache::state_write_data(llama_io_write_i & io, const cell_ranges_t
const uint64_t v_size_row = ggml_row_size(v->type, n_embd_v_gqa);
io.write(&v_size_row, sizeof(v_size_row));
// Read each range of cells of v_size length each into tmp_buf and write out
// Read each range of cells of v_size length and write out
for (const auto & range : cr.data) {
const size_t range_size = range.second - range.first;
const size_t buf_size = range_size * v_size_row;
@ -1678,6 +1840,9 @@ void llama_kv_cache::state_write_data(llama_io_write_i & io, const cell_ranges_t
const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
auto * v = layer.v_stream[cr.strm];
if (!v) {
continue;
}
// Write value type
const int32_t v_type_i = (int32_t) v->type;
@ -1692,7 +1857,7 @@ void llama_kv_cache::state_write_data(llama_io_write_i & io, const cell_ranges_t
// For each row, we get the element values of each cell
for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
// Read each range of cells of v_size_el length each into tmp_buf and write out
// Read each range of cells of v_size_el length and write out
for (const auto & range : cr.data) {
const size_t range_size = range.second - range.first;
const size_t src_offset = (range.first + j * kv_size) * v_size_el;
@ -1730,6 +1895,14 @@ bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32
return false;
}
if (hparams.n_pos_per_embd() > 1) {
llama_kv_cell_ext ext;
io.read_to(&ext, sizeof(ext));
ubatch.pos[i + ubatch.n_tokens] = ext.y;
ubatch.pos[i + ubatch.n_tokens*2] = ext.x;
}
// read the sequence id, but directly discard it - we will use dest_seq_id instead
{
llama_seq_id seq_id;
@ -1780,6 +1953,12 @@ bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32
cells.pos_set(i, pos);
if (hparams.n_pos_per_embd() > 1) {
llama_kv_cell_ext ext;
io.read_to(&ext, sizeof(ext));
cells.ext_set(i, ext);
}
for (uint32_t j = 0; j < n_seq_id; ++j) {
llama_seq_id seq_id;
io.read_to(&seq_id, sizeof(seq_id));
@ -1881,6 +2060,9 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
auto * v = layer.v_stream[strm];
if (!v) {
continue;
}
// Read type of value
int32_t v_type_i_ref;
@ -1922,6 +2104,9 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
auto * v = layer.v_stream[strm];
if (!v) {
continue;
}
// Read type of value
int32_t v_type_i_ref;

5
examples/talk-llama/llama-kv-cache.h
View File

@ -257,8 +257,6 @@ private:
size_t size_k_bytes() const;
size_t size_v_bytes() const;
bool is_masked_swa(llama_pos p0, llama_pos p1) const;
ggml_tensor * build_rope_shift(
const llama_cparams & cparams,
ggml_context * ctx,
@ -266,7 +264,8 @@ private:
ggml_tensor * shift,
ggml_tensor * factors,
float freq_base,
float freq_scale) const;
float freq_scale,
uint32_t il) const;
ggml_cgraph * build_graph_shift(
llm_graph_result * res,

275
examples/talk-llama/llama-memory-hybrid-iswa.cpp Normal file
View File

@ -0,0 +1,275 @@
#include "llama-memory-hybrid-iswa.h"
#include "llama-impl.h"
#include "llama-model.h"
#include "llama-context.h"
//
// llama_memory_hybrid_iswa
//
llama_memory_hybrid_iswa::llama_memory_hybrid_iswa(
const llama_model & model,
/* attn */
ggml_type type_k,
ggml_type type_v,
bool v_trans,
bool swa_full,
uint32_t kv_size,
uint32_t n_ubatch,
uint32_t n_pad,
/* recurrent */
ggml_type type_r,
ggml_type type_s,
uint32_t rs_size,
/* common */
uint32_t n_seq_max,
bool offload,
bool unified,
/* layer filters */
const layer_filter_cb & filter_attn,
const layer_filter_cb & filter_recr) :
hparams(model.hparams),
mem_attn(new llama_kv_cache_iswa(
model,
type_k,
type_v,
v_trans,
offload,
swa_full,
unified,
kv_size,
n_seq_max,
n_ubatch,
n_pad,
filter_attn == nullptr ?
[&](int32_t il) { return !hparams.is_recurrent(il); }
: filter_attn,
nullptr
)),
mem_recr(new llama_memory_recurrent(
model,
type_r,
type_s,
offload,
rs_size,
n_seq_max,
filter_recr == nullptr ?
[&](int32_t il) { return hparams.is_recurrent(il); }
: filter_recr
)) {}
llama_memory_context_ptr llama_memory_hybrid_iswa::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) {
do {
balloc.split_reset();
// follow the recurrent pattern for creating the ubatch splits
std::vector<llama_ubatch> ubatches;
while (true) {
llama_ubatch ubatch;
if (embd_all) {
// if all tokens are output, split by sequence
ubatch = balloc.split_seq(n_ubatch);
} else {
// TODO: non-sequential equal split can be done if using unified KV cache
// for simplicity, we always use sequential equal split for now
ubatch = balloc.split_equal(n_ubatch, true);
}
if (ubatch.n_tokens == 0) {
break;
}
ubatches.push_back(std::move(ubatch)); // NOLINT
}
if (balloc.get_n_used() < balloc.get_n_tokens()) {
// failed to find a suitable split
break;
}
// prepare the recurrent batches first
if (!mem_recr->prepare(ubatches)) {
// TODO: will the recurrent cache be in an undefined context at this point?
LLAMA_LOG_ERROR("%s: failed to prepare recurrent ubatches\n", __func__);
return std::make_unique<llama_memory_hybrid_iswa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
}
// prepare the attention cache (iswa version returns both base and swa slot infos)
auto sinfos_base = mem_attn->get_base()->prepare(ubatches);
if (sinfos_base.empty()) {
LLAMA_LOG_ERROR("%s: failed to prepare attention base ubatches\n", __func__);
return std::make_unique<llama_memory_hybrid_iswa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
}
auto sinfos_swa = mem_attn->get_swa()->prepare(ubatches);
if (sinfos_swa.empty()) {
LLAMA_LOG_ERROR("%s: failed to prepare attention swa ubatches\n", __func__);
return std::make_unique<llama_memory_hybrid_iswa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
}
return std::make_unique<llama_memory_hybrid_iswa_context>(
this, std::move(sinfos_base), std::move(sinfos_swa), std::move(ubatches));
} while(false);
return std::make_unique<llama_memory_hybrid_iswa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
}
llama_memory_context_ptr llama_memory_hybrid_iswa::init_full() {
return std::make_unique<llama_memory_hybrid_iswa_context>(this);
}
llama_memory_context_ptr llama_memory_hybrid_iswa::init_update(llama_context * lctx, bool optimize) {
return std::make_unique<llama_memory_hybrid_iswa_context>(this, lctx, optimize);
}
bool llama_memory_hybrid_iswa::get_can_shift() const {
// Shifting is trivially supported for recurrent
return mem_attn->get_can_shift();
}
void llama_memory_hybrid_iswa::clear(bool data) {
mem_attn->clear(data);
mem_recr->clear(data);
}
bool llama_memory_hybrid_iswa::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
// Try removing from the recurrent cache first since it may fail. If it does
// fail, the cache will not have been mutated.
if (!mem_recr->seq_rm(seq_id, p0, p1)) {
return false;
}
return mem_attn->seq_rm(seq_id, p0, p1);
}
void llama_memory_hybrid_iswa::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
mem_attn->seq_cp(seq_id_src, seq_id_dst, p0, p1);
mem_recr->seq_cp(seq_id_src, seq_id_dst, p0, p1);
}
void llama_memory_hybrid_iswa::seq_keep(llama_seq_id seq_id) {
mem_attn->seq_keep(seq_id);
mem_recr->seq_keep(seq_id);
}
void llama_memory_hybrid_iswa::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
mem_attn->seq_add(seq_id, p0, p1, shift);
mem_recr->seq_add(seq_id, p0, p1, shift);
}
void llama_memory_hybrid_iswa::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
mem_attn->seq_div(seq_id, p0, p1, d);
mem_recr->seq_div(seq_id, p0, p1, d);
}
llama_pos llama_memory_hybrid_iswa::seq_pos_min(llama_seq_id seq_id) const {
// the min of the total cache is the max of the two caches' min values
return std::max(mem_attn->seq_pos_min(seq_id), mem_recr->seq_pos_min(seq_id));
}
llama_pos llama_memory_hybrid_iswa::seq_pos_max(llama_seq_id seq_id) const {
// the max of the total cache is the min of the two caches' max values
return std::min(mem_attn->seq_pos_max(seq_id), mem_recr->seq_pos_max(seq_id));
}
std::map<ggml_backend_buffer_type_t, size_t> llama_memory_hybrid_iswa::memory_breakdown() const {
std::map<ggml_backend_buffer_type_t, size_t> mb = mem_attn->memory_breakdown();
for (const auto & buft_size : mem_recr->memory_breakdown()) {
mb[buft_size.first] += buft_size.second;
}
return mb;
}
void llama_memory_hybrid_iswa::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
mem_attn->state_write(io, seq_id, flags);
mem_recr->state_write(io, seq_id, flags);
}
void llama_memory_hybrid_iswa::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
mem_attn->state_read(io, seq_id, flags);
mem_recr->state_read(io, seq_id, flags);
}
llama_kv_cache_iswa * llama_memory_hybrid_iswa::get_mem_attn() const {
return mem_attn.get();
}
llama_memory_recurrent * llama_memory_hybrid_iswa::get_mem_recr() const {
return mem_recr.get();
}
//
// llama_memory_hybrid_iswa_context
//
llama_memory_hybrid_iswa_context::llama_memory_hybrid_iswa_context(llama_memory_status status) : status(status) {}
llama_memory_hybrid_iswa_context::llama_memory_hybrid_iswa_context(llama_memory_hybrid_iswa * mem) :
ctx_attn(mem->get_mem_attn()->init_full()),
ctx_recr(mem->get_mem_recr()->init_full()),
status(llama_memory_status_combine(ctx_attn->get_status(), ctx_recr->get_status())) {
}
llama_memory_hybrid_iswa_context::llama_memory_hybrid_iswa_context(
llama_memory_hybrid_iswa * mem,
llama_context * lctx,
bool optimize) :
ctx_attn(mem->get_mem_attn()->init_update(lctx, optimize)),
ctx_recr(mem->get_mem_recr()->init_update(lctx, optimize)),
status(llama_memory_status_combine(ctx_attn->get_status(), ctx_recr->get_status())) {
}
llama_memory_hybrid_iswa_context::llama_memory_hybrid_iswa_context(
llama_memory_hybrid_iswa * mem,
slot_info_vec_t sinfos_base,
slot_info_vec_t sinfos_swa,
std::vector<llama_ubatch> ubatches) :
ubatches(std::move(ubatches)),
// note: here we copy the ubatches. not sure if this is ideal
ctx_attn(new llama_kv_cache_iswa_context(mem->get_mem_attn(), std::move(sinfos_base), std::move(sinfos_swa), this->ubatches)),
ctx_recr(new llama_memory_recurrent_context(mem->get_mem_recr(), this->ubatches)),
status(llama_memory_status_combine(ctx_attn->get_status(), ctx_recr->get_status())) {
}
bool llama_memory_hybrid_iswa_context::next() {
assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
ctx_attn->next();
ctx_recr->next();
if (++i_next >= ubatches.size()) {
return false;
}
return true;
}
bool llama_memory_hybrid_iswa_context::apply() {
assert(!llama_memory_status_is_fail(status));
bool res = true;
res = res & ctx_attn->apply();
res = res & ctx_recr->apply();
return res;
}
llama_memory_status llama_memory_hybrid_iswa_context::get_status() const {
return status;
}
const llama_ubatch & llama_memory_hybrid_iswa_context::get_ubatch() const {
assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
return ubatches[i_next];
}
const llama_kv_cache_iswa_context * llama_memory_hybrid_iswa_context::get_attn() const {
return static_cast<const llama_kv_cache_iswa_context *>(ctx_attn.get());
}
const llama_memory_recurrent_context * llama_memory_hybrid_iswa_context::get_recr() const {
return static_cast<const llama_memory_recurrent_context *>(ctx_recr.get());
}

140
examples/talk-llama/llama-memory-hybrid-iswa.h Normal file
View File

@ -0,0 +1,140 @@
#pragma once
#include "llama-batch.h"
#include "llama-graph.h"
#include "llama-kv-cache-iswa.h"
#include "llama-memory.h"
#include "llama-memory-recurrent.h"
#include <memory>
#include <vector>
//
// llama_memory_hybrid_iswa
//
// utilizes instances of llama_memory_recurrent and llama_kv_cache_iswa to
// support models where each layer may be either attention-based (with SWA support) or recurrent
class llama_memory_hybrid_iswa : public llama_memory_i {
public:
llama_memory_hybrid_iswa(
const llama_model & model,
/* attn */
ggml_type type_k,
ggml_type type_v,
bool v_trans,
bool swa_full,
uint32_t kv_size,
uint32_t n_ubatch,
uint32_t n_pad,
/* recurrent */
ggml_type type_r,
ggml_type type_s,
uint32_t rs_size,
/* common */
uint32_t n_seq_max,
bool offload,
bool unified,
/* layer filters */
const layer_filter_cb & filter_attn = nullptr,
const layer_filter_cb & filter_recr = nullptr);
~llama_memory_hybrid_iswa() = default;
//
// llama_memory_i
//
llama_memory_context_ptr init_batch(
llama_batch_allocr & balloc,
uint32_t n_ubatch,
bool embd_all) override;
llama_memory_context_ptr init_full() override;
llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) override;
bool get_can_shift() const override;
void clear(bool data) override;
bool seq_rm (llama_seq_id seq_id, llama_pos p0, llama_pos p1) override;
void seq_cp (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) override;
void seq_keep(llama_seq_id seq_id) override;
void seq_add (llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) override;
void seq_div (llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) override;
llama_pos seq_pos_min(llama_seq_id seq_id) const override;
llama_pos seq_pos_max(llama_seq_id seq_id) const override;
std::map<ggml_backend_buffer_type_t, size_t> memory_breakdown() const override;
// state write/load
void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const override;
void state_read (llama_io_read_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) override;
//
// llama_memory_hybrid_iswa specific API
//
llama_kv_cache_iswa * get_mem_attn() const;
llama_memory_recurrent * get_mem_recr() const;
private:
const llama_hparams & hparams;
const std::unique_ptr<llama_kv_cache_iswa> mem_attn;
const std::unique_ptr<llama_memory_recurrent> mem_recr;
};
class llama_memory_hybrid_iswa_context : public llama_memory_context_i {
public:
using slot_info_vec_t = llama_kv_cache::slot_info_vec_t;
// init failure
explicit llama_memory_hybrid_iswa_context(llama_memory_status status);
// init full
explicit llama_memory_hybrid_iswa_context(llama_memory_hybrid_iswa * mem);
// init update
explicit llama_memory_hybrid_iswa_context(
llama_memory_hybrid_iswa * mem,
llama_context * lctx,
bool optimize);
// init success
llama_memory_hybrid_iswa_context(
llama_memory_hybrid_iswa * mem,
slot_info_vec_t sinfos_base,
slot_info_vec_t sinfos_swa,
std::vector<llama_ubatch> ubatches);
~llama_memory_hybrid_iswa_context() = default;
bool next() override;
bool apply() override;
llama_memory_status get_status() const override;
const llama_ubatch & get_ubatch() const override;
//
// llama_memory_hybrid_iswa_context
//
const llama_kv_cache_iswa_context * get_attn() const;
const llama_memory_recurrent_context * get_recr() const;
private:
// the index of the next ubatch to process
size_t i_next = 0;
std::vector<llama_ubatch> ubatches;
const llama_memory_context_ptr ctx_attn;
const llama_memory_context_ptr ctx_recr;
const llama_memory_status status;
};

24
examples/talk-llama/llama-memory-recurrent.cpp
View File

@ -163,7 +163,7 @@ bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
const auto & cell = cells[tail_id];
// partial intersection is invalid if it includes the final pos
if (0 < p0 && p0 <= cell.pos && p1 > cell.pos) {
//printf("[DEBUG] inside `llama_memory_recurrent::seq_rm`: partial intersection is invalid, so returning false\n");
//printf("[DEBUG] inside `llama_memory_recurrent::seq_rm`: partial intersection is invalid, so returning false, p0 = %d, cell.pos = %d, p1 = %d\n", p0, cell.pos, p1);
return false;
}
// invalidate tails which will be cleared
@ -785,23 +785,21 @@ void llama_memory_recurrent::state_write_data(llama_io_write_i & io, const std::
io.write(&s_trans, sizeof(s_trans));
io.write(&n_layer, sizeof(n_layer));
std::vector<uint8_t> tmp_buf;
// Iterate and write all the keys first, each row is a cell
// Iterate and write all the R tensors first, each row is a cell
// Get whole range at a time
for (uint32_t il = 0; il < n_layer; ++il) {
// skip null layers (read_data will handle this by checking "r_l" and "s_l" for null)
if (r_l[il] == nullptr) continue;
// Write key type
// Write R tensor type
const int32_t r_type_i = (int32_t)r_l[il]->type;
io.write(&r_type_i, sizeof(r_type_i));
// Write row size of key
// Write row size of R tensor
const uint64_t r_size_row = ggml_row_size(r_l[il]->type, hparams.n_embd_r());
io.write(&r_size_row, sizeof(r_size_row));
// Read each range of cells of k_size length each into tmp_buf and write out
// Write each range of cells of r_size_row length
for (const auto & range : cell_ranges) {
const size_t range_size = range.second - range.first;
const size_t buf_size = range_size * r_size_row;
@ -814,15 +812,15 @@ void llama_memory_recurrent::state_write_data(llama_io_write_i & io, const std::
// skip null layers (read_data will handle this by checking "r_l" and "s_l" for null)
if (s_l[il] == nullptr) continue;
// Write value type
// Write S tensor type
const int32_t s_type_i = (int32_t)s_l[il]->type;
io.write(&s_type_i, sizeof(s_type_i));
// Write row size of value
// Write row size of S tensor
const uint64_t s_size_row = ggml_row_size(s_l[il]->type, hparams.n_embd_s());
io.write(&s_size_row, sizeof(s_size_row));
// Read each range of cells of s_size length each into tmp_buf and write out
// Write each range of S tensor rows
for (const auto & range : cell_ranges) {
const size_t range_size = range.second - range.first;
const size_t buf_size = range_size * s_size_row;
@ -830,7 +828,7 @@ void llama_memory_recurrent::state_write_data(llama_io_write_i & io, const std::
}
}
} else {
// When v is transposed, we also need the element size and get the element ranges from each row
// When S tensor is transposed, we also need the element size and get the element ranges from each row
const uint32_t mem_size = size;
for (uint32_t il = 0; il < n_layer; ++il) {
// skip null layers (read_data will handle this by checking "r_l" and "s_l" for null)
@ -838,7 +836,7 @@ void llama_memory_recurrent::state_write_data(llama_io_write_i & io, const std::
const uint32_t n_embd_s = hparams.n_embd_s();
// Write value type
// Write S tensor type
const int32_t s_type_i = (int32_t)s_l[il]->type;
io.write(&s_type_i, sizeof(s_type_i));
@ -851,7 +849,7 @@ void llama_memory_recurrent::state_write_data(llama_io_write_i & io, const std::
// For each row, we get the element values of each cell
for (uint32_t j = 0; j < n_embd_s; ++j) {
// Read each range of cells of v_size_el length each into tmp_buf and write out
// Write each range of cells of s_size_el length
for (const auto & range : cell_ranges) {
const size_t range_size = range.second - range.first;
const size_t src_offset = (range.first + j * mem_size) * s_size_el;

39
examples/talk-llama/llama-mmap.cpp
View File

@ -244,11 +244,14 @@ struct llama_file::impl {
}
errno = 0;
if (fd == -1) {
std::size_t ret = std::fread(ptr, len, 1, fp);
const size_t curr_off = tell();
const size_t to_read = std::min(len, size - curr_off);
std::size_t ret = std::fread(ptr, to_read, 1, fp);
if (ferror(fp)) {
throw std::runtime_error(format("read error: %s", strerror(errno)));
}
if (ret != 1) {
if (to_read > 0 && ret != 1) {
throw std::runtime_error("unexpectedly reached end of file");
}
} else {
@ -262,7 +265,8 @@ struct llama_file::impl {
continue; // Interrupted by signal, retry
}
// Fallback to std::fread in case the DMA controller cannot access the buffer
if (errno == EFAULT) {
if (errno == EFAULT || errno == EINVAL) {
LLAMA_LOG_WARN("%s: Falling back to buffered IO due to %s\n", __func__, strerror(errno));
auto curr_off = tell();
close(fd);
fd = -1;
@ -381,6 +385,9 @@ int llama_file::file_id() const {
#ifdef _WIN32
return _fileno(pimpl->fp);
#else
if (pimpl->fd != -1) {
return pimpl->fd;
}
#if defined(fileno)
return fileno(pimpl->fp);
#else
@ -497,6 +504,8 @@ struct llama_mmap::impl {
}
}
#elif defined(_WIN32)
HANDLE hMapping = nullptr;
impl(struct llama_file * file, size_t prefetch, bool numa) {
GGML_UNUSED(numa);
@ -504,7 +513,7 @@ struct llama_mmap::impl {
HANDLE hFile = (HANDLE) _get_osfhandle(file->file_id());
HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
if (hMapping == NULL) {
DWORD error = GetLastError();
@ -513,9 +522,9 @@ struct llama_mmap::impl {
addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0);
DWORD error = GetLastError();
CloseHandle(hMapping);
if (addr == NULL) {
CloseHandle(hMapping);
throw std::runtime_error(format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str()));
}
@ -547,9 +556,17 @@ struct llama_mmap::impl {
}
~impl() {
if (!UnmapViewOfFile(addr)) {
LLAMA_LOG_WARN("warning: UnmapViewOfFile failed: %s\n",
llama_format_win_err(GetLastError()).c_str());
if (hMapping) {
if (addr) {
if (!UnmapViewOfFile(addr)) {
LLAMA_LOG_WARN("warning: UnmapViewOfFile failed: %s\n",
llama_format_win_err(GetLastError()).c_str());
}
}
if (!CloseHandle(hMapping)) {
LLAMA_LOG_WARN("warning: CloseHandle failed: %s\n",
llama_format_win_err(GetLastError()).c_str());
}
}
}
#else
@ -611,9 +628,9 @@ struct llama_mlock::impl {
char* errmsg = std::strerror(errno);
bool suggest = (errno == ENOMEM);
#if defined(TARGET_OS_VISION) || defined(TARGET_OS_TV) || defined(_AIX)
// visionOS/tvOS dont't support RLIMIT_MEMLOCK
// Skip resource limit checks on visionOS/tvOS
#if defined(TARGET_OS_VISION) || defined(TARGET_OS_TV) || defined(_AIX) || defined(__HAIKU__)
// visionOS/tvOS/Haiku don't support RLIMIT_MEMLOCK
// Skip resource limit checks on these platforms
suggest = false;
#else
struct rlimit lock_limit;

