Skip to content

SOTA Open Source TTS

Add AMD ROCm Docker support (RDNA3/RDNA4) (#1292)

Kaihui-AMDe5e292632cb1
diff --git a/compose.rocm.yml b/compose.rocm.yml
new file mode 100644
index 0000000..cf0bfec
--- /dev/null
+++ b/compose.rocm.yml
@@ -0,0 +1,58 @@
+name: fish-speech-rocm
+
+# AMD ROCm compose for Fish Speech (RDNA3 / RDNA4).
+# Mount your checkpoints into ./checkpoints before running.
+#
+# docker compose -f compose.rocm.yml --profile webui up --build
+# docker compose -f compose.rocm.yml --profile server up --build
+
+services:
+ webui:
+ build:
+ context: .
+ dockerfile: docker/Dockerfile.rocm
+ target: webui
+ image: fish-speech-webui:rocm
+ profiles: ["webui"]
+ ports:
+ - "${GRADIO_PORT:-7860}:7860"
+ volumes:
+ - ./checkpoints:/app/checkpoints
+ - ./references:/app/references
+ environment:
+ - ROCBLAS_USE_HIPBLASLT=0
+ - COMPILE=${COMPILE:-1}
+ devices:
+ - /dev/kfd
+ - /dev/dri
+ group_add:
+ - video
+ - render
+ shm_size: "16g"
+ tty: true
+ stdin_open: true
+
+ server:
+ build:
+ context: .
+ dockerfile: docker/Dockerfile.rocm
+ target: server
+ image: fish-speech-server:rocm
+ profiles: ["server"]
+ ports:
+ - "${API_PORT:-8080}:8080"
+ volumes:
+ - ./checkpoints:/app/checkpoints
+ - ./references:/app/references
+ environment:
+ - ROCBLAS_USE_HIPBLASLT=0
+ - COMPILE=${COMPILE:-1}
+ devices:
+ - /dev/kfd
+ - /dev/dri
+ group_add:
+ - video
+ - render
+ shm_size: "16g"
+ tty: true
+ stdin_open: true
diff --git a/docker/Dockerfile.rocm b/docker/Dockerfile.rocm
new file mode 100644
index 0000000..aca6fc5
--- /dev/null
+++ b/docker/Dockerfile.rocm
@@ -0,0 +1,106 @@
+# docker/Dockerfile.rocm
+#
+# Fish Speech on AMD ROCm (RDNA3 / RDNA4).
+# The checkpoints are NOT bundled — mount them at /app/checkpoints.
+#
+# Build:
+# docker build -f docker/Dockerfile.rocm --target webui -t fish-speech-webui:rocm .
+# docker build -f docker/Dockerfile.rocm --target server -t fish-speech-server:rocm .
+#
+# Run (webui):
+# docker run --device=/dev/kfd --device=/dev/dri \
+# --group-add video --group-add render \
+# -e ROCBLAS_USE_HIPBLASLT=0 \
+# -v ./checkpoints:/app/checkpoints \
+# -p 7860:7860 fish-speech-webui:rocm
+
+ARG ROCM_VERSION=7.2.3
+ARG BASE_IMAGE=rocm/pytorch:rocm${ROCM_VERSION}_ubuntu24.04_py3.12_pytorch_release_2.9.1
+
+FROM ${BASE_IMAGE} AS app-base
+
+ENV DEBIAN_FRONTEND=noninteractive \
+ PYTHONDONTWRITEBYTECODE=1 \
+ PYTHONUNBUFFERED=1 \
+ ROCBLAS_USE_HIPBLASLT=0
+
+RUN apt-get update \
+ && apt-get install -y --no-install-recommends \
+ git ffmpeg libsox-dev build-essential cmake \
+ libasound-dev portaudio19-dev libportaudio2 libportaudiocpp0 \
+ && apt-get clean \
+ && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /app
+
+COPY . /app
+
+# Install runtime dependencies WITHOUT torch/torchaudio — the ROCm base image
+# already ships a gfx-tuned torch (2.9.1+rocm7.2.3). Then install the package
+# itself with --no-deps so pip does not try to pull a CUDA/CPU torch.
+RUN pip install --no-cache-dir --upgrade pip setuptools wheel \
+ && pip install --no-cache-dir \
+ numpy "transformers<=4.57.3" datasets lightning pytorch_lightning \
+ hydra-core natsort einops librosa rich "gradio>5.0.0" wandb grpcio kui \
+ uvicorn loguru loralib pyrootutils resampy "einx[torch]==0.2.2" zstandard \
+ pydub "modelscope==1.17.1" "opencc-python-reimplemented==0.1.7" \
+ silero-vad ormsgpack tiktoken "pydantic==2.9.2" cachetools \
+ descript-audio-codec safetensors soundfile vector_quantize_pytorch \
+ && pip install --no-cache-dir --no-build-isolation pyaudio \
+ && pip install --no-cache-dir --no-deps -e . \
+ # descript-audiotools pins protobuf<3.20, but fish-speech's generated proto
