Spark-X2.6-1.7B WebGPU Runtime
Packed Q4 weights and tokenizer for the model-specific WebGPU runtime in
mohdmahmodi/Spark-X2.6-1.7B-WebGPU.
The bundle is produced by tools/pack_weights.py from the official
mohdmahmodi/Spark-X2.6-1.7B-ONNX-WebGPU
Q4f16 export. Quantized payloads are copied byte-for-byte; nothing is requantized.
Contents
runtime/manifest.json— tensor offsets, layer plan, KV capacitiesruntime/layer_00.bin…runtime/layer_27.bin— per-layer q4/q8 tensorsruntime/lm_head_0.bin…runtime/lm_head_3.bin— tied embedding/lm_head q8 chunksruntime/lm_head_scales.binruntime/rope.bin,runtime/final_norm.binruntime/pack-report.json— SHA-256 of every packed filetokenizer.json,tokenizer_config.json,generation_config.json
The app loads this repository at:
https://huggingface.co/leventkoldas/Spark-X2.6-1.7B-WebGPU-Runtime/resolve/main
and reads runtime/manifest.json, then the weight files and tokenizer relative to that base.
Model
- 28 layers, hidden 2048, intermediate 6656
- 8 query heads, 2 KV heads, head dimension 256
- 131,072 vocabulary, tied embedding/lm_head
- 512-token sliding window on 21 layers; full attention on 7
- 32,768-token context
- q4f16 runtime requires
shader-f16
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support