662
examples/talk-llama/llama-model-loader.cpp
View File

@ -1,11 +1,17 @@
#include "llama-model-loader.h"
#include "ggml-alloc.h"
#include "ggml.h"
#include "gguf.h"
#include "llama-hparams.h"
#include <algorithm>
#include <array>
#include <cinttypes>
#include <cstdint>
#include <cstring>
#include <future>
#include <regex>
static const size_t kiB = 1024;
static const size_t MiB = 1024*kiB;
@ -36,6 +42,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
case LLAMA_FTYPE_MOSTLY_Q5_1: return "Q5_1";
case LLAMA_FTYPE_MOSTLY_Q8_0: return "Q8_0";
case LLAMA_FTYPE_MOSTLY_MXFP4_MOE: return "MXFP4 MoE";
case LLAMA_FTYPE_MOSTLY_NVFP4: return "NVFP4";
case LLAMA_FTYPE_MOSTLY_Q2_K: return "Q2_K - Medium";
case LLAMA_FTYPE_MOSTLY_Q2_K_S: return "Q2_K - Small";
case LLAMA_FTYPE_MOSTLY_Q3_K_S: return "Q3_K - Small";
@ -262,7 +269,7 @@ namespace GGUFMeta {
template<typename T>
typename std::enable_if<std::is_integral<T>::value, bool>::type
llama_model_loader::get_arr_n(const std::string & key, T & result, bool required) {
const int kid = gguf_find_key(meta.get(), key.c_str());
const int kid = gguf_find_key(metadata, key.c_str());
if (kid < 0) {
if (required) {
@ -272,7 +279,7 @@ namespace GGUFMeta {
}
struct GGUFMeta::ArrayInfo arr_info =
GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(meta.get(), kid);
GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(metadata, kid);
result = arr_info.length;
@ -289,7 +296,7 @@ namespace GGUFMeta {
template<typename T>
bool llama_model_loader::get_arr(const std::string & key, std::vector<T> & result, bool required) {
const gguf_context * ctx = meta.get();
const gguf_context * ctx = metadata;
const int kid = gguf_find_key(ctx, key.c_str());
if (kid < 0 || gguf_get_kv_type(ctx, kid) != GGUF_TYPE_ARRAY) {
@ -330,7 +337,7 @@ namespace GGUFMeta {
template<typename T, size_t N_MAX>
bool llama_model_loader::get_arr(const std::string & key, std::array<T, N_MAX> & result, bool required) {
const gguf_context * ctx = meta.get();
const gguf_context * ctx = metadata;
const int kid = gguf_find_key(ctx, key.c_str());
if (kid < 0 || gguf_get_kv_type(ctx, kid) != GGUF_TYPE_ARRAY) {
@ -344,6 +351,7 @@ namespace GGUFMeta {
GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(ctx, kid);
switch (arr_info.gt) {
case GGUF_TYPE_BOOL:
case GGUF_TYPE_UINT32:
case GGUF_TYPE_INT32: GGML_ASSERT((std::is_same<T, int32_t>::value) ||
(std::is_same<T, uint32_t>::value)); break;
@ -365,7 +373,13 @@ namespace GGUFMeta {
result[i] = value;
}
} else {
std::copy((const T*)arr_info.data, (const T *)arr_info.data + arr_info.length, result.begin());
if (arr_info.gt == GGUF_TYPE_BOOL) {
std::transform((const bool *)arr_info.data, (const bool *)arr_info.data + arr_info.length, result.begin(), [](bool x) {
return static_cast<T>(x);
});
} else {
std::copy((const T*)arr_info.data, (const T *)arr_info.data + arr_info.length, result.begin());
}
}
return true;
@ -385,7 +399,7 @@ namespace GGUFMeta {
const struct llama_model_kv_override * override =
it != kv_overrides.end() ? &it->second : nullptr;
const bool found = GGUFMeta::GKV<T>::set(meta.get(), key, result, override);
const bool found = GGUFMeta::GKV<T>::set(metadata, key, result, override);
if (required && !found) {
throw std::runtime_error(format("key not found in model: %s", key.c_str()));
@ -419,7 +433,7 @@ namespace GGUFMeta {
// get array of n <= N_MAX elements, or a single element repeated n times
template<typename T, size_t N_MAX>
bool llama_model_loader::get_key_or_arr(const std::string & key, std::array<T, N_MAX> & result, uint32_t n, bool required) {
const int kid = gguf_find_key(meta.get(), key.c_str());
const int kid = gguf_find_key(metadata, key.c_str());
if (kid < 0) {
if (required) {
@ -432,9 +446,9 @@ namespace GGUFMeta {
throw std::runtime_error(format("n > N_MAX: %u > %u for key %s", (uint32_t) n, (uint32_t) N_MAX, key.c_str()));
}
if (gguf_get_kv_type(meta.get(), kid) == GGUF_TYPE_ARRAY) {
if (gguf_get_kv_type(metadata, kid) == GGUF_TYPE_ARRAY) {
struct GGUFMeta::ArrayInfo arr_info =
GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(meta.get(), kid);
GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(metadata, kid);
if (n != arr_info.length) {
throw std::runtime_error(format("key %s has wrong array length; expected %u, got %u", key.c_str(), n, (uint32_t) arr_info.length));
@ -465,7 +479,7 @@ namespace GGUFMeta {
bool llama_model_loader::get_key_or_arr(enum llm_kv kid, uint32_t & result, bool required) {
const std::string key = llm_kv(kid);
const int id = gguf_find_key(meta.get(), key.c_str());
const int id = gguf_find_key(metadata, key.c_str());
if (id < 0) {
if (required) {
@ -475,7 +489,7 @@ namespace GGUFMeta {
}
// throw and error if type is an array
if (gguf_get_kv_type(meta.get(), id) == GGUF_TYPE_ARRAY) {
if (gguf_get_kv_type(metadata, id) == GGUF_TYPE_ARRAY) {
if (required) {
throw std::runtime_error(format("expected scalar, found array for key: %s", key.c_str()));
}
@ -492,6 +506,9 @@ namespace GGUFMeta {
llama_model_loader::llama_model_loader(
struct gguf_context * meta,
llama_model_set_tensor_data_t set_tensor_data,
void * set_tensor_data_ud,
const std::string & fname,
std::vector<std::string> & splits,
bool use_mmap,
@ -499,7 +516,8 @@ llama_model_loader::llama_model_loader(
bool check_tensors,
bool no_alloc,
const llama_model_kv_override * param_overrides_p,
const llama_model_tensor_buft_override * param_tensor_buft_overrides_p) {
const llama_model_tensor_buft_override * param_tensor_buft_overrides_p)
: metadata(meta), set_tensor_data(set_tensor_data), set_tensor_data_ud(set_tensor_data_ud) {
int trace = 0;
if (getenv("LLAMA_TRACE")) {
trace = atoi(getenv("LLAMA_TRACE"));
@ -513,130 +531,142 @@ llama_model_loader::llama_model_loader(
tensor_buft_overrides = param_tensor_buft_overrides_p;
// Load the main GGUF
struct ggml_context * ctx = NULL;
struct gguf_init_params params = {
/*.no_alloc = */ true,
/*.ctx = */ &ctx,
};
if (!fname.empty()) {
// Load the main GGUF
struct ggml_context * ctx = NULL;
struct gguf_init_params params = {
/*.no_alloc = */ true,
/*.ctx = */ &ctx,
};
meta.reset(gguf_init_from_file(fname.c_str(), params));
if (!meta) {
throw std::runtime_error(format("%s: failed to load model from %s", __func__, fname.c_str()));
}
get_key(llm_kv(LLM_KV_GENERAL_ARCHITECTURE), arch_name, false);
llm_kv = LLM_KV(llm_arch_from_string(arch_name));
files.emplace_back(new llama_file(fname.c_str(), "rb", use_direct_io));
contexts.emplace_back(ctx);
use_direct_io = use_direct_io && files.back()->has_direct_io();
// Disable mmap in case Direct I/O is enabled and available
if (use_direct_io && use_mmap) {
use_mmap = false;
LLAMA_LOG_WARN("%s: direct I/O is enabled, disabling mmap\n", __func__);
}
// Save tensors data offset of the main file.
// For subsidiary files, `meta` tensor data offset must not be used,
// so we build a unified tensors index for weights.
for (ggml_tensor * cur = ggml_get_first_tensor(ctx); cur; cur = ggml_get_next_tensor(ctx, cur)) {
std::string tensor_name = std::string(cur->name);
// make sure there is no duplicated tensor names
if (weights_map.find(tensor_name) != weights_map.end()) {
throw std::runtime_error(format("invalid model: tensor '%s' is duplicated", ggml_get_name(cur)));
}
n_elements += ggml_nelements(cur);
n_bytes += ggml_nbytes(cur);
weights_map.emplace(tensor_name, llama_tensor_weight(files.back().get(), 0, meta.get(), cur));
}
uint16_t n_split = 0;
get_key(llm_kv(LLM_KV_SPLIT_COUNT), n_split, false);
// Load additional GGML contexts
if (n_split > 1) {
// make sure the main file is loaded first
uint16_t idx = 0;
const std::string kv_split_no = llm_kv(LLM_KV_SPLIT_NO);
get_key(kv_split_no, idx);
if (idx != 0) {
throw std::runtime_error(format("illegal split file idx: %d (file: %s), model must be loaded with the first split", idx, fname.c_str()));
metadata_ptr.reset(gguf_init_from_file(fname.c_str(), params));
metadata = metadata_ptr.get();
if (metadata == nullptr) {
throw std::runtime_error(format("%s: failed to load model from %s", __func__, fname.c_str()));
}
// generate list of splits if needed
if (splits.empty()) {
splits = llama_get_list_splits(fname, idx, n_split);
get_key(llm_kv(LLM_KV_GENERAL_ARCHITECTURE), arch_name, false);
llm_kv = LLM_KV(llm_arch_from_string(arch_name));
files.emplace_back(new llama_file(fname.c_str(), "rb", use_direct_io));
contexts.emplace_back(ctx);
if (use_mmap && use_direct_io) {
if (files.back()->has_direct_io()) {
LLAMA_LOG_WARN("%s: direct I/O is enabled, disabling mmap\n", __func__);
use_mmap = false;
} else {
LLAMA_LOG_WARN("%s: direct I/O is not available, using mmap\n", __func__);
use_direct_io = false;
// reopen file using std::fopen for mmap
files.pop_back();
files.emplace_back(new llama_file(fname.c_str(), "rb", false));
}
}
// in case user give a custom list of splits, check if it matches the expected number
if (n_split != (uint16_t)splits.size()) {
throw std::runtime_error(format("invalid split count, given: %zu splits, but expected %d", splits.size(), n_split));
// Save tensors data offset of the main file.
// For subsidiary files, `meta` tensor data offset must not be used,
// so we build a unified tensors index for weights.
for (ggml_tensor * cur = ggml_get_first_tensor(ctx); cur; cur = ggml_get_next_tensor(ctx, cur)) {
std::string tensor_name = std::string(cur->name);
// make sure there is no duplicated tensor names
if (weights_map.find(tensor_name) != weights_map.end()) {
throw std::runtime_error(format("invalid model: tensor '%s' is duplicated", ggml_get_name(cur)));
}
n_elements += ggml_nelements(cur);
n_bytes += ggml_nbytes(cur);
weights_map.emplace(tensor_name, llama_tensor_weight(files.back().get(), 0, metadata, cur));
}
uint16_t n_split = 0;
get_key(llm_kv(LLM_KV_SPLIT_COUNT), n_split, false);
if (trace > 0) {
LLAMA_LOG_INFO("%s: loading additional %d GGUFs\n", __func__, n_split);
}
// load other splits
for (idx = 1; idx < n_split; idx++) {
const char * fname_split = splits[idx].c_str();
struct gguf_init_params split_params = {
/*.no_alloc = */ true,
/*.ctx = */ &ctx,
};
gguf_context_ptr ctx_gguf { gguf_init_from_file(fname_split, split_params) };
if (!ctx_gguf) {
throw std::runtime_error(format("%s: failed to load GGUF split from %s", __func__, fname_split));
// Load additional GGML contexts
if (n_split > 1) {
// make sure the main file is loaded first
uint16_t idx = 0;
const std::string kv_split_no = llm_kv(LLM_KV_SPLIT_NO);
get_key(kv_split_no, idx);
if (idx != 0) {
throw std::runtime_error(format("illegal split file idx: %d (file: %s), model must be loaded with the first split", idx, fname.c_str()));
}
// check idx
// generate list of splits if needed
if (splits.empty()) {
splits = llama_get_list_splits(fname, idx, n_split);
}
// in case user give a custom list of splits, check if it matches the expected number
if (n_split != (uint16_t)splits.size()) {
throw std::runtime_error(format("invalid split count, given: %zu splits, but expected %d", splits.size(), n_split));
}
if (trace > 0) {
LLAMA_LOG_INFO("%s: loading additional %d GGUFs\n", __func__, n_split);
}
// load other splits
for (idx = 1; idx < n_split; idx++) {
const char * fname_split = splits[idx].c_str();
struct gguf_init_params split_params = {
/*.no_alloc = */ true,
/*.ctx = */ &ctx,
};
gguf_context_ptr ctx_gguf { gguf_init_from_file(fname_split, split_params) };
if (!ctx_gguf) {
throw std::runtime_error(format("%s: failed to load GGUF split from %s", __func__, fname_split));
}
// check idx
{
const int kid = gguf_find_key(ctx_gguf.get(), kv_split_no.c_str());
if (kid < 0) {
throw std::runtime_error(format("missing key %s in GGUF split %s", kv_split_no.c_str(), fname_split));
}
int idx_gguf = gguf_get_val_u16(ctx_gguf.get(), kid);
if (idx_gguf != idx) {
throw std::runtime_error(format("invalid split file idx: %d (file: %s), expected %d", idx_gguf, fname_split, idx));
}
}
files.emplace_back(new llama_file(fname_split, "rb", use_direct_io));
contexts.emplace_back(ctx);
// Save tensors data offset info of the shard.
for (ggml_tensor * cur = ggml_get_first_tensor(ctx); cur; cur = ggml_get_next_tensor(ctx, cur)) {
std::string tensor_name = std::string(cur->name);
// make sure there is no duplicated tensor names
if (weights_map.find(tensor_name) != weights_map.end()) {
throw std::runtime_error(format("invalid model: tensor '%s' is duplicated", ggml_get_name(cur)));
}
n_elements += ggml_nelements(cur);
n_bytes += ggml_nbytes(cur);
weights_map.emplace(tensor_name, llama_tensor_weight(files.back().get(), idx, ctx_gguf.get(), cur));
}
}
get_key(llm_kv(LLM_KV_SPLIT_TENSORS_COUNT), n_tensors);
// sanity check
{
const int kid = gguf_find_key(ctx_gguf.get(), kv_split_no.c_str());
if (kid < 0) {
throw std::runtime_error(format("missing key %s in GGUF split %s", kv_split_no.c_str(), fname_split));
}
int idx_gguf = gguf_get_val_u16(ctx_gguf.get(), kid);
if (idx_gguf != idx) {
throw std::runtime_error(format("invalid split file idx: %d (file: %s), expected %d", idx_gguf, fname_split, idx));
const int n_tensors_loaded = (int) weights_map.size();
if (n_tensors != n_tensors_loaded) {
throw std::runtime_error(format("corrupted model: %d tensors expected but %d found", n_tensors, n_tensors_loaded));
}
}
files.emplace_back(new llama_file(fname_split, "rb", use_direct_io));
contexts.emplace_back(ctx);
// Save tensors data offset info of the shard.
for (ggml_tensor * cur = ggml_get_first_tensor(ctx); cur; cur = ggml_get_next_tensor(ctx, cur)) {
std::string tensor_name = std::string(cur->name);
// make sure there is no duplicated tensor names
if (weights_map.find(tensor_name) != weights_map.end()) {
throw std::runtime_error(format("invalid model: tensor '%s' is duplicated", ggml_get_name(cur)));
}
n_elements += ggml_nelements(cur);
n_bytes += ggml_nbytes(cur);
weights_map.emplace(tensor_name, llama_tensor_weight(files.back().get(), idx, ctx_gguf.get(), cur));
}
LLAMA_LOG_INFO("%s: additional %d GGUFs metadata loaded.\n", __func__, n_split - 1);
}
get_key(llm_kv(LLM_KV_SPLIT_TENSORS_COUNT), n_tensors);
// sanity check
{
const int n_tensors_loaded = (int) weights_map.size();
if (n_tensors != n_tensors_loaded) {
throw std::runtime_error(format("corrupted model: %d tensors expected but %d found", n_tensors, n_tensors_loaded));
}
}
LLAMA_LOG_INFO("%s: additional %d GGUFs metadata loaded.\n", __func__, n_split - 1);
} else {
get_key(llm_kv(LLM_KV_GENERAL_ARCHITECTURE), arch_name, false);
llm_kv = LLM_KV(llm_arch_from_string(arch_name));
}
n_kv = gguf_get_n_kv(meta.get());
n_kv = gguf_get_n_kv(metadata);
n_tensors = weights_map.size();
fver = (enum llama_fver) gguf_get_version(meta.get());
fver = (enum llama_fver) gguf_get_version(metadata);
LLAMA_LOG_INFO("%s: loaded meta data with %d key-value pairs and %d tensors from %s (version %s)\n",
__func__, n_kv, n_tensors, fname.c_str(), llama_file_version_name(fver));
@ -695,6 +725,7 @@ llama_model_loader::llama_model_loader(
case GGML_TYPE_IQ4_NL: ftype = LLAMA_FTYPE_MOSTLY_IQ4_NL; break;
case GGML_TYPE_IQ4_XS: ftype = LLAMA_FTYPE_MOSTLY_IQ4_XS; break;
case GGML_TYPE_IQ3_S: ftype = LLAMA_FTYPE_MOSTLY_IQ3_S; break;
case GGML_TYPE_NVFP4: ftype = LLAMA_FTYPE_MOSTLY_NVFP4; break;
default:
{
LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max));
@ -715,14 +746,14 @@ llama_model_loader::llama_model_loader(
LLAMA_LOG_INFO("%s: Dumping metadata keys/values. Note: KV overrides do not apply in this output.\n", __func__);
for (int i = 0; i < n_kv; i++) {
const char * name = gguf_get_key(meta.get(), i);
const enum gguf_type type = gguf_get_kv_type(meta.get(), i);
const char * name = gguf_get_key(metadata, i);
const enum gguf_type type = gguf_get_kv_type(metadata, i);
const std::string type_name =
type == GGUF_TYPE_ARRAY
? format("%s[%s,%zu]", gguf_type_name(type), gguf_type_name(gguf_get_arr_type(meta.get(), i)), gguf_get_arr_n(meta.get(), i))
? format("%s[%s,%zu]", gguf_type_name(type), gguf_type_name(gguf_get_arr_type(metadata, i)), gguf_get_arr_n(metadata, i))
: gguf_type_name(type);
std::string value = gguf_kv_to_str(meta.get(), i);
std::string value = gguf_kv_to_str(metadata, i);
const size_t MAX_VALUE_LEN = 40;
if (value.size() > MAX_VALUE_LEN) {
value = format("%s...", value.substr(0, MAX_VALUE_LEN - 3).c_str());
@ -824,15 +855,382 @@ const struct ggml_tensor * llama_model_loader::check_tensor_dims(const std::stri
return cur;
}
struct ggml_tensor * llama_model_loader::create_tensor(struct ggml_context * ctx, const std::string & name, const std::initializer_list<int64_t> & ne, int flags) {
LLAMA_LOG_DEBUG("%s: loading tensor %s\n", __func__, name.c_str());
const struct ggml_tensor * cur = check_tensor_dims(name, ne, !(flags & TENSOR_NOT_REQUIRED));
// checks if the weight tensor can be used with the specified buffer type and device
static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w, ggml_op op, ggml_backend_buffer_type_t buft, ggml_backend_dev_t dev) {
GGML_ASSERT(w != nullptr);
if (op == GGML_OP_NONE) {
return true;
}
ggml_init_params params = {
/*.mem_size =*/ ggml_tensor_overhead()*8,
/*.mem_buffer =*/ NULL,
/*.no_alloc =*/ true,
};
ggml_context_ptr ctx_ptr { ggml_init(params) };
if (!ctx_ptr) {
throw std::runtime_error(format("failed to create ggml context"));
}
ggml_context * ctx = ctx_ptr.get();
ggml_tensor * op_tensor = nullptr;
switch (op) {
case GGML_OP_GET_ROWS:
{
ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 512);
op_tensor = ggml_get_rows(ctx, w, b);
} break;
case GGML_OP_MUL_MAT:
{
ggml_tensor * b = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, w->ne[0], 512, w->ne[2], w->ne[3]);
op_tensor = ggml_mul_mat(ctx, w, b);
} break;
case GGML_OP_MUL_MAT_ID:
{
const int n_expert_used = hparams.n_expert_used;
GGML_ASSERT(n_expert_used > 0);
ggml_tensor * b = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, w->ne[0], n_expert_used, 512);
ggml_tensor * ids = ggml_new_tensor_2d(ctx, GGML_TYPE_I32, n_expert_used, 512);
op_tensor = ggml_mul_mat_id(ctx, w, b, ids);
} break;
case GGML_OP_ADD:
{
ggml_tensor * a = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, w->ne[0], w->ne[1], w->ne[2], w->ne[3]);
op_tensor = ggml_add(ctx, a, w);
} break;
case GGML_OP_ADD_ID:
{
const int n_expert_used = hparams.n_expert_used;
GGML_ASSERT(n_expert_used > 0);
ggml_tensor * a = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, w->ne[0], n_expert_used, 512);
ggml_tensor * c = ggml_new_tensor_2d(ctx, GGML_TYPE_I32, n_expert_used, 512);
op_tensor = ggml_add_id(ctx, a, w, c);
} break;
case GGML_OP_MUL:
{
ggml_tensor * a = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, w->ne[0], w->ne[1], w->ne[2], w->ne[3]);
op_tensor = ggml_mul(ctx, a, w);
} break;
case GGML_OP_DIV:
{
ggml_tensor * a = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, w->ne[0]);
op_tensor = ggml_div(ctx, a, w);
} break;
case GGML_OP_ROPE:
{
const int n_embd_head = hparams.n_embd_head_v();
const int n_head = hparams.n_head();
ggml_tensor * a = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, n_embd_head, n_head, 512);
ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 512);
op_tensor = ggml_rope_ext(
ctx, a, b, w,
0, 0, 0, 0, 0,
0, 0, 0, 0
);
} break;
case GGML_OP_SSM_CONV:
{
const int64_t n_seq_tokens = 512;
const int64_t n_seqs = 3;
ggml_tensor * conv_x = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, w->ne[0] - 1 + n_seq_tokens, w->ne[1], n_seqs);
op_tensor = ggml_ssm_conv(ctx, conv_x, w);
} break;
case GGML_OP_SSM_SCAN:
{
// w is ssm_a, which is used to distinguish Mamba-1 and Mamba-2
const int64_t d_state = w->ne[0] == 1 ? hparams.ssm_d_state : w->ne[0];
const int64_t n_head = w->ne[1];
const int64_t head_dim = hparams.ssm_d_inner / n_head;
const int64_t n_group = hparams.ssm_n_group ? hparams.ssm_n_group : 1;
const int64_t n_seq_tokens = 512;
const int64_t n_seqs = 3;
ggml_tensor * s = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, d_state, head_dim, n_head, n_seqs);
ggml_tensor * x = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, head_dim, n_head, n_seq_tokens, n_seqs);
ggml_tensor * dt = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, n_head, n_seq_tokens, n_seqs);
ggml_tensor * B = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, d_state, n_group, n_seq_tokens, n_seqs);
ggml_tensor * C = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, d_state, n_group, n_seq_tokens, n_seqs);
ggml_tensor * ids = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, n_seqs);
op_tensor = ggml_ssm_scan(ctx, s, x, dt, w, B, C, ids);
} break;
case GGML_OP_RWKV_WKV6:
{
// FIXME
const int64_t S = 123;
const int64_t H = 123;
const int64_t n_tokens = 123;
const int64_t n_seqs = 123;
ggml_tensor * k = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, S, H, n_tokens);
ggml_tensor * v = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, S, H, n_tokens);
ggml_tensor * r = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, S, H, n_tokens);
ggml_tensor * tf = w;
ggml_tensor * td = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, S, H, n_tokens);
ggml_tensor * state = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, S, n_seqs, S, H);
op_tensor = ggml_rwkv_wkv6(ctx, k, v, r, tf, td, state);
} break;
case GGML_OP_IM2COL:
{
const int n_embd_inp = hparams.n_embd_inp();
ggml_tensor * b = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, n_embd_inp, w->ne[1], 1, 1);
op_tensor = ggml_im2col(ctx, w, b, 1, 0, 0, 0, 1, 0, false, GGML_TYPE_F16);
} break;
case GGML_OP_SCALE:
{
op_tensor = ggml_scale(ctx, w, 1.0f);
} break;
default:
GGML_ABORT("%s: missing test for op %s for tensor %s", __func__, ggml_op_name(op), w->name);
}
// create a temporary dummy buffer for the weight so that supports_op can check the buffer type
GGML_ASSERT(w->buffer == nullptr);
w->buffer = ggml_backend_buft_alloc_buffer(buft, 0);
bool op_supported = ggml_backend_dev_supports_op(dev, op_tensor);
ggml_backend_buffer_free(w->buffer);
w->buffer = nullptr;
return op_supported;
}
// find the first buffer type in the list that can use the tensor
static ggml_backend_buffer_type_t select_weight_buft(const llama_hparams & hparams, ggml_tensor * tensor, ggml_op op, const buft_list_t * buft_list) {
GGML_ASSERT(!buft_list->empty());
for (const auto & cur : *buft_list) {
ggml_backend_dev_t cur_dev = cur.first;
ggml_backend_buffer_type_t cur_buft = cur.second;
if (weight_buft_supported(hparams, tensor, op, cur_buft, cur_dev)) {
return cur_buft;
}
}
return nullptr;
}
struct ggml_tensor * llama_model_loader::create_tensor(
const llama_hparams & hparams, const buft_list_t * buft_list_cpu, const buft_list_t * buft_list_input, const buft_list_t * buft_list_output,
const buft_list_t * buft_list_layer, const LLM_TN_IMPL & tn, const std::initializer_list<int64_t> & ne, int flags) {
auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * {
auto it = ctx_map.find(buft);
if (it == ctx_map.end()) {
// one ggml context per buffer type
int max_n_tensors = n_tensors;
max_n_tensors += 1; // duplicated output tensor
max_n_tensors += hparams.n_layer*2; // duplicated rope freq tensors
if (files.empty()) {
max_n_tensors += hparams.n_layer*256; // this should be well above what any model actually uses
}
const size_t ctx_size = ggml_tensor_overhead()*max_n_tensors;
ggml_init_params params = {
/*.mem_size =*/ ctx_size,
/*.mem_buffer =*/ NULL,
/*.no_alloc =*/ true,
};
ggml_context * ctx = ggml_init(params);
if (!ctx) {
throw std::runtime_error(format("failed to create ggml context"));
}
ctx_map.emplace(buft, ctx);
return ctx;
}
return it->second.get();
};
auto buft_for_tensor = [&](ggml_tensor * t_meta) -> ggml_backend_buffer_type_t {
if (!t_meta) {
if (flags & TENSOR_NOT_REQUIRED) {
return nullptr;
}
throw std::runtime_error(format("missing tensor '%s'", tn.str().c_str()));
}
// some models use the token embedding tensor as the output, but since these are used in different layers and with different ops
// the tensor is duplicated
// to handle this, we check if the tensor is duplicated, and if so, we assume that it is being loaded as the output tensor
llm_tensor tn_tensor = tn.tensor;
if (tn.tensor == LLM_TENSOR_TOKEN_EMBD && (flags & TENSOR_DUPLICATED)) {
tn_tensor = LLM_TENSOR_OUTPUT;
}
llm_tensor_info info;
try {
info = llm_tensor_info_for(tn_tensor);
} catch (const std::out_of_range & e) {
throw std::runtime_error(format("missing tensor info mapping for %s", tn.str().c_str()));
}
// skip unused tensors
if (info.op == GGML_OP_NONE || (flags & TENSOR_SKIP)) {
const size_t nbytes = ggml_nbytes(t_meta);
LLAMA_LOG_WARN("model has unused tensor %s (size = %zu bytes) -- ignoring\n", tn.str().c_str(), nbytes);
size_data -= nbytes;
n_created++;
return nullptr;
}
// tensors with "bias" suffix are always used with GGML_OP_ADD or GGML_OP_ADD_ID
ggml_op op;
bool bias = tn.suffix != nullptr && strcmp(tn.suffix, "bias") == 0;
if (bias) {
if (info.op == GGML_OP_MUL_MAT_ID) {
op = GGML_OP_ADD_ID;
} else {
op = GGML_OP_ADD;
}
} else {
op = info.op;
}
// sanity checks
if (info.layer == LLM_TENSOR_LAYER_INPUT || info.layer == LLM_TENSOR_LAYER_OUTPUT) {
if (tn.bid != -1) {
GGML_ABORT("input/output layer tensor %s used with a layer number", tn.str().c_str());
}
} else {
if (tn.bid == -1) {
GGML_ABORT("repeating layer tensor %s used without a layer number", tn.str().c_str());
}
}
// select the buffer type for this tensor
const buft_list_t * buft_list;
switch (info.layer) {
case LLM_TENSOR_LAYER_INPUT:
buft_list = buft_list_input;
break;
case LLM_TENSOR_LAYER_OUTPUT:
buft_list = buft_list_output;
break;
case LLM_TENSOR_LAYER_REPEATING:
GGML_ASSERT(buft_list_layer != nullptr);
buft_list = buft_list_layer;
break;
default:
GGML_ABORT("invalid layer %d for tensor %s", info.layer, tn.str().c_str());
}
ggml_backend_buffer_type_t buft = nullptr;
// check overrides
if (tensor_buft_overrides) {
std::string tensor_name = tn.str();
for (const auto * overrides = tensor_buft_overrides; overrides->pattern != nullptr; ++overrides) {
std::regex pattern(overrides->pattern);
if (std::regex_search(tensor_name, pattern)) {
if (overrides->buft == ggml_backend_cpu_buffer_type()) {
// when overriding to a CPU buffer, consider the extra buffer types
buft = select_weight_buft(hparams, t_meta, op, buft_list_cpu);
} else {
buft = overrides->buft;
}
LLAMA_LOG_DEBUG("tensor %s (%zu MiB %s) buffer type overridden to %s\n",
tensor_name.c_str(),
ggml_nbytes(t_meta) / 1024 / 1024, ggml_type_name(t_meta->type),
ggml_backend_buft_name(buft));
break;
}
}
}
if (!buft) {
buft = select_weight_buft(hparams, t_meta, op, buft_list);
if (!buft) {
throw std::runtime_error(format("failed to find a compatible buffer type for tensor %s", tn.str().c_str()));
}
}
// avoid using a host buffer when using mmap
auto * buft_dev = ggml_backend_buft_get_device(buft);
if (use_mmap && buft_dev && buft == ggml_backend_dev_host_buffer_type(buft_dev)) {
auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
if (!cpu_dev) {
throw std::runtime_error("no CPU backend found");
}
buft = ggml_backend_dev_buffer_type(cpu_dev);
}
if (buft != buft_list->front().second) {
if (n_tensors_moved == 0) {
first_tensor_moved_name = t_meta->name;
first_tensor_moved_type_name = ggml_type_name(t_meta->type);
first_moved_from_buft = buft_list->front().second;
first_moved_to_buft = buft;
}
n_tensors_moved++;
}
return buft;
};
if (files.empty()) {
if (flags & TENSOR_SKIP_IF_VIRTUAL) {
return nullptr;
}
ggml_type type = GGML_TYPE_F32;
const int64_t tid = gguf_find_tensor(metadata, tn.str().c_str());
if (tid != -1) {
type = gguf_get_tensor_type(metadata, tid);
}
// for tensors that are not required some of the dimensions can be invalid:
if (flags & TENSOR_NOT_REQUIRED) {
for (size_t dim = 0; dim < ne.size(); dim++) {
if (ne.begin()[dim] <= 0) {
return nullptr;
}
}
}
ggml_tensor t_meta;
memset(&t_meta, 0, sizeof(ggml_tensor));
t_meta.type = type;
for (size_t dim = 0; dim < GGML_MAX_DIMS; dim++) {
t_meta.ne[dim] = dim < ne.size() ? ne.begin()[dim] : 1;
GGML_ASSERT(t_meta.ne[dim] >= 1);
t_meta.nb[dim] = dim == 0 ? ggml_type_size(type) : t_meta.ne[dim-1]*t_meta.nb[dim-1];
GGML_ASSERT(t_meta.nb[dim] >= 1);
}
ggml_set_name(&t_meta, tn.str().c_str());
ggml_backend_buffer_type_t buft = buft_for_tensor(&t_meta);
GGML_ASSERT(buft != nullptr);
ggml_context * ctx = ctx_for_buft(buft);
ggml_tensor * ret = ggml_dup_tensor(ctx, &t_meta);
ggml_set_name(ret, tn.str().c_str());
return ret;
}
ggml_tensor * t_meta = get_tensor_meta(tn.str().c_str());
ggml_backend_buffer_type_t buft = buft_for_tensor(t_meta);
if (buft == nullptr) {
return nullptr; // return type is ggml_tensor *
}
ggml_context * ctx = ctx_for_buft(buft);
// if duplicated, check if the original tensor was allocated in the same buffer type context and avoid creating a new one
if (flags & TENSOR_DUPLICATED) {
ggml_tensor * t = ggml_get_tensor(ctx, tn.str().c_str());
if (t) {
return t;
}
}
LLAMA_LOG_DEBUG("%s: loading tensor %s\n", __func__, tn.str().c_str());
const struct ggml_tensor * cur = check_tensor_dims(tn.str(), ne, !(flags & TENSOR_NOT_REQUIRED));
if (cur == NULL) {
return NULL;
}
bool duplicated = flags & TENSOR_DUPLICATED;
const bool duplicated = flags & TENSOR_DUPLICATED;
struct ggml_tensor * tensor = ggml_dup_tensor(ctx, cur);
ggml_set_name(tensor, ggml_get_name(cur));
@ -844,7 +1242,6 @@ struct ggml_tensor * llama_model_loader::create_tensor(struct ggml_context * ctx
}
return tensor;
}
struct ggml_tensor * llama_model_loader::create_tensor_as_view(struct ggml_context * ctx, struct ggml_tensor * base, const std::string & name, const std::initializer_list<int64_t> & ne, size_t offset, bool required) {
@ -879,6 +1276,11 @@ void llama_model_loader::done_getting_tensors() const {
if (n_created != n_tensors) {
throw std::runtime_error(format("%s: wrong number of tensors; expected %d, got %d", __func__, n_tensors, n_created));
}
if (n_tensors_moved > 0) {
LLAMA_LOG_DEBUG("%s: tensor '%s' (%s) (and %zu others) cannot be used with preferred buffer type %s, using %s instead\n",
__func__, first_tensor_moved_name.c_str(), first_tensor_moved_type_name.c_str(), n_tensors_moved - 1,
ggml_backend_buft_name(first_moved_from_buft), ggml_backend_buft_name(first_moved_to_buft));
}
}
void llama_model_loader::init_mappings(bool prefetch, llama_mlocks * mlock_mmaps) {
@ -960,6 +1362,12 @@ bool llama_model_loader::load_all_data(
llama_mlocks * lmlocks,
llama_progress_callback progress_callback,
void * progress_callback_user_data) {
if (files.empty()) {
for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != nullptr; t = ggml_get_next_tensor(ctx, t)) {
set_tensor_data(t, set_tensor_data_ud);
}
return true;
}
GGML_ASSERT(size_data != 0 && "call init_mappings() first");
std::vector<no_init<uint8_t>> read_buf;