+ # code needs >=3.20. Override after install (mirrors pyproject's uv override).
+ && pip install --no-cache-dir --no-deps --upgrade "protobuf>=4.25,<6.0"
+
+EXPOSE 7860 8080
+
+# torch.compile is enabled by default (verified working on gfx1201/RDNA4).
+# Set COMPILE=0 to disable.
+ENV COMPILE=1
+
+##############################################################
+# Gradio WebUI
+##############################################################
+FROM app-base AS webui
+
+ARG GRADIO_SERVER_NAME="0.0.0.0"
+ARG GRADIO_SERVER_PORT=7860
+ENV GRADIO_SERVER_NAME=${GRADIO_SERVER_NAME} \
+ GRADIO_SERVER_PORT=${GRADIO_SERVER_PORT}
+
+RUN printf '%s\n' \
+ '#!/bin/bash' \
+ 'set -e' \
+ 'ARGS=()' \
+ 'if [ "${COMPILE:-0}" = "1" ] || [ "${COMPILE:-}" = "true" ]; then ARGS+=(--compile); fi' \
+ 'exec python tools/run_webui.py \' \
+ ' --llama-checkpoint-path checkpoints/s2-pro \' \
+ ' --decoder-checkpoint-path checkpoints/s2-pro/codec.pth \' \
+ ' --decoder-config-name modded_dac_vq "${ARGS[@]}"' \
+ > /app/start_webui.sh && chmod +x /app/start_webui.sh
+
+ENTRYPOINT ["/app/start_webui.sh"]
+
+##############################################################
+# API Server
+##############################################################
+FROM app-base AS server
+
+ARG API_SERVER_NAME="0.0.0.0"
+ARG API_SERVER_PORT=8080
+ENV API_SERVER_NAME=${API_SERVER_NAME} \
+ API_SERVER_PORT=${API_SERVER_PORT}
+
+RUN printf '%s\n' \
+ '#!/bin/bash' \
+ 'set -e' \
+ 'ARGS=()' \
+ 'if [ "${COMPILE:-0}" = "1" ] || [ "${COMPILE:-}" = "true" ]; then ARGS+=(--compile); fi' \
+ 'exec python tools/api_server.py \' \
+ ' --listen 0.0.0.0:8080 \' \
+ ' --llama-checkpoint-path checkpoints/s2-pro \' \
+ ' --decoder-checkpoint-path checkpoints/s2-pro/codec.pth \' \
+ ' --decoder-config-name modded_dac_vq "${ARGS[@]}"' \
+ > /app/start_server.sh && chmod +x /app/start_server.sh
+
+ENTRYPOINT ["/app/start_server.sh"]
diff --git a/docs/en/install.md b/docs/en/install.md
index 1d090b5..dc76474 100644
--- a/docs/en/install.md
+++ b/docs/en/install.md
@@ -189,3 +189,40 @@ Both methods require mounting these directories:
!!! warning
GPU support requires NVIDIA Docker runtime. For CPU-only deployment, remove the `--gpus all` flag and use CPU images.
+
+### AMD ROCm support
+
+Fish Speech runs on AMD GPUs via ROCm. The ROCm image is based on the official `rocm/pytorch` image, which already ships a gfx-tuned PyTorch, so no separate torch install is needed. Verified on RDNA4 (Radeon AI PRO R9700 / gfx1201) with ROCm 7.2.3; RDNA3 (gfx1100/gfx1101) should also work.
+
+**Prerequisites:**
+
+- AMD GPU with ROCm support (RDNA3 / RDNA4)
+- ROCm drivers installed on the host
+- Docker with GPU passthrough (`/dev/kfd` and `/dev/dri`)
+
+**Using Docker Compose:**
+
+```bash
+# WebUI
+docker compose -f compose.rocm.yml --profile webui up --build
+
+# API server
+docker compose -f compose.rocm.yml --profile server up --build
+```
+
+**Manual build and run:**
+
+```bash
+docker build -f docker/Dockerfile.rocm --target webui -t fish-speech-webui:rocm .
+
+docker run \
+ --device=/dev/kfd --device=/dev/dri \
+ --group-add video --group-add render \
+ -e ROCBLAS_USE_HIPBLASLT=0 \
+ -v ./checkpoints:/app/checkpoints \
+ -p 7860:7860 \
+ fish-speech-webui:rocm
+```
+
+!!! note
+ `ROCBLAS_USE_HIPBLASLT=0` is set by default for RDNA4 (gfx1201) stability; RDNA3 users may not need it. Fish Speech uses `scaled_dot_product_attention`, which dispatches to ROCm's AOTriton flash-attention backend automatically — no custom kernel build is required. The first run is slower while MIOpen auto-tunes kernels. `torch.compile` is enabled by default (`COMPILE=1`); set `COMPILE=0` to disable.