Hakureirm/rwkv7-sglang-w4gptq-7.2b

RWKV-7 G1 7.2B — int4 GPTQ for rwkv-sglang (fits a 16 GB card)

Hand-written weight-only int4 (GPTQ-calibrated, symmetric, group 64) quantization of BlinkDL's RWKV-7 "Goose" G1 7.2B, for the rwkv-sglang serving overlay.

• Accuracy: near-lossless at this size. MATH500 avg@64 (500 problems × 64 rollouts): 61.08% vs fp16's 64.18% (−3.1pt) — the same symmetric GPTQ scheme that loses 25.6pt at 1.5B costs three points at 7.2B; bigger models quantize better, measured not assumed. Greedy decoding is token-for-token EXACT on the oracle fixture (8/8); 192/192 projections GPTQ-quantized (streamed Hessian calibration, no RTN fallback).
• Runs a 7.2B model on a 16 GB GPU: checkpoint 4.6 GB (vs 14.4 GB fp16); the int4 path serves 7.2B in well under 16 GB total. On a real T4 (16 GB) the sibling RTN build measured greedy 8/8 EXACT, 32.9 tok/s bsz1, peak 6.7 GB.
• Speed: int4 decode faster than fp16 at small batch on every tested arch.

Where this fits

This checkpoint loads through the rwkv-sglang overlay. Native RWKV-7 support is being upstreamed into SGLang — see sgl-project/sglang#30115; once that lands the overlay is no longer required.

If you want RWKV-7 without quantization or a custom runtime, the same base weights are also published in standard HuggingFace layout — plain safetensors, ordinary config.json, no trust_remote_code:

0.1B	rwkv7-0.1b-hf
0.4B	rwkv7-0.4b-hf
1.5B	rwkv7-1.5b-hf
2.9B	rwkv7-2.9b-hf
7.2B (G1)	rwkv7-g1h-7.2b-hf
Pile 168M	rwkv7-168m-pile-hf

Transformers support for the architecture itself is open as huggingface/transformers#47780.

Format & loading (important)

Not a drop-in HuggingFace checkpoint. Group-wise (GROUP=64) symmetric int4 (.qweight + .scale); loads only through the rwkv-sglang overlay:

bash scripts/deploy.sh   # from github.com/Hakureirm/rwkv-sglang, built against sglang v0.5.10.post1 — newer releases need the overlay rebased
RWKV_W4=1 python -m sglang.launch_server --model-path <this-dir> --dtype float16 \
  --trust-remote-code --disable-radix-cache --mem-fraction-static 0.8

LoRA/norm/embedding/head stay full precision. Base model © BlinkDL (Bo Peng), Apache-2.0.