40
examples/talk-llama/llama-model-loader.h
View File

@ -4,17 +4,22 @@
#include "llama-impl.h"
#include "llama-arch.h"
#include "llama-hparams.h"
#include "llama-mmap.h"
#include "ggml-cpp.h"
#include <cstddef>
#include <cstring>
#include <map>
#include <stdexcept>
#include <unordered_map>
using llama_buf_map = std::unordered_map<uint32_t, ggml_backend_buffer_t>;
// lists of buffer types used for each layer
using buft_list_t = std::vector<std::pair<ggml_backend_dev_t, ggml_backend_buffer_type_t>>;
enum llama_fver {
GGUF_FILE_VERSION_V1 = 1,
GGUF_FILE_VERSION_V2 = 2,
@ -58,9 +63,10 @@ struct llama_model_loader {
}
};
static const int TENSOR_NOT_REQUIRED = 1 << 0;
static const int TENSOR_DUPLICATED = 1 << 1;
static const int TENSOR_SKIP = 1 << 2;
static const int TENSOR_NOT_REQUIRED = 1 << 0;
static const int TENSOR_DUPLICATED = 1 << 1;
static const int TENSOR_SKIP = 1 << 2;
static const int TENSOR_SKIP_IF_VIRTUAL = 1 << 3;
int n_kv = 0;
int n_tensors = 0;
@ -84,7 +90,10 @@ struct llama_model_loader {
std::unordered_map<std::string, llama_model_kv_override> kv_overrides;
const llama_model_tensor_buft_override * tensor_buft_overrides;
gguf_context_ptr meta;
gguf_context_ptr metadata_ptr;
struct gguf_context * metadata; // either metadata_ptr.get() or externally set
llama_model_set_tensor_data_t set_tensor_data;
void * set_tensor_data_ud;
std::vector<ggml_context_ptr> contexts;
std::string arch_name;
@ -94,7 +103,26 @@ struct llama_model_loader {
size_t size_data = 0;
std::vector<std::pair<size_t, size_t>> mmaps_used;
// define a comparator for the buft -> ctx map to ensure that the order is well-defined:
struct ggml_backend_buft_comparator {
bool operator()(const ggml_backend_buffer_type_t & lhs, const ggml_backend_buffer_type_t & rhs) const {
return strcmp(ggml_backend_buft_name(lhs), ggml_backend_buft_name(rhs)) < 0;
}
};
std::map<ggml_backend_buffer_type_t, ggml_context_ptr, ggml_backend_buft_comparator> ctx_map;
// track tensors that had to be moved for debugging:
size_t n_tensors_moved = 0;
std::string first_tensor_moved_name;
std::string first_tensor_moved_type_name;
ggml_backend_buffer_type_t first_moved_from_buft = nullptr;
ggml_backend_buffer_type_t first_moved_to_buft = nullptr;
llama_model_loader(
struct gguf_context * metadata,
llama_model_set_tensor_data_t set_tensor_data,
void * set_tensor_data_ud,
const std::string & fname,
std::vector<std::string> & splits, // optional, only need if the split does not follow naming scheme
bool use_mmap,
@ -149,7 +177,9 @@ struct llama_model_loader {
const struct ggml_tensor * check_tensor_dims(const std::string & name, const std::vector<int64_t> & ne, bool required) const;
struct ggml_tensor * create_tensor(struct ggml_context * ctx, const std::string & name, const std::initializer_list<int64_t> & ne, int flags = 0);
struct ggml_tensor * create_tensor(
const llama_hparams & hparams, const buft_list_t * buft_list_cpu, const buft_list_t * buft_list_input, const buft_list_t * buft_list_output,
const buft_list_t * buft_list_layer, const LLM_TN_IMPL & tn, const std::initializer_list<int64_t> & ne, int flags);
struct ggml_tensor * create_tensor_as_view(struct ggml_context * ctx, struct ggml_tensor * base, const std::string & name, const std::initializer_list<int64_t> & ne, size_t offset, bool required = true);

