phoenix-oss/llama-stack-mirror
0
0
Fork 1
mirror of https://github.com/meta-llama/llama-stack.git synced 2025-07-31 16:01:46 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Convert TGI

Browse source
This commit is contained in:
Ashwin Bharambe 2024-11-17 14:49:41 -08:00
parent 9bb07ce298
commit 028530546f
14 changed files with 485 additions and 160 deletions
Download patch file Download diff file Expand all files Collapse all files

10
distributions/remote-vllm/compose.yaml
View file

@ -1,13 +1,3 @@
# NOTES:
#
# This Docker Compose (and the associated run.yaml) assumes you will be
# running in the default "bridged" network mode.
#
# If you need "host" network mode, please uncomment
# - network_mode: "host"
#
# Similarly change "host.docker.internal" to "localhost" in the run.yaml file
#
services:
vllm-inference:
image: vllm/vllm-openai:latest

6
distributions/remote-vllm/run-with-safety.yaml
View file

@ -1,14 +1,14 @@
version: '2'
built_at: 2024-11-17 14:07:24.568750
built_at: 2024-11-17 14:48:55.487270
image_name: remote-vllm
docker_image: llamastack/distribution-remote-vllm:test-0.0.52rc3
conda_env: null
apis:
- safety
- agents
- telemetry
- safety
- inference
- memory
- inference
providers:
inference:
- provider_id: vllm-inference

6
distributions/remote-vllm/run.yaml
View file

@ -1,14 +1,14 @@
version: '2'
built_at: 2024-11-17 14:07:24.563541
built_at: 2024-11-17 14:48:55.476058
image_name: remote-vllm
docker_image: llamastack/distribution-remote-vllm:test-0.0.52rc3
conda_env: null
apis:
- safety
- agents
- telemetry
- safety
- inference
- memory
- inference
providers:
inference:
- provider_id: vllm-inference

