BananaMind-CodeQ-1.3-2B · ONNX q4f16 (WebGPU)
Browser-ready ONNX export of BananaMind/BananaMind-CodeQ-1.3-2B (a LoRA merge on MiniCPM5-2B), used by the BananaMind-CodeQ WebGPU Pi Space.
How it was made
CodeQ-1.3-2B has exactly the same architecture as openbmb/MiniCPM5-2B, so its BF16 weights were written into the graph of
Mike0021/MiniCPM5-2B-ONNX (rev 04a6c49), keeping that export's optimizations
(fused QKV, GroupQueryAttention, last-position LM head, embedding split into 4 column shards, 7 external-data chunks).
- Linear layers: symmetric int4, block size 32 (
MatMulNBits), fp16 scales — the quantizer reproduces the original ONNX Runtime GenAI output bit-for-bit on the untouchedlm_head. - Embeddings, norms: fp16. RoPE caches unchanged (same
rope_theta). - Tokenizer and corrected chat template taken from the MiniCPM5 ONNX package (tokenizer is byte-identical to CodeQ's).
manifest.jsonlists sizes and SHA-256 of every file.
Sanity check (ONNX Runtime CPU, greedy): asked for a Python factorial function, it answers with a typed, documented implementation. Int4 quantization loses some quality versus BF16.