<!--
SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
SPDX-License-Identifier: Apache-2.0
-->
# aiconfigurator
> [!WARNING]
> **AIConfigurator is in a maintenance-only transition to
> [AISimulate](https://github.com/ai-dynamo/aisimulate).** New features, model
> coverage, and other active development belong in the AISimulate repository.
> AIC accepts only bug, security, and migration-blocking fixes while its 0.12
> compatibility window is completed. Existing releases and repository history
> remain available; follow the
> [AISimulate migration guide](docs/aisimulate_migration.md) after the stable
> replacement artifacts are published.
[](https://deepwiki.com/ai-dynamo/aiconfigurator)
[](https://discord.gg/mRJ2KNzwYE)
Explore the [AIC Developer Universe](https://ai-dynamo.github.io/aiconfigurator/universe/), an interactive map of AIConfigurator and its Dynamo integration.
In disaggregated serving, configuring an effective deployment is challenging: you need to decide how many prefill and decode
workers to run, and the parallelism for each worker. Combined with SLA targets for TTFT (Time to First Token) and
TPOT (Time per Output Token), optimizing throughput at a given latency becomes even more complex.
`aiconfigurator` helps you find a strong starting configuration for disaggregated serving. Given your model, GPU
count, and GPU type, it searches the configuration space and generates configuration files you can use for deployment with Dynamo or llm-d.
For a technical deep dive into the design and methodology of AIConfigurator, please refer to our paper:
[**AIConfigurator: Lightning-Fast Configuration Optimization for Multi-Framework LLM Serving**](https://arxiv.org/abs/2601.06288).
The tool models LLM inference using collected data for a target machine and framework. It evaluates thousands of
configurations and runs anywhere via the CLI.
## Build and Install
### Compatibility install from PyPI
New development and new integrations should target AISimulate. Until the
stable AISimulate 0.12 artifacts are published, existing AIC users may continue
installing the latest supported AIC release:
> **Public PyPI support: Linux x86-64 only.** The required
> `aiconfigurator-core` wheel bundles a native Rust/PyO3 extension and is
> published to PyPI as a `manylinux_2_28_x86_64` wheel (Linux x86-64,
> glibc >= 2.28).
>
> **Release wheel coverage:** The
> [platform-wheel workflow](.github/workflows/build-platform-wheels.yml) also
> builds, installs, and verifies `manylinux_2_28_aarch64` and
> `macosx_11_0_arm64` core wheels. Release-branch runs stage the complete wheel
> set to Artifactory when platform-wheel staging is enabled. These Artifactory
> artifacts are separate from public PyPI: Linux aarch64 and macOS arm64 users
> with access to the release artifacts can install the matching wheel set.
> Each staged set publishes `_MANIFEST.json` with immutable wheel SHA-256 and
> size metadata, then writes `_COMPLETE.json` last with the manifest checksum.
> Consumers must verify both files before installing a wheel.
> Copied pull-request workflows validate wheels only on isolated GitHub-hosted
> runners. They do not use the persistent release runners, receive Artifactory
> credentials, or publish wheel artifacts.
> Windows has no supported installation path.
```bash
python -m pip install aiconfigurator
```
After the stable AISimulate 0.12 artifacts are published on PyPI and the
corresponding release is announced in
[AISimulate releases](https://github.com/ai-dynamo/aisimulate/releases), migrate
the installed distribution while retaining the established `aiconfigurator`
command:
```bash
python -m pip uninstall -y aiconfigurator aiconfigurator-core
python -m pip install "aisimulate==0.12.0"
```
Do not substitute an AISimulate development build for the stable transition
artifact. See the [migration guide](docs/aisimulate_migration.md) for the
publication gate, supported platforms, CLI behavior, and API changes.
The upper `aiconfigurator` wheel contains the CLI, generator, and versioned
server-config adapter.
