Modelpublic

Arm/tinyllama-1-1b-chat-onnx-genai-int4-kquantlast-emb-int8-graviton-g4

sourceHugging Faceapache-2.0updated 21d agoView on Hugging Face

0likes501downloads

Model Card

TinyLlama-1.1B-Chat-v1.0 optimized for Arm-based Cloud CPU

A decoder-only chat language model for text generation, provided in ONNX for the ONNX Runtime GenAI runtime and quantized to INT4 groupwise asymmetric weights with INT8 per-token dynamic activations, INT8 per-row embeddings and an INT8 LM head, targeting Arm-based Cloud CPU systems.

Summary

This repository contains an Arm-optimized version of TinyLlama/TinyLlama-1.1B-Chat-v1.0 for text generation. The model is provided in ONNX for the ONNX Runtime GenAI runtime, targeting Cloud CPU systems. The transformer weights are quantized with GPTQ to INT4, asymmetric per-group with group size 32, activations are INT8 per-token dynamic, and the embedding table and LM head are INT8.

This version is intended to demonstrate efficient inference on Arm-based platforms while preserving the original model's intended behavior. Arm has evaluated this model on HellaSwag and measured performance on a representative evaluation target.

Key results

Area	Result
Model format	ONNX
Target device class	Cloud CPU
Reference device	AWS Graviton G4 (Neoverse-V2, Linux Ubuntu 24.04.4 LTS)
Primary performance result	66.88 tokens/sec decode throughput, 407.29 ms time to first token
Accuracy result	HellaSwag 59.0 %
Size / memory result	727.86 MB with INT4 weights, INT8 activations, INT8 embeddings and INT8 LM head (5.77x smaller than the FP32 baseline); peak memory 864.55 MB

Original model

Field	Value
Original model	TinyLlama/TinyLlama-1.1B-Chat-v1.0
Original source	Hugging Face
Original developer	The TinyLlama project
Original model card	TinyLlama/TinyLlama-1.1B-Chat-v1.0
Original license	Apache-2.0

Model files

File	Description
`model.onnx`	Arm-optimized INT4-weight model graph for deployment; weights ship in the `model.onnx.data` side-car
`example.py`	Minimal inference example
`pyproject.toml`	Pinned runtime dependencies for `example.py`, resolved with uv
`uv.lock`	Locked dependency resolution for `pyproject.toml`
`config.yaml`	Model I/O contract used by the example
`benchmarks/`	FP32 baseline and Arm-optimized benchmark records

Performance

Performance was measured on the reference configuration below. Results are intended to make the optimization reproducible but do not guarantee identical performance on every Arm-based system.

Reference configuration

Field	Value
Device / platform	AWS Graviton G4
Instance	c8g.12xlarge
CPU / accelerator	Neoverse-V2, aarch64, CPU execution backend
OS	Linux Ubuntu 24.04.4 LTS
Runtime	ONNX Runtime 1.26.0
Backend / delegate	MLAS, KleidiAI
Batch size	1
Threads	4 intra-op threads
Precision	INT4 groupwise asymmetric weights (GPTQ, group size 32), INT8 per-token dynamic activations, INT8 per-row embeddings, INT8 LM head (scheme W4A8dynemb_int8)
Runs	10 warmup runs + 50 measured runs

Performance results

Metric	Original / baseline	Arm-optimized	Improvement
End-to-end latency p50 (ms)	4376.03	1307.46	3.35x faster
End-to-end latency p90 (ms)	4408.70	1318.86	3.34x faster
End-to-end latency p99 (ms)	4453.49	1338.23	3.33x faster
Decode throughput (tokens/sec)	17.71	66.88	3.78x
Time to first token (ms)	980.35	407.29	2.41x faster
Model load time (ms)	2074.28	974.83	2.13x faster
Time to first inference (ms)	4364.16	1312.96	3.32x faster
Model size (MB)	4196.99	727.86	5.77x smaller
Peak memory (MB)	4080.35	864.55	4.72x less
Average memory (MB)	4058.90	854.32	4.75x less

Accuracy

Accuracy was evaluated using the same preprocessing, input resolution, and evaluation protocol described below. Where possible, the optimized model is compared against the original model under the same evaluation conditions.

Evaluation setup

Field	Value
Dataset	HellaSwag
Split	validation (0-shot)
Number of samples	8000
Metric(s)	HellaSwag accuracy, acc_norm convention (character-length-normalized)
Evaluation runtime	ONNX Runtime 1.26.0

Accuracy results

Metric	Original / baseline	Arm-optimized	Change
HellaSwag accuracy (%)	59.8	59.0	-0.8 pp

Accuracy was measured using the evaluation setup described above. Users should re-evaluate the model on their own data before production use.

Arm optimization approach

Arm optimized this model for efficient inference on Arm-based platforms using a hardware-aware conversion and validation flow.

For this release, Arm used:

Optimization area	Applied?	Notes
Model conversion	Yes	Converted to ONNX in the ONNX Runtime GenAI graph layout, with attention fused into a single GroupQueryAttention op and the KV cache wired internally
Quantization	Yes	GPTQ weight-only quantization to INT4, asymmetric per-group with group size 32, calibrated on 256 wikitext samples; INT8 per-token dynamic activations; INT8 per-row embedding table; INT8 LM head (scheme W4A8dynemb_int8)
Runtime/backend selection	Yes	ONNX Runtime 1.26.0 CPU execution provider, with MLAS and KleidiAI kernels
Graph/runtime compatibility updates	Yes	Performed as part of the shared PT2E export pipeline, with ONNX-specific graph translation
Accuracy validation	Yes	Compared against the original model or published baseline
Performance validation	Yes	Measured on the reference Arm platform

The goal of this process is to improve deployment characteristics such as latency, memory use, model size, and runtime compatibility while preserving the model's intended behavior. Detailed conversion scripts, calibration configuration, or backend-specific implementation details may be provided separately where appropriate.

