Downloads · 30 days
0
Hakureirm/rwkv7-sglang-w4gptq-1.5b
rwkv7-sglang-w4gptq-1.5b is a text generation model from Hakureirm. Use it when you need the model to write or continue text. It is set up for rwkv. The card lists the license as apache-2.0.
Hand-written weight-only int4 (GPTQ-calibrated, symmetric, group 64) quantization of BlinkDL's RWKV-7 "Goose" G1 1.5B, for the rwkv-sglang serving overlay.
Downloads · 30 days
0
Access
Public
Updated Aug 6, 2026
Parameters
942M
1.3 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors1.3 GB · 100%
How the weights are stored.
U8604M · 64%
From the Hugging Face model README
Hand-written weight-only int4 (GPTQ-calibrated, symmetric, group 64) quantization of BlinkDL's RWKV-7 "Goose" G1 1.5B, for the rwkv-sglang serving overlay.
Perplexity-style metrics understate int4's damage to multi-step reasoning at this model size. On MATH500 (avg@64, 32,000 rollouts) this checkpoint scores 14.98% vs fp16's 40.60% (−25.6pt) — the quantized model tends to lose the thread mid-derivation and run to the token cap (57.7% truncation vs fp16's 14.2%). This is a 1.5B-specific fragility, not a property of the scheme itself: the identical symmetric GPTQ at 7.2B costs only −3.1pt on the same ruler (see rwkv7-sglang-w4gptq-7.2b).
Recommendation: treat this checkpoint as a memory-footprint tool for non-reasoning workloads. For reasoning-heavy use at 1.5B, use the int8 w8g64 tier instead (greedy-exact, no measurable accuracy cost) — int4 is not the lossless tier at this size.
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.
Not a drop-in HuggingFace checkpoint. Weights are group-wise (GROUP=64) symmetric int4
(.qweight + .scale); they load 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
LoRA/norm/embedding/head stay full precision. Base model © BlinkDL (Bo Peng), Apache-2.0.