From bcbac5c8b3ca01fa216fe3fca0884a8f21c1fd3a Mon Sep 17 00:00:00 2001
From: Adolfo Delorenzo <adelorenzo@oe74.net>
Date: Tue, 3 Jun 2025 09:16:41 -0600
Subject: [PATCH] Add multi-GPU support for Docker deployments
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- Created separate docker-compose override files for different GPU types:
  - docker-compose.nvidia.yml for NVIDIA GPUs
  - docker-compose.amd.yml for AMD GPUs with ROCm
  - docker-compose.apple.yml for Apple Silicon
- Updated README with GPU-specific Docker configurations
- Updated deployment instructions to use appropriate override files
- Added detailed configurations for each GPU type including:
  - Device mappings and drivers
  - Environment variables
  - Platform specifications
  - Memory and resource limits

This allows users to easily deploy Talk2Me with their specific GPU hardware.

🤖 Generated with [Claude Code](https://claude.ai/code)

Co-Authored-By: Claude <noreply@anthropic.com>
---
 README.md                 | 95 +++++++++++++++++++++++++++++++++++++--
 docker-compose.amd.yml    | 19 ++++++++
 docker-compose.apple.yml  | 11 +++++
 docker-compose.nvidia.yml | 16 +++++++
 4 files changed, 138 insertions(+), 3 deletions(-)
 create mode 100644 docker-compose.amd.yml
 create mode 100644 docker-compose.apple.yml
 create mode 100644 docker-compose.nvidia.yml

diff --git a/README.md b/README.md
index d9dd834..677ab12 100644
--- a/README.md
+++ b/README.md
@@ -280,18 +280,31 @@ curl -X POST -H "X-Admin-Token: $ADMIN_TOKEN" \
 ### Docker Deployment
 
 ```bash
-# Build and run with Docker Compose
+# Build and run with Docker Compose (CPU only)
 docker-compose up -d
 
+# With NVIDIA GPU support
+docker-compose -f docker-compose.yml -f docker-compose.nvidia.yml up -d
+
+# With AMD GPU support (ROCm)
+docker-compose -f docker-compose.yml -f docker-compose.amd.yml up -d
+
+# With Apple Silicon support
+docker-compose -f docker-compose.yml -f docker-compose.apple.yml up -d
+
 # Scale web workers
-docker-compose up -d --scale web=4
+docker-compose up -d --scale talk2me=4
 
 # View logs
-docker-compose logs -f web
+docker-compose logs -f talk2me
 ```
 
 ### Docker Compose Configuration
 
+Choose the appropriate configuration based on your GPU:
+
+#### NVIDIA GPU Configuration
+
 ```yaml
 version: '3.8'
 services:
@@ -316,6 +329,82 @@ services:
               capabilities: [gpu]
 ```
 
+#### AMD GPU Configuration (ROCm)
+
+```yaml
+version: '3.8'
+services:
+  web:
+    build: .
+    ports:
+      - "5005:5005"
+    environment:
+      - GUNICORN_WORKERS=4
+      - GUNICORN_THREADS=2
+      - HSA_OVERRIDE_GFX_VERSION=10.3.0  # Adjust for your GPU
+    volumes:
+      - ./logs:/app/logs
+      - whisper-cache:/root/.cache/whisper
+      - /dev/kfd:/dev/kfd  # ROCm KFD interface
+      - /dev/dri:/dev/dri  # Direct Rendering Interface
+    devices:
+      - /dev/kfd
+      - /dev/dri
+    group_add:
+      - video
+      - render
+    deploy:
+      resources:
+        limits:
+          memory: 4G
+```
+
+#### Apple Silicon Configuration
+
+```yaml
+version: '3.8'
+services:
+  web:
+    build: .
+    platform: linux/arm64/v8  # For M1/M2 Macs
+    ports:
+      - "5005:5005"
+    environment:
+      - GUNICORN_WORKERS=4
+      - GUNICORN_THREADS=2
+      - PYTORCH_ENABLE_MPS_FALLBACK=1  # Enable MPS fallback
+    volumes:
+      - ./logs:/app/logs
+      - whisper-cache:/root/.cache/whisper
+    deploy:
+      resources:
+        limits:
+          memory: 4G
+```
+
+#### CPU-Only Configuration
+
+```yaml
+version: '3.8'
+services:
+  web:
+    build: .
+    ports:
+      - "5005:5005"
+    environment:
+      - GUNICORN_WORKERS=4
+      - GUNICORN_THREADS=2
+      - OMP_NUM_THREADS=4  # OpenMP threads for CPU
+    volumes:
+      - ./logs:/app/logs
+      - whisper-cache:/root/.cache/whisper
+    deploy:
+      resources:
+        limits:
+          memory: 4G
+          cpus: '4.0'
+```
+
 ### Nginx Configuration
 
 ```nginx
diff --git a/docker-compose.amd.yml b/docker-compose.amd.yml
new file mode 100644
index 0000000..2c947c1
--- /dev/null
+++ b/docker-compose.amd.yml
@@ -0,0 +1,19 @@
+version: '3.8'
+
+# Docker Compose override for AMD GPU support (ROCm)
+# Usage: docker-compose -f docker-compose.yml -f docker-compose.amd.yml up
+
+services:
+  talk2me:
+    environment:
+      - HSA_OVERRIDE_GFX_VERSION=10.3.0  # Adjust based on your GPU model
+      - ROCR_VISIBLE_DEVICES=0  # Use first GPU
+    volumes:
+      - /dev/kfd:/dev/kfd  # ROCm KFD interface
+      - /dev/dri:/dev/dri  # Direct Rendering Interface
+    devices:
+      - /dev/kfd
+      - /dev/dri
+    group_add:
+      - video  # Required for GPU access
+      - render # Required for GPU access
\ No newline at end of file
diff --git a/docker-compose.apple.yml b/docker-compose.apple.yml
new file mode 100644
index 0000000..2841eac
--- /dev/null
+++ b/docker-compose.apple.yml
@@ -0,0 +1,11 @@
+version: '3.8'
+
+# Docker Compose override for Apple Silicon
+# Usage: docker-compose -f docker-compose.yml -f docker-compose.apple.yml up
+
+services:
+  talk2me:
+    platform: linux/arm64/v8  # For M1/M2/M3 Macs
+    environment:
+      - PYTORCH_ENABLE_MPS_FALLBACK=1  # Enable Metal Performance Shaders fallback
+      - PYTORCH_MPS_HIGH_WATERMARK_RATIO=0.7  # Memory management for MPS
\ No newline at end of file
diff --git a/docker-compose.nvidia.yml b/docker-compose.nvidia.yml
new file mode 100644
index 0000000..c702e05
--- /dev/null
+++ b/docker-compose.nvidia.yml
@@ -0,0 +1,16 @@
+version: '3.8'
+
+# Docker Compose override for NVIDIA GPU support
+# Usage: docker-compose -f docker-compose.yml -f docker-compose.nvidia.yml up
+
+services:
+  talk2me:
+    environment:
+      - CUDA_VISIBLE_DEVICES=0  # Use first GPU
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1
+              capabilities: [gpu]
\ No newline at end of file