feat(huri): kube local example with TTS + EMAGE + RAG

.gitignore

@@ -176,6 +176,12 @@ cython_debug/
 # PyPI configuration file
 .pypirc
 
+# Helm
+**/charts/*.tgz
+
 # Others
 .trash
-docs
+docs
+
+# HuRI client outputs
+*.wav

config/client_full.yaml

@@ -0,0 +1,44 @@
+huri_url: ws://localhost:8000/session
+
+topic_list: [transcript, question, token, motion]
+
+senders:
+  audio:
+    name: audio
+    args:
+      sample_rate: 16000
+      frame_duration: 0.030
+
+modules:
+  mic:
+    name: mic
+    args:
+      vad_agressiveness: 3
+      silence_duration: 1.5
+      block_duration: ${senders.audio.args.frame_duration}
+    logging: INFO
+  stt:
+    name: stt
+    args:
+      language: en
+      block_duration: ${senders.audio.args.frame_duration}
+    logging: INFO
+  tag:
+    name: tag
+    logging: INFO
+  rag:
+    name: rag
+    args:
+      language: en
+      tone: formal
+      response_format: paragraph
+      max_length: 1024
+    logging: INFO
+  tts:
+    name: tts
+    args:
+      min_clause_chars: 20
+    logging: INFO
+  gesture:
+    name: gesture
+    logging: INFO

deploy/Dockerfile.amd

@@ -0,0 +1,79 @@
+FROM rayproject/ray:2.55.1-py312
+WORKDIR /app
+USER root
+RUN apt-get update && apt-get install -y \
+    build-essential \
+    curl \
+    gnupg2 \
+    && rm -rf /var/lib/apt/lists/*
+
+# Add ROCm 7.2.3 repository (Ubuntu 22.04 Jammy)
+RUN curl -fsSL https://repo.radeon.com/rocm/rocm.gpg.key | gpg --dearmor -o /etc/apt/keyrings/rocm.gpg \
+    && printf 'deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/7.2.3 jammy main\ndeb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/7.2 jammy main\n' \
+        > /etc/apt/sources.list.d/rocm.list \
+    && printf 'Package: *\nPin: release o=repo.radeon.com\nPin-Priority: 600\n' \
+        > /etc/apt/preferences.d/rocm-pin-600
+
+# Install ROCm runtime + libraries
+RUN apt-get update && apt-get install -y \
+    rocm-hip-runtime \
+    rocm-hip-libraries \
+    rocm-device-libs \
+    rocm-smi-lib \
+    rocblas \
+    hipblas \
+    miopen-hip \
+    rccl \
+    rocsolver \
+    rocfft \
+    rocrand \
+    hipsparse \
+    && rm -rf /var/lib/apt/lists/*
+
+# Set ROCm environment
+ENV ROCM_PATH=/opt/rocm
+ENV PATH="${ROCM_PATH}/bin:${PATH}"
+ENV LD_LIBRARY_PATH="${ROCM_PATH}/lib:${LD_LIBRARY_PATH}"
+
+USER ray
+
+
+COPY serve_requirements.txt /app
+RUN pip install --no-cache-dir -r serve_requirements.txt
+
+# 1. AMD's PyTorch built for ROCm (NOT the PyPI one — it's built for ROCm 6.2 and will silently break)
+ARG ROCM_VERSION=7.2
+ARG PYTHON_VERSION=cp312
+ARG TRITON_VERSION=3.4.0+rocm7.2.0.git0cace8d2
+
+RUN pip install --no-cache-dir \
+    "https://repo.radeon.com/rocm/manylinux/rocm-rel-${ROCM_VERSION}/triton-${TRITON_VERSION}-${PYTHON_VERSION}-${PYTHON_VERSION}-linux_x86_64.whl"
+
+RUN pip install --no-cache-dir \
+    --extra-index-url https://repo.radeon.com/rocm/pypi/ \
+    "https://repo.radeon.com/rocm/manylinux/rocm-rel-${ROCM_VERSION}/torch-2.8.0+rocm${ROCM_VERSION}.0.lw.gitbf943426-${PYTHON_VERSION}-${PYTHON_VERSION}-linux_x86_64.whl" \
+    "https://repo.radeon.com/rocm/manylinux/rocm-rel-${ROCM_VERSION}/torchaudio-2.8.0+rocm${ROCM_VERSION}.0.git6e1c7fe9-${PYTHON_VERSION}-${PYTHON_VERSION}-linux_x86_64.whl"
+
+RUN pip install --no-cache-dir filelock sympy networkx jinja2 fsspec numpy
+
+USER root
+
+# 3. Official CTranslate2 ROCm wheel (it's inside a zip on the releases page)
+RUN apt-get update && apt-get install -y unzip curl \
+ && curl -L https://github.com/OpenNMT/CTranslate2/releases/download/v4.7.1/rocm-python-wheels-Linux.zip \
+    -o /tmp/ct2-rocm.zip \
+ && unzip -j /tmp/ct2-rocm.zip 'temp-linux/ctranslate2-4.7.1-cp312-*manylinux*x86_64.whl' -d /tmp/ct2 \
+ && pip install --no-cache-dir /tmp/ct2/ctranslate2-4.7.1-cp312-*.whl \
+ && rm -rf /tmp/ct2 /tmp/ct2-rocm.zip
+
+USER ray
+
+# 4. faster-whisper
+RUN pip install --no-cache-dir faster-whisper
+
+# 5. RAG / LLM extras (httpx, qdrant-client, sentence-transformers, …)
+#    Installed last so the ROCm torch wheel installed above is the resolved one.
+COPY requirements-amd.txt /app
+RUN pip install --no-cache-dir -r requirements-amd.txt
+
+COPY src /app/src