86
distributions/tgi/compose.yaml
View file

@ -1,51 +1,89 @@
services:
text-generation-inference:
tgi-inference:
image: ghcr.io/huggingface/text-generation-inference:latest
network_mode: "host"
volumes:
- $HOME/.cache/huggingface:/data
network_mode: ${NETWORK_MODE:-bridged}
ports:
- "5009:5009"
- "${TGI_INFERENCE_PORT:-8080}:${TGI_INFERENCE_PORT:-8080}"
devices:
- nvidia.com/gpu=all
environment:
- CUDA_VISIBLE_DEVICES=0
- CUDA_VISIBLE_DEVICES=${TGI_INFERENCE_GPU:-0}
- HF_TOKEN=$HF_TOKEN
- HF_HOME=/data
- HF_DATASETS_CACHE=/data
- HF_MODULES_CACHE=/data
- HF_HUB_CACHE=/data
command: ["--dtype", "bfloat16", "--usage-stats", "on", "--sharded", "false", "--model-id", "meta-llama/Llama-3.1-8B-Instruct", "--port", "5009", "--cuda-memory-fraction", "0.3"]
command: >
--dtype bfloat16
--usage-stats off
--sharded false
--model-id ${TGI_INFERENCE_MODEL:-meta-llama/Llama-3.2-3B-Instruct}
--port ${TGI_INFERENCE_PORT:-8080}
--cuda-memory-fraction 0.75
healthcheck:
test: ["CMD", "curl", "-f", "http://tgi-inference:${TGI_INFERENCE_PORT:-8080}/health"]
interval: 5s
timeout: 5s
retries: 30
deploy:
resources:
reservations:
devices:
- driver: nvidia
# that's the closest analogue to --gpus; provide
# an integer amount of devices or 'all'
count: 1
# Devices are reserved using a list of capabilities, making
# capabilities the only required field. A device MUST
# satisfy all the requested capabilities for a successful
# reservation.
capabilities: [gpu]
runtime: nvidia
tgi-${TGI_SAFETY_MODEL:+safety}:
image: ghcr.io/huggingface/text-generation-inference:latest
volumes:
- $HOME/.cache/huggingface:/data
network_mode: ${NETWORK_MODE:-bridged}
ports:
- "${TGI_SAFETY_PORT:-8081}:${TGI_SAFETY_PORT:-8081}"
devices:
- nvidia.com/gpu=all
environment:
- CUDA_VISIBLE_DEVICES=${TGI_SAFETY_GPU:-1}
- HF_TOKEN=$HF_TOKEN
- HF_HOME=/data
- HF_DATASETS_CACHE=/data
- HF_MODULES_CACHE=/data
- HF_HUB_CACHE=/data
command: >
--dtype bfloat16
--usage-stats off
--sharded false
--model-id ${TGI_SAFETY_MODEL:-meta-llama/Llama-Guard-3-1B}
--port ${TGI_SAFETY_PORT:-8081}
--cuda-memory-fraction 0.75
healthcheck:
test: ["CMD", "curl", "-f", "http://text-generation-inference:5009/health"]
test: ["CMD", "curl", "-f", "http://tgi-safety:${TGI_SAFETY_PORT:-8081}/health"]
interval: 5s
timeout: 5s
retries: 30
deploy:
resources:
reservations:
devices:
- driver: nvidia
capabilities: [gpu]
runtime: nvidia
llamastack:
depends_on:
text-generation-inference:
tgi-inference:
condition: service_healthy
image: llamastack/distribution-tgi
network_mode: "host"
tgi-${TGI_SAFETY_MODEL:+safety}:
condition: service_healthy
image: llamastack/distribution-tgi:test-0.0.52rc3
network_mode: ${NETWORK_MODE:-bridged}
volumes:
- ~/.llama:/root/.llama
# Link to TGI run.yaml file
- ./run.yaml:/root/my-run.yaml
- ./run${TGI_SAFETY_MODEL:+-with-safety}.yaml:/root/my-run.yaml
ports:
- "5000:5000"
- "${LLAMA_STACK_PORT:-5001}:${LLAMA_STACK_PORT:-5001}"
# Hack: wait for TGI server to start before starting docker
entrypoint: bash -c "sleep 60; python -m llama_stack.distribution.server.server --yaml_config /root/my-run.yaml"
restart_policy:
@ -53,3 +91,13 @@ services:
delay: 3s
max_attempts: 5
window: 60s
environment:
- TGI_URL=http://tgi-inference:${TGI_INFERENCE_PORT:-8080}
- SAFETY_TGI_URL=http://tgi-safety:${TGI_SAFETY_PORT:-8081}
- INFERENCE_MODEL=${INFERENCE_MODEL:-meta-llama/Llama-3.2-3B-Instruct}
- SAFETY_MODEL=${SAFETY_MODEL:-meta-llama/Llama-Guard-3-1B}
volumes:
tgi-inference:
tgi-safety:
llamastack:

67
distributions/tgi/run-with-safety.yaml Normal file
View file

@ -0,0 +1,67 @@
version: '2'
built_at: 2024-11-17 14:48:56.991119
image_name: tgi
docker_image: llamastack/distribution-remote-tgi:test-0.0.52rc3
conda_env: null
apis:
- safety
- agents
- telemetry
- memory
- inference
providers:
inference:
- provider_id: tgi-inference
provider_type: remote::tgi
config:
url: ${env.TGI_URL}
- provider_id: tgi-safety
provider_type: remote::tgi
config:
url: ${env.SAFETY_TGI_URL}
memory:
- provider_id: faiss
provider_type: inline::faiss
config:
kvstore:
type: sqlite
namespace: null
db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/tgi}/faiss_store.db
safety:
- provider_id: llama-guard
provider_type: inline::llama-guard
config: {}
agents:
- provider_id: meta-reference
provider_type: inline::meta-reference
config:
persistence_store:
type: sqlite
namespace: null
db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/tgi}/agents_store.db
telemetry:
- provider_id: meta-reference
provider_type: inline::meta-reference
config: {}
metadata_store:
namespace: null
type: sqlite
db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/tgi}/registry.db
models:
- metadata: {}
model_id: ${env.INFERENCE_MODEL}
provider_id: tgi-inference
provider_model_id: null
- metadata: {}
model_id: ${env.SAFETY_MODEL}
provider_id: tgi-safety
provider_model_id: null
shields:
- params: null
shield_id: ${env.SAFETY_MODEL}
provider_id: null
provider_shield_id: null
memory_banks: []
datasets: []
scoring_fns: []
eval_tasks: []

