4-bit Walsh–Hadamard + Cholesky GPTQ (W4A16-HGPTQ, group size 64) quantization of fastino/GLiDE-0.8B (GLiDE code-readout-v1 single-pass decision model). It compresses the 1.40 GiB BF16 checkpoint (readout.safetensors + backbone) to 0.72 GiB (1.94× smaller) while preserving 100.00% top-1 decision agreement and 0.998195 readout logit cosine similarity against the BF16 reference.
0.72 GiB total across 156 quantized text linears (text-only decision checkpoint), plus the exact BF16 readout.safetensors (255 × 1024 code-readout head).H_64) + MSE-initialized Cholesky GPTQ (group_size=64, actorder="static", dampening_frac=0.01), calibrated layer-by-layer under noncausal_full_attention attention.24 layers (hidden_size=1024). Single-pass hybrid GatedDeltaNet + full-attention forward pass.model.language_model.* (and model.visual.*) prefixes with "quantization_status": "compressed", loading cleanly with missing: 0, unexpected: 0 in Qwen3_5Model.from_pretrained, Glide.from_pretrained, and warp-34.fastino/GLiDE-0.8B)Single-pass code-readout-v1 evaluation across multi-question choice, yes_no, multi_select, and scale decision tasks (with forward + reverse option-permutation averaging):
| Evaluation | BF16 Base (fastino/GLiDE-0.8B) |
GLiDE-0.8B-W4A16-HGPTQ |
Notes |
|---|---|---|---|
Safetensors Size (incl. readout.safetensors) |
1.40 GiB (1,505,349,928 B) |
0.72 GiB (775,384,816 B) |
1.94× compression (−0.68 GiB) |
Decision Top-1 Agreement (12 questions) ↑ |
100.00% |
100.00% |
Forward + reverse permutation-averaged readout |
| Option Probability KL Divergence (mean / max nats) ↓ | 0.000000 |
0.000421 / 0.004050 |
Per-question normalized option distribution KL |
| 255-Code Readout Logit Cosine Similarity ↑ | 1.000000 |
0.998195 |
Full 255-code softmax KL 0.002742 nats |
156 linear layers in W4A16-HGPTQ):156 linears): All 24 hybrid GatedDeltaNet + full-attention blocks (q_proj, k_proj, v_proj, o_proj, in_proj_qkv, in_proj_z, out_proj, gate_proj, up_proj, down_proj) with online block-Hadamard rotation (head_dim=64) + Cholesky GPTQ (group_size=64, actorder="static").0 linears): None (text-only checkpoint).readout.safetensors (weight [255, 1024]), embed_tokens, linear_attn.in_proj_a, linear_attn.in_proj_b, norms, biases.from glide import Glide
model = Glide.from_pretrained("nchapman/GLiDE-0.8B-W4A16-HGPTQ", device="cuda:0")
result = model.decide(
state="Customer ordered item #4412, delivered on time, requested an invoice copy.",
questions={
"intent": {
"type": "choice",
"instructions": "What does the customer need?",
"criteria": {"invoice": "Copy of invoice", "refund": "Return and refund"},
}
},
)
print(result)
model-00001-of-00001.safetensors 775 MBreadout.safetensors 522.3 kBfastino/GLiDE-0.8B