deploy/Dockerfile.base

@@ -0,0 +1,16 @@
+FROM rayproject/ray:2.55.1-py312
+
+WORKDIR /app
+
+
+USER root
+RUN apt-get update && apt-get install -y \
+build-essential \
+&& rm -rf /var/lib/apt/lists/*
+
+USER ray
+
+COPY serve_requirements.txt /app
+RUN pip install --no-cache-dir -r serve_requirements.txt
+
+COPY src /app/src

deploy/Dockerfile.nvidia

@@ -0,0 +1,31 @@
+FROM rayproject/ray:2.55.1-py312-gpu
+
+WORKDIR /app
+
+# Full CUDA 12.1 dependency stack (CosyVoice2, faster-whisper, TensorRT, …).
+# PyTorch cu121 wheels live on the PyTorch index; TensorRT wheels on the NGC index.
+COPY requirements-nvidia.txt /app
+RUN pip install --no-cache-dir \
+    --extra-index-url https://download.pytorch.org/whl/cu121 \
+    --extra-index-url https://pypi.ngc.nvidia.com \
+    -r requirements-nvidia.txt
+
+COPY serve_requirements.txt /app
+RUN pip install --no-cache-dir -r serve_requirements.txt
+
+USER root
+
+RUN apt-get update && apt-get install -y --no-install-recommends git \
+    && rm -rf /var/lib/apt/lists/*
+
+USER ray
+
+# CosyVoice2 has no setup.py/pyproject.toml so it cannot be pip-installed.
+# Clone at a pinned commit for supply-chain integrity and expose it via PYTHONPATH.
+RUN git clone https://github.com/FunAudioLLM/CosyVoice.git /app/cosyvoice \
+    && git -C /app/cosyvoice checkout 074ca6dc9e80a2f424f1f74b48bdd7d3fea531cc \
+    && git -C /app/cosyvoice submodule update --init --recursive
+
+ENV PYTHONPATH="/app/cosyvoice:${PYTHONPATH:-}"
+
+COPY src /app/src