It depends on the exact matching `aiconfigurator-core` wheel, which independently
owns the SDK, model/system data, and native extension. Installing
`aiconfigurator` therefore installs the complete product, while core-only
consumers can install `aiconfigurator-core` without pulling in the upper layer.
`Task` and the orchestration APIs live in the application wheel only:
```python
from aiconfigurator.sdk.task_v2 import Task
```
The core wheel intentionally does not expose `task_v2`; the standalone core
never depends back on the application package.
#### Upgrading an existing AIC environment from 0.9
Version 0.9 shipped core files inside `aiconfigurator`. Package installers cannot
safely transfer those same paths to the new dependency during a normal in-place
upgrade because dependencies are installed before dependents. Remove the old
owner first when crossing this package boundary. This compatibility path applies
only after the matching AIC 0.12 artifact is published; it is not the starting
point for new development or integrations:
```bash
python3 -m pip uninstall -y aiconfigurator aiconfigurator-core
python3 -m pip install 'aiconfigurator==0.12.0'
```
If a normal upgrade was already attempted, repair the core payload with:
```bash
python3 -m pip install --force-reinstall --no-deps 'aiconfigurator-core==0.12.0'
```
### Build and Install from Source
```bash
# 1. Clone the repo
git clone https://github.com/ai-dynamo/aiconfigurator.git
cd aiconfigurator
# 2. Create and activate a virtual environment
python3 -m venv myenv && source myenv/bin/activate # (requires Python 3.11-3.13)
# 3. Install the standalone core, then the upper package
pip3 install ./aic-core
pip3 install .
```
Current performance profiles are checked-in Parquet files, so normal builds
and usage do not require Git LFS. Install Git LFS and run `git lfs pull` only
when working with retained legacy `*.txt` perf assets or their compatibility
tests.
### Build with Docker
```bash
# This creates disjoint upper AIC and standalone core wheels
docker build -f docker/Dockerfile --no-cache --target build -t aiconfigurator:latest .
docker create --name aic aiconfigurator:latest && docker cp aic:/workspace/dist dist/ && docker rm aic
```
## Run
### CLI
```bash
aiconfigurator cli default --model Qwen/Qwen3-32B-FP8 --total-gpus 32 --system h200_sxm
aiconfigurator cli recommend --model Qwen/Qwen3-32B-FP8 --system h200_sxm --target-request-rate 50 --ttft 2000 --tpot 30
aiconfigurator cli exp --yaml-path exp.yaml
aiconfigurator cli generate --model-path Qwen/Qwen3-32B-FP8 --total-gpus 8 --system h200_sxm
aiconfigurator cli support --model-path Qwen/Qwen3-32B-FP8 --system h200_sxm
```
- We have six modes: `default`, `estimate`, `recommend`, `exp`, `generate`, and `support`.
- Use `default` to find the estimated best deployment by searching the configuration space.
- The experimental Spica smart sweeper now lives in Dynamo's standalone
[AI Simulate distribution](https://docs.nvidia.com/dynamo/dev/knowledge-base/modular-components/ai-simulate/spica/overview).
From a matching Dynamo checkout, install it with `python -m pip install ./aisimulate`, then use
`python -m aisimulate.spica` for Spica searches. Runnable configurations and tools live under
`examples/aisimulate/spica`.
- Use `exp` to run customized experiments defined in a YAML file.
- Use `generate` to quickly create a naive configuration without a parameter sweep.
- Use `recommend` to find the minimum GPU count and optimal deployment configuration needed to meet a performance target. This mode is designed as a procurement sizing tool -- specify exactly one load target (`--target-request-rate` or `--target-concurrency` — mutually exclusive) along with SLA constraints, and the system calculates the minimum GPUs required. You can also omit `--total-gpus` in default mode with a load target for the same behavior.
- Use `support` to verify if AIC supports a model/hardware combination for agg and disagg modes.
- `--model` is an alias for `--model-path` in the CLI.