100
examples/talk-llama/llama-model-saver.cpp
View File

@ -7,14 +7,19 @@
#include "llama-model.h"
#include "llama-vocab.h"
#include <cstdint>
#include <string>
llama_model_saver::llama_model_saver(const struct llama_model & model) : model(model), llm_kv(model.arch) {
gguf_ctx = gguf_init_empty();
}
llama_model_saver::llama_model_saver(const struct llama_model * model) :
gguf_ctx(gguf_init_empty()), gguf_ctx_owned(true), model(model), llm_kv(model->arch) {}
llama_model_saver::llama_model_saver(enum llm_arch arch, struct gguf_context * gguf_ctx) :
gguf_ctx(gguf_ctx == nullptr ? gguf_init_empty() : gguf_ctx), gguf_ctx_owned(gguf_ctx == nullptr), model(nullptr), llm_kv(arch) {}
llama_model_saver::~llama_model_saver() {
gguf_free(gguf_ctx);
if (gguf_ctx_owned) {
gguf_free(gguf_ctx);
}
}
void llama_model_saver::add_kv(const enum llm_kv key, const uint32_t value) {
@ -46,7 +51,8 @@ void llama_model_saver::add_kv(const enum llm_kv key, const char value) {
template <typename Container>
void llama_model_saver::add_kv(const enum llm_kv key, const Container & value, const bool per_layer) {
const size_t n_values = per_layer ? size_t(model.hparams.n_layer) : value.size();
GGML_ASSERT(model != nullptr || !per_layer);
const size_t n_values = per_layer ? size_t(model->hparams.n_layer) : value.size();
GGML_ASSERT(n_values <= value.size());
if (n_values == 0) {
@ -83,6 +89,8 @@ void llama_model_saver::add_kv(const enum llm_kv key, const Container & value, c
GGML_ABORT("fatal error");
}
}
// instantiate for external usage:
template void llama_model_saver::add_kv<std::vector<uint32_t>>(const enum llm_kv, const std::vector<uint32_t> &, const bool);
void llama_model_saver::add_kv(const enum llm_kv key, const std::vector<std::string> & value) {
std::vector<const char *> tmp(value.size());
@ -104,37 +112,39 @@ void llama_model_saver::add_tensor(const struct ggml_tensor * tensor) {
}
void llama_model_saver::add_kv_from_model() {
const llama_hparams & hparams = model.hparams;
const llama_vocab & vocab = model.vocab;
const llama_hparams & hparams = model->hparams;
const llama_vocab & vocab = model->vocab;
const int32_t n_vocab = vocab.n_tokens();
std::vector<std::string> tokens(n_vocab);
std::vector<float> scores(n_vocab);
std::vector<int32_t> token_types(n_vocab);
for (int32_t id = 0; id < n_vocab; ++id) {
const llama_vocab::token_data & token_data = vocab.get_token_data(id);
if (vocab.get_type() != LLAMA_VOCAB_TYPE_NONE) {
for (int32_t id = 0; id < n_vocab; ++id) {
const llama_vocab::token_data & token_data = vocab.get_token_data(id);
tokens[id] = token_data.text;
scores[id] = token_data.score;
tokens[id] = token_data.text;
scores[id] = token_data.score;
switch(token_data.attr) {
case LLAMA_TOKEN_ATTR_UNKNOWN: token_types[id] = LLAMA_TOKEN_TYPE_UNKNOWN; break;
case LLAMA_TOKEN_ATTR_UNUSED: token_types[id] = LLAMA_TOKEN_TYPE_UNUSED; break;
case LLAMA_TOKEN_ATTR_NORMAL: token_types[id] = LLAMA_TOKEN_TYPE_NORMAL; break;
case LLAMA_TOKEN_ATTR_CONTROL: token_types[id] = LLAMA_TOKEN_TYPE_CONTROL; break;
case LLAMA_TOKEN_ATTR_USER_DEFINED: token_types[id] = LLAMA_TOKEN_TYPE_USER_DEFINED; break;
case LLAMA_TOKEN_ATTR_BYTE: token_types[id] = LLAMA_TOKEN_TYPE_BYTE; break;
case LLAMA_TOKEN_ATTR_UNDEFINED:
default: token_types[id] = LLAMA_TOKEN_TYPE_UNDEFINED; break;
switch(token_data.attr) {
case LLAMA_TOKEN_ATTR_UNKNOWN: token_types[id] = LLAMA_TOKEN_TYPE_UNKNOWN; break;
case LLAMA_TOKEN_ATTR_UNUSED: token_types[id] = LLAMA_TOKEN_TYPE_UNUSED; break;
case LLAMA_TOKEN_ATTR_NORMAL: token_types[id] = LLAMA_TOKEN_TYPE_NORMAL; break;
case LLAMA_TOKEN_ATTR_CONTROL: token_types[id] = LLAMA_TOKEN_TYPE_CONTROL; break;
case LLAMA_TOKEN_ATTR_USER_DEFINED: token_types[id] = LLAMA_TOKEN_TYPE_USER_DEFINED; break;
case LLAMA_TOKEN_ATTR_BYTE: token_types[id] = LLAMA_TOKEN_TYPE_BYTE; break;
case LLAMA_TOKEN_ATTR_UNDEFINED:
default: token_types[id] = LLAMA_TOKEN_TYPE_UNDEFINED; break;
}
}
}
// add_kv(LLM_KV_GENERAL_TYPE, ???);
add_kv(LLM_KV_GENERAL_ARCHITECTURE, model.arch_name());
add_kv(LLM_KV_GENERAL_ARCHITECTURE, model->arch_name());
// add_kv(LLM_KV_GENERAL_QUANTIZATION_VERSION, ???);
// add_kv(LLM_KV_GENERAL_ALIGNMENT, ???);
add_kv(LLM_KV_GENERAL_NAME, model.name);
add_kv(LLM_KV_GENERAL_NAME, model->name);
// add_kv(LLM_KV_GENERAL_AUTHOR, ???);
// add_kv(LLM_KV_GENERAL_VERSION, ???);
// add_kv(LLM_KV_GENERAL_URL, ???);
@ -146,8 +156,8 @@ void llama_model_saver::add_kv_from_model() {
add_kv(LLM_KV_VOCAB_SIZE, vocab.n_tokens());
add_kv(LLM_KV_CONTEXT_LENGTH, hparams.n_ctx_train);
add_kv(LLM_KV_EMBEDDING_LENGTH, hparams.n_embd);
if (hparams.n_embd_out > 0) {
add_kv(LLM_KV_EMBEDDING_LENGTH_OUT, hparams.n_embd_out);
if (hparams.n_embd_out_impl > 0) {
add_kv(LLM_KV_EMBEDDING_LENGTH_OUT, hparams.n_embd_out_impl);
}
add_kv(LLM_KV_BLOCK_COUNT, hparams.n_layer);
add_kv(LLM_KV_LEADING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead);
@ -176,8 +186,10 @@ void llama_model_saver::add_kv_from_model() {
add_kv(LLM_KV_ATTENTION_HEAD_COUNT_KV, hparams.n_head_kv_arr, true);
add_kv(LLM_KV_ATTENTION_MAX_ALIBI_BIAS, hparams.f_max_alibi_bias);
add_kv(LLM_KV_ATTENTION_CLAMP_KQV, hparams.f_clamp_kqv);
add_kv(LLM_KV_ATTENTION_KEY_LENGTH, hparams.n_embd_head_k);
add_kv(LLM_KV_ATTENTION_VALUE_LENGTH, hparams.n_embd_head_v);
add_kv(LLM_KV_ATTENTION_KEY_LENGTH, hparams.n_embd_head_k_full);
add_kv(LLM_KV_ATTENTION_VALUE_LENGTH, hparams.n_embd_head_v_full);
add_kv(LLM_KV_ATTENTION_KEY_LENGTH_SWA, hparams.n_embd_head_k_swa);
add_kv(LLM_KV_ATTENTION_VALUE_LENGTH_SWA, hparams.n_embd_head_v_swa);
add_kv(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
add_kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
add_kv(LLM_KV_ATTENTION_CAUSAL, hparams.causal_attn);
@ -189,7 +201,8 @@ void llama_model_saver::add_kv_from_model() {
const float rope_scaling_factor = hparams.rope_freq_scale_train == 1.0f ? 0.0f : 1.0f/hparams.rope_freq_scale_train;
add_kv(LLM_KV_ROPE_DIMENSION_COUNT, hparams.n_rot);
add_kv(LLM_KV_ROPE_DIMENSION_COUNT, hparams.n_rot_full);
add_kv(LLM_KV_ROPE_DIMENSION_COUNT_SWA, hparams.n_rot_swa);
add_kv(LLM_KV_ROPE_FREQ_BASE, hparams.rope_freq_base_train);
// add_kv(LLM_KV_ROPE_SCALE_LINEAR, rope_scaling_factor); // old name
add_kv(LLM_KV_ROPE_SCALING_TYPE, llama_rope_scaling_type_name(hparams.rope_scaling_type_train));
@ -255,24 +268,25 @@ void llama_model_saver::add_kv_from_model() {
}
void llama_model_saver::add_tensors_from_model() {
if (std::string(model.output->name) != std::string(model.tok_embd->name)) {
add_tensor(model.tok_embd); // some models use the same tensor for tok_embd and output
if (std::string(model->output->name) != std::string(model->tok_embd->name)) {
add_tensor(model->tok_embd); // some models use the same tensor for tok_embd and output
}
add_tensor(model.type_embd);
add_tensor(model.pos_embd);
add_tensor(model.tok_norm);
add_tensor(model.tok_norm_b);
add_tensor(model.output_norm);
add_tensor(model.output_norm_b);
add_tensor(model.output);
add_tensor(model.output_b);
add_tensor(model.output_norm_enc);
add_tensor(model.cls);
add_tensor(model.cls_b);
add_tensor(model.cls_out);
add_tensor(model.cls_out_b);
add_tensor(model->type_embd);
add_tensor(model->pos_embd);
add_tensor(model->tok_norm);
add_tensor(model->tok_norm_b);
add_tensor(model->output_norm);
add_tensor(model->output_norm_b);
add_tensor(model->output);
add_tensor(model->output_b);
add_tensor(model->output_norm_enc);
add_tensor(model->cls);
add_tensor(model->cls_b);
add_tensor(model->cls_out);
add_tensor(model->cls_out_b);
add_tensor(model->cls_norm);
for (const struct llama_layer & layer : model.layers) {
for (const struct llama_layer & layer : model->layers) {
for (size_t i = 0; i < sizeof(layer)/sizeof(struct ggml_tensor *); ++i) {
add_tensor(reinterpret_cast<const struct ggml_tensor * const *>(&layer)[i]);
}

7
examples/talk-llama/llama-model-saver.h
View File

@ -1,5 +1,6 @@
#pragma once
#include "gguf.h"
#include "llama.h"
#include "llama-arch.h"
@ -7,10 +8,12 @@
struct llama_model_saver {
struct gguf_context * gguf_ctx = nullptr;
const struct llama_model & model;
const bool gguf_ctx_owned;
const struct llama_model * model;
const struct LLM_KV llm_kv;
llama_model_saver(const struct llama_model & model);
llama_model_saver(const struct llama_model * model);
llama_model_saver(enum llm_arch arch, struct gguf_context * gguf_ctx);
~llama_model_saver();
void add_kv(enum llm_kv key, uint32_t value);

1950
examples/talk-llama/llama-model.cpp
View File

File diff suppressed because it is too large Load Diff

91
examples/talk-llama/llama-model.h
View File

@ -11,6 +11,7 @@
#include <memory>
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <vector>
struct llama_cparams;
@ -53,6 +54,7 @@ enum llm_type {
LLM_TYPE_0_3B,
LLM_TYPE_0_5B,
LLM_TYPE_0_6B,
LLM_TYPE_0_8B,
LLM_TYPE_1B,
LLM_TYPE_1_2B,
LLM_TYPE_1_3B,
@ -115,17 +117,25 @@ enum llm_type {
LLM_TYPE_8B_A1B, // lfm2moe
LLM_TYPE_16B_A1B,
LLM_TYPE_21B_A3B, // Ernie MoE small
LLM_TYPE_24B_A2B, // lfm2moe
LLM_TYPE_30B_A3B,
LLM_TYPE_31B_A3_5B,
LLM_TYPE_35B_A3B, // Qwen3.5
LLM_TYPE_48B_A3B, // Kimi Linear
LLM_TYPE_80B_A3B, // Qwen3 Next
LLM_TYPE_100B_A6B,
LLM_TYPE_102B_A12B, // Solar-Open
LLM_TYPE_106B_A12B, // GLM-4.5-Air
LLM_TYPE_120B_A12B, // Nemotron 3 Super
LLM_TYPE_122B_A10B, // Qwen3.5
LLM_TYPE_196B_A11B, // Step3.5-Flash
LLM_TYPE_230B_A10B, // Minimax M2
LLM_TYPE_235B_A22B,
LLM_TYPE_300B_A47B, // Ernie MoE big
LLM_TYPE_310B_A15B, // /MiMo-V2-Flash
LLM_TYPE_355B_A32B, // GLM-4.5
LLM_TYPE_397B_A17B, // Qwen3.5
LLM_TYPE_744B_A40B, // GLM-5
LLM_TYPE_E2B,
LLM_TYPE_E4B,
};
@ -274,14 +284,25 @@ struct llama_layer {
struct ggml_tensor * ffn_up_enc = nullptr;
// ff MoE
struct ggml_tensor * ffn_gate_inp = nullptr;
struct ggml_tensor * ffn_gate_exps = nullptr;
struct ggml_tensor * ffn_down_exps = nullptr;
struct ggml_tensor * ffn_up_exps = nullptr;
struct ggml_tensor * ffn_gate_inp_b = nullptr;
struct ggml_tensor * ffn_gate_exps_b = nullptr;
struct ggml_tensor * ffn_down_exps_b = nullptr;
struct ggml_tensor * ffn_up_exps_b = nullptr;
struct ggml_tensor * ffn_gate_inp = nullptr;
struct ggml_tensor * ffn_gate_exps = nullptr;
struct ggml_tensor * ffn_down_exps = nullptr;
struct ggml_tensor * ffn_up_exps = nullptr;
struct ggml_tensor * ffn_gate_up_exps = nullptr;
struct ggml_tensor * ffn_gate_inp_b = nullptr;
struct ggml_tensor * ffn_gate_exps_b = nullptr;
struct ggml_tensor * ffn_down_exps_b = nullptr;
struct ggml_tensor * ffn_up_exps_b = nullptr;
struct ggml_tensor * ffn_gate_up_exps_b = nullptr;
// ff MoE per-expert scales (NVFP4 per-tensor scale2)
struct ggml_tensor * ffn_gate_exps_s = nullptr;
struct ggml_tensor * ffn_down_exps_s = nullptr;
struct ggml_tensor * ffn_up_exps_s = nullptr;
// ff MoE latent proj
struct ggml_tensor * ffn_latent_down = nullptr;
struct ggml_tensor * ffn_latent_up = nullptr;
// ff shared expert (shexp)
struct ggml_tensor * ffn_gate_inp_shexp = nullptr;
@ -319,6 +340,9 @@ struct llama_layer {
// qwen3next
struct ggml_tensor * ssm_beta_alpha = nullptr;
// qwen3.5
struct ggml_tensor * ssm_alpha = nullptr;
// rwkv
struct ggml_tensor * time_mix_w1 = nullptr;
struct ggml_tensor * time_mix_w2 = nullptr;
@ -373,13 +397,21 @@ struct llama_layer {
struct ggml_tensor * rope_freqs = nullptr;
// bitnet scale
struct ggml_tensor * wq_scale = nullptr;
struct ggml_tensor * wk_scale = nullptr;
struct ggml_tensor * wv_scale = nullptr;
struct ggml_tensor * wo_scale = nullptr;
struct ggml_tensor * ffn_gate_scale = nullptr;
struct ggml_tensor * ffn_up_scale = nullptr;
struct ggml_tensor * ffn_down_scale = nullptr;
struct ggml_tensor * wq_s = nullptr;
struct ggml_tensor * wk_s = nullptr;
struct ggml_tensor * wv_s = nullptr;
struct ggml_tensor * wo_s = nullptr;
struct ggml_tensor * wqkv_s = nullptr;
struct ggml_tensor * wqkv_gate_s = nullptr;
struct ggml_tensor * ffn_gate_s = nullptr;
struct ggml_tensor * ffn_up_s = nullptr;
struct ggml_tensor * ffn_down_s = nullptr;
struct ggml_tensor * ffn_gate_shexp_s = nullptr;
struct ggml_tensor * ffn_up_shexp_s = nullptr;
struct ggml_tensor * ffn_down_shexp_s = nullptr;
struct ggml_tensor * ssm_out_s = nullptr;
struct ggml_tensor * ssm_alpha_s = nullptr;
struct ggml_tensor * ssm_beta_s = nullptr;
// altup & laurel
struct ggml_tensor * per_layer_inp_gate = nullptr;
@ -410,6 +442,25 @@ struct llama_layer {
struct ggml_tensor * ffn_act_beta = nullptr;
struct ggml_tensor * ffn_act_eps = nullptr;
// Kimi Linear KDA (using ssm_ prefix for consistency)
// Note: ssm_dt_b already exists above (mamba bias), reused for Kimi dt_bias
struct ggml_tensor * ssm_q_conv = nullptr;
struct ggml_tensor * ssm_k_conv = nullptr;
struct ggml_tensor * ssm_v_conv = nullptr;
struct ggml_tensor * ssm_f_a = nullptr;
struct ggml_tensor * ssm_f_b = nullptr;
struct ggml_tensor * ssm_beta = nullptr;
struct ggml_tensor * ssm_g_a = nullptr;
struct ggml_tensor * ssm_g_b = nullptr;
struct ggml_tensor * ssm_o_norm = nullptr;
// DSA (deepseek sparse attention)
struct ggml_tensor * indexer_k_norm = nullptr;
struct ggml_tensor * indexer_k_norm_b = nullptr;
struct ggml_tensor * indexer_proj = nullptr;
struct ggml_tensor * indexer_attn_k = nullptr;
struct ggml_tensor * indexer_attn_q_b = nullptr; // note: for lora a/b, not bias
struct llama_layer_posnet posnet;
struct llama_layer_convnext convnext;
@ -448,6 +499,7 @@ struct llama_model {
struct ggml_tensor * cls_b = nullptr;
struct ggml_tensor * cls_out = nullptr;
struct ggml_tensor * cls_out_b = nullptr;
struct ggml_tensor * cls_norm = nullptr;
struct ggml_tensor * conv1d = nullptr;
struct ggml_tensor * conv1d_b = nullptr;
@ -464,8 +516,9 @@ struct llama_model {
//Dense linear projections for SentenceTransformers models like embeddinggemma
// For Sentence Transformers models structure see
// https://sbert.net/docs/sentence_transformer/usage/custom_models.html#structure-of-sentence-transformer-models
struct ggml_tensor * dense_2_out_layers = nullptr;
struct ggml_tensor * dense_3_out_layers = nullptr;
struct ggml_tensor * dense_2_out_layers = nullptr;
struct ggml_tensor * dense_2_out_layers_b = nullptr;
struct ggml_tensor * dense_3_out_layers = nullptr;
// gguf metadata
std::unordered_map<std::string, std::string> gguf_kv;
@ -476,8 +529,8 @@ struct llama_model {
// for quantize-stats only
std::vector<std::pair<std::string, struct ggml_tensor *>> tensors_by_name;
// for keeping track of extra nodes used by lora adapters
uint32_t n_lora_nodes = 0;
// for keeping track of associated LoRA adapters
std::unordered_set<llama_adapter_lora *> loras;
int64_t t_load_us = 0;
int64_t t_start_us = 0;

920
examples/talk-llama/llama-quant.cpp
View File

File diff suppressed because it is too large Load Diff

266
examples/talk-llama/llama-sampling.cpp → examples/talk-llama/llama-sampler.cpp
View File