Using this model

Install dependencies

Dependencies are declared in pyproject.toml, which ships with this repository. Resolve and install them into a local virtual environment with uv:

bash

uv python install
uv sync --frozen

This installs Python 3.14, pinned in .python-version, with onnxruntime 1.26.0, onnxruntime-genai 0.14.1, tokenizers 0.23.1 and jinja2 3.1.6, exactly as recorded in uv.lock. The environment was resolved and validated on the target itself: AWS Graviton G4, Ubuntu 24.04.4 LTS, glibc 2.39, aarch64. Every dependency installs as a prebuilt aarch64 wheel; nothing is built from source, and no system packages are required beyond Python and uv. The example needs the CPU execution provider, which ONNX Runtime always provides on this platform.

Run the example

bash

uv run example.py

The example writes predictions.json next to example.py. Decoding is greedy, so repeated runs on the same machine and runtime build reproduce the same text.

Expected input

Property	Value
Input shape	[1, T], where T is the runtime sequence length, variable up to 2048
Input type	int64, input_ids
Input range	Token ids in the range 0 to 31999, for a vocabulary size of 32000
Preprocessing	Apply the chat template in `chat_template.jinja`, then tokenize with `tokenizer.json` with special tokens added. position_ids is not exposed as a graph input, and KV-cache I/O is wired up internally by the ONNX Runtime GenAI Generator.

Expected output

Property	Value
Output shape	One token id per decode step, streamed by the ONNX Runtime GenAI Generator
Output type	Token ids, integer indices into the 32000-entry vocabulary
Postprocessing	Decode with `tokenizer.json`, stopping on the EOS token or when max_length is reached

Intended use

This model is intended for developers evaluating text generation workloads on Arm-based platforms. It is suitable as a reference implementation for benchmarking, prototyping, and integration exploration.

Limitations

—Performance depends on the target device, runtime version, backend/delegate support, memory configuration, and system load.
—Accuracy was evaluated on HellaSwag and may not generalize to all domains.
—This release preserves the original model's intended task and behavior, but users should validate it for their own application, data, and deployment environment.
—This repository is not a replacement for the original model documentation.
—Greedy decoding is deterministic for a given machine and runtime build, but not portable between them. At this model size and weight precision the top two candidate tokens are sometimes separated by a very small logit margin, so a different CPU or a different ONNX Runtime build can select the other candidate and diverge for the rest of the continuation. Treat a committed generation as one known-good sample rather than a bit-exact conformance target.

Additional notes

—The ONNX Runtime GenAI bundle requires the literal file names model.onnx and model.onnx.data, because genai_config.json hard-references them.
—The context window is inherited unchanged from the base model at 2048 tokens.
—The quantization recipe is selective: the last projection in each transformer block is kept at higher fidelity (kquantlast), so some layers were skipped from INT4 quantization to keep accuracy within acceptable ranges.
—Accuracy evaluation scope is HellaSwag zero-shot acc_norm only.

About this version

Original Model: TinyLlama/TinyLlama-1.1B-Chat-v1.0 by The TinyLlama project - Repository

Optimization/conversion: Arm-Optimized version for execution on Arm-based platforms.

Converted/optimized by: Arm

License: The Original Model and the Optimized Model are subject to Apache-2.0.

This repository contains a converted or optimized version of the Original Model (the “Optimized Model”). The Original Model has been converted or optimized as described above for execution on Arm-based platforms.

No retraining or fine-tuning of the Original Model was performed as part of the conversion or optimization. The conversion or optimization was not intended to change the Original Model’s behavior or intended use.

Original Model and Documentation

For information about the Original Model, including its development, training data, intended uses, limitations and other relevant information, please refer to the Original Model repository. Information in that repository was provided by the original developer or other third parties and, unless expressly stated otherwise, has not been independently verified by Arm.

Licenses and Third-Party Terms

Use of the Original Model and the Optimized Model is subject to the applicable licenses, usage restrictions and other terms identified above and in the relevant repositories. Publication of the Optimized Model does not grant any rights beyond those provided under the applicable license terms.

You are responsible for reviewing those terms and ensuring that your use of the Original Model and the Optimized Model is permitted.

Purpose of this Release

The Optimized Model is provided as a reference implementation to demonstrate and evaluate execution and performance on Arm-based systems. It is not a production-ready or supported solution.

Arm’s publication of the Optimized Model does not constitute an endorsement or certification of the Original Model or a representation that the Optimized Model is suitable for production use or any particular purpose.

To the fullest extent permitted by applicable law (i) the Optimized Model is provided “as is.” Arm makes no representations or warranties that the Original Model, the Optimized Model or their outputs are accurate, safe, secure, non-infringing, legally compliant, suitable for production use or fit for any particular purpose; and (ii) Arm will not be liable for any loss or damage arising from or in connection with the Optimized Model, its use or its outputs.

You are responsible for independently evaluating the Optimized Model, its outputs and its suitability for your intended use, including compliance with applicable legal, regulatory, safety and security requirements.

Arm does not commit to provide ongoing support, maintenance or updates for the Optimized Model. Any use of or reliance on the Optimized Model or its outputs is at your own risk.