- Use `--backend` to specify the inference backend: `trtllm` (default), `vllm`, or `sglang`.
- Use `--deployment-target` to specify the artifact platform: `dynamo-j2` (default, typed Dynamo manifests), `dynamo-python`, `llm-d-helm`, `llm-d-kustomize`, or `fpm`. FPM V1 supports one aggregated vLLM worker group and emits exactly three artifacts: a reusable keepalive Pod, LeaderWorkerSet, or Grove PodCliqueSet in `k8s_deploy.yaml`, `fpm_env.sh` (the collection-facts contract file), and a launch-only `run.sh`; see the [Generator overview](docs/generator_overview.md#fpm-v1-target).
- Use `exp`, pass in exp.yaml by `--yaml-path` to customize your experiments and even a heterogenous one.
- Use `--save-dir DIR` to generate deployment artifacts for the selected target (Dynamo manifests, llm-d values/overlays, or an FPM resource workload + script).
- Use `--database-mode` to control performance estimation mode: `SILICON` (default, uses collected silicon data), `HYBRID` (uses silicon data when available, otherwise SOL+empirical), `EMPIRICAL` (SOL+empirical for all), or `SOL` (speed-of-light only). Please be careful, only `SILICON` mode's result is reproducible. Other modes are for research purpose
- Use `--systems-paths` to override where system YAMLs and data are loaded from (comma-separated; `default` maps to the built-in systems path). First match wins for identical system/backend/version.
- Use `-h` for more options and customization.
- SLA constraints:
- `--ttft` and `--tpot` filter configurations that exceed either bound; omit a flag to leave that constraint unset.
- `--request-latency` applies an end-to-end per-request limit. The CLI searches for all configurations whose estimated
latency stays within that budget, optionally honoring a provided `--ttft`.
When this flag is set, `--tpot` becomes implicit and is ignored.
Quantization defaults are inferred from the Hugging Face model config (`config.json` plus optional `hf_quant_config.json`).
For low-precision models, use a quantized HF ID (for example, `Qwen/Qwen3-32B-FP8`) or a local model directory containing those files.
Any quantization set via `profiles` or YAML `config` overrides the HF defaults.
For a full end-to-end walkthrough (support check, sweep, deploy, benchmark), see the [CLI User Guide -- End-to-End Workflow](docs/cli_user_guide.md#end-to-end-workflow).
Refer to [CLI User Guide](docs/cli_user_guide.md)
### Python API
You can also use `aiconfigurator` programmatically in Python:
```python
from aiconfigurator.cli import cli_default, cli_exp, cli_generate, cli_recommend, cli_support
# 1. Run default agg vs disagg comparison
result = cli_default(model_path="Qwen/Qwen3-32B-FP8", total_gpus=32, system="h200_sxm")
print(result.best_configs["disagg"].head())
# 2. Run experiments from a YAML file or a dictionary config
result = cli_exp(yaml_path="my_experiments.yaml")
# Or use a dictionary config directly
result = cli_exp(config={
"my_exp": {
"serving_mode": "disagg",
"model_path": "Qwen/Qwen3-32B-FP8",
"total_gpus": 32,
"system_name": "h200_sxm",
"isl": 4000,
"osl": 1000,
}
})
# 3. Find minimum GPUs for a performance target (procurement sizing)
result = cli_recommend(model_path="Qwen/Qwen3-32B-FP8", system="h200_sxm", target_request_rate=50.0, ttft=2000, tpot=30)
for mode, df in result.best_configs.items():
print(f"{mode}: {df[['total_gpus_needed', 'replicas_needed', 'tp', 'tpot']].head()}")
# 4. Generate a naive configuration
result = cli_generate(model_path="Qwen/Qwen3-32B-FP8", total_gpus=8, system="h200_sxm")
print(result["parallelism"]) # {'tp': 1, 'pp': 1, 'replicas': 8, 'gpus_used': 8}
# 5. Check support for a model/system combination
agg, disagg = cli_support(model_path="Qwen/Qwen3-32B-FP8", system="h200_sxm")
print(f"Agg supported: {agg}, Disagg supported: {disagg}")
```
Serving configs can be adapted into validated estimate requests without running
an estimate. InferenceX DB records, DynamoGraphDeployments, and concrete
`dynamo-ci` SGLang benchmark recipes are supported:
```python
from pathlib import Path
from aiconfigurator.sdk.config_adapter import (
AdapterOverrides,
DynamoRecipeSource,
adapt_config,
to_cli_estimate_kwargs,
)
report = adapt_config(
DynamoRecipeSource(Path("deploy.yaml"), Path("perf.yaml")),
AdapterOverrides(system_name="h200_sxm"),
)
for request in report.requests:
kwargs = to_cli_estimate_kwargs(request)
```
See the [Config Adapter Guide](docs/config_adapter.md) for schema, precedence,
supported recipe shapes, diagnostics, and explicit estimate execution.