@ -1,4 +1,4 @@
#include "llama-sampling.h"
#include "llama-sampler.h"
#include "llama-impl.h"
#include "llama-vocab.h"
@ -1025,11 +1025,7 @@ struct llama_sampler_dist : public llama_sampler_backend {
std::mt19937 rng;
// backend input
struct ggml_tensor * inp_uniform;
ggml_context_ptr inp_ctx;
ggml_backend_buffer_ptr inp_buf;
ggml_tensor * inp_uniform;
};
static const char * llama_sampler_dist_name(const struct llama_sampler * smpl) {
@ -1138,37 +1134,10 @@ static bool llama_sampler_dist_backend_init(
ggml_backend_buffer_type_t buft) {
auto * sctx = (llama_sampler_dist *) smpl->ctx;
// allocate inputs
{
ggml_init_params params = {
/*.mem_size =*/ ggml_tensor_overhead(),
/*.mem_buffer =*/ nullptr,
/*.no_alloc =*/ true,
};
sctx->inp_ctx.reset(ggml_init(params));
// Create the uniform random scalar input tensor. This will be set by
// llama_sampler_dist_backend_set_input after this graph is built.
sctx->inp_uniform = ggml_new_tensor_1d(sctx->inp_ctx.get(), GGML_TYPE_F32, 1);
ggml_set_name (sctx->inp_uniform, "uniform");
ggml_set_input(sctx->inp_uniform);
// Allocate all tensors from our context to the backend
sctx->inp_buf.reset(ggml_backend_alloc_ctx_tensors_from_buft(sctx->inp_ctx.get(), buft));
ggml_backend_buffer_clear(sctx->inp_buf.get(), 0);
}
const bool res = llama_sampler_backend_support(smpl, buft);
sctx->init(res);
if (!res) {
sctx->inp_ctx.reset(nullptr);
sctx->inp_buf.reset(nullptr);
}
return res;
}
@ -1178,8 +1147,13 @@ static void llama_sampler_dist_backend_apply(
struct ggml_cgraph * gf,
struct llama_sampler_data * data) {
GGML_UNUSED(gf);
auto * sctx = (llama_sampler_dist *) smpl->ctx;
sctx->inp_uniform = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
ggml_set_name (sctx->inp_uniform, "uniform");
ggml_set_input(sctx->inp_uniform);
struct ggml_tensor * probs = ggml_soft_max(ctx, data->logits);
ggml_set_name(probs, "dist_probs");
@ -1226,6 +1200,7 @@ static void llama_sampler_dist_backend_apply(
static void llama_sampler_dist_backend_set_input(struct llama_sampler * smpl) {
auto * sctx = (llama_sampler_dist *) smpl->ctx;
GGML_ASSERT(sctx->inp_uniform != nullptr);
// We sample in double precision and cast to float to match rnd numbers of
@ -1262,8 +1237,6 @@ struct llama_sampler * llama_sampler_init_dist(uint32_t seed) {
/* .seed_cur = */ seed_cur,
/* .rng = */ std::mt19937(seed_cur),
/* .inp_uniform = */ nullptr,
/* .inp_ctx = */ nullptr,
/* .inp_buf = */ nullptr,
}
);
}
@ -1513,12 +1486,9 @@ static void llama_sampler_top_p_backend_apply(
mask_reshaped = ggml_set_rows(ctx, mask_reshaped, ones, ggml_cast(ctx, idxf, GGML_TYPE_I32));
mask = ggml_reshape_1d(ctx, mask_reshaped, mask->ne[0]);
// Use ggml_scale_bias (output = (a * s) + b) which in this case becomes:
// top_p_bias = (mask * 1e9f) - 1e9f.
// So entries in the mask that we want to discard will become -1e9f, and
// others will be 0 (meaning that will not effect the logits).
const float large_val = 1e9f;
struct ggml_tensor * top_p_bias = ggml_scale_bias(ctx, mask, large_val, -large_val);
// Apply -INFINITY bias for masked-out tokens
// log(1) = 0 (keep), log(0) = -INF (discard)
struct ggml_tensor * top_p_bias = ggml_log(ctx, mask);
ggml_set_name(top_p_bias, "top_p_bias");
data->logits = ggml_add(ctx, sorted_logits, top_p_bias);
@ -1673,15 +1643,11 @@ static void llama_sampler_min_p_backend_apply(
struct ggml_tensor * mask = ggml_step(ctx, sub);
ggml_set_name(mask, "min_p_mask");
// Use ggml_scale_bias (output = (a * s) + b) which in this case becomes:
// min_p_bias = (mask * 1e9f) - 1e9f.
// So entries in the mask that we want to discard will become -1e9f, and
// others will be 0 (meaning that will not effect the logits).
const float large_val = 1e9f;
struct ggml_tensor * min_p_bias = ggml_scale_bias(ctx, mask, large_val, -large_val);
// Apply -INFINITY bias for masked-out tokens
// log(1) = 0 (keep), log(0) = -INF (discard)
struct ggml_tensor * min_p_bias = ggml_log(ctx, mask);
ggml_set_name(min_p_bias, "min_p_bias");
// Add the min_p bias to the logits.
data->logits = ggml_add(ctx, data->logits, min_p_bias);
ggml_set_name(data->logits, "min_p_logits");
@ -3293,6 +3259,170 @@ struct llama_sampler * llama_sampler_init_dry_testing(int32_t context_size, floa
return result;
}
// adaptive-p sampler state
//
// maintains an exponential moving average of the *ORIGINAL* probabilities
// of selected tokens, used to compute an adapted target at each sampling step.
//
// see llama.h for a full description of the sampler
//
// ref: https://github.com/ggml-org/llama.cpp/pull/17927
//
struct llama_sampler_adaptive_p {
const float target; // target probability (0.0 - 1.0; negative = disabled)
const float decay; // EMA decay; history ~= 1/(1-decay) tokens (0.0 - 0.99)
const uint32_t seed; // original RNG seed
uint32_t seed_cur; // actual RNG seed
std::mt19937 rng; // RNG state
float weighted_sum; // sum(p_i * decay^i)
float total_weight; // sum(decay^i), converges to 1/(1-decay)
std::vector<float> original_probs; // pre-transform probs, cached for EMA update
llama_token pending_token_id; // token ID of selected token
int32_t pending_token_idx; // index of orig. prob. of selected token in original_probs
};
// adaptive probability transformation constants
static constexpr float DISTRIBUTION_WIDTH = 0.3f;
static constexpr float PEAK_LOGIT_VALUE = 5.0f;
static constexpr float SHARPNESS = 10.0f;
static constexpr float INV_WIDTH = 1.0f / DISTRIBUTION_WIDTH;
static const char * llama_sampler_adaptive_p_name(const struct llama_sampler * /*smpl*/) {
return "adaptive-p";
}
static void llama_sampler_adaptive_p_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
auto * ctx = (llama_sampler_adaptive_p *) smpl->ctx;
llama_sampler_softmax_impl(cur_p, false);
if (ctx->target < 0.0f) {
// at negative target values, adaptive-p is no-op
// we simply sample from the existing distribution
cur_p->selected = llama_sample_dist(cur_p, ctx->rng);
return;
}
// store the original probabilities
ctx->original_probs.resize(cur_p->size);
for (size_t i = 0; i < cur_p->size; ++i) {
ctx->original_probs[i] = cur_p->data[i].p;
}
// using the EMA, compute the adapted target probability for the current sampling step
auto target = std::clamp(ctx->target, 0.0f, 1.0f);
float adapted_target = std::clamp(
ctx->total_weight == 0.0f ? target : 2.0f * target - (ctx->weighted_sum / ctx->total_weight),
0.0f, 1.0f
);
// adaptive probability transform
//
// quadratic near target for fine differentiation, transitioning to linear decay in the
// tails. unbounded negative logits ensure proper suppression of far-from-target tokens
// after the softmax.
//
for (size_t i = 0; i < cur_p->size; ++i) {
if (cur_p->data[i].logit == -INFINITY) {
// don't transform logits that are -INFINITY
// (as masked out by e.g. min-p and top-p when using backend sampling)
continue;
}
float dist = std::abs((cur_p->data[i].p - adapted_target) * INV_WIDTH);
cur_p->data[i].logit = PEAK_LOGIT_VALUE - SHARPNESS * dist * dist / (1.0f + dist);
}
// softmax and sample from the transformed distribution
llama_sampler_softmax_impl(cur_p, false);
const int idx = llama_sample_dist(cur_p, ctx->rng);
cur_p->selected = idx;
// store the selected token ID for acceptance later
ctx->pending_token_id = cur_p->data[idx].id;
ctx->pending_token_idx = idx;
}
static void llama_sampler_adaptive_p_accept(struct llama_sampler * smpl, llama_token token) {
auto * ctx = (llama_sampler_adaptive_p *) smpl->ctx;
if (ctx->pending_token_id == token) {
GGML_ASSERT(ctx->pending_token_id != LLAMA_TOKEN_NULL);
GGML_ASSERT(ctx->pending_token_idx != -1);
// update EMA with the original probability of the selected token
ctx->weighted_sum = ctx->original_probs[ctx->pending_token_idx] + ctx->decay * ctx->weighted_sum;
ctx->total_weight = 1.0f + ctx->decay * ctx->total_weight;
}
ctx->pending_token_id = LLAMA_TOKEN_NULL;
ctx->pending_token_idx = -1;
}
static void llama_sampler_adaptive_p_reset(struct llama_sampler * smpl) {
auto * ctx = (llama_sampler_adaptive_p *) smpl->ctx;
// ctx->target and ctx->decay never change after init, so it's safe to keep them as is.
// original_probs is completely overwritten on every call to _apply.
// so we only need to reset the EMA state and pending token.
ctx->weighted_sum = ctx->target / (1.0f - ctx->decay);
ctx->total_weight = 1.0f / (1.0f - ctx->decay);
ctx->pending_token_id = LLAMA_TOKEN_NULL;
ctx->pending_token_idx = -1;
ctx->seed_cur = get_rng_seed(ctx->seed);
ctx->rng.seed(ctx->seed_cur);
}
static struct llama_sampler * llama_sampler_adaptive_p_clone(const struct llama_sampler * smpl) {
const auto * ctx = (const llama_sampler_adaptive_p *) smpl->ctx;
auto * result = llama_sampler_init_adaptive_p(ctx->target, ctx->decay, ctx->seed);
auto * result_ctx = (llama_sampler_adaptive_p *) result->ctx;
// copy everything (target, decay, seed, and RNG are already set)
result_ctx->weighted_sum = ctx->weighted_sum;
result_ctx->total_weight = ctx->total_weight;
result_ctx->pending_token_id = ctx->pending_token_id;
result_ctx->pending_token_idx = ctx->pending_token_idx;
return result;
}
static void llama_sampler_adaptive_p_free(struct llama_sampler * smpl) {
delete (llama_sampler_adaptive_p *) smpl->ctx;
}
static struct llama_sampler_i llama_sampler_adaptive_p_i = {
/* .name = */ llama_sampler_adaptive_p_name,
/* .accept = */ llama_sampler_adaptive_p_accept,
/* .apply = */ llama_sampler_adaptive_p_apply,
/* .reset = */ llama_sampler_adaptive_p_reset,
/* .clone = */ llama_sampler_adaptive_p_clone,
/* .free = */ llama_sampler_adaptive_p_free,
/* .backend_init = */ nullptr,
/* .backend_accept = */ nullptr,
/* .backend_apply = */ nullptr,
/* .backend_set_input = */ nullptr,
};
struct llama_sampler * llama_sampler_init_adaptive_p(
float target,
float decay,
uint32_t seed
) {
auto seed_cur = get_rng_seed(seed);
float clamped_decay = std::clamp(decay, 0.0f, 0.99f);
return llama_sampler_init(
/* .iface = */ &llama_sampler_adaptive_p_i,
/* .ctx = */ new llama_sampler_adaptive_p {
/* .target = */ target,
/* .decay = */ clamped_decay,
/* .seed = */ seed,
/* .seed_cur = */ seed_cur,
/* .rng = */ std::mt19937(seed_cur),
/* .weighted_sum = */ target / (1.0f - clamped_decay),
/* .total_weight = */ 1.0f / (1.0f - clamped_decay),
/* .original_probs = */ {},
/* .pending_token_id = */ LLAMA_TOKEN_NULL,
/* .pending_token_idx = */ -1
}
);
}
// logit-bias
struct llama_sampler_logit_bias : public llama_sampler_backend {
@ -3304,9 +3434,6 @@ struct llama_sampler_logit_bias : public llama_sampler_backend {
struct ggml_tensor * inp_logit_bias;
struct ggml_tensor * inp_logit_idxs;
ggml_context_ptr inp_ctx;
ggml_backend_buffer_ptr inp_buf;
};
static const char * llama_sampler_logit_bias_name(const struct llama_sampler * smpl) {
@ -3369,6 +3496,16 @@ static void llama_sampler_logit_bias_backend_apply(
return;
}
const size_t n = sctx->logit_bias.size();
sctx->inp_logit_bias = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 1, n);
ggml_set_name(sctx->inp_logit_bias, "logit_bias");
ggml_set_input(sctx->inp_logit_bias);
sctx->inp_logit_idxs = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, n);
ggml_set_name(sctx->inp_logit_idxs, "logit_idxs");
ggml_set_input(sctx->inp_logit_idxs);
ggml_tensor * cur = ggml_fill(ctx, data->logits, 0.0f);
cur = ggml_reshape_2d(ctx, cur, 1, ggml_nelements(cur));
@ -3405,6 +3542,8 @@ static void llama_sampler_logit_bias_backend_set_input(struct llama_sampler * sm
static bool llama_sampler_logit_bias_backend_init(
struct llama_sampler * smpl,
ggml_backend_buffer_type_t buft) {
GGML_UNUSED(buft);
auto * sctx = (llama_sampler_logit_bias *) smpl->ctx;
sctx->init(true);
@ -3413,29 +3552,6 @@ static bool llama_sampler_logit_bias_backend_init(
return true;
}
ggml_init_params params = {
/*.mem_size =*/ 2*ggml_tensor_overhead(),
/*.mem_buffer =*/ nullptr,
/*.no_alloc =*/ true,
};
sctx->inp_ctx.reset(ggml_init(params));
const size_t n = sctx->logit_bias.size();
sctx->inp_logit_bias = ggml_new_tensor_2d(sctx->inp_ctx.get(), GGML_TYPE_F32, 1, n);
ggml_set_name(sctx->inp_logit_bias, "logit_bias");
ggml_set_input(sctx->inp_logit_bias);
sctx->inp_logit_idxs = ggml_new_tensor_1d(sctx->inp_ctx.get(), GGML_TYPE_I32, n);
ggml_set_name(sctx->inp_logit_idxs, "logit_idxs");
ggml_set_input(sctx->inp_logit_idxs);
// Allocate all tensors from our context to the backend
sctx->inp_buf.reset(ggml_backend_alloc_ctx_tensors_from_buft(sctx->inp_ctx.get(), buft));
ggml_backend_buffer_clear(sctx->inp_buf.get(), 0);
return true;
}
@ -3471,8 +3587,6 @@ struct llama_sampler * llama_sampler_init_logit_bias(
/* .to_search = */ {},
/* .inp_logit_bias = */ nullptr,
/* .inp_logit_idxs = */ nullptr,
/* .inp_ctx = */ nullptr,
/* .inp_buf = */ nullptr,
}
);
}

2
examples/talk-llama/llama-sampling.h → examples/talk-llama/llama-sampler.h
View File

@ -1,7 +1,5 @@
#pragma once
// TODO: rename llama-sampling.h/.cpp to llama-sampler.h/.cpp ?
#include "llama.h"
#include <vector>

168
examples/talk-llama/llama-vocab.cpp
View File

@ -90,7 +90,7 @@ static_assert(std::is_trivially_copyable<llm_symbol>::value, "llm_symbol is not
//
// SPM tokenizer
// original implementation:
// https://github.com/ggerganov/llama.cpp/commit/074bea2eb1f1349a0118239c4152914aecaa1be4
// https://github.com/ggml-org/llama.cpp/commit/074bea2eb1f1349a0118239c4152914aecaa1be4
//
struct llm_bigram_spm {
@ -285,10 +285,19 @@ struct llm_tokenizer_bpe : llm_tokenizer {
// original regex from tokenizer.json
//"(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
// adapted: https://github.com/ggerganov/llama.cpp/pull/6920#issuecomment-2080233989
// adapted: https://github.com/ggml-org/llama.cpp/pull/6920#issuecomment-2080233989
"(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
};
break;
case LLAMA_VOCAB_PRE_TYPE_JAIS2:
regex_exprs = {
// original regex from tokenizer.json
//"(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s{512}(?!\\S)|\\s{256}(?!\\S)|\\s{128}(?!\\S)|\\s{64}(?!\\S)|\\s{32}(?!\\S)|\\s{16}(?!\\S)|\\s{8}(?!\\S)|\\s{4}(?!\\S)|\\s{1,2}(?!\\S)|\\s{1}",
// adapted: same as llama3 but with cascading whitespace pattern
"(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s{512}(?!\\S)|\\s{256}(?!\\S)|\\s{128}(?!\\S)|\\s{64}(?!\\S)|\\s{32}(?!\\S)|\\s{16}(?!\\S)|\\s{8}(?!\\S)|\\s{4}(?!\\S)|\\s{1,2}(?!\\S)|\\s{1}",
};
break;
case LLAMA_VOCAB_PRE_TYPE_DBRX:
case LLAMA_VOCAB_PRE_TYPE_SMAUG:
regex_exprs = {
@ -308,6 +317,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
break;
case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM:
case LLAMA_VOCAB_PRE_TYPE_HUNYUAN_DENSE:
case LLAMA_VOCAB_PRE_TYPE_JOYAI_LLM:
regex_exprs = {
"\\p{N}{1,3}",
"[一-龥぀-ゟ゠-ヿ]+",
@ -368,6 +378,13 @@ struct llm_tokenizer_bpe : llm_tokenizer {
"(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
};
break;
case LLAMA_VOCAB_PRE_TYPE_QWEN35:
regex_exprs = {
// original regex from tokenizer.json
// "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+"
"(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
};
break;
case LLAMA_VOCAB_PRE_TYPE_PORO:
case LLAMA_VOCAB_PRE_TYPE_BLOOM:
case LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH:
@ -415,6 +432,14 @@ struct llm_tokenizer_bpe : llm_tokenizer {
"[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))*((?=[\\p{L}])([^A-Z]))+(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))+((?=[\\p{L}])([^A-Z]))*(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
};
break;
case LLAMA_VOCAB_PRE_TYPE_TINY_AYA:
regex_exprs = {
// original regex from tokenizer.json: "\\d{1,3}(?=(?:\\d{3})*\\b)"
"\\d{1,3}(?=(?:\\d{3})*\\b)",
// original regex from tokenizer.json: "[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]*[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]+(?i:'s|'t|'re|'ve|'m|'ll|'d)?|[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]+[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]*(?i:'s|'t|'re|'ve|'m|'ll|'d)?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+"
"[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]*[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]+(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]+[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]*(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
};
break;
case LLAMA_VOCAB_PRE_TYPE_KIMI_K2:
regex_exprs = {
// K2 trigger pattern - this will activate the custom K2 handler in unicode.cpp
@ -461,6 +486,13 @@ struct llm_tokenizer_bpe : llm_tokenizer {
"[!\"#$%&'()*+,\\-./:;<=>?@\\[\\\\\\]^_`{|}~][A-Za-z]+|[^\\r\\n\\p{L}\\p{P}\\p{S}]?[\\p{L}\\p{M}]+| ?[\\p{P}\\p{S}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
};
break;
case LLAMA_VOCAB_PRE_TYPE_EXAONE_MOE:
regex_exprs = {
// original regex from tokenizer.json
// "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?(?:\\p{L}\\p{M}*(?: \\p{L}\\p{M}*)*)+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]?|\\s*[\\r\\n]|\\s+(?!\\S)|\\s+"
"(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?(?:\\p{L}\\p{M}*(?: \\p{L}\\p{M}*)*)+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]?|\\s*[\\r\\n]|\\s+(?!\\S)|\\s+",
};
break;
default:
// default regex for BPE tokenization pre-processing
regex_exprs = {
@ -1687,7 +1719,7 @@ private:
};
void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
struct gguf_context * ctx = ml.meta.get();
struct gguf_context * ctx = ml.metadata;
// determine vocab type
{
@ -1745,26 +1777,33 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
// read bpe merges and populate bpe ranks
const int merges_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_MERGES).c_str());
// Kimi-K2 uses custom tokenization without traditional BPE merges
const bool is_kimi_k2 = (tokenizer_pre == "kimi-k2");
if (merges_keyidx == -1) {
throw std::runtime_error("cannot find tokenizer merges in model file\n");
}
const int n_merges = gguf_get_arr_n(ctx, merges_keyidx);
for (int i = 0; i < n_merges; i++) {
const std::string word = gguf_get_arr_str(ctx, merges_keyidx, i);
//GGML_ASSERT(unicode_cpts_from_utf8(word).size() > 0);
std::string first;
std::string second;
const size_t pos = word.find(' ', 1);
if (pos != std::string::npos) {
first = word.substr(0, pos);
second = word.substr(pos + 1);
if (!is_kimi_k2) {
throw std::runtime_error("cannot find tokenizer merges in model file\n");
}
// Kimi-K2 doesn't need merges, skip
LLAMA_LOG_INFO("%s: Kimi-K2 tokenizer detected, skipping BPE merges\n", __func__);
} else {
const int n_merges = gguf_get_arr_n(ctx, merges_keyidx);
for (int i = 0; i < n_merges; i++) {
const std::string word = gguf_get_arr_str(ctx, merges_keyidx, i);
//GGML_ASSERT(unicode_cpts_from_utf8(word).size() > 0);
bpe_ranks.emplace(std::make_pair(first, second), i);
std::string first;
std::string second;
const size_t pos = word.find(' ', 1);
if (pos != std::string::npos) {
first = word.substr(0, pos);
second = word.substr(pos + 1);
}
bpe_ranks.emplace(std::make_pair(first, second), i);
}
}
// default special tokens
@ -1794,7 +1833,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
const char * pc = (const char *) gguf_get_arr_data(ctx, precompiled_charsmap_keyidx);
precompiled_charsmap.assign(pc, pc + n_precompiled_charsmap);
#if defined(__BYTE_ORDER__) && defined(__ORDER_BIG_ENDIAN__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
// correct endiannes of data in precompiled_charsmap binary blob
// correct endianness of data in precompiled_charsmap binary blob
uint32_t * xcda_blob_size = (uint32_t *) &precompiled_charsmap[0];
*xcda_blob_size = __builtin_bswap32(*xcda_blob_size);
assert(*xcda_blob_size + sizeof(uint32_t) < n_precompiled_charsmap);
@ -1851,7 +1890,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
tokenizer_pre == "falcon-h1" ||
tokenizer_pre == "pixtral" ||
tokenizer_pre == "midm-2.0" ||
tokenizer_pre == "lfm2") {
tokenizer_pre == "lfm2" ||
tokenizer_pre == "jina-v5-nano") {
pre_type = LLAMA_VOCAB_PRE_TYPE_LLAMA3;
ignore_merges = true;
add_bos = true;
@ -1891,8 +1931,11 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
tokenizer_pre == "jina-v2-de" ||
tokenizer_pre == "a.x-4.0" ||
tokenizer_pre == "mellum" ||
tokenizer_pre == "modern-bert" ) {
tokenizer_pre == "modern-bert") {
pre_type = LLAMA_VOCAB_PRE_TYPE_GPT2;
} else if (
tokenizer_pre == "jais-2") {
pre_type = LLAMA_VOCAB_PRE_TYPE_JAIS2;
} else if (
tokenizer_pre == "jina-v1-en" ||
tokenizer_pre == "jina-v2-code" ||
@ -1912,6 +1955,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
tokenizer_pre == "kormo") {
pre_type = LLAMA_VOCAB_PRE_TYPE_QWEN2;
clean_spaces = false;
} else if (
tokenizer_pre == "qwen35") {
pre_type = LLAMA_VOCAB_PRE_TYPE_QWEN35;
clean_spaces = false;
} else if (
tokenizer_pre == "stablelm2") {
pre_type = LLAMA_VOCAB_PRE_TYPE_STABLELM2;
@ -1965,6 +2012,9 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
} else if (
tokenizer_pre == "exaone4") {
pre_type = LLAMA_VOCAB_PRE_TYPE_GPT2;
} else if (
tokenizer_pre == "exaone-moe") {
pre_type = LLAMA_VOCAB_PRE_TYPE_EXAONE_MOE;
} else if (
tokenizer_pre == "chameleon") {
pre_type = LLAMA_VOCAB_PRE_TYPE_CHAMELEON;
@ -1977,10 +2027,15 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
tokenizer_pre == "megrez") {
pre_type = LLAMA_VOCAB_PRE_TYPE_QWEN2;
} else if (
tokenizer_pre == "gpt-4o" ||
tokenizer_pre == "llama4") {
tokenizer_pre == "gpt-4o" ||
tokenizer_pre == "llama4" ||
tokenizer_pre == "kanana2") {
pre_type = LLAMA_VOCAB_PRE_TYPE_GPT4O;
clean_spaces = false;
} else if (
tokenizer_pre == "tiny_aya") {
pre_type = LLAMA_VOCAB_PRE_TYPE_TINY_AYA;
clean_spaces = false;
} else if (
tokenizer_pre == "superbpe") {
pre_type = LLAMA_VOCAB_PRE_TYPE_SUPERBPE;
@ -2011,6 +2066,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
tokenizer_pre == "hunyuan-dense") {
pre_type = LLAMA_VOCAB_PRE_TYPE_HUNYUAN_DENSE;
clean_spaces = false;
} else if (
tokenizer_pre == "joyai-llm") {
pre_type = LLAMA_VOCAB_PRE_TYPE_JOYAI_LLM;
clean_spaces = false;
} else if (
tokenizer_pre == "kimi-k2") {
pre_type = LLAMA_VOCAB_PRE_TYPE_KIMI_K2;
@ -2216,6 +2275,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
|| t.first == "<|end_of_text|>" // granite
|| t.first == "<EOT>"
|| t.first == "_<EOT>"
|| t.first == "[EOT]" // Kimi-K2
|| t.first == "<｜end▁of▁sentence｜>" // DeepSeek
|| t.first == "<end_of_utterance>" // smoldocling
) {
@ -2252,6 +2312,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
|| t.first == "<PRE>"
|| t.first == "▁<PRE>" // CodeLlama
|| t.first == "<|code_prefix|>" // GLM-4.5
|| t.first == "<|prefix|>" // Falcon-H1-Tiny-Coder
) {
special_fim_pre_id = t.second;
if ((attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@ -2272,6 +2333,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
|| t.first == "<SUF>"
|| t.first == "▁<SUF>" // CodeLlama
|| t.first == "<|code_suffix|>" // GLM-4.5
|| t.first == "<|suffix|>" // Falcon-H1-Tiny-Coder
) {
special_fim_suf_id = t.second;
if ((attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@ -2292,6 +2354,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
|| t.first == "<MID>"
|| t.first == "▁<MID>" // CodeLlama
|| t.first == "<|code_middle|>" // GLM-4.5
|| t.first == "<|middle|>" // Falcon-H1-Tiny-Coder
) {
special_fim_mid_id = t.second;
if ((attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@ -2309,6 +2372,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
|| t.first == "<fim-pad>"
|| t.first == "<fim_pad>" // Granite
|| t.first == "<PAD>"
|| t.first == "[PAD]" // Kimi-K2
) {
special_fim_pad_id = t.second;
if ((attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@ -2380,7 +2444,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
// maintain a list of tokens that cause end-of-generation
// this is currently determined based on the token text, which is obviously not ideal
// ref: https://github.com/ggerganov/llama.cpp/issues/9606
// ref: https://github.com/ggml-org/llama.cpp/issues/9606
special_eog_ids.clear();
if (special_fim_pad_id != LLAMA_TOKEN_NULL && special_eog_ids.count(special_fim_pad_id) == 0) {
@ -2408,9 +2472,12 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
|| t.first == "<|calls|>" // solar-open
|| t.first == "<end_of_turn>"
|| t.first == "<|endoftext|>"
|| t.first == "</s>" // paddleocr
|| t.first == "<|eom_id|>"
|| t.first == "<EOT>"
|| t.first == "_<EOT>"
|| t.first == "[EOT]" // Kimi-K2
|| t.first == "[EOS]" // Kimi-K2
|| t.first == "<|end_of_text|>"
|| t.first == "<end_of_utterance>" // smoldocling
) {
@ -2436,7 +2503,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
auto & attr = id_to_token[t.second].attr;
if (t.first == "<|channel|>" || t.first == "<|message|>" || t.first == "<|start|>" || t.first == "<|constrain|>") {
attr = (llama_token_attr) (attr | LLAMA_TOKEN_ATTR_USER_DEFINED);
LLAMA_LOG_WARN("%s: setting token '%s' (%d) attribute to USER_DEFINED (%u), old attributes: %u\n",
__func__, t.first.c_str(), t.second, LLAMA_TOKEN_ATTR_USER_DEFINED, attr);
attr = LLAMA_TOKEN_ATTR_USER_DEFINED;
}
}
@ -2489,7 +2559,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
special_eog_ids.erase(end_id);
auto & attr = id_to_token[end_id].attr;
attr = (llama_token_attr) (attr | LLAMA_TOKEN_ATTR_USER_DEFINED);
attr = LLAMA_TOKEN_ATTR_USER_DEFINED;
LLAMA_LOG_WARN("%s: special_eog_ids contains both '<|return|>' and '<|call|>', or '<|calls|>' and '<|flush|>' tokens, removing '<|end|>' token from EOG list\n", __func__);
}
@ -3066,7 +3136,7 @@ std::vector<llama_token> llama_vocab::impl::tokenize(
}
int32_t llama_vocab::impl::token_to_piece(llama_token token, char * buf, int32_t length, int32_t lstrip, bool special) const {
// ref: https://github.com/ggerganov/llama.cpp/pull/7587#discussion_r1620983843
// ref: https://github.com/ggml-org/llama.cpp/pull/7587#discussion_r1620983843
static const int attr_special = LLAMA_TOKEN_ATTR_UNKNOWN | LLAMA_TOKEN_ATTR_CONTROL;
const llama_token_attr attr = token_get_attr(token);
if (!special && (attr & attr_special)) {
@ -3289,34 +3359,34 @@ int32_t llama_vocab::impl::detokenize(
}
void llama_vocab::impl::print_info() const {
LLAMA_LOG_INFO("%s: vocab type = %s\n", __func__, type_name().c_str());
LLAMA_LOG_INFO("%s: n_vocab = %u\n", __func__, vocab.n_tokens());
LLAMA_LOG_INFO("%s: n_merges = %u\n", __func__, (uint32_t) bpe_ranks.size());
LLAMA_LOG_INFO("%s: vocab type = %s\n", __func__, type_name().c_str());
LLAMA_LOG_INFO("%s: n_vocab = %u\n", __func__, vocab.n_tokens());
LLAMA_LOG_INFO("%s: n_merges = %u\n", __func__, (uint32_t) bpe_ranks.size());
// special tokens
if (special_bos_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: BOS token = %d '%s'\n", __func__, special_bos_id, id_to_token.at(special_bos_id).text.c_str() ); }
if (special_eos_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: EOS token = %d '%s'\n", __func__, special_eos_id, id_to_token.at(special_eos_id).text.c_str() ); }
if (special_eot_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: EOT token = %d '%s'\n", __func__, special_eot_id, id_to_token.at(special_eot_id).text.c_str() ); }
if (special_eom_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: EOM token = %d '%s'\n", __func__, special_eom_id, id_to_token.at(special_eom_id).text.c_str() ); }
if (special_unk_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: UNK token = %d '%s'\n", __func__, special_unk_id, id_to_token.at(special_unk_id).text.c_str() ); }
if (special_sep_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: SEP token = %d '%s'\n", __func__, special_sep_id, id_to_token.at(special_sep_id).text.c_str() ); }
if (special_pad_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: PAD token = %d '%s'\n", __func__, special_pad_id, id_to_token.at(special_pad_id).text.c_str() ); }
if (special_mask_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: MASK token = %d '%s'\n", __func__, special_mask_id, id_to_token.at(special_mask_id).text.c_str() ); }
if (special_bos_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: BOS token = %d '%s'\n", __func__, special_bos_id, id_to_token.at(special_bos_id).text.c_str() ); }
if (special_eos_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: EOS token = %d '%s'\n", __func__, special_eos_id, id_to_token.at(special_eos_id).text.c_str() ); }
if (special_eot_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: EOT token = %d '%s'\n", __func__, special_eot_id, id_to_token.at(special_eot_id).text.c_str() ); }
if (special_eom_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: EOM token = %d '%s'\n", __func__, special_eom_id, id_to_token.at(special_eom_id).text.c_str() ); }
if (special_unk_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: UNK token = %d '%s'\n", __func__, special_unk_id, id_to_token.at(special_unk_id).text.c_str() ); }
if (special_sep_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: SEP token = %d '%s'\n", __func__, special_sep_id, id_to_token.at(special_sep_id).text.c_str() ); }
if (special_pad_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: PAD token = %d '%s'\n", __func__, special_pad_id, id_to_token.at(special_pad_id).text.c_str() ); }
if (special_mask_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: MASK token = %d '%s'\n", __func__, special_mask_id, id_to_token.at(special_mask_id).text.c_str() ); }
if (linefeed_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: LF token = %d '%s'\n", __func__, linefeed_id, id_to_token.at(linefeed_id).text.c_str() ); }
if (linefeed_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: LF token = %d '%s'\n", __func__, linefeed_id, id_to_token.at(linefeed_id).text.c_str() ); }
if (special_fim_pre_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM PRE token = %d '%s'\n", __func__, special_fim_pre_id, id_to_token.at(special_fim_pre_id).text.c_str() ); }
if (special_fim_suf_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM SUF token = %d '%s'\n", __func__, special_fim_suf_id, id_to_token.at(special_fim_suf_id).text.c_str() ); }
if (special_fim_mid_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM MID token = %d '%s'\n", __func__, special_fim_mid_id, id_to_token.at(special_fim_mid_id).text.c_str() ); }
if (special_fim_pad_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM PAD token = %d '%s'\n", __func__, special_fim_pad_id, id_to_token.at(special_fim_pad_id).text.c_str() ); }
if (special_fim_rep_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM REP token = %d '%s'\n", __func__, special_fim_rep_id, id_to_token.at(special_fim_rep_id).text.c_str() ); }
if (special_fim_sep_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM SEP token = %d '%s'\n", __func__, special_fim_sep_id, id_to_token.at(special_fim_sep_id).text.c_str() ); }
if (special_fim_pre_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM PRE token = %d '%s'\n", __func__, special_fim_pre_id, id_to_token.at(special_fim_pre_id).text.c_str() ); }
if (special_fim_suf_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM SUF token = %d '%s'\n", __func__, special_fim_suf_id, id_to_token.at(special_fim_suf_id).text.c_str() ); }
if (special_fim_mid_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM MID token = %d '%s'\n", __func__, special_fim_mid_id, id_to_token.at(special_fim_mid_id).text.c_str() ); }
if (special_fim_pad_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM PAD token = %d '%s'\n", __func__, special_fim_pad_id, id_to_token.at(special_fim_pad_id).text.c_str() ); }
if (special_fim_rep_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM REP token = %d '%s'\n", __func__, special_fim_rep_id, id_to_token.at(special_fim_rep_id).text.c_str() ); }
if (special_fim_sep_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM SEP token = %d '%s'\n", __func__, special_fim_sep_id, id_to_token.at(special_fim_sep_id).text.c_str() ); }
for (const auto & id : special_eog_ids) {
LLAMA_LOG_INFO( "%s: EOG token = %d '%s'\n", __func__, id, id_to_token.at(id).text.c_str() );
LLAMA_LOG_INFO( "%s: EOG token = %d '%s'\n", __func__, id, id_to_token.at(id).text.c_str() );
}
LLAMA_LOG_INFO("%s: max token length = %d\n", __func__, max_token_len);
LLAMA_LOG_INFO("%s: max token length = %d\n", __func__, max_token_len);
}
llama_vocab::llama_vocab() : pimpl(new impl(*this)) {