66
distributions/tgi/run.yaml
View file

@ -1,45 +1,55 @@
version: '2'
built_at: '2024-10-08T17:40:45.325529'
image_name: local
docker_image: null
conda_env: local
built_at: 2024-11-17 14:48:56.975663
image_name: tgi
docker_image: llamastack/distribution-remote-tgi:test-0.0.52rc3
conda_env: null
apis:
- shields
- agents
- models
- memory
- memory_banks
- inference
- safety
- agents
- telemetry
- memory
- inference
providers:
inference:
- provider_id: tgi0
- provider_id: tgi-inference
provider_type: remote::tgi
config:
url: http://127.0.0.1:5009
safety:
- provider_id: meta0
provider_type: inline::llama-guard
config:
model: Llama-Guard-3-1B
excluded_categories: []
- provider_id: meta1
provider_type: inline::prompt-guard
config:
model: Prompt-Guard-86M
url: ${env.TGI_URL}
memory:
- provider_id: meta0
provider_type: inline::meta-reference
- provider_id: faiss
provider_type: inline::faiss
config:
kvstore:
type: sqlite
namespace: null
db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/tgi}/faiss_store.db
safety:
- provider_id: llama-guard
provider_type: inline::llama-guard
config: {}
agents:
- provider_id: meta0
- provider_id: meta-reference
provider_type: inline::meta-reference
config:
persistence_store:
namespace: null
type: sqlite
db_path: ~/.llama/runtime/kvstore.db
namespace: null
db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/tgi}/agents_store.db
telemetry:
- provider_id: meta0
- provider_id: meta-reference
provider_type: inline::meta-reference
config: {}
metadata_store:
namespace: null
type: sqlite
db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/tgi}/registry.db
models:
- metadata: {}
model_id: ${env.INFERENCE_MODEL}
provider_id: tgi-inference
provider_model_id: null
shields: []
memory_banks: []
datasets: []
scoring_fns: []
eval_tasks: []