An example here,
```bash
aiconfigurator cli default --model-path Qwen/Qwen3-32B-FP8 --total-gpus 32 --system h200_sxm --isl 4000 --osl 500 --prefix 500 --ttft 300 --tpot 10
```
```text
********************************************************************************
* AIConfigurator Final Results *
********************************************************************************
----------------------------------------------------------------------------
Input Configuration & SLA Target:
Model: Qwen/Qwen3-32B-FP8 (is_moe: False)
Total GPUs: 32
Best Experiment Chosen: disagg at 684.79 tokens/s/gpu (disagg 1.67x better)
----------------------------------------------------------------------------
Overall Best Configuration:
- Best Throughput: 21,913.22 tokens/s
- Per-GPU Throughput: 684.79 tokens/s/gpu
- Per-User Throughput: 100.31 tokens/s/user
- TTFT: 295.71ms
- TPOT: 9.97ms
- Request Latency: 5270.24ms
----------------------------------------------------------------------------
Pareto Frontier:
Qwen/Qwen3-32B-FP8 Pareto Frontier: tokens/s/gpu_cluster vs tokens/s/user
┌────────────────────────────────────────────────────────────────────────┐
1250.0┤ •• agg │
│ ff disagg │
│ xx disagg best │
│ │
1041.7┤ │
│ f │
│ fffffffff │
│ fff │
833.3┤ ffff │
│ f │
│ • ff │
│ •• fxfff │
625.0┤ ••••• f │
│ • f │
│ •••••••••••• f │
│ ••• f │
416.7┤ ••••ff │
│ ••ff │
│ •fffffffffffff │
│ ••••••••ff• │
208.3┤ ff••• │
│ ff •••• │
│ fff ••• │
│ • │
0.0┤ │
└┬─────────────────┬─────────────────┬────────────────┬─────────────────┬┘
0 60 120 180 240
tokens/s/gpu_cluster tokens/s/user
----------------------------------------------------------------------------
Deployment Details:
(p) stands for prefill, (d) stands for decode, bs stands for batch size, a replica stands for the smallest scalable unit xPyD of the disagg system
Some math: total gpus used = replicas * gpus/replica
gpus/replica = (p)gpus/worker * (p)workers + (d)gpus/worker * (d)workers; for Agg, gpus/replica = gpus/worker
gpus/worker = tp * pp * dp = etp * ep * pp for MoE models; tp * pp for dense models (underlined numbers are the actual values in math)
agg Top Configurations: (Sorted by tokens/s/gpu)
+------+---------+--------------+---------------+--------+-----------------+-------------+-------------------+----------+--------------+-------------+----------+----+
| Rank | backend | tokens/s/gpu | tokens/s/user | TTFT | request_latency | concurrency | total_gpus (used) | replicas | gpus/replica | gpus/worker | parallel | bs |
+------+---------+--------------+---------------+--------+-----------------+-------------+-------------------+----------+--------------+-------------+----------+----+
| 1 | trtllm | 410.22 | 108.48 | 251.10 | 4850.91 | 128 (=16x8) | 32 (32=8x4) | 8 | 4 | 4 (=4x1x1) | tp4pp1 | 16 |
