Merge pull request #620 from Paladinium/docker_cuda_12_8

Docker: Updating to CUDA 12.8, pytorch 2.7 and DeepSpeed 0.17.2

DOCKER_README.md

@@ -3,7 +3,7 @@ The Docker image currently works on Windows and Linux supporting NVIDIA GPUs.
 
 ## System requirements
 - AMD compatible Windows or Linux system
-- CUDA 12.6
+- CUDA 12.8
 - Nvidia Docker Container Toolkit
 - At least 25GB of free disk space
 
@@ -34,7 +34,7 @@ visit `http://localhost:7851/`. See instructions below for passing more argument
 4. Download CUDA toolkit keyring: `wget https://developer.download.nvidia.com/compute/cuda/repos/wsl-ubuntu/x86_64/cuda-keyring_1.1-1_all.deb`
 5. Install keyring: `sudo dpkg -i cuda-keyring_1.1-1_all.deb`
 6. Update package list: `sudo apt-get update`
-7. Install CUDA toolkit: `sudo apt-get -y install cuda-toolkit-12-6`
+7. Install CUDA toolkit: `sudo apt-get -y install cuda-toolkit-12-8`
 8. Install Docker Desktop using WSL2 as the backend
 9. Restart
 10. If you wish to monitor the terminal remotely via SSH, follow [this guide](https://www.hanselman.com/blog/how-to-ssh-into-wsl2-on-windows-10-from-an-external-machine).
@@ -108,3 +108,8 @@ reason, you may want to use `docker-build.sh` with the following arguments:
     - Example: `docker-build.sh --clean`
 
 
+# Older images
+- CUDA 12.6:
+  - XTTS  Dockerhub hash: sha256:fd4741714af0c82e4e7a66bf9ae47c98f3e51ebefdb3af5795b47d41ca5fb159
+  - VITS Dockerhub hash: sha256:9f411e8fdb997486bc217fcd316bf6b163c39736655ea3400028ad4fcc5c60f9
+  - Piper:  Dockerhub hash: sha256:6ae2520051bb00a0356dc4ed80a33bf6784cb567b59759af46046ddcf356b765

Dockerfile

@@ -6,7 +6,7 @@ FROM ${DOCKER_REPOSITORY}alltalk_tts_environment:${DOCKER_TAG}
 ARG TTS_MODEL="xtts"
 ENV TTS_MODEL=$TTS_MODEL
 
-ARG DEEPSPEED_VERSION=0.16.2
+ARG DEEPSPEED_VERSION=0.17.2
 ENV DEEPSPEED_VERSION=$DEEPSPEED_VERSION
 
 ENV GRADIO_SERVER_NAME="0.0.0.0"
@@ -94,7 +94,7 @@ RUN <<EOR
     # Download DeepSpeed wheel if it was not built locally:
     if [ -z "${DEEPSPEED_WHEEL}" ] || [ ! -f $DEEPSPEED_WHEEL ] ; then
       echo "Downloading pre-built DeepSpeed wheel"
-      CURL_ERROR=$( { curl --output-dir /tmp/deepspeed -fLO "https://github.com/erew123/alltalk_tts/releases/download/DeepSpeed-for-docker/deepspeed-0.16.2+b344c04d-cp311-cp311-linux_x86_64.whl" ; } 2>&1 )
+      CURL_ERROR=$( { curl --output-dir /tmp/deepspeed -fLO "https://github.com/erew123/alltalk_tts/releases/download/DeepSpeed-for-docker/deepspeed-0.17.2+15f054d9-cp311-cp311-linux_x86_64.whl" ; } 2>&1 )
       if [ $? -ne 0 ] ; then
         echo "Failed to download DeepSpeed: $CURL_ERROR"
         exit 1

