Add files using upload-large-folder tool

.gitattributes

@@ -69,3 +69,9 @@ GLM-5.2-REAP-504B-Q4_K_XL-00001-of-00008.gguf filter=lfs diff=lfs merge=lfs -tex
 GLM-5.2-REAP-504B-Q4_K_XL-00006-of-00008.gguf filter=lfs diff=lfs merge=lfs -text
 GLM-5.2-REAP-504B-Q3_K_XL-00003-of-00006.gguf filter=lfs diff=lfs merge=lfs -text
 GLM-5.2-REAP-504B-Q4_K_XL-00003-of-00008.gguf filter=lfs diff=lfs merge=lfs -text
+GLM-5.2-REAP-504B-Q3_K_XL-00006-of-00006.gguf filter=lfs diff=lfs merge=lfs -text
+GLM-5.2-REAP-504B-Q2_K_XL-00005-of-00005.gguf filter=lfs diff=lfs merge=lfs -text
+GLM-5.2-REAP-504B-Q2_K_XL-00002-of-00005.gguf filter=lfs diff=lfs merge=lfs -text
+GLM-5.2-REAP-504B-Q2_K_XL-00004-of-00005.gguf filter=lfs diff=lfs merge=lfs -text
+GLM-5.2-REAP-504B-Q2_K_XL-00001-of-00005.gguf filter=lfs diff=lfs merge=lfs -text
+GLM-5.2-REAP-504B-Q2_K_XL-00003-of-00005.gguf filter=lfs diff=lfs merge=lfs -text

GLM-5.2-REAP-504B-Q2_K_XL-00001-of-00005.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:58417ae9e82666624eb153be37c1b693826b7c35bc3a638a9b58e956e2f49e3b
+size 44769604288

GLM-5.2-REAP-504B-Q2_K_XL-00002-of-00005.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:55479b34a090bf33ef3796b0f5d434aaab7013c9410ce6263b8f26199efaa493
+size 44168328864

GLM-5.2-REAP-504B-Q2_K_XL-00003-of-00005.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:10f037e9011a3df669c401dd126ba4b9755ab3c54d0cd096710aba5dcf41a55f
+size 44564908096

GLM-5.2-REAP-504B-Q2_K_XL-00004-of-00005.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d5f400699af5bcf50dad51dead657ee8d4c2c23e657f86bcb1c7f9ae77d9c2f3
+size 44876437408

GLM-5.2-REAP-504B-Q2_K_XL-00005-of-00005.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:87b0a466f6ff75690abaaac94a5f1eae32337a49766a0081f7f6314a927cbe4b
+size 27009792000

GLM-5.2-REAP-504B-Q3_K_XL-00006-of-00006.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fa0ddda9eb8c66b1e7ab15fc7d0947a3b9a631e9fe9769a76a4ea874c4299a28
+size 37565276768