| 2 | trtllm | 361.33 | 107.43 | 224.48 | 4869.40 | 112 (=28x4) | 32 (32=4x8) | 4 | 8 | 8 (=8x1x1) | tp8pp1 | 28 |
| 3 | trtllm | 117.92 | 122.25 | 292.72 | 4374.38 | 32 (=2x16) | 32 (32=16x2) | 16 | 2 | 2 (=2x1x1) | tp2pp1 | 2 |
+------+---------+--------------+---------------+--------+-----------------+-------------+-------------------+----------+--------------+-------------+----------+----+
disagg Top Configurations: (Sorted by tokens/s/gpu)
+------+---------+--------------+---------------+--------+-----------------+--------------+-------------------+----------+---------------+------------+----------------+-------------+-------+------------+----------------+-------------+-------+
| Rank | backend | tokens/s/gpu | tokens/s/user | TTFT | request_latency | concurrency | total_gpus (used) | replicas | gpus/replica | (p)workers | (p)gpus/worker | (p)parallel | (p)bs | (d)workers | (d)gpus/worker | (d)parallel | (d)bs |
+------+---------+--------------+---------------+--------+-----------------+--------------+-------------------+----------+---------------+------------+----------------+-------------+-------+------------+----------------+-------------+-------+
| 1 | trtllm | 684.79 | 100.31 | 295.71 | 5270.24 | 272 (=68x4) | 32 (32=4x8) | 4 | 8 (=2x2+1x4) | 2 | 2 (=2x1) | tp2pp1 | 1 | 1 | 4 (=4x1) | tp4pp1 | 68 |
| 2 | trtllm | 684.79 | 100.16 | 295.71 | 5277.73 | 240 (=120x2) | 32 (32=2x16) | 2 | 16 (=4x2+1x8) | 4 | 2 (=2x1) | tp2pp1 | 1 | 1 | 8 (=8x1) | tp8pp1 | 120 |
| 3 | trtllm | 404.71 | 100.35 | 295.71 | 5268.25 | 140 (=140x1) | 32 (24=1x24) | 1 | 24 (=5x2+7x2) | 5 | 2 (=2x1) | tp2pp1 | 1 | 7 | 2 (=2x1) | tp2pp1 | 20 |
+------+---------+--------------+---------------+--------+-----------------+--------------+-------------------+----------+---------------+------------+----------------+-------------+-------+------------+----------------+-------------+-------+
********************************************************************************
2026-02-08 23:10:21,413 - aiconfigurator.cli.main - INFO - All experiments completed in 6.50 seconds
```
These results indicate that deploying Qwen3-32B-FP8 on h200_sxm in FP8 can achieve **1.67x** higher tokens/s/gpu for disaggregated versus aggregated deployment **under the SLA targets TTFT ≤ 300 ms and TPOT ≤ 10 ms**, with ISL:OSL of 4000:500 (with prefix len: 500).
Try different ISL:OSL values and SLA limits to fit your use case, for example:
```bash
aiconfigurator cli default --model-path Qwen/Qwen3-32B-FP8 --total-gpus 32 --system h200_sxm --ttft 200 --tpot 10 --isl 8000 --osl 200 --prefix 500
```
You will get different results.
### Customized Configuration for aiconfigurator
The `default` mode will create two experiments, one is `agg` and another one is `disagg` and then compare the results.
To further customize (including the search space and per-component quantization), parameters are defined in a YAML file.