80
docs/source/getting_started/distributions/self_hosted_distro/tgi.md
View file

@ -2,35 +2,55 @@
The `llamastack/distribution-tgi` distribution consists of the following provider configurations.
| **API** | **Inference** | **Agents** | **Memory** | **Safety** | **Telemetry** |
|----------------- |--------------- |---------------- |-------------------------------------------------- |---------------- |---------------- |
| **Provider(s)** | remote::tgi | meta-reference | meta-reference, remote::pgvector, remote::chroma | meta-reference | meta-reference |
Provider Configuration
┏━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ API ┃ Provider(s) ┃
┡━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ agents │ `inline::meta-reference` │
│ inference │ `remote::tgi` │
│ memory │ `inline::faiss`, `remote::chromadb`, `remote::pgvector` │
│ safety │ `inline::llama-guard` │
│ telemetry │ `inline::meta-reference` │
└───────────┴─────────────────────────────────────────────────────────┘
### Docker: Start the Distribution (Single Node GPU)
You can use this distribution if you have GPUs and want to run an independent TGI server container for running inference.### Environment Variables
> [!NOTE]
> This assumes you have access to GPU to start a TGI server with access to your GPU.
The following environment variables can be configured:
- `LLAMASTACK_PORT`: Port for the Llama Stack distribution server (default: `5001`)
- `INFERENCE_MODEL`: Inference model loaded into the TGI server (default: `meta-llama/Llama-3.2-3B-Instruct`)
- `TGI_URL`: URL of the TGI server with the main inference model (default: `http://host.docker.internal:8080}/v1`)
- `SAFETY_TGI_URL`: URL of the TGI server with the safety model (default: `http://host.docker.internal:8081/v1`)
- `SAFETY_MODEL`: Name of the safety (Llama-Guard) model to use (default: `meta-llama/Llama-Guard-3-1B`)
### Models
The following models are configured by default:
- `${env.INFERENCE_MODEL}`
- `${env.SAFETY_MODEL}`
```
$ cd distributions/tgi && docker compose up
## Using Docker Compose
You can use `docker compose` to start a TGI container and Llama Stack server container together.
```bash
$ cd distributions/tgi; docker compose up
```
The script will first start up TGI server, then start up Llama Stack distribution server hooking up to the remote TGI provider for inference. You should be able to see the following outputs --
```
```bash
[text-generation-inference] | 2024-10-15T18:56:33.810397Z INFO text_generation_router::server: router/src/server.rs:1813: Using config Some(Llama)
[text-generation-inference] | 2024-10-15T18:56:33.810448Z WARN text_generation_router::server: router/src/server.rs:1960: Invalid hostname, defaulting to 0.0.0.0
[text-generation-inference] | 2024-10-15T18:56:33.864143Z INFO text_generation_router::server: router/src/server.rs:2353: Connected
INFO: Started server process [1]
INFO: Waiting for application startup.
INFO: Application startup complete.
INFO: Uvicorn running on http://[::]:5000 (Press CTRL+C to quit)
INFO: Uvicorn running on http://[::]:5001 (Press CTRL+C to quit)
```
To kill the server
```
```bash
docker compose down
```
@ -42,8 +62,12 @@ If you wish to separately spin up a TGI server, and connect with Llama Stack, yo
#### Start TGI server locally
- Please check the [TGI Getting Started Guide](https://github.com/huggingface/text-generation-inference?tab=readme-ov-file#get-started) to get a TGI endpoint.
```
docker run --rm -it -v $HOME/.cache/huggingface:/data -p 5009:5009 --gpus all ghcr.io/huggingface/text-generation-inference:latest --dtype bfloat16 --usage-stats on --sharded false --model-id meta-llama/Llama-3.1-8B-Instruct --port 5009
```bash
docker run --rm -it -v $HOME/.cache/huggingface:/data \
-p 5009:5009 --gpus all \
ghcr.io/huggingface/text-generation-inference:latest \
--dtype bfloat16 --usage-stats on --sharded false \
--model-id meta-llama/Llama-3.2-3B-Instruct --port 5009
```
#### Start Llama Stack server pointing to TGI server
@ -57,12 +81,15 @@ llama stack run ./gpu/run.yaml
```
**Via Docker**
```
docker run --network host -it -p 5000:5000 -v ./run.yaml:/root/my-run.yaml --gpus=all llamastack/distribution-tgi --yaml_config /root/my-run.yaml
```bash
docker run --network host -it -p 5001:5001 \
-v ./run.yaml:/root/my-run.yaml --gpus=all \
llamastack/distribution-tgi \
--yaml_config /root/my-run.yaml
```
Make sure in you `run.yaml` file, you inference provider is pointing to the correct TGI server endpoint. E.g.
```
We have provided a template `run.yaml` file in the `distributions/tgi` directory. Make sure in your `run.yaml` file, you inference provider is pointing to the correct TGI server endpoint. E.g.
```yaml
inference:
- provider_id: tgi0
provider_type: remote::tgi
@ -76,17 +103,24 @@ To serve a new model with `tgi`, change the docker command flag `--model-id <mod
This can be done by edit the `command` args in `compose.yaml`. E.g. Replace "Llama-3.2-1B-Instruct" with the model you want to serve.
```
command: ["--dtype", "bfloat16", "--usage-stats", "on", "--sharded", "false", "--model-id", "meta-llama/Llama-3.2-1B-Instruct", "--port", "5009", "--cuda-memory-fraction", "0.3"]
```yaml
command: >
--dtype bfloat16 --usage-stats on --sharded false
--model-id meta-llama/Llama-3.2-1B-Instruct
--port 5009 --cuda-memory-fraction 0.7
```
or by changing the docker run command's `--model-id` flag
```
docker run --rm -it -v $HOME/.cache/huggingface:/data -p 5009:5009 --gpus all ghcr.io/huggingface/text-generation-inference:latest --dtype bfloat16 --usage-stats on --sharded false --model-id meta-llama/Llama-3.2-1B-Instruct --port 5009
```bash
docker run --rm -it -v $HOME/.cache/huggingface:/data \
-p 5009:5009 --gpus all \
ghcr.io/huggingface/text-generation-inference:latest \
--dtype bfloat16 --usage-stats off --sharded false \
--model-id meta-llama/Llama-3.2-3B-Instruct --port 5009
```
In `run.yaml`, make sure you point the correct server endpoint to the TGI server endpoint serving your model.
```
```yaml
inference:
- provider_id: tgi0
provider_type: remote::tgi

