Modelpublic

amd/tiny-qwen3-moe-w8a8-int8

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face

0likes1.8kdownloads

Model Card

Model Overview

—Model Architecture: Qwen3MoeForCausalLM (tiny, randomly initialized)
—Input: Text
—Output: Text
—Supported Hardware Microarchitecture: AMD MI300 / MI350 / MI355 (gfx942 / gfx950), NVIDIA (SM ≥ 7.5)
—Inference Engine: vLLM
—Model Optimizer: AMD-Quark
—Weight quantization: INT8, per-channel, static, symmetric
—Activation quantization: INT8, per-token, dynamic, symmetric
—Quantized layers: attention (self_attn) and the routed MoE experts. The MoE router (mlp.gate) and lm_head are kept in the original precision.

This is a tiny, randomly-initialized Qwen3-MoE model quantized to W8A8-INT8, used purely as vLLM CI coverage for the Quark INT8 fused-MoE path (QuarkW8A8Int8MoEMethod), which dispatches through the modular Triton INT8 MoE kernel. It is not intended to produce meaningful text.

Note: the router (mlp.gate) is intentionally left in bf16. Its weight is [num_experts, hidden] and num_experts (8) is not a multiple of 16, which would trip AITER's INT8 linear-kernel shape assertion. Routers are normally left in high precision anyway.

Model Creation

Built and quantized with AMD-Quark: per-channel INT8 weights (static) + per-token INT8 activations (dynamic), symmetric, on a tiny qwen3_moe model. Attention and routed experts are quantized; mlp.gate and lm_head are excluded. vocab_size matches the tokenizer so token ids stay within the embedding table.

Usage in vLLM

bash

vllm serve amd/tiny-qwen3-moe-w8a8-int8 --enforce-eager

Because the weights are random, outputs are not meaningful — this model is a structural / smoke-test fixture only.

License

Apache-2.0. The tiny model is randomly initialized and derives no weights from any base model.