5
examples/talk-llama/llama-vocab.h
View File

@ -53,6 +53,11 @@ enum llama_vocab_pre_type {
LLAMA_VOCAB_PRE_TYPE_AFMOE = 42,
LLAMA_VOCAB_PRE_TYPE_SOLAR_OPEN = 43,
LLAMA_VOCAB_PRE_TYPE_YOUTU = 44,
LLAMA_VOCAB_PRE_TYPE_EXAONE_MOE = 45,
LLAMA_VOCAB_PRE_TYPE_QWEN35 = 46,
LLAMA_VOCAB_PRE_TYPE_TINY_AYA = 47,
LLAMA_VOCAB_PRE_TYPE_JOYAI_LLM = 48,
LLAMA_VOCAB_PRE_TYPE_JAIS2 = 49,
};
struct LLM_KV;

98
examples/talk-llama/llama.cpp
View File

@ -1,5 +1,6 @@
#include "llama.h"
#include "ggml-cpp.h"
#include "llama-impl.h"
#include "llama-chat.h"
@ -12,6 +13,7 @@
#include "ggml.h"
#include "ggml-backend.h"
#include "gguf.h"
#include <algorithm>
#include <cassert>
@ -311,8 +313,12 @@ static void llama_params_fit_impl(
__func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
}
} else {
LLAMA_LOG_INFO("%s: default model context size is %" PRIu32 " which is <= the min. context size of %" PRIu32 " -> no change\n",
__func__, hp_nct, n_ctx_min);
if (n_ctx_min == UINT32_MAX) {
LLAMA_LOG_INFO("%s: user has requested full context size of %" PRIu32 " -> no change\n", __func__, hp_nct);
} else {
LLAMA_LOG_INFO("%s: default model context size is %" PRIu32 " which is <= the min. context size of %" PRIu32 " -> no change\n",
__func__, hp_nct, n_ctx_min);
}
}
} else {
LLAMA_LOG_INFO("%s: context size set by user to %" PRIu32 " -> no change\n", __func__, cparams->n_ctx);
@ -821,7 +827,8 @@ int64_t llama_time_us(void) {
}
// Returns 0 on success, -1 on error, and -2 on cancellation via llama_progress_callback
static int llama_model_load(const std::string & fname, std::vector<std::string> & splits, llama_model & model, llama_model_params & params) {
static int llama_model_load(struct gguf_context * metadata, llama_model_set_tensor_data_t set_tensor_data, void * set_tensor_data_ud,
const std::string & fname, std::vector<std::string> & splits, llama_model & model, llama_model_params & params) {
// loading time will be recalculated after the first eval, so
// we take page faults deferred by mmap() into consideration
model.t_load_us = 0;
@ -830,7 +837,8 @@ static int llama_model_load(const std::string & fname, std::vector<std::string>
model.t_start_us = tm.t_start_us;
try {
llama_model_loader ml(fname, splits, params.use_mmap, params.use_direct_io, params.check_tensors, params.no_alloc, params.kv_overrides, params.tensor_buft_overrides);
llama_model_loader ml(metadata, set_tensor_data, set_tensor_data_ud, fname, splits, params.use_mmap, params.use_direct_io,
params.check_tensors, params.no_alloc, params.kv_overrides, params.tensor_buft_overrides);
ml.print_info();
@ -876,9 +884,13 @@ static int llama_model_load(const std::string & fname, std::vector<std::string>
}
static struct llama_model * llama_model_load_from_file_impl(
struct gguf_context * metadata,
llama_model_set_tensor_data_t set_tensor_data,
void * set_tensor_data_ud,
const std::string & path_model,
std::vector<std::string> & splits,
struct llama_model_params params) {
GGML_ASSERT((metadata == nullptr) != path_model.empty() && "exactly one out of metadata and path_model needs to be defined");
ggml_time_init();
if (!params.vocab_only && ggml_backend_reg_count() == 0) {
@ -999,7 +1011,7 @@ static struct llama_model * llama_model_load_from_file_impl(
props.memory_free/1024/1024);
}
const int status = llama_model_load(path_model, splits, *model, params);
const int status = llama_model_load(metadata, set_tensor_data, set_tensor_data_ud, path_model, splits, *model, params);
GGML_ASSERT(status <= 0);
if (status < 0) {
if (status == -1) {
@ -1015,6 +1027,18 @@ static struct llama_model * llama_model_load_from_file_impl(
return model;
}
struct llama_model * llama_model_init_from_user(
struct gguf_context * metadata,
llama_model_set_tensor_data_t set_tensor_data,
void * set_tensor_data_ud,
struct llama_model_params params) {
GGML_ASSERT(metadata != nullptr);
std::string path_model;
std::vector<std::string> splits = {};
params.use_mmap = false;
params.use_extra_bufts = false;
return llama_model_load_from_file_impl(metadata, set_tensor_data, set_tensor_data_ud, path_model, splits, params);
}
// deprecated
struct llama_model * llama_load_model_from_file(
const char * path_model,
@ -1026,7 +1050,7 @@ struct llama_model * llama_model_load_from_file(
const char * path_model,
struct llama_model_params params) {
std::vector<std::string> splits = {};
return llama_model_load_from_file_impl(path_model, splits, params);
return llama_model_load_from_file_impl(nullptr, nullptr, nullptr, path_model, splits, params);
}
struct llama_model * llama_model_load_from_splits(
@ -1042,11 +1066,11 @@ struct llama_model * llama_model_load_from_splits(
for (size_t i = 0; i < n_paths; ++i) {
splits.push_back(paths[i]);
}
return llama_model_load_from_file_impl(splits.front(), splits, params);
return llama_model_load_from_file_impl(nullptr, nullptr, nullptr, splits.front(), splits, params);
}
void llama_model_save_to_file(const struct llama_model * model, const char * path_model) {
llama_model_saver ms(*model);
llama_model_saver ms(model);
ms.add_kv_from_model();
ms.add_tensors_from_model();
ms.save(path_model);
@ -1091,25 +1115,55 @@ int32_t llama_chat_apply_template(
// model split
//
int llama_split_path(char * split_path, size_t maxlen, const char * path_prefix, int split_no, int split_count) {
int32_t llama_split_path(
char * split_path,
size_t maxlen,
const char * path_prefix,
int32_t split_no,
int32_t split_count) {
static const char * const SPLIT_PATH_FORMAT = "%s-%05d-of-%05d.gguf";
if (snprintf(split_path, maxlen, SPLIT_PATH_FORMAT, path_prefix, split_no + 1, split_count)) {
return strlen(split_path);
const int written = snprintf(
split_path,
maxlen,
SPLIT_PATH_FORMAT,
path_prefix,
split_no + 1,
split_count
);
if (written < 0 || (size_t) written >= maxlen) {
return 0;
}
return 0;
return (int32_t) written;
}
int llama_split_prefix(char * split_prefix, size_t maxlen, const char * split_path, int split_no, int split_count) {
std::string str_split_path(split_path);
char postfix[32];
snprintf(postfix, 32, "-%05d-of-%05d.gguf", split_no + 1, split_count);
std::string str_postfix(postfix);
int32_t llama_split_prefix(
char * split_prefix,
size_t maxlen,
const char * split_path,
int32_t split_no,
int32_t split_count) {
// check if split_prefix ends with postfix
int size_prefix = str_split_path.size() - str_postfix.size();
if (size_prefix > 0 && str_split_path.find(str_postfix, size_prefix) != std::string::npos) {
snprintf(split_prefix, std::min((size_t) size_prefix + 1, maxlen), "%s", split_path);
return size_prefix;
const std::string str_split_path(split_path);
char postfix[32];
snprintf(postfix, sizeof(postfix), "-%05d-of-%05d.gguf", split_no + 1, split_count);
const std::string str_postfix(postfix);
if (str_split_path.size() <= str_postfix.size()) {
return 0;
}
const size_t size_prefix = str_split_path.size() - str_postfix.size();
if (str_split_path.compare(size_prefix, std::string::npos, str_postfix) == 0) {
const size_t copy_len = std::min(size_prefix + 1, maxlen);
snprintf(split_prefix, copy_len, "%s", split_path);
return (int32_t) size_prefix;
}
return 0;

