RASMUS/MiniCPM5-2B-ONNX

MiniCPM5-2B — ONNX (q4f16) for WebGPU

An ONNX export of openbmb/MiniCPM5-2B that runs in a browser on WebGPU through Transformers.js.

At the time this was built, no export of the 2B existed that a browser could load. The official ONNX release covers the 1B only, and it is a CPU-shaped fp16 build (MultiHeadAttention, no MatMulNBits, no config.json) in a layout Transformers.js does not resolve.

Contents

File	Size
onnx/model_q4f16.onnx	316 KB (graph)
onnx/model_q4f16.onnx_data	1.83 GB (weights)

Single variant, q4f16: int4 weights (MatMulNBits) with fp16 graph I/O. A GPU without the shader-f16 feature cannot run this.

The embedding table is not quantised, because MiniCPM5-2B sets tie_word_embeddings: false and the model builder quantises that table only for tied models. Those 267 M parameters cost about 535 MB of the total.

Usage

import { pipeline } from '@huggingface/transformers';

const generator = await pipeline('text-generation', 'RASMUS/MiniCPM5-2B-ONNX', {
  device: 'webgpu',
  dtype: 'q4f16',
});

const output = await generator(
  [{ role: 'user', content: 'Explain WebGPU in two sentences.' }],
  { max_new_tokens: 256 },
);

Measured performance

RTX 4080 Laptop, Chromium 152 on Windows, @huggingface/transformers 4.2.0 over onnxruntime-web WebGPU. Warmed runs, first run at each prompt length discarded.

Measurement	Value
Prefill, 1024-token prompt	0.395 ms / prompt token
Prefill, 512-token prompt	0.411 ms / prompt token
Prefill, 128-token prompt	0.611 ms / prompt token
Decode	24.1 ms / token (41.6 tok/s)
First token, 64-token prompt	199 ms
Session load, warm cache	6.5 s

These are one GPU's numbers, not a general claim.

How it was built

python -m onnxruntime_genai.models.builder \
  -i <local openbmb/MiniCPM5-2B> -o out -p int4 -e webgpu -c cache

onnxruntime-genai 0.15.2. -e webgpu is what selects the GroupQueryAttention branch — the same build on the cpu provider silently emits MultiHeadAttention instead, which is the defect in the published 1B export. The graph is 473 nodes: MatMulNBits x211, GroupQueryAttention x42, no Scan/Loop, and no position_ids input, because GQA fuses RoPE.

Three changes were then needed to make the output loadable by Transformers.js. Each is recorded because each fails in a way that does not name its cause.

1. The KV cache head dimension was made concrete. The builder declares it as the symbol kv_cache_dim. Transformers.js builds the first empty cache from the session's input metadata and resolves any symbol it does not know to 0, so GroupQueryAttention rejected the tensor on the first forward: Input 'past_key' dimension 3 should be same as head_size, got 0 expected 128.
2. The chat template was moved into tokenizer_config.json. Transformers.js reads chat_template.jinja only for multimodal processors; a text tokenizer reads tokenizerConfig.chat_template and nothing else.
3. One dead line was removed from the chat template. It assigned min_count = [...]|min, and @huggingface/jinja has no min filter for arrays, so every assistant turn carrying tool_calls failed to render. The variable is never read. chat_template.jinja here carries the patched text, identical to the copy inside tokenizer_config.json.

transformers.js_config in config.json declares one external-data chunk and an fp16 KV cache, both of which Transformers.js requires and neither of which the builder writes.

Limitations

• English and Chinese. Other languages are not claimed by the base model.
• Quantised. int4 weights change outputs relative to the bf16 original. Spot-checked against the bf16 model on CPU, including a refusal that both produce identically. Not benchmarked for quality.
• Tool calls are XML, not the Pythonic form some other models emit. A parser written for one will not read the other.

Attribution

Base model © OpenBMB, Apache-2.0. This repository redistributes a converted and quantised copy of those weights under the same licence. No weights were retrained or fine-tuned.