17
llama_stack/providers/remote/inference/tgi/config.py
View file

@ -12,19 +12,20 @@ from pydantic import BaseModel, Field
@json_schema_type
class TGIImplConfig(BaseModel):
host: str = "localhost"
port: int = 8080
protocol: str = "http"
@property
def url(self) -> str:
return f"{self.protocol}://{self.host}:{self.port}"
url: str = Field(
description="The URL for the TGI serving endpoint",
)
api_token: Optional[str] = Field(
default=None,
description="A bearer token if your TGI endpoint is protected.",
)
@classmethod
def sample_run_config(cls, url: str = "${env.TGI_URL}"):
return {
"url": url,
}
@json_schema_type
class InferenceEndpointImplConfig(BaseModel):

35
llama_stack/providers/remote/inference/tgi/docker_compose.yaml
View file

@ -1,35 +0,0 @@
services:
${SERVICE_NAME:-tgi}:
image: ghcr.io/huggingface/text-generation-inference:2.3.1
network_mode: "host"
volumes:
- $HOME/.cache/huggingface:/data
ports:
- ${TGI_PORT:-8000}:${TGI_PORT:-8000}
devices:
- nvidia.com/gpu=all
environment:
- CUDA_VISIBLE_DEVICES=${CUDA_VISIBLE_DEVICES:-0}
- HF_HOME=/data
- HF_DATASETS_CACHE=/data
- HF_MODULES_CACHE=/data
- HF_HUB_CACHE=/data
command: >
--dtype bfloat16
--usage-stats off
--sharded false
--model-id ${TGI_MODEL:-meta-llama/Llama-3.2-3B-Instruct}
--port ${TGI_PORT:-8000}
--cuda-memory-fraction ${TGI_CUDA_MEMORY_FRACTION:-0.8}
deploy:
resources:
reservations:
devices:
- driver: nvidia
capabilities: [gpu]
runtime: nvidia
healthcheck:
test: ["CMD", "curl", "-f", "http://${SERVICE_NAME:-tgi}:${TGI_PORT:-8000}/health"]
interval: 5s
timeout: 5s
retries: 30

26
llama_stack/providers/remote/inference/vllm/docker_compose.yaml
View file

@ -1,26 +0,0 @@
services:
${SERVICE_NAME:-vllm}:
image: vllm/vllm-openai:latest
ports:
- ${VLLM_PORT:-5100}:${VLLM_PORT:-5100}
volumes:
- $HOME/.cache/huggingface:/root/.cache/huggingface
devices:
- nvidia.com/gpu=all
deploy:
resources:
reservations:
devices:
- driver: nvidia
capabilities: [gpu]
runtime: nvidia
environment:
- CUDA_VISIBLE_DEVICES=${CUDA_VISIBLE_DEVICES:-0}
- HUGGING_FACE_HUB_TOKEN=${HF_TOKEN}
command: >
--gpu-memory-utilization 0.75
--model ${VLLM_MODEL:-meta-llama/Llama-3.2-3B-Instruct}
--enforce-eager
--max-model-len 8192
--max-num-seqs 16
--port ${VLLM_PORT:-5100}