docker/base/Dockerfile

@@ -1,6 +1,6 @@
 FROM continuumio/miniconda3:24.7.1-0 AS build
 
-ARG CUDA_VERSION=12.6.0
+ARG CUDA_VERSION=12.8.1
 ENV CUDA_VERSION=$CUDA_VERSION
 
 ARG PYTHON_VERSION=3.11.11
@@ -53,7 +53,8 @@ RUN <<EOR
       cuda-command-line-tools=${CUDA_VERSION} cuda-compiler=${CUDA_VERSION} cuda-runtime=${CUDA_VERSION} \
       cuda-libraries=${CUDA_VERSION} cuda-libraries-dev=${CUDA_VERSION} cuda-libraries-static=${CUDA_VERSION} \
       cuda-cudart=${CUDA_SHORT_VERSION} cuda-cudart_linux-64=${CUDA_SHORT_VERSION} \
-      cudnn=9.3 \
+      conda-forge::nccl=2.27.5.1 \
+      cudnn=9.10 \
       py-cpuinfo=9 \
       conda-forge::ffmpeg=7.1.0 \
       conda-forge::portaudio=19.7.0 \
@@ -76,7 +77,7 @@ RUN <<EOR
   conda activate alltalk
   mkdir -p ${ALLTALK_DIR}/pip_cache
   pip install --no-cache-dir --cache-dir=${ALLTALK_DIR}/pip_cache \
-      torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu${CUDA_SHORT_VERSION_NO_DOT}
+      torch==2.7.1 torchvision==0.22.1 torchaudio==2.7.1 --index-url https://download.pytorch.org/whl/cu${CUDA_SHORT_VERSION_NO_DOT}
 
   if [ $? -ne 0 ] ; then
     echo "Failed to install pip dependencies: $RESULT"

docker/deepspeed/Dockerfile

@@ -2,7 +2,7 @@ ARG DOCKER_REPOSITORY
 ARG DOCKER_TAG=latest
 FROM ${DOCKER_REPOSITORY}alltalk_tts_environment:${DOCKER_TAG}
 
-ARG DEEPSPEED_VERSION=0.16.3
+ARG DEEPSPEED_VERSION=0.17.2
 ENV DEEPSPEED_VERSION=$DEEPSPEED_VERSION
 
 WORKDIR ${ALLTALK_DIR}
@@ -27,15 +27,6 @@ RUN <<EOR
   cd ${ALLTALK_DIR}/DeepSpeed
   git checkout .
   git checkout "tags/v${DEEPSPEED_VERSION}" -b "v${DEEPSPEED_VERSION}"
-
-  # Patching some files due to bug https://github.com/microsoft/DeepSpeed/issues/6709
-  START_LINE=`awk '/#ifndef BF16_AVAILABLE/{ print NR; exit }' csrc/transformer/inference/csrc/gelu.cu`
-  TO_LINE=$((START_LINE + 2))
-  sed -i "${START_LINE},${TO_LINE}d" ./csrc/transformer/inference/csrc/gelu.cu
-
-  START_LINE=`awk '/#ifndef BF16_AVAILABLE/{ print NR; exit }' csrc/transformer/inference/csrc/transform.cu`
-  TO_LINE=$((START_LINE + 2))
-  sed -i "${START_LINE},${TO_LINE}d" ./csrc/transformer/inference/csrc/transform.cu
 EOR
 
 ##############################################################################

docker/variables.sh

@@ -1,9 +1,9 @@
 #!/usr/bin/env bash
 
 declare -A ALLTALK_VARS
-ALLTALK_VARS["CUDA_VERSION"]=12.6.0
+ALLTALK_VARS["CUDA_VERSION"]=12.8.1
 ALLTALK_VARS["PYTHON_VERSION"]=3.11.11
-ALLTALK_VARS["DEEPSPEED_VERSION"]=0.16.2
+ALLTALK_VARS["DEEPSPEED_VERSION"]=0.17.2
 
 # Export single variables (needed by Docker locally)
 for key in "${!ALLTALK_VARS[@]}"

system/tts_engines/rvc/lib/utils.py

@@ -1,20 +1,20 @@
 import numpy as np
 import re
 import unicodedata
-from fairseq import checkpoint_utils
-from fairseq.data import Dictionary
-import torch
 
+# Next 2 lines are a fix for fairseq to be compatible with pytorch > 2.6
+import sys
+sys.path = [str(p) for p in sys.path]
+
+from fairseq import checkpoint_utils
+import torch
 import logging
 
 logging.getLogger("fairseq").setLevel(logging.WARNING)
 import sys
-import os
 import numpy as np
-import asyncio
 import ffmpeg
 from pathlib import Path
-import subprocess
 
 # Get the parent directory of the current file
 this_dir = Path(__file__).resolve().parent.parent.parent.parent.parent