README.md

@@ -0,0 +1,91 @@
+---
+license: mit
+base_model: zai-org/GLM-5.2
+tags:
+  - moe
+  - reap
+  - pruning
+  - gguf
+  - llama.cpp
+  - glm
+---
+
+# GLM-5.2-REAP-504B-GGUF (keep-168, Router-KD)
+
+GGUF builds of a **34%-pruned** GLM-5.2 — 168 of 256 routed experts kept per layer (incl. the MTP
+layer), **~504B params**, recovered with **gate-only Router-KD** to the unpruned teacher. This is
+the keep-168 "floor": it reaches **parity with the full unpruned GLM-5.2** on real-world prompts
+while being a third smaller.
+
+## 🙏 Sponsor
+
+All pruning, distillation, and evaluation ran on **8× NVIDIA B200 sponsored by [Lambda](https://lambda.ai)**. **Thank you, Lambda.** 🙏
+
+## Files
+
+| file | bits | size (approx) | notes |
+|---|---|---|---|
+| `GLM-5.2-REAP-504B-BF16`        | 16   | ~933 GB | full precision — for fine-tuning / re-quantizing |
+| `GLM-5.2-REAP-504B-UD-Q4_K_XL`  | ~4.5 | ~270 GB | dynamic — **recommended** quality/size balance |
+| `GLM-5.2-REAP-504B-UD-Q3_K_XL`  | ~3.5 | ~210 GB | dynamic — smaller, strong quality |
+| `GLM-5.2-REAP-504B-UD-Q2_K_XL`  | ~2.7 | ~160 GB | dynamic — smallest usable |
+
+**`UD` = Unsloth-style Dynamic.** Quantized with an **importance matrix (imatrix)** calibration pass,
+and **per-tensor dynamic precision**: attention, the shared expert, token-embedding and output tensors
+are kept at higher bits while the routed experts are pushed low. This is dramatically better than flat
+K-quants at 2–3 bit. The `imatrix.dat` used is included in the repo for reproducibility.
+
+Large files are split into ≤45 GB shards (`...-00001-of-000NN.gguf`); llama.cpp loads them directly
+from the first shard.
+
+## What this is
+
+- **Arch:** `GlmMoeDsaForCausalLM` — 78 layers (3 dense + 75 MoE) + 1 MTP layer, DeepSeek Sparse
+  Attention, sigmoid router (top-8), 1 shared expert, hidden 6144.
+- **Prune:** REAP (saliency = `gate × ‖expert_output‖`) → top-168/layer, consistent across all MoE
+  layers **and** the MTP layer; `n_routed_experts: 168`.
+- **Recover:** freeze experts + backbone, train only the 75 router gates (~0.016% of params) to
+  KL-match the **unpruned** GLM-5.2 teacher's next-token distribution (plain uniform weighting,
+  lr 5e-5, 2999 sequences). This variant minimizes repetition / thought-loops.
+
+## Eval
+
+Held-out real-world prompts (50 probes, raw sampling, **no max_tokens / no timeout** — loops are
+*detected*, not truncated):
+
+| metric | keep-168 + Router-KD | unpruned teacher |
+|---|---|---|
+| attractor / loop rate | **0.08** (4/50) | 0.10 (6/50) |
+| natural-EOS rate | **0.92** | 0.88 |
+| output diversity (distinct-4) | **0.878** | — |
+| median output length | 995 tok | — |
+
+On this eval the pruned+recovered model is **at parity with the full teacher** — the 0.08 vs 0.10
+gap is within noise on n=50. Metrics are measured on the recovered model; lower-bit quants (esp.
+Q2_K) inherit the recipe but were not individually loop-tested and will degrade somewhat.
+
+## Serving (llama.cpp)
+
+```bash
+# loads all shards from the first one
+llama-cli    -m GLM-5.2-REAP-504B-Q4_K_M-00001-of-000NN.gguf -p "Hello"
+llama-server -m GLM-5.2-REAP-504B-Q4_K_M-00001-of-000NN.gguf --host 0.0.0.0 --port 8080
+```
+
+**Tip — brevity prompt:** GLM-5.2 (pruned or not) can over-think on some agent prompts. An optional
+system prompt like *"Be concise. Think only as much as the task needs, then answer and stop."* cuts
+median length roughly in half and removes some loops, at no retraining cost.
+
+## Known limitations
+
+- A few residual `</think>`-restart loops are **inherent to GLM-5.2** — the *unpruned* teacher loops
+  on the same prompts — so they are not an artifact of pruning and are not fully fixable via routing.
+- The dynamic low-bit quants inherit the recipe but were not individually loop-tested; prefer `UD-Q4_K_XL` or BF16 where quality matters most.
+
+## More
+
+- Siblings: [`0xSero/GLM-5.2-481B`](https://huggingface.co/0xSero/GLM-5.2-481B) (keep-160), [`0xSero/GLM-5.2-469B`](https://huggingface.co/0xSero/GLM-5.2-469B) (keep-156).
+- Writeup + code: [github.com/0xSero/glm52-reap-suite](https://github.com/0xSero/glm52-reap-suite)
+
+---
+*Compute sponsored by **[Lambda](https://lambda.ai)** — thank you. 🙏*