7
llama_stack/templates/tgi/__init__.py Normal file
View file

@ -0,0 +1,7 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from .tgi import get_distribution_template # noqa: F401

17
llama_stack/templates/tgi/build.yaml
View file

@ -1,12 +1,19 @@
version: '2'
name: tgi
distribution_spec:
description: Use TGI for running LLM inference
description: Use (an external) TGI server for running LLM inference
docker_image: llamastack/distribution-remote-tgi:test-0.0.52rc3
providers:
inference: remote::tgi
inference:
- remote::tgi
memory:
- inline::faiss
- remote::chromadb
- remote::pgvector
safety: inline::llama-guard
agents: inline::meta-reference
telemetry: inline::meta-reference
safety:
- inline::llama-guard
agents:
- inline::meta-reference
telemetry:
- inline::meta-reference
image_type: conda

125
llama_stack/templates/tgi/doc_template.md Normal file
View file

@ -0,0 +1,125 @@
# TGI Distribution
The `llamastack/distribution-{{ name }}` distribution consists of the following provider configurations.
{{ providers_table }}
You can use this distribution if you have GPUs and want to run an independent TGI server container for running inference.
{%- if docker_compose_env_vars %}
### Environment Variables
The following environment variables can be configured:
{% for var, (default_value, description) in docker_compose_env_vars.items() %}
- `{{ var }}`: {{ description }} (default: `{{ default_value }}`)
{% endfor %}
{% endif %}
{%- if default_models %}
### Models
The following models are configured by default:
{% for model in default_models %}
- `{{ model.model_id }}`
{% endfor %}
{% endif %}
## Using Docker Compose
You can use `docker compose` to start a TGI container and Llama Stack server container together.
```bash
$ cd distributions/{{ name }}; docker compose up
```
The script will first start up TGI server, then start up Llama Stack distribution server hooking up to the remote TGI provider for inference. You should be able to see the following outputs --
```bash
[text-generation-inference] | 2024-10-15T18:56:33.810397Z INFO text_generation_router::server: router/src/server.rs:1813: Using config Some(Llama)
[text-generation-inference] | 2024-10-15T18:56:33.810448Z WARN text_generation_router::server: router/src/server.rs:1960: Invalid hostname, defaulting to 0.0.0.0
[text-generation-inference] | 2024-10-15T18:56:33.864143Z INFO text_generation_router::server: router/src/server.rs:2353: Connected
INFO: Started server process [1]
INFO: Waiting for application startup.
INFO: Application startup complete.
INFO: Uvicorn running on http://[::]:5001 (Press CTRL+C to quit)
```
To kill the server
```bash
docker compose down
```
### Conda: TGI server + llama stack run
If you wish to separately spin up a TGI server, and connect with Llama Stack, you may use the following commands.
#### Start TGI server locally
- Please check the [TGI Getting Started Guide](https://github.com/huggingface/text-generation-inference?tab=readme-ov-file#get-started) to get a TGI endpoint.
```bash
docker run --rm -it -v $HOME/.cache/huggingface:/data \
-p 5009:5009 --gpus all \
ghcr.io/huggingface/text-generation-inference:latest \
--dtype bfloat16 --usage-stats on --sharded false \
--model-id meta-llama/Llama-3.2-3B-Instruct --port 5009
```
#### Start Llama Stack server pointing to TGI server
**Via Conda**
```bash
llama stack build --template {{ name }} --image-type conda
# -- start a TGI server endpoint
llama stack run ./gpu/run.yaml
```
**Via Docker**
```bash
docker run --network host -it -p 5001:5001 \
-v ./run.yaml:/root/my-run.yaml --gpus=all \
llamastack/distribution-{{ name }} \
--yaml_config /root/my-run.yaml
```
We have provided a template `run.yaml` file in the `distributions/{{ name }}` directory. Make sure in your `run.yaml` file, you inference provider is pointing to the correct TGI server endpoint. E.g.
```yaml
inference:
- provider_id: tgi0
provider_type: remote::tgi
config:
url: http://127.0.0.1:5009
```
### (Optional) Update Model Serving Configuration
To serve a new model with `tgi`, change the docker command flag `--model-id <model-to-serve>`.
This can be done by edit the `command` args in `compose.yaml`. E.g. Replace "Llama-3.2-1B-Instruct" with the model you want to serve.
```yaml
command: >
--dtype bfloat16 --usage-stats on --sharded false
--model-id meta-llama/Llama-3.2-1B-Instruct
--port 5009 --cuda-memory-fraction 0.7
```
or by changing the docker run command's `--model-id` flag
```bash
docker run --rm -it -v $HOME/.cache/huggingface:/data \
-p 5009:5009 --gpus all \
ghcr.io/huggingface/text-generation-inference:latest \
--dtype bfloat16 --usage-stats off --sharded false \
--model-id meta-llama/Llama-3.2-3B-Instruct --port 5009
```
In `run.yaml`, make sure you point the correct server endpoint to the TGI server endpoint serving your model.
```yaml
inference:
- provider_id: tgi0
provider_type: remote::tgi
config:
url: http://127.0.0.1:5009
```