93
examples/talk-llama/llama.h
View File

@ -5,6 +5,7 @@
#include "ggml-cpu.h"
#include "ggml-backend.h"
#include "ggml-opt.h"
#include "gguf.h"
#include <stddef.h>
#include <stdint.h>
@ -152,6 +153,7 @@ extern "C" {
LLAMA_FTYPE_MOSTLY_TQ1_0 = 36, // except 1d tensors
LLAMA_FTYPE_MOSTLY_TQ2_0 = 37, // except 1d tensors
LLAMA_FTYPE_MOSTLY_MXFP4_MOE = 38, // except 1d tensors
LLAMA_FTYPE_MOSTLY_NVFP4 = 39, // except 1d tensors
LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
};
@ -309,7 +311,7 @@ extern "C" {
// Keep the booleans together to avoid misalignment during copy-by-value.
bool vocab_only; // only load the vocabulary, no weights
bool use_mmap; // use mmap if possible
bool use_direct_io; // use direct io, takes precedence over use_mmap
bool use_direct_io; // use direct io, takes precedence over use_mmap when supported
bool use_mlock; // force system to keep model in RAM
bool check_tensors; // validate model tensor data
bool use_extra_bufts; // use extra buffer types (used for weight repacking)
@ -389,6 +391,7 @@ extern "C" {
bool only_copy; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
bool pure; // quantize all tensors to the default type
bool keep_split; // quantize to the same number of shards
bool dry_run; // calculate and show the final quantization size without performing quantization
void * imatrix; // pointer to importance matrix data
void * kv_overrides; // pointer to vector containing overrides
void * tensor_types; // pointer to vector containing tensor types
@ -439,19 +442,30 @@ extern "C" {
LLAMA_API void llama_detach_threadpool(struct llama_context * ctx);
typedef void (*llama_model_set_tensor_data_t)(struct ggml_tensor * tensor, void * userdata);
// Create a new model from GGUF metadata as well as a function to set the tensor data
// - tensors are created as GGML_TYPE_F32 by default,
// override by adding a tensor with the same name but a different name to the context
LLAMA_API struct llama_model * llama_model_init_from_user(
struct gguf_context * metadata,
llama_model_set_tensor_data_t set_tensor_data, // function to initialize tensor data with
void * set_tensor_data_ud, // userdata for function
struct llama_model_params params);
DEPRECATED(LLAMA_API struct llama_model * llama_load_model_from_file(
const char * path_model,
struct llama_model_params params),
"use llama_model_load_from_file instead");
// Load the model from a file
// Load a model from a file
// If the file is split into multiple parts, the file name must follow this pattern: <name>-%05d-of-%05d.gguf
// If the split file name does not follow this pattern, use llama_model_load_from_splits
LLAMA_API struct llama_model * llama_model_load_from_file(
const char * path_model,
struct llama_model_params params);
// Load the model from multiple splits (support custom naming scheme)
// Load a model from multiple splits (support custom naming scheme)
// The paths must be in the correct order
LLAMA_API struct llama_model * llama_model_load_from_splits(
const char ** paths,
@ -482,13 +496,14 @@ extern "C" {
enum llama_params_fit_status {
LLAMA_PARAMS_FIT_STATUS_SUCCESS = 0, // found allocations that are projected to fit
LLAMA_PARAMS_FIT_STATUS_FAILURE = 1, // could not find allocations that are projected to fit
LLAMA_PARAMS_FIT_STATUS_ERROR = 2, // a hard error occured, e.g. because no model could be found at the specified path
LLAMA_PARAMS_FIT_STATUS_ERROR = 2, // a hard error occurred, e.g. because no model could be found at the specified path
};
// fits mparams and cparams to free device memory (assumes system memory is unlimited)
// - returns true if the parameters could be successfully modified to fit device memory
// - this function is NOT thread safe because it modifies the global llama logger state
// - only parameters that have the same value as in llama_default_model_params are modified
// with the exception of the context size which is modified if and only if equal to 0
LLAMA_API enum llama_params_fit_status llama_params_fit(
const char * path_model,
struct llama_model_params * mparams,
@ -646,7 +661,8 @@ extern "C" {
// Manually free a LoRA adapter
// NOTE: loaded adapters will be free when the associated model is deleted
LLAMA_API void llama_adapter_lora_free(struct llama_adapter_lora * adapter);
LLAMA_API DEPRECATED(void llama_adapter_lora_free(struct llama_adapter_lora * adapter),
"adapters are now freed together with the associated model");
// Get the invocation tokens if the current lora is an alora
LLAMA_API uint64_t llama_adapter_get_alora_n_invocation_tokens(const struct llama_adapter_lora * adapter);
@ -654,21 +670,12 @@ extern "C" {
// The following functions operate on a llama_context, hence the naming: llama_verb_...
// Add a loaded LoRA adapter to given context
// This will not modify model's weight
LLAMA_API int32_t llama_set_adapter_lora(
// Set LoRa adapters on the context. Will only modify if the adapters currently in context are different.
LLAMA_API int32_t llama_set_adapters_lora(
struct llama_context * ctx,
struct llama_adapter_lora * adapter,
float scale);
// Remove a specific LoRA adapter from given context
// Return -1 if the adapter is not present in the context
LLAMA_API int32_t llama_rm_adapter_lora(
struct llama_context * ctx,
struct llama_adapter_lora * adapter);
// Remove all LoRA adapters from given context
LLAMA_API void llama_clear_adapter_lora(struct llama_context * ctx);
struct llama_adapter_lora ** adapters,
size_t n_adapters,
float * scales);
// Apply a loaded control vector to a llama_context, or if data is NULL, clear
// the currently loaded vector.
@ -676,7 +683,7 @@ extern "C" {
// to an n_embd x n_layers buffer starting from layer 1.
// il_start and il_end are the layer range the vector should apply to (both inclusive)
// See llama_control_vector_load in common to load a control vector.
LLAMA_API int32_t llama_apply_adapter_cvec(
LLAMA_API int32_t llama_set_adapter_cvec(
struct llama_context * ctx,
const float * data,
size_t len,
@ -979,7 +986,7 @@ extern "C" {
// Logits for the ith token. For positive indices, Equivalent to:
// llama_get_logits(ctx) + ctx->output_ids[i]*n_vocab
// Negative indicies can be used to access logits in reverse order, -1 is the last logit.
// Negative indices can be used to access logits in reverse order, -1 is the last logit.
// returns NULL for invalid ids.
LLAMA_API float * llama_get_logits_ith(struct llama_context * ctx, int32_t i);
@ -994,7 +1001,7 @@ extern "C" {
// Get the embeddings for the ith token. For positive indices, Equivalent to:
// llama_get_embeddings(ctx) + ctx->output_ids[i]*n_embd
// Negative indicies can be used to access embeddings in reverse order, -1 is the last embedding.
// Negative indices can be used to access embeddings in reverse order, -1 is the last embedding.
// shape: [n_embd] (1-dimensional)
// returns NULL for invalid ids.
LLAMA_API float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i);
@ -1014,9 +1021,9 @@ extern "C" {
// Returns LLAMA_TOKEN_NULL if no token was sampled.
LLAMA_API llama_token llama_get_sampled_token_ith(struct llama_context * ctx, int32_t i);
// Get the backend sampled probabilites for the ith token
// Get the backend sampled probabilities for the ith token
// The index matches llama_get_sampled_token_ith().
// Returns NULL if no probabilites were generated.
// Returns NULL if no probabilities were generated.
LLAMA_API float * llama_get_sampled_probs_ith (struct llama_context * ctx, int32_t i);
LLAMA_API uint32_t llama_get_sampled_probs_count_ith(struct llama_context * ctx, int32_t i);
@ -1148,9 +1155,9 @@ extern "C" {
//
/// Apply chat template. Inspired by hf apply_chat_template() on python.
/// Both "model" and "custom_template" are optional, but at least one is required. "custom_template" has higher precedence than "model"
///
/// NOTE: This function does not use a jinja parser. It only support a pre-defined list of template. See more: https://github.com/ggml-org/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template
/// @param tmpl A Jinja template to use for this chat. If this is nullptr, the model’s default chat template will be used instead.
/// @param tmpl A Jinja template to use for this chat.
/// @param chat Pointer to a list of multiple llama_chat_message
/// @param n_msg Number of llama_chat_message in this chat
/// @param add_ass Whether to end the prompt with the token(s) that indicate the start of an assistant message.
@ -1255,7 +1262,6 @@ extern "C" {
// [EXPERIMENTAL]
// attach a sampler to the context
// note: prefer initializing the context with llama_context_params.samplers when possible
// note: changing the samplers of a context can cause graph reallocations and degraded performance
LLAMA_API bool llama_set_sampler(struct llama_context * ctx, llama_seq_id seq_id, struct llama_sampler * smpl);
// mirror of llama_sampler_i:
@ -1344,7 +1350,7 @@ extern "C" {
float tau,
float eta);
/// @details Intializes a GBNF grammar, see grammars/README.md for details.
/// @details Initializes a GBNF grammar, see grammars/README.md for details.
/// @param vocab The vocabulary that this grammar will be used with.
/// @param grammar_str The production rules for the grammar, encoded as a string. Returns an empty grammar if empty. Returns NULL if parsing of grammar_str fails.
/// @param grammar_root The name of the start symbol for the grammar.
@ -1395,6 +1401,33 @@ extern "C" {
const char ** seq_breakers,
size_t num_breakers);
/// adaptive-p: select tokens near a configurable target probability over time.
///
/// the adaptive-p sampler transforms the token probability distribution to favor tokens
/// that fall near a user-configurable probability target.
///
/// internally, the sampler maintains an exponential moving average of the *ORIGINAL*
/// probabilities of selected tokens at each sampling step. it uses this EMA to compute an
/// adapted target probability at each sampling step, thus maintaining the desired target
/// probability over time.
///
/// adaptive-p selects a token ID rather than just mutating candidates, so it must be last
/// in the sampler chain (like mirostat, dist, greedy).
///
/// only mild truncation before this sampler is recommended. we suggest applying min-p
/// before adaptive-p as the only other active sampler in the chain.
///
/// @param target select tokens near this probability (valid range 0.0 to 1.0; negative = disabled)
/// @param decay EMA decay for adaptation; history ≈ 1/(1-decay) tokens (valid range 0.0 - 0.99)
/// @param seed RNG seed
///
/// ref: https://github.com/ggml-org/llama.cpp/pull/17927
///
LLAMA_API struct llama_sampler * llama_sampler_init_adaptive_p(
float target,
float decay,
uint32_t seed);
LLAMA_API struct llama_sampler * llama_sampler_init_logit_bias(
int32_t n_vocab,
int32_t n_logit_bias,
@ -1448,12 +1481,12 @@ extern "C" {
/// @details Build a split GGUF final path for this chunk.
/// llama_split_path(split_path, sizeof(split_path), "/models/ggml-model-q4_0", 2, 4) => split_path = "/models/ggml-model-q4_0-00002-of-00004.gguf"
// Returns the split_path length.
LLAMA_API int llama_split_path(char * split_path, size_t maxlen, const char * path_prefix, int split_no, int split_count);
LLAMA_API int32_t llama_split_path(char * split_path, size_t maxlen, const char * path_prefix, int32_t split_no, int32_t split_count);
/// @details Extract the path prefix from the split_path if and only if the split_no and split_count match.
/// llama_split_prefix(split_prefix, 64, "/models/ggml-model-q4_0-00002-of-00004.gguf", 2, 4) => split_prefix = "/models/ggml-model-q4_0"
// Returns the split_prefix length.
LLAMA_API int llama_split_prefix(char * split_prefix, size_t maxlen, const char * split_path, int split_no, int split_count);
LLAMA_API int32_t llama_split_prefix(char * split_prefix, size_t maxlen, const char * split_path, int32_t split_no, int32_t split_count);
// Print system information
LLAMA_API const char * llama_print_system_info(void);

5
examples/talk-llama/models/afmoe.cpp
View File

@ -1,8 +1,8 @@
#include "models.h"
llm_build_afmoe::llm_build_afmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
const int64_t n_embd_head = hparams.n_embd_head_v();
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
ggml_tensor * cur;
ggml_tensor * inpL;
@ -127,7 +127,6 @@ llm_build_afmoe::llm_build_afmoe(const llama_model & model, const llm_graph_para
n_expert, n_expert_used,
LLM_FFN_SILU,
hparams.expert_weights_norm, // norm_w (route_norm=True)
hparams.expert_weights_scale, // scale_w
hparams.expert_weights_scale, // w_scale (route_scale=2.826)
(llama_expert_gating_func_type) hparams.expert_gating_func,
il);

6
examples/talk-llama/models/apertus.cpp
View File

@ -3,10 +3,10 @@
llm_build_apertus::llm_build_apertus(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_head = hparams.n_embd_head_v();
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_rot);
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
GGML_ASSERT(n_embd_head == n_rot);
ggml_tensor * cur;
ggml_tensor * inpL;

6
examples/talk-llama/models/arcee.cpp
View File

@ -2,10 +2,10 @@
llm_build_arcee::llm_build_arcee(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_head = hparams.n_embd_head_v();
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_rot);
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
GGML_ASSERT(n_embd_head == n_rot);
ggml_tensor * cur;
ggml_tensor * inpL;

9
examples/talk-llama/models/arctic.cpp
View File

@ -1,11 +1,10 @@
#include "models.h"
llm_build_arctic::llm_build_arctic(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_head = hparams.n_embd_head_v();
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_rot);
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
GGML_ASSERT(n_embd_head == n_rot);
ggml_tensor * cur;
ggml_tensor * inpL;
@ -104,7 +103,7 @@ llm_build_arctic::llm_build_arctic(const llama_model & model, const llm_graph_pa
nullptr,
n_expert, n_expert_used,
LLM_FFN_SILU, true,
false, 0.0,
hparams.expert_weights_scale,
LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
il);
cb(cur, "ffn_moe_out", il);

7
examples/talk-llama/models/baichuan.cpp
View File

@ -2,10 +2,10 @@
llm_build_baichuan::llm_build_baichuan(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_head = hparams.n_embd_head_v();
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_rot);
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
GGML_ASSERT(n_embd_head == n_rot);
ggml_tensor * cur;
ggml_tensor * inpL;
@ -56,6 +56,7 @@ llm_build_baichuan::llm_build_baichuan(const llama_model & model, const llm_grap
);
break;
case LLM_TYPE_13B:
case LLM_TYPE_UNKNOWN:
break;
default:
GGML_ABORT("fatal error");

3
examples/talk-llama/models/bailingmoe.cpp
View File

@ -1,6 +1,5 @@
#include "models.h"
llm_build_bailingmoe::llm_build_bailingmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
ggml_tensor * cur;
ggml_tensor * inpL;
@ -97,7 +96,7 @@ llm_build_bailingmoe::llm_build_bailingmoe(const llama_model & model, const llm_
nullptr,
n_expert, n_expert_used,
LLM_FFN_SILU, hparams.expert_weights_norm,
false, hparams.expert_weights_scale,
hparams.expert_weights_scale,
LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
il);
cb(moe_out, "ffn_moe_out", il);

8
examples/talk-llama/models/bailingmoe2.cpp
View File

@ -1,13 +1,11 @@
#include "models.h"
llm_build_bailingmoe2::llm_build_bailingmoe2(const llama_model & model, const llm_graph_params & params) :
llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_head = hparams.n_embd_head_v();
const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
ggml_tensor * cur;
ggml_tensor * inpL;
@ -90,7 +88,7 @@ llm_build_bailingmoe2::llm_build_bailingmoe2(const llama_model & model, const ll
model.layers[il].ffn_exp_probs_b,
n_expert, n_expert_used,
LLM_FFN_SILU, hparams.expert_weights_norm,
true, hparams.expert_weights_scale,
hparams.expert_weights_scale,
(llama_expert_gating_func_type) hparams.expert_gating_func,
il);
cb(moe_out, "ffn_moe_out", il);

20
examples/talk-llama/models/bert.cpp
View File

@ -1,12 +1,10 @@
#include "models.h"
llm_build_bert::llm_build_bert(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_head = hparams.n_embd_head_v();
const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
ggml_tensor * cur;
ggml_tensor * inpL;
@ -129,9 +127,17 @@ llm_build_bert::llm_build_bert(const llama_model & model, const llm_graph_params
// feed-forward network
if (hparams.moe_every_n_layers > 0 && il % hparams.moe_every_n_layers == 1) {
// MoE branch
cur = build_moe_ffn(cur, model.layers[il].ffn_gate_inp, model.layers[il].ffn_up_exps, nullptr,
model.layers[il].ffn_down_exps, nullptr, hparams.n_expert, hparams.n_expert_used,
LLM_FFN_GELU, false, false, 0.0f, LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il);
cur = build_moe_ffn(cur,
model.layers[il].ffn_gate_inp,
model.layers[il].ffn_up_exps,
nullptr,
model.layers[il].ffn_down_exps,
nullptr,
hparams.n_expert, hparams.n_expert_used,
LLM_FFN_GELU, false,
hparams.expert_weights_scale,
LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
il);
cb(cur, "ffn_moe_out", il);
} else if (model.arch == LLM_ARCH_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE ||
model.arch == LLM_ARCH_JINA_BERT_V3) {

33
examples/talk-llama/models/bitnet.cpp
View File

@ -2,9 +2,9 @@
llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_head = hparams.n_embd_head_v();
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
ggml_tensor * cur;
ggml_tensor * inpL;
@ -29,10 +29,7 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
// self-attention
{
// compute Q and K and RoPE them
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
if (model.layers[il].wq_scale) {
Qcur = ggml_mul(ctx0, Qcur, model.layers[il].wq_scale);
}
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur, model.layers[il].wq_s);
cb(Qcur, "Qcur", il);
if (model.layers[il].bq) {
Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
@ -40,10 +37,7 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
}
// B1.K
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
if (model.layers[il].wk_scale) {
Kcur = ggml_mul(ctx0, Kcur, model.layers[il].wk_scale);
}
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur, model.layers[il].wk_s);
cb(Kcur, "Kcur", il);
if (model.layers[il].bk) {
Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
@ -51,10 +45,7 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
}
// B1.V
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
if (model.layers[il].wv_scale) {
Vcur = ggml_mul(ctx0, Vcur, model.layers[il].wv_scale);
}
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur, model.layers[il].wv_s);
cb(Vcur, "Vcur", il);
if (model.layers[il].bv) {
Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
@ -90,10 +81,7 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
LLM_NORM_RMS, il);
cb(cur, "attn_sub_norm", il);
cur = build_lora_mm(model.layers[il].wo, cur);
if (model.layers[il].wo_scale) {
cur = ggml_mul(ctx0, cur, model.layers[il].wo_scale);
}
cur = build_lora_mm(model.layers[il].wo, cur, model.layers[il].wo_s);
if (model.layers[il].bo) {
cur = ggml_add(ctx0, cur, model.layers[il].bo);
}
@ -115,8 +103,8 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
cb(cur, "ffn_norm", il);
cur = build_ffn(cur,
model.layers[il].ffn_up, NULL, model.layers[il].ffn_up_scale,
model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_scale,
model.layers[il].ffn_up, NULL, model.layers[il].ffn_up_s,
model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_s,
NULL, NULL, NULL,
NULL,
LLM_FFN_SILU, LLM_FFN_PAR, il);
@ -127,10 +115,7 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
LLM_NORM_RMS, il);
cb(cur, "ffn_sub_norm", il);
cur = build_lora_mm(model.layers[il].ffn_down, cur);
if (model.layers[il].ffn_down_scale) {
cur = ggml_mul(ctx0, cur, model.layers[il].ffn_down_scale);
}
cur = build_lora_mm(model.layers[il].ffn_down, cur, model.layers[il].ffn_down_s);
cb(cur, "ffn_down", il);
cur = ggml_add(ctx0, cur, ffn_inp);

4
examples/talk-llama/models/bloom.cpp
View File

@ -1,10 +1,10 @@
#include "models.h"
llm_build_bloom::llm_build_bloom(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_head = hparams.n_embd_head_v();
const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
ggml_tensor * cur;
ggml_tensor * inpL;

6
examples/talk-llama/models/chameleon.cpp
View File

@ -3,10 +3,10 @@
#include <float.h>
llm_build_chameleon::llm_build_chameleon(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_head = hparams.n_embd_head_v();
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_rot);
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
GGML_ASSERT(n_embd_head == n_rot);
ggml_tensor * cur;
ggml_tensor * inpL;

4
examples/talk-llama/models/chatglm.cpp
View File

@ -2,10 +2,10 @@
llm_build_chatglm::llm_build_chatglm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_head = hparams.n_embd_head_v();
const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
ggml_tensor * cur;
ggml_tensor * inpL;

6
examples/talk-llama/models/codeshell.cpp
View File

@ -1,11 +1,11 @@
#include "models.h"
llm_build_codeshell::llm_build_codeshell(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_head = hparams.n_embd_head_v();
const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_rot);
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
GGML_ASSERT(n_embd_head == n_rot);
ggml_tensor * cur;
ggml_tensor * inpL;

6
examples/talk-llama/models/cogvlm.cpp
View File

@ -2,11 +2,11 @@
llm_build_cogvlm::llm_build_cogvlm(const llama_model & model, const llm_graph_params & params) :
llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_head = hparams.n_embd_head_v();
const float kq_scale = 1.0f / sqrtf(float(n_embd_head));
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_rot);
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
GGML_ASSERT(n_embd_head == n_rot);
ggml_tensor * inpL;
ggml_tensor * cur;

4
examples/talk-llama/models/cohere2-iswa.cpp
View File

@ -1,9 +1,9 @@
#include "models.h"
llm_build_cohere2_iswa::llm_build_cohere2_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_head = hparams.n_embd_head_v();
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
const float f_logit_scale = hparams.f_logit_scale;

4
examples/talk-llama/models/command-r.cpp
View File

@ -4,9 +4,9 @@
llm_build_command_r::llm_build_command_r(const llama_model & model, const llm_graph_params & params) :
llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_head = hparams.n_embd_head_v();
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
const float f_logit_scale = hparams.f_logit_scale;

9
examples/talk-llama/models/dbrx.cpp
View File

@ -1,12 +1,11 @@
#include "models.h"
llm_build_dbrx::llm_build_dbrx(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_head = hparams.n_embd_head_v();
const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_rot);
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
GGML_ASSERT(n_embd_head == n_rot);
ggml_tensor * cur;
ggml_tensor * inpL;
@ -89,7 +88,7 @@ llm_build_dbrx::llm_build_dbrx(const llama_model & model, const llm_graph_params
nullptr,
n_expert, n_expert_used,
LLM_FFN_SILU, true,
false, 0.0,
hparams.expert_weights_scale,
LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
il);
cb(cur, "ffn_moe_out", il);

6
examples/talk-llama/models/deci.cpp
View File

@ -3,10 +3,10 @@
llm_build_deci::llm_build_deci(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_head = hparams.n_embd_head_v();
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_rot);
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
GGML_ASSERT(n_embd_head == n_rot);
ggml_tensor * cur;
ggml_tensor * inpL;

10
examples/talk-llama/models/deepseek.cpp
View File

@ -1,13 +1,11 @@
#include "models.h"
llm_build_deepseek::llm_build_deepseek(const llama_model & model, const llm_graph_params & params) :
llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_head = hparams.n_embd_head_v();
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_rot);
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
GGML_ASSERT(n_embd_head == n_rot);
ggml_tensor * cur;
ggml_tensor * inpL;
@ -100,7 +98,7 @@ llm_build_deepseek::llm_build_deepseek(const llama_model & model, const llm_grap
nullptr,
n_expert, n_expert_used,
LLM_FFN_SILU, false,
false, hparams.expert_weights_scale,
hparams.expert_weights_scale,
LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
il);
cb(moe_out, "ffn_moe_out", il);

45
examples/talk-llama/models/deepseek2.cpp
View File

@ -2,22 +2,19 @@
llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_graph_params & params) :
llm_graph_context(params) {
// lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B
bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26);
const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0);
const bool is_mla = hparams.is_mla();
// note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
const int64_t n_embd_head_k = is_mla ? hparams.n_embd_head_k_mla : hparams.n_embd_head_k;
const int64_t n_embd_head_v = is_mla ? hparams.n_embd_head_v_mla : hparams.n_embd_head_v;
const int64_t n_embd_head_k = hparams.n_embd_head_k_mla();
const int64_t n_embd_head_v = hparams.n_embd_head_v_mla();
const int64_t n_embd_head_qk_rope = hparams.n_rot;
const int64_t n_embd_head_qk_rope = hparams.n_rot();
const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope;
const uint32_t kv_lora_rank = hparams.n_lora_kv;
// We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly.
// See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation.
// See https://github.com/ggml-org/llama.cpp/discussions/7416 for detailed explanation.
// And also: https://github.com/ggml-org/llama.cpp/pull/17945 [TAG_DEEPSEEK2_YARN_LOG_MUL_FIX]
// first cancel the adjustment from llama_hparams::yarn_attn_factor_adjust to get the original attn_factor
@ -43,11 +40,13 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
// inp_pos - contains the positions
ggml_tensor * inp_pos = build_inp_pos();
auto * inp_attn = build_attn_inp_kv();
auto * inp_attn_kv = !is_mla ? build_attn_inp_kv() : nullptr;
auto * inp_attn_k = is_mla ? build_attn_inp_k() : nullptr;
ggml_tensor * inp_out_ids = build_inp_out_ids();
for (int il = 0; il < n_layer; ++il) {
int effective_n_layers = hparams.n_layer - hparams.nextn_predict_layers;
for (int il = 0; il < effective_n_layers; ++il) {
ggml_tensor * inpSA = inpL;
// norm
@ -57,6 +56,9 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
// self_attention
{
ggml_tensor * q = NULL;
const bool is_lite = model.layers[il].wq;
if (!is_lite) {
q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur);
cb(q, "q", il);
@ -124,14 +126,14 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
// {n_embd_head_qk_rope + kv_lora_rank, n_head, n_tokens}
// note: rope must go first for in-place context shifting in build_rope_shift()
ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope_absorbed, 0);
ggml_tensor * Qcur = ggml_concat(ctx0, q_nope_absorbed, q_pe, 0);
cb(Qcur, "Qcur", il);
kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens);
cb(kv_cmpr, "kv_cmpr_reshape", il);
// {n_embd_head_qk_rope + kv_lora_rank, 1, n_tokens}
ggml_tensor * Kcur = ggml_concat(ctx0, k_pe, kv_cmpr, 0);
ggml_tensor * Kcur = ggml_concat(ctx0, kv_cmpr, k_pe, 0);
cb(Kcur, "Kcur", il);
// {kv_lora_rank, 1, n_tokens}
@ -144,8 +146,8 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
cb(Qcur, "Qcur_attn_temp_scaled", il);
}
// note: MLA with the absorption optimzation converts into MQA (ie: GQA with 1 group)
cur = build_attn(inp_attn,
// note: MLA with the absorption optimization converts into MQA (ie: GQA with 1 group)
cur = build_attn(inp_attn_k,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, kq_scale, il);
} else {
@ -169,11 +171,10 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
Vcur = ggml_cont(ctx0, Vcur);
cb(Vcur, "Vcur_cont", il);
// note: rope must go first for in-place context shifting in build_rope_shift()
ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope, 0);
ggml_tensor * Qcur = ggml_concat(ctx0, q_nope, q_pe, 0);
cb(Qcur, "Qcur", il);
ggml_tensor * Kcur = ggml_concat(ctx0, ggml_repeat(ctx0, k_pe, q_pe), k_nope, 0);
ggml_tensor * Kcur = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0);
cb(Kcur, "Kcur", il);
if (inp_attn_scale) {
@ -183,12 +184,12 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
}
// note: MLA without the absorption optimization converts into MHA (ie: GQA with full n_head groups)
cur = build_attn(inp_attn,
cur = build_attn(inp_attn_kv,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
}
}
if (il == n_layer - 1 && inp_out_ids) {
if (il == effective_n_layers - 1 && inp_out_ids) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
}
@ -215,9 +216,11 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
model.layers[il].ffn_exp_probs_b,
n_expert, n_expert_used,
LLM_FFN_SILU, hparams.expert_weights_norm,
hparams.expert_weights_scale, hparams.expert_weights_scale,
hparams.expert_weights_scale,
(llama_expert_gating_func_type) hparams.expert_gating_func,
il);
il,
nullptr,
model.layers[il].ffn_gate_up_exps);
cb(moe_out, "ffn_moe_out", il);
// FFN shared expert