deploy/examples/local_nvidia_amd/Chart.lock

@@ -0,0 +1,6 @@
+dependencies:
+- name: kuberay-operator
+  repository: https://ray-project.github.io/kuberay-helm/
+  version: 1.6.0
+digest: sha256:b9057481d9a5e2d8b8798488b0b321bbd3f6e43dcb5a9dea18b181641a63b400
+generated: "2026-05-22T17:28:37.5934885+02:00"

deploy/examples/local_nvidia_amd/Chart.yaml

@@ -0,0 +1,20 @@
+apiVersion: v2
+name: huri
+description: HuRI service powered by Ray Serve on KubeRay
+type: application
+version: 0.1.0
+appVersion: "2.52.0"
+keywords:
+  - ray
+  - kuberay
+  - ray-serve
+  - robotics
+  - hri
+maintainers:
+  - name: Sentience Robotics
+
+dependencies:
+  - name: kuberay-operator
+    version: "1.6.0"
+    repository: "https://ray-project.github.io/kuberay-helm/"
+    condition: kuberay.install

deploy/examples/local_nvidia_amd/templates/_helpers.tpl

@@ -0,0 +1,72 @@
+{{/*
+Expand the name of the chart.
+*/}}
+{{- define "huri.name" -}}
+{{- default .Chart.Name .Values.nameOverride | trunc 63 | trimSuffix "-" }}
+{{- end }}
+
+{{/*
+Create a default fully-qualified app name.
+Truncated at 63 chars because some Kubernetes name fields have this limit.
+*/}}
+{{- define "huri.fullname" -}}
+{{- if .Values.fullnameOverride }}
+{{- .Values.fullnameOverride | trunc 63 | trimSuffix "-" }}
+{{- else }}
+{{- $name := default .Chart.Name .Values.nameOverride }}
+{{- if contains $name .Release.Name }}
+{{- .Release.Name | trunc 63 | trimSuffix "-" }}
+{{- else }}
+{{- printf "%s-%s" .Release.Name $name | trunc 63 | trimSuffix "-" }}
+{{- end }}
+{{- end }}
+{{- end }}
+
+{{/*
+Chart label: <name>-<version>.
+*/}}
+{{- define "huri.chart" -}}
+{{- printf "%s-%s" .Chart.Name .Chart.Version | replace "+" "_" | trunc 63 | trimSuffix "-" }}
+{{- end }}
+
+{{/*
+Common labels applied to every resource.
+*/}}
+{{- define "huri.labels" -}}
+helm.sh/chart: {{ include "huri.chart" . }}
+{{ include "huri.selectorLabels" . }}
+app.kubernetes.io/version: {{ .Chart.AppVersion | quote }}
+app.kubernetes.io/managed-by: {{ .Release.Service }}
+{{- end }}
+
+{{/*
+Selector labels (used in matchLabels / ingress backends).
+*/}}
+{{- define "huri.selectorLabels" -}}
+app.kubernetes.io/name: {{ include "huri.name" . }}
+app.kubernetes.io/instance: {{ .Release.Name }}
+{{- end }}
+
+{{/*
+Name of the KubeRay-managed serve service.
+KubeRay appends "-serve-svc" to the RayService name.
+*/}}
+{{- define "huri.serveSvcName" -}}
+{{- printf "%s-serve-svc" (include "huri.fullname" .) }}
+{{- end }}
+
+{{/*
+Name of the KubeRay-managed head service.
+KubeRay appends "-head-svc" to the RayService name.
+*/}}
+{{- define "huri.headSvcName" -}}
+{{- printf "%s-head-svc" (include "huri.fullname" .) }}
+{{- end }}
+
+{{/*
+Name of the stable dashboard service managed by this chart.
+Selects the head pod via stable labels, avoiding KubeRay's random-suffix service.
+*/}}
+{{- define "huri.headDashboardSvcName" -}}
+{{- printf "%s-head-dashboard-svc" (include "huri.fullname" .) }}
+{{- end }}