97
llama_stack/templates/tgi/tgi.py Normal file
View file

@ -0,0 +1,97 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from pathlib import Path
from llama_stack.distribution.datatypes import ModelInput, Provider, ShieldInput
from llama_stack.providers.remote.inference.tgi import TGIImplConfig
from llama_stack.templates.template import DistributionTemplate, RunConfigSettings
def get_distribution_template() -> DistributionTemplate:
providers = {
"inference": ["remote::tgi"],
"memory": ["inline::faiss", "remote::chromadb", "remote::pgvector"],
"safety": ["inline::llama-guard"],
"agents": ["inline::meta-reference"],
"telemetry": ["inline::meta-reference"],
}
inference_provider = Provider(
provider_id="tgi-inference",
provider_type="remote::tgi",
config=TGIImplConfig.sample_run_config(
url="${env.TGI_URL}",
),
)
inference_model = ModelInput(
model_id="${env.INFERENCE_MODEL}",
provider_id="tgi-inference",
)
safety_model = ModelInput(
model_id="${env.SAFETY_MODEL}",
provider_id="tgi-safety",
)
return DistributionTemplate(
name="tgi",
distro_type="self_hosted",
description="Use (an external) TGI server for running LLM inference",
docker_image="llamastack/distribution-remote-tgi:test-0.0.52rc3",
template_path=Path(__file__).parent / "doc_template.md",
providers=providers,
default_models=[inference_model, safety_model],
run_configs={
"run.yaml": RunConfigSettings(
provider_overrides={
"inference": [inference_provider],
},
default_models=[inference_model],
),
"run-with-safety.yaml": RunConfigSettings(
provider_overrides={
"inference": [
inference_provider,
Provider(
provider_id="tgi-safety",
provider_type="remote::tgi",
config=TGIImplConfig.sample_run_config(
url="${env.SAFETY_TGI_URL}",
),
),
],
},
default_models=[
inference_model,
safety_model,
],
default_shields=[ShieldInput(shield_id="${env.SAFETY_MODEL}")],
),
},
docker_compose_env_vars={
"LLAMASTACK_PORT": (
"5001",
"Port for the Llama Stack distribution server",
),
"INFERENCE_MODEL": (
"meta-llama/Llama-3.2-3B-Instruct",
"Inference model loaded into the TGI server",
),
"TGI_URL": (
"http://host.docker.internal:8080}/v1",
"URL of the TGI server with the main inference model",
),
"SAFETY_TGI_URL": (
"http://host.docker.internal:8081/v1",
"URL of the TGI server with the safety model",
),
"SAFETY_MODEL": (
"meta-llama/Llama-Guard-3-1B",
"Name of the safety (Llama-Guard) model to use",
),
},
)