Built-in YAML files are under `src/aiconfigurator/cli/example.yaml` and `src/aiconfigurator/cli/exps/*.yaml`
Refer to the YAML file and modify as needed. Pass your customized YAML file to `exp` mode:
```bash
aiconfigurator cli exp --yaml-path customized_config.yaml
```
We can use `exp` mode to compare multiple results, including disagg vs. agg, homogeneous vs. heterogeneous, and more than 2 experiments.
We've crafted several examples in `src/aiconfigurator/cli/exps/*.yaml`
For the full guide, refer to [CLI User Guide](docs/cli_user_guide.md).
### Deploying to llm-d Platform
AIConfigurator supports deploying to the llm-d platform using Helm values. Use `--deployment-target llm-d-helm` with vLLM or SGLang backends:
```bash
# vLLM on llm-d
aiconfigurator cli default \
--model-path Qwen/Qwen3-32B \
--total-gpus 32 \
--system h200_sxm \
--backend vllm \
--deployment-target llm-d-helm \
--save-dir ./output
# SGLang on llm-d
aiconfigurator cli default \
--model-path Qwen/Qwen3-32B \
--total-gpus 32 \
--system h200_sxm \
--backend sglang \
--deployment-target llm-d-helm \
--save-dir ./output
```
This generates `llm-d-values.yaml` files compatible with the llm-d-modelservice Helm chart. The generated Helm values include model artifacts, parallelism settings, and container configurations optimized for your workload.
You can customize llm-d-specific settings using generator overrides:
```bash
--generator-set LlmdConfig.vllm_image=vllm/vllm-openai:v0.6.0 \
--generator-set LlmdConfig.model_cache_size=200Gi \
--generator-set LlmdConfig.routing_proxy_enabled=true
```
### Generate Configurations for Dynamo and Reproduce the results
Please refer to the [Deployment Guide](docs/dynamo_deployment_guide.md) for details about deployment and reproduction especially about the benchmark methodology.
To simplify the deployment and reproduction, in the `aiconfigurator` CLI, if you specify `--save-dir`, the tool generates configuration files for your chosen deployment target.
The folder structure varies based on `--deployment-target`:
**For Dynamo deployments** (`--deployment-target dynamo-j2` or `dynamo-python`):
```text
results/QWEN3_32B_FP8_h200_sxm_trtllm_isl4000_osl1000_ttft1000_tpot20_904495
├── agg
│ ├── best_config_topn.csv
│ ├── config.yaml
│ ├── pareto.csv
│ ├── top1
│ │ ├── agg
│ │ │ ├── agg_config.yaml
│ │ │ ├── bench_run.sh # aiperf benchmark sweep script (bare-metal)
│ │ │ ├── k8s_bench.yaml # aiperf benchmark sweep Job (Kubernetes)
│ │ │ ├── k8s_deploy.yaml
│ │ │ └── node_0_run.sh
│ │ └── generator_config.yaml
│ ...
├── disagg
│ ├── best_config_topn.csv
│ ├── config.yaml
│ ├── pareto.csv
│ ├── top1
│ │ ├── disagg
│ │ │ ├── bench_run.sh # aiperf benchmark sweep script (bare-metal)
│ │ │ ├── decode_config.yaml
│ │ │ ├── k8s_bench.yaml # aiperf benchmark sweep Job (Kubernetes)
│ │ │ ├── k8s_deploy.yaml
│ │ │ ├── node_0_run.sh
│ │ │ └── prefill_config.yaml
│ │ └── generator_config.yaml
│ ...
└── pareto_frontier.png
```
**For llm-d Helm deployments** (`--deployment-target llm-d-helm`):
```text
results/QWEN3_32B_h200_sxm_vllm_isl4000_osl1000_ttft1000_tpot20_904495
├── disagg
│ ├── best_config_topn.csv
│ ├── config.yaml
│ ├── pareto.csv
│ ├── top1
│ │ ├── disagg
│ │ │ ├── decode_config.yaml
│ │ │ ├── llm-d-values.yaml # Helm values for llm-d-modelservice chart
│ │ │ ├── node_0_run.sh
│ │ │ └── prefill_config.yaml
│ │ └── generator_config.yaml
│ ...
