Add measured KL divergence: 0.506 nats
Browse files
README.md
CHANGED
|
@@ -34,6 +34,7 @@ kept experts renumbered, and the routers + MTP layer sliced to match.
|
|
| 34 |
| Base | [cyankiwi/GLM-5.3-AWQ-INT4](https://huggingface.co/cyankiwi/GLM-5.3-AWQ-INT4) (compressed-tensors INT4 / AWQ) |
|
| 35 |
| Prune | REAP, keep-168/256 (max-over-domain routed-mass saliency), all 75 MoE layers + MTP |
|
| 36 |
| Format | INT4 W4A16, compressed-tensors (Marlin MoE kernel in vLLM) |
|
|
|
|
| 37 |
|
| 38 |
Criterion and fidelity: [`0xSero/glm-5.3-reap-fidelity-study`](https://huggingface.co/datasets/0xSero/glm-5.3-reap-fidelity-study).
|
| 39 |
This is the Hopper-servable sibling of [`GLM-5.3-504B-EXL3-3.0bpw`](https://huggingface.co/0xSero/GLM-5.3-504B-EXL3-3.0bpw).
|
|
|
|
| 34 |
| Base | [cyankiwi/GLM-5.3-AWQ-INT4](https://huggingface.co/cyankiwi/GLM-5.3-AWQ-INT4) (compressed-tensors INT4 / AWQ) |
|
| 35 |
| Prune | REAP, keep-168/256 (max-over-domain routed-mass saliency), all 75 MoE layers + MTP |
|
| 36 |
| Format | INT4 W4A16, compressed-tensors (Marlin MoE kernel in vLLM) |
|
| 37 |
+
| KL vs BF16 | **0.506 nats** (95% CI 0.403-0.607; 25-window full-vocabulary panel vs GLM-5.3-BF16) |
|
| 38 |
|
| 39 |
Criterion and fidelity: [`0xSero/glm-5.3-reap-fidelity-study`](https://huggingface.co/datasets/0xSero/glm-5.3-reap-fidelity-study).
|
| 40 |
This is the Hopper-servable sibling of [`GLM-5.3-504B-EXL3-3.0bpw`](https://huggingface.co/0xSero/GLM-5.3-504B-EXL3-3.0bpw).
|