From 5abe0089cb7bf67b639ee241be3f824d58e86cee Mon Sep 17 00:00:00 2001
From: Jannik Streidl <jannik.stdl@gmail.com>
Date: Mon, 18 Mar 2024 17:08:34 +0100
Subject: [PATCH] cuda support

---
 Dockerfile | 32 +++++++++++++++++++++-----------
 1 file changed, 21 insertions(+), 11 deletions(-)

diff --git a/Dockerfile b/Dockerfile
index 1ccfdb62..33656bba 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -37,7 +37,7 @@ ENV OPENAI_API_KEY="" \
     SCARF_NO_ANALYTICS=true \
     DO_NOT_TRACK=true
 
-#### Preloaded models ##########################################################
+#### Preloaded models #########################################################
 ## whisper TTS Settings ##
 ENV WHISPER_MODEL="base" \
     WHISPER_MODEL_DIR="/app/backend/data/cache/whisper/models"
@@ -48,19 +48,32 @@ ENV WHISPER_MODEL="base" \
 # for better persormance and multilangauge support use "intfloat/multilingual-e5-large" (~2.5GB) or "intfloat/multilingual-e5-base" (~1.5GB)
 # IMPORTANT: If you change the default model (all-MiniLM-L6-v2) and vice versa, you aren't able to use RAG Chat with your previous documents loaded in the WebUI! You need to re-embed them.
 ENV RAG_EMBEDDING_MODEL="all-MiniLM-L6-v2" \
-    # device type for whisper tts and embbeding models - "cpu" (default), "cuda" (nvidia gpu and CUDA required) or "mps" (apple silicon) - choosing this right can lead to better performance
-    RAG_EMBEDDING_MODEL_DEVICE_TYPE="cpu" \
     RAG_EMBEDDING_MODEL_DIR="/app/backend/data/cache/embedding/models" \
-    SENTENCE_TRANSFORMERS_HOME="/app/backend/data/cache/embedding/models"
+    SENTENCE_TRANSFORMERS_HOME="/app/backend/data/cache/embedding/models" \
+    # device type for whisper tts and embbeding models - "cpu" (default) or "mps" (apple silicon) - choosing this right can lead to better performance
+    # Important:
+    #  If you want to use CUDA you need to install the nvidia-container-toolkit (https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html) 
+    #  you can set this to "cuda" but its recomended to use --build-arg CUDA_ENABLED=true flag when building the image
+    RAG_EMBEDDING_MODEL_DEVICE_TYPE="cuda"
+# device type for whisper tts and embbeding models - "cpu" (default), "cuda" (nvidia gpu and CUDA required) or "mps" (apple silicon) - choosing this right can lead to better performance
 #### Preloaded models ##########################################################
 
 WORKDIR /app/backend
-
 # install python dependencies
 COPY ./backend/requirements.txt ./requirements.txt
 
-RUN pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu --no-cache-dir \
-    && pip3 install -r requirements.txt --no-cache-dir
+RUN pip3 install -r requirements.txt --no-cache-dir
+
+RUN if [ "$RAG_EMBEDDING_MODEL_DEVICE_TYPE" = "cuda" ]; then \
+        echo "CUDA enabled" && \
+        pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu117 --no-cache-dir; \
+    else \
+        pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu --no-cache-dir && \
+        python -c "import os; from chromadb.utils import embedding_functions; sentence_transformer_ef = embedding_functions.SentenceTransformerEmbeddingFunction(model_name=os.environ['RAG_EMBEDDING_MODEL'], device=os.environ['RAG_EMBEDDING_MODEL_DEVICE_TYPE'])"; \
+    fi
+
+# preload tts model
+RUN python -c "import os; from faster_whisper import WhisperModel; WhisperModel(os.environ['WHISPER_MODEL'], device='auto', compute_type='int8', download_root=os.environ['WHISPER_MODEL_DIR'])"
 
 #  install required packages
 RUN apt-get update \
@@ -71,10 +84,7 @@ RUN apt-get update \
     # cleanup
     && rm -rf /var/lib/apt/lists/*
 
-# preload embedding model
-RUN python -c "import os; from chromadb.utils import embedding_functions; sentence_transformer_ef = embedding_functions.SentenceTransformerEmbeddingFunction(model_name=os.environ['RAG_EMBEDDING_MODEL'], device=os.environ['RAG_EMBEDDING_MODEL_DEVICE_TYPE'])"
-# preload tts model
-RUN python -c "import os; from faster_whisper import WhisperModel; WhisperModel(os.environ['WHISPER_MODEL'], device='auto', compute_type='int8', download_root=os.environ['WHISPER_MODEL_DIR'])"
+
 
 # copy embedding weight from build
 # RUN mkdir -p /root/.cache/chroma/onnx_models/all-MiniLM-L6-v2