dasilva333/rwkv7-g1-webgpu-prefabs

RWKV-7 G1 WebGPU Pre-Quantized Prefabs

Pre-quantized CBOR .prefab models for WebGPU inference in the browser via @cryscan/web-rwkv-wasm and AIRI.

These prefabs are pre-quantized offline directly from DanielClough/rwkv7-g1-safetensors for fast, zero-degradation browser execution.

Available Models

Model	Variant	Size	Format	Direct HF Download Link
RWKV-7 G1 0.4B	Int8	581 MB	.prefab	rwkv7-g1d-0.4b-int8.prefab
RWKV-7 G1 0.4B	NF4	437 MB	.prefab	rwkv7-g1d-0.4b-nf4.prefab
RWKV-7 G1 1.5B	Int8	1.88 GB	.prefab	rwkv7-g1d-1.5b-int8.prefab
RWKV-7 G1 1.5B	NF4	1.28 GB	.prefab	rwkv7-g1d-1.5b-nf4.prefab

Why .prefab instead of SafeTensors?

• Instant browser startup: Deserializes directly into WebGPU buffers via Session.from_prefab(data, SessionType.Chat) in ~5.8 seconds (vs 60s+ for in-browser quantization).
• Zero in-browser compute degradation: Completely bypasses the on-the-fly shader quantization bug in browser WebGPU runtimes.
• Smaller downloads:
  • rwkv7-g1d-0.4b-int8.prefab: 581 MB (-35% vs 902 MB raw FP16).
  • rwkv7-g1d-0.4b-nf4.prefab: 437 MB (-51% vs 902 MB raw FP16).
  • rwkv7-g1d-1.5b-int8.prefab: 1.88 GB (-38% vs 3.06 GB raw FP16).
  • rwkv7-g1d-1.5b-nf4.prefab: 1.28 GB (-58% vs 3.06 GB raw FP16).

Usage with @cryscan/web-rwkv-wasm

import init, { Session, SessionType } from '@cryscan/web-rwkv-wasm'

await init()
const res = await fetch('https://huggingface.co/dasilva333/rwkv7-g1-webgpu-prefabs/resolve/main/rwkv7-g1d-1.5b-nf4.prefab')
const buffer = new Uint8Array(await res.arrayBuffer())

const session = await Session.from_prefab(buffer, SessionType.Chat)