└── pareto_frontier.png
```
Note: llm-d deployments generate `llm-d-values.yaml` instead of `k8s_deploy.yaml` and `k8s_bench.yaml`.
Use `--generator-config path/to/file.yaml` to load a YAML payload with `ServiceConfig`, `K8sConfig`, `DynConfig`, `WorkerConfig`, and `Workers.<role>` sections, or specify inline overrides with `--generator-set KEY=VALUE` (repeatable). Examples:
- `--generator-set ServiceConfig.model_path=Qwen/Qwen3-32B-FP8`
- `--generator-set K8sConfig.k8s_namespace=dynamo \`
Run `aiconfigurator cli default --generator-help` to print information that is sourced directly from `src/aiconfigurator/generator/config/deployment_config.yaml` and `backend_config_mapping.yaml`.
## Tuning with Advanced Features
There are many features, such as different quantizations and parallelism strategies, to tune performance beyond the default configurations.
These apply to the CLI. Refer to [Advanced Tuning](docs/advanced_tuning.md) for details.
## How It Works
### Modeling and Mechanism
LLM inference performance is dominated by:
1. Compute cost (such as GEMM and attention).
2. Communication cost (such as all-reduce for tensor parallel and P2P for pipeline parallel).
To estimate performance, we take the following steps:
1. Break down LLM inference into operations: GEMM, attention, communication, embedding, element-wise operations, and others.
2. Collect operation execution times on the target hardware.
3. Estimate end-to-end execution time for a configuration by composing operation times using interpolation and extrapolation.
4. Model in-flight batching (aggregated) and disaggregated serving on top of that.
5. Search thousands of combinations to find strong configurations and generate Dynamo or llm-d configuration files based on the results.
### Supported Features
- **Models**:
- GPT
- LLAMA (2, 3)
- MOE
- QWEN
- DEEPSEEK_V3
- Support using huggingface model id if falls into these model family and not MoE models.
- **Operations**:
- Attention
- MHA/GQA (FP8, BF16)
- MLA (FP8, BF16)
- KV Cache (BF16, FP8, INT8)
- GEMM (BF16, FP8, FP8-Block, FP8-OOTB, SQ, INT8 WO, INT4 WO, NVFP4)
- CustomAllReduce (BF16)
- Embedding
- P2P
- ElementWise
- NCCL (all_reduce, all_gather, all-to-all, reduce_scatter)
- MoE (BF16, FP8, FP8-Block, W4A-FP8, INT4 WO, NVFP4)
- MLA BMM (BF16, FP8)
- **Parallel modes**:
- Tensor-parallel
- Pipeline-parallel
- Expert Tensor-parallel/Expert-parallel
- Attention DP (for DEEPSEEK and MoE)
- **Scheduling**:
- Static
- Aggregated serving (continuous batching)
- Disaggregated serving
- MTP (for DEEPSEEK)
- **Inference Backends**:
- TensorRT-LLM (trtllm)
- vLLM
- SGLang
### Data Collection
Data collection is a standalone process for building the database used by aiconfigurator. By default, you do not need to collect data yourself.
Small changes to the database may not materially change performance estimates. For example, you can use 1.0.0rc3 data of `trtllm` on `h200_sxm` and deploy the generated configuration with Dynamo and a `trtllm` 1.0.0rc4 worker.
To go through the process, refer to the [guidance](collector/README.md) under the `collector` folder.
**New:** The collector now supports optional GPU power monitoring during kernel execution. Use the `--measure_power` flag to collect power consumption data alongside performance metrics. See the [collector README](collector/README.md#power-monitoring-optional) for details.