445
examples/talk-llama/models/delta-net-base.cpp Normal file
View File

@ -0,0 +1,445 @@
#include "models.h"
#include "llama-impl.h"
// utility to get one slice from the third dimension
// input dim: [x, y, c, b]
// output dim: [x, y, 1, b]
static ggml_tensor * get_slice_2d(ggml_context * ctx0, ggml_tensor * t, int64_t c) {
return ggml_view_4d(ctx0, t, t->ne[0], t->ne[1], 1, t->ne[3],
t->nb[1], t->nb[2], t->nb[3], t->nb[2] * c);
}
llm_build_delta_net_base::llm_build_delta_net_base(const llm_graph_params & params) : llm_graph_context(params) {}
std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_net_chunking(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * b,
ggml_tensor * s,
int il) {
const int64_t S_k = q->ne[0];
const int64_t H_k = q->ne[1];
const int64_t n_tokens = q->ne[2];
const int64_t n_seqs = q->ne[3];
const int64_t S_v = v->ne[0];
const int64_t H_v = v->ne[1];
const bool kda = (g->ne[0] == S_k && g->ne[1] == H_k);
GGML_ASSERT(S_k == S_v);
GGML_ASSERT(H_v % H_k == 0);
GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
GGML_ASSERT(v->ne[0] == S_v && v->ne[1] == H_v && v->ne[2] == n_tokens && v->ne[3] == n_seqs);
GGML_ASSERT(g->ne[0] == 1 || g->ne[0] == S_v);
GGML_ASSERT( g->ne[1] == H_v && g->ne[2] == n_tokens && g->ne[3] == n_seqs);
GGML_ASSERT(b->ne[0] == 1 && b->ne[1] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v && s->ne[3] == n_seqs);
const float scale = 1.0f / sqrtf(S_k);
q = ggml_scale(ctx0, q, scale);
cb(q, "q_in", il);
cb(k, "k_in", il);
cb(v, "v_in", il);
cb(b, "b_in", il);
cb(g, "g_in", il);
q = ggml_permute(ctx0, q, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs]
k = ggml_permute(ctx0, k, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs]
v = ggml_permute(ctx0, v, 0, 2, 1, 3); // [S_v, n_tokens, H_v, n_seqs]
g = ggml_permute(ctx0, g, 0, 2, 1, 3); // [g_0, n_tokens, H_v, n_seqs]
b = ggml_permute(ctx0, b, 0, 2, 1, 3); // [ 1, n_tokens, H_v, n_seqs]
const int CS = kda ? 16 : 64; // chunk size
const int pad = (CS - n_tokens % CS) % CS;
const int n_chunks = (n_tokens + pad) / CS;
q = ggml_pad(ctx0, q, 0, pad, 0, 0);
k = ggml_pad(ctx0, k, 0, pad, 0, 0);
v = ggml_pad(ctx0, v, 0, pad, 0, 0);
g = ggml_pad(ctx0, g, 0, pad, 0, 0);
b = ggml_pad(ctx0, b, 0, pad, 0, 0);
ggml_tensor * v_b = ggml_mul(ctx0, v, b);
ggml_tensor * k_b = ggml_mul(ctx0, k, b);
cb(v_b, "v_b", il);
cb(k_b, "k_b", il);
q = ggml_reshape_4d(ctx0, q, S_k, CS, n_chunks, H_k * n_seqs);
k = ggml_reshape_4d(ctx0, k, S_k, CS, n_chunks, H_k * n_seqs);
k_b = ggml_reshape_4d(ctx0, k_b, S_k, CS, n_chunks, H_v * n_seqs);
v = ggml_reshape_4d(ctx0, v, S_v, CS, n_chunks, H_v * n_seqs);
v_b = ggml_reshape_4d(ctx0, v_b, S_v, CS, n_chunks, H_v * n_seqs);
g = ggml_reshape_4d(ctx0, g, g->ne[0], CS, n_chunks, H_v * n_seqs);
b = ggml_reshape_4d(ctx0, b, 1, CS, n_chunks, H_v * n_seqs);
// [CS, g_0, n_chunks, H_v * n_seqs]
// TODO: extend ggml_cumsum with axis parameter to avoid transpose
ggml_tensor * g_cs = ggml_cumsum(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, g)));
cb(g_cs, "g_cs", il);
ggml_tensor * kb = nullptr;
ggml_tensor * kq = nullptr;
if (kda) {
const int64_t CHB = n_chunks * H_k * n_seqs;
ggml_tensor * g_cs_i = ggml_reshape_4d(ctx0, g_cs, CS, 1, S_k, CHB); // [chunk_size, 1, S_k, CHB]
ggml_tensor * g_cs_j = ggml_reshape_4d(ctx0, g_cs, 1, CS, S_k, CHB); // [1, chunk_size, S_k, CHB]
g_cs_j = ggml_repeat_4d(ctx0, g_cs_j, CS, CS, S_k, CHB); // [1, chunk_size, S_k, CHB] -> [chunk_size, chunk_size, S_k, CHB]
// decay_mask [chunk_size,chunk_size,S_k,CHB]
ggml_tensor * decay_mask;
decay_mask = ggml_sub(ctx0, g_cs_j, g_cs_i);
decay_mask = ggml_tri(ctx0, decay_mask, GGML_TRI_TYPE_LOWER_DIAG);
decay_mask = ggml_exp(ctx0, decay_mask);
cb(decay_mask, "decay_mask", il);
// decay_mask [S_k,BT_j,BT_i,CHB] *Note* second and third chunk_sizes are switched
decay_mask = ggml_cont_4d(ctx0, ggml_permute(ctx0, decay_mask, 2, 1, 0, 3), S_k, CS, CS, CHB);
ggml_tensor * k_b_i = ggml_reshape_4d(ctx0, k_b, S_k, CS, 1, CHB);
ggml_tensor * k_j = ggml_reshape_4d(ctx0, k, S_k, 1, CS, CHB);
ggml_tensor * q_i = ggml_reshape_4d(ctx0, q, S_k, CS, 1, CHB);
ggml_tensor * decay_k_b_i = ggml_mul(ctx0, decay_mask, k_b_i);
ggml_tensor * decay_q_i = ggml_mul(ctx0, decay_mask, q_i);
// decay_k_b_i [S,BT,BT,CHB] @ k_j [S,1,BT,CHB] = Akk [BT,1,BT,CHB]
kb = ggml_mul_mat(ctx0, decay_k_b_i, k_j);
kq = ggml_mul_mat(ctx0, decay_q_i, k_j);
kb = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_4d(ctx0, kb, CS, CS, n_chunks, H_v * n_seqs)));
kq = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_4d(ctx0, kq, CS, CS, n_chunks, H_v * n_seqs)));
} else {
ggml_tensor * g_cs_i = g_cs;
ggml_tensor * g_cs_j = ggml_reshape_4d(ctx0, g_cs, 1, CS, n_chunks, H_v * n_seqs);
g_cs_j = ggml_repeat_4d(ctx0, g_cs_j, CS, CS, n_chunks, H_v * n_seqs);
// [CS, CS, n_chunks, H_v * n_seqs]
ggml_tensor * decay_mask;
decay_mask = ggml_sub(ctx0, g_cs_j, g_cs_i);
decay_mask = ggml_tri(ctx0, decay_mask, GGML_TRI_TYPE_LOWER_DIAG);
decay_mask = ggml_exp(ctx0, decay_mask);
cb(decay_mask, "decay_mask", il);
// [CS, CS, n_chunks, H_k * n_seqs]
kb = ggml_mul_mat(ctx0, k, k_b);
kb = ggml_mul (ctx0, kb, decay_mask);
// [CS, CS, n_chunks, H_k * n_seqs]
kq = ggml_mul_mat(ctx0, k, q);
kq = ggml_mul(ctx0, kq, decay_mask);
}
kq = ggml_tri(ctx0, kq, GGML_TRI_TYPE_LOWER_DIAG);
cb(kq, "kq", il);
// [CS, CS, n_chunks, H_k * n_seqs]
ggml_tensor * attn;
attn = ggml_tri(ctx0, kb, GGML_TRI_TYPE_LOWER);
cb(attn, "attn", il);
ggml_tensor * identity;
identity = ggml_view_1d(ctx0, attn, CS, 0);
identity = ggml_fill (ctx0, identity, 1.0f);
identity = ggml_diag (ctx0, identity);
ggml_tensor * lhs = ggml_add(ctx0, attn, identity);
cb(lhs, "dnet_add_ch_lhs", il);
attn = ggml_neg(ctx0, attn);
cb(attn, "attn_pre_solve", il);
ggml_tensor * lin_solve = ggml_solve_tri(ctx0, lhs, attn, true, true, false);
attn = ggml_add(ctx0, lin_solve, identity);
cb(attn, "dnet_add_ch_attn_solved", il); // [CS, CS, n_chunks, H_k * n_seqs]
// [S_v, CS, n_chunks, H_v * n_seqs]
v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_b)), attn);
// [CS, 1, n_chunks, H_v * n_seqs] KDA: [CS, S_k, n_chunks, H_v * n_seqs]
ggml_tensor * g_exp = ggml_exp(ctx0, g_cs);
k_b = ggml_cont(ctx0, ggml_transpose(ctx0, k_b));
// [CS, S_k, n_chunks, H_k * n_seqs]
ggml_tensor * kbg = ggml_mul(ctx0, k_b, g_exp);
cb(kbg, "k_beta_g_exp", il);
// [S_k, CS, n_chunks, H_k * n_seqs]
ggml_tensor * k_cd = ggml_mul_mat(ctx0, kbg, attn);
cb(k_cd, "k_cumdecay", il);
// [1, CS, n_chunks, H_k * n_seqs] KDA: [S_k, CS, n_chunks, H_k * n_seqs]
ggml_tensor * g_exp_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_exp));
ggml_tensor * q_g_exp = ggml_mul(ctx0, q, g_exp_t);
// vectorized calculation of key_gdiff
// improved from the chunked version:
// g_last = torch.clamp(g_cum[:, :, -1], max=50.0).exp().unsqueeze(-1).unsqueeze(-1)
// g_diff = torch.clamp(g_cum[:, :, -1:] - g_cum, max=50.0).exp()
// key_gdiff = key * g_diff.unsqueeze(-1)
// kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
// last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
// get last element in g_cumsum along CS dimension (ne0)
// example: [[x, y, z, ..., last], ...] -> [[last], ...]
// [1, 1, n_chunks, H_v * n_seqs] KDA: [1, S_k, n_chunks, H_v * n_seqs]
ggml_tensor * g_last = ggml_view_4d(ctx0, g_cs, 1, g_cs->ne[1], g_cs->ne[2], g_cs->ne[3],
g_cs->nb[1],
g_cs->nb[2],
g_cs->nb[3],
ggml_row_size(g_cs->type, g_cs->ne[0] - 1));
cb(g_last, "g_last", il);
// TODO: remove this cont when CUDA supports non-cont unary ops
g_last = ggml_cont(ctx0, g_last);
// [1, 1, n_chunks, H_v * n_seqs] KDA: [S_k, 1, n_chunks, H_v * n_seqs]
ggml_tensor * g_last_exp_t = ggml_transpose(ctx0, ggml_exp(ctx0, g_last));
cb(g_last_exp_t, "g_last_exp_t", il);
// [CS, 1, n_chunks, H_v * n_seqs] KDA: [CS, S_k, n_chunks, H_v * n_seqs]
ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cs, g_last));
cb(g_diff, "g_diff", il);
ggml_tensor * g_diff_exp_t = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_exp(ctx0, g_diff)));
// [S_k, CS, n_chunks, H_v * n_seqs]
ggml_tensor * kg = ggml_mul(ctx0, k, g_diff_exp_t);
cb(kg, "key_gdiff", il);
// [CS, S_k, n_chunks, H_v * n_seqs]
ggml_tensor * kg_t = ggml_cont(ctx0, ggml_transpose(ctx0, kg));
cb(kg_t, "key_gdiff_t", il);
s = ggml_reshape_4d(ctx0, s, S_v, S_v, 1, H_v * n_seqs);
cb(s, "dnet_add_ch_state", il);
// [CS, S_v, n_chunks, H_v * n_seqs]
ggml_tensor * v_t = ggml_cont(ctx0, ggml_transpose(ctx0, v));
for (int64_t chunk = 0; chunk < n_chunks; chunk++) {
ggml_tensor * ch_k_cd = get_slice_2d(ctx0, k_cd, chunk); // [S_k, CS, 1, H_k * n_seqs]
ggml_tensor * ch_v_t = get_slice_2d(ctx0, v_t, chunk); // [ CS, S_v, 1, H_v * n_seqs]
ggml_tensor * ch_kq = get_slice_2d(ctx0, kq, chunk); // [ CS, CS, 1, H_k * n_seqs]
ggml_tensor * ch_q_g_exp = get_slice_2d(ctx0, q_g_exp, chunk); // [S_k, CS, 1, H_k * n_seqs]
ggml_tensor * ch_kg_t = get_slice_2d(ctx0, kg_t, chunk); // [ CS, S_k, 1, H_v * n_seqs]
// [CS, S_v, 1, H_v * n_seqs]
ggml_tensor * v_t_p = ggml_mul_mat(ctx0, ch_k_cd, s);
cb(v_t_p, "v_prime", il);
// [CS, S_v, 1, H_v * n_seqs]
ggml_tensor * v_t_new = ggml_sub(ctx0, ch_v_t, v_t_p);
cb(v_t_new, "v_t_new", il);
// [S_v, CS, 1, H_v * n_seqs]
ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_t_new, ch_kq);
cb(v_attn, "v_attn", il);
// [S_v, CS, 1, H_v * n_seqs]
ggml_tensor * attn_inter = ggml_mul_mat(ctx0, s, ch_q_g_exp);
cb(attn_inter, "attn_inter", il);
// [S_v, CS, 1, H_v * n_seqs]
ggml_tensor * o_ch = ggml_add(ctx0, attn_inter, v_attn);
cb(o_ch, "dnet_add_ch_attn_out", il);
v = ggml_set_inplace(ctx0, v, o_ch, v->nb[1], v->nb[2], v->nb[3], chunk * v->nb[2]);
// kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
// TODO: head broadcast might not work here - probably will need a transpose
ggml_tensor * kgv = ggml_mul_mat(ctx0, ch_kg_t, v_t_new); // [S_k, S_v, 1, H_k * n_seqs]
// last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
ggml_tensor * ch_g_last_exp_t = get_slice_2d(ctx0, g_last_exp_t, chunk);
s = ggml_mul(ctx0, s, ch_g_last_exp_t);
s = ggml_add(ctx0, s, kgv);
cb(s, "dnet_add_ch_state", il);
}
// truncate padded tokens
ggml_tensor * o = ggml_view_4d(ctx0, v,
S_v, n_tokens, H_v, n_seqs,
ggml_row_size(v->type, S_v),
ggml_row_size(v->type, S_v * CS * n_chunks),
ggml_row_size(v->type, S_v * CS * n_chunks * H_v), 0);
o = ggml_permute (ctx0, o, 0, 2, 1, 3); // [S_v, H_v, n_tokens, n_seqs]
s = ggml_reshape_4d(ctx0, s, S_v, S_v, H_v, n_seqs);
cb(s, "output_state", il);
return {o, s};
}
std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_net_autoregressive(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * b, // beta
ggml_tensor * s, // state
int il) {
const int64_t S_k = q->ne[0];
const int64_t H_k = q->ne[1];
const int64_t n_tokens = q->ne[2];
const int64_t n_seqs = q->ne[3];
const int64_t S_v = v->ne[0];
const int64_t H_v = v->ne[1];
GGML_ASSERT(n_tokens == 1);
GGML_ASSERT(S_k == S_v);
GGML_ASSERT(H_v % H_k == 0);
GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
GGML_ASSERT(v->ne[0] == S_v && v->ne[1] == H_v && v->ne[2] == n_tokens && v->ne[3] == n_seqs);
GGML_ASSERT(g->ne[0] == 1 || g->ne[0] == S_v);
GGML_ASSERT( g->ne[1] == H_v && g->ne[2] == n_tokens && g->ne[3] == n_seqs);
GGML_ASSERT(b->ne[0] == 1 && b->ne[1] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v && s->ne[3] == n_seqs);
const float scale = 1.0f / sqrtf(S_k);
q = ggml_scale(ctx0, q, scale);
q = ggml_permute(ctx0, q, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs]
k = ggml_permute(ctx0, k, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs]
v = ggml_permute(ctx0, v, 0, 2, 1, 3); // [S_v, n_tokens, H_v, n_seqs]
cb(q, "q_in", il);
cb(k, "k_in", il);
cb(v, "v_in", il);
cb(b, "b_in", il);
cb(g, "g_in", il);
// GDA: [1, 1, H_v, n_seqs]
// KDA: [1, S_k, H_v, n_seqs]
g = ggml_reshape_4d(ctx0, g, 1, g->ne[0], H_v, n_seqs);
b = ggml_reshape_4d(ctx0, b, 1, 1, H_v, n_seqs);
// [S_v, S_v, H_v, n_seqs]
g = ggml_exp(ctx0, g);
s = ggml_mul(ctx0, s, g);
// [1, S_v, H_v, n_seqs]
ggml_tensor * sk;
sk = ggml_mul (ctx0, s, k);
sk = ggml_sum_rows(ctx0, sk);
// [S_v, 1, H_v, n_seqs]
ggml_tensor * d;
d = ggml_sub(ctx0, v, ggml_transpose(ctx0, sk));
d = ggml_mul(ctx0, d, b);
// [1, S_v, H_v, n_seqs]
ggml_tensor * d_t;
d_t = ggml_transpose(ctx0, d);
// [S_v, S_v, H_v, n_seqs]
ggml_tensor * kd;
k = ggml_repeat(ctx0, k, s);
kd = ggml_mul (ctx0, k, d_t);
s = ggml_add(ctx0, s, kd);
cb(s, "dnet_add_ar_state", il);
ggml_tensor * s_q = ggml_mul (ctx0, s, q);
ggml_tensor * o = ggml_sum_rows(ctx0, s_q);
o = ggml_permute (ctx0, o, 2, 0, 1, 3); // [S_v, H_v, n_tokens, n_seqs]
return {o, s};
}
std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_net_fused(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * b,
ggml_tensor * s,
int il) {
const int64_t S_k = q->ne[0];
const int64_t H_k = q->ne[1];
const int64_t n_tokens = q->ne[2];
const int64_t n_seqs = q->ne[3];
const int64_t S_v = v->ne[0];
const int64_t H_v = v->ne[1];
GGML_ASSERT(S_k == S_v);
GGML_ASSERT(H_v % H_k == 0);
GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
GGML_ASSERT(v->ne[0] == S_v && v->ne[1] == H_v && v->ne[2] == n_tokens && v->ne[3] == n_seqs);
GGML_ASSERT(g->ne[0] == 1 || g->ne[0] == S_v);
GGML_ASSERT( g->ne[1] == H_v && g->ne[2] == n_tokens && g->ne[3] == n_seqs);
GGML_ASSERT(b->ne[0] == 1 && b->ne[1] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v && s->ne[3] == n_seqs);
ggml_tensor * result = ggml_gated_delta_net(ctx0, q, k, v, g, b, s);
if (n_tokens == 1) {
cb(result, LLAMA_TENSOR_NAME_FGDN_AR, il);
} else {
cb(result, LLAMA_TENSOR_NAME_FGDN_CH, il);
}
ggml_tensor * output = ggml_view_4d(ctx0, result,
S_v, H_v, n_tokens, n_seqs,
ggml_row_size(result->type, S_v),
ggml_row_size(result->type, S_v * H_v),
ggml_row_size(result->type, S_v * H_v * n_tokens), 0);
ggml_tensor * new_state = ggml_view_4d(ctx0, result,
S_v, S_v, H_v, n_seqs,
ggml_row_size(result->type, S_v),
ggml_row_size(result->type, S_v * S_v),
ggml_row_size(result->type, S_v * S_v * H_v),
ggml_row_size(result->type, S_v * H_v * n_tokens * n_seqs));
return {output, new_state};
}
std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_net(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * b,
ggml_tensor * s,
int il) {
const int64_t n_seq_tokens = q->ne[2];
if (n_seq_tokens == 1) {
if (cparams.fused_gdn_ar) {
return build_delta_net_fused(q, k, v, g, b, s, il);
}
return build_delta_net_autoregressive(q, k, v, g, b, s, il);
}
if (cparams.fused_gdn_ch) {
return build_delta_net_fused(q, k, v, g, b, s, il);
}
return build_delta_net_chunking(q, k, v, g, b, s, il);
}

10
examples/talk-llama/models/dots1.cpp
View File

@ -1,13 +1,11 @@
#include "models.h"
llm_build_dots1::llm_build_dots1(const llama_model & model, const llm_graph_params & params) :
llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_head = hparams.n_embd_head_v();
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_rot);
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
GGML_ASSERT(n_embd_head == n_rot);
ggml_tensor * cur;
ggml_tensor * inpL;
@ -91,7 +89,7 @@ llm_build_dots1::llm_build_dots1(const llama_model & model, const llm_graph_para
model.layers[il].ffn_exp_probs_b,
n_expert, n_expert_used,
LLM_FFN_SILU, hparams.expert_weights_norm,
true, hparams.expert_weights_scale,
hparams.expert_weights_scale,
(llama_expert_gating_func_type) hparams.expert_gating_func,
il);
cb(moe_out, "ffn_moe_out", il);

Some files were not shown because too many files have changed in this diff Show More