GLM-5.2-504B-K / README.md
0xSero's picture
Add files using upload-large-folder tool
381eb63 verified
|
Raw
History Blame
2.07 kB
metadata
license: mit
base_model: zai-org/GLM-5.2
pipeline_tag: text-generation
tags:
  - moe
  - reap
  - pruning
  - router-kd
  - knowledge
  - nvfp4
  - glm
  - glm-5.2

GLM-5.2-504B-K — knowledge-augmented REAP keep-168 (full-data Router-KD, NVFP4)

The "K-cut" sibling of 0xSero/GLM-5.2-504B: the same 504B / keep-168 budget, but the expert selection is biased toward knowledge & reasoning — the winning top-160 core (kept bit-for-bit) plus the 8 highest-priority knowledge-exclusive experts per layer that coding-saliency pruning drops. Recovered with gate-only **Router-KD trained on the FULL calibration set (18.6k real traces)** — 6x the data of the first-pass cuts.

Sponsor

8x NVIDIA B200 sponsored by Lambda. Thank you.

Why this variant exists

REAP saliency computed from coding traces under-weights experts that fire mainly on reasoning/knowledge. The K-cut deliberately re-includes them — trading a sliver of coding-saliency coverage for broader knowledge coverage. Reach for this on knowledge/reasoning-heavy workloads; use the plain GLM-5.2-504B otherwise.

Eval (n=2000 held-out real prompts, raw, no max_tokens / no timeout)

metric GLM-5.2-504B-K (this) GLM-5.2-504B (plain floor)
attractor / loop rate 0.078 0.072
natural-EOS rate 0.923 0.928
distinct-4 0.881 0.880
median tokens 1232 1267

On the loop metric this is at parity with (or better than) the plain-cut floor. The residual loops are inherent to GLM-5.2 itself (the unpruned teacher loops on the same prompts), so neither cut "fixes" them — this one buys knowledge-expert coverage.

Serving (vLLM)

vllm serve 0xSero/GLM-5.2-504B-K --tensor-parallel-size 8 \
  --quantization modelopt_fp4 --kv-cache-dtype fp8 --trust-remote-code --max-model-len 262144

REAP knowledge-augmented cut + full-data Router-KD. Compute sponsored by Lambda.