### System Data Support Matrix
| System | Framework(Version) | Status |
|--------|-------------------|--------|
| h100_sxm | TRTLLM(1.0.0rc3, 1.2.0rc5), SGLang(0.5.6.post2), vLLM(0.12.0) | ✅ |
| h200_sxm | TRTLLM(1.0.0rc3, 1.2.0rc5), SGLang(0.5.6.post2), vLLM(0.12.0) | ✅ |
| b200_sxm | TRTLLM(1.0.0rc3, 1.2.0rc5), SGLang(0.5.6.post2) | ✅ |
| gb200 | TRTLLM(1.0.0rc3, 1.2.0rc5) | ✅ |
| a100_sxm | TRTLLM(1.0.0), vLLM(0.12.0) | ✅ |
| h100_pcie, a100_pcie, l4, a30 | Estimate-only system specs for `generate` and non-SILICON database modes (`SOL`, `EMPIRICAL`, `HYBRID`) | ⚠️ |
(last updated: 2026/02/02)
> **Note**: b200 and gb200 are under dev. Results are to be aligned. For preview now.
> `h100_pcie`, `a100_pcie`, `l4`, and `a30` do not include built-in silicon performance databases yet. Use them for naive sizing or rough SOL/EMPIRICAL estimates, and use `--systems-paths` to provide measured data for production-quality predictions.
#### Detailed Support Matrix
For a comprehensive, interactive view of which model/system/backend/version combinations are supported in both aggregated and disaggregated modes, visit the **[Support Matrix on GitHub Pages](https://ai-dynamo.github.io/aiconfigurator/support-matrix/)**. The page fetches the split support matrix CSV files directly from GitHub at load time and supports filtering by system, mode, model search, and switching between branches.
The raw data is also available as [per-system CSV files](aic-core/src/aiconfigurator_core/systems/support_matrix).
You can also check support via the CLI:
```bash
aiconfigurator cli support --model-path Qwen/Qwen3-32B-FP8 --system h100_sxm --backend-version 1.2.0rc5
```
## Contributing and Development
We welcome contributions from the community! Check out the below resources to get started:
- [DEVELOPMENT.md](DEVELOPMENT.md) - Set up your development environment, run tests, and follow our coding standards
- [CONTRIBUTING.md](CONTRIBUTING.md) - Contribution guidelines and requirements
- [Discord](https://discord.gg/mRJ2KNzwYE) - Chat with team and community
### How To Add A New Model
Adding a new model will require modifying the source code and perhaps collecting new data for the model. Please refer to [How to Add a New Model](docs/add_a_new_model.md).
## Citation
If you use AIConfigurator for your research, please cite our paper:
```bibtex
@article{xu2026aiconfigurator,
title={AIConfigurator: Lightning-Fast Configuration Optimization for Multi-Framework LLM Serving},
author={Tianhao Xu and Yiming Liu and Xianglong Lu and Yijia Zhao and Xuting Zhou and Aichen Feng and Yiyi Chen and Yi Shen and Qin Zhou and Xumeng Chen and Ilya Sherstyuk and Haorui Li and Rishi Thakkar and Ben Hamm and Yuanzhe Li and Xue Huang and Wenpeng Wu and Anish Shanbhag and Harry Kim and Chuan Chen and Junjie Lai},
journal={arXiv preprint arXiv:2601.06288},
year={2026}
}
```
## Known Issues
1. Memory estimation for the backends needs to be studied more.
2. Results can be overly optimistic in the low-speed, high-throughput region.
3. **vLLM and SGLang support is currently being evaluated**. While both backends are functional and available for use, we are still completing comprehensive performance evaluations and alignment testing. We recommend validating results with real benchmarks for production use.
> **Note**: The results are not final or absolute. They can be inaccurate due to modeling gaps or indicate performance improvement opportunities. The tool aims to align with the framework's current implementation and to provide configuration suggestions. Verify results in real benchmarks with the generated configurations and perform follow-up tuning.