GLM-5.2 research kitchen

Self-contained research bundle for GLM-5.2 work — the tracing framework, the mixed-precision quantization scripts, and the layer-level structured pruning artifacts that produced the ShortGPT-pruned spaced-N=12 model variant (36 GB / 15.4% saved, all baselines passing).

Originally lived scattered inside the gguf2mlx converter repo (scripts/, src/gguf2mlx/tracing/, reports/, traces/, prompts/). Extraction into this kitchen makes the work independently portable + installable and lets the converter stay focused on GGUF→MLX translation only.

Vendored dependencies (git submodules)

Path	Branch	Provides
vendor/gguf2mlx	Deviad/gguf2mlx @ feature/update_for_glm5.2_cooking	from gguf2mlx import convert — end-to-end GGUF→MLX translator, used for the Phase 8 MLX conversion step
vendor/llama.cpp	Deviad/llama.cpp @ feature/patch_used_to_create_mixed_quantization_of_glm5.2	patched llama-cli, llama-quantize, llama-trace-moe binaries; trace-moe example + BI computation needed by every trace run + the mixed-precision quant_glm52_mixed.sh
vendor/mlx-lm	ml-explore/mlx-lm @ main (2c008fd…, v0.31.3-13-g2c008fd)	reference-only upstream mlx-lm Apple MLX LM library; mlx_lm/models/glm_moe_dsa.py (53-line subclass of deepseek_v32.Model) + mlx_lm/models/deepseek_v32.py (the actual MLA+DSA+IndexShare forward graph; site of the documented IndexShare blocker in REAP37_EXPERIMENTS.md). Not built/imported by the kitchen — vendored so the GLM-5.2 MLX-side code is greppable alongside the other sources.
vendor/jangq	jjang-ai/jangq @ main (e70f220…, depth 1)	JANGTQ TurboQuant runtime — jang_tools.load_jangtq, per-model JANGTQ converters, MXTQ Metal kernels (jang-runtime/Sources/JANGCoreMetal/). Vendored so the JANGTQ_K converter + vMLX speed/quality work is scannable/patchable; read-only upstream (fork only if upstreaming a patch).
vendor/vmlx	jjang-ai/vmlx @ main (b7da1b8…, v1.5.69, depth 1)	vMLX engine — vmlx_engine.cli/server/jang_loader + model_configs.py (glm5 family + cache=mla registration). The server used to serve the JANGTQ_K MLX bundle (vmlx-serve serve ... --port 8082). Read-only upstream.

The two Deviad forks are pinned to the tip of their respective fork feature branches; after a fresh clone, run:

git submodule update --init --recursive      # ~5 min for the llama.cpp clone

Layout

zai-glm-kitchen/
├── glm52_kitchen/                     # Python package (pip install -e .)
│   └── tracing/                       # MoE routing / activation / retrieval tracing
├── tests/                             # 4 tracing-related test modules
├── common/
│   ├── scripts/                       # 9 tracing CLI scripts (+ bash wrappers)
│   ├── baselines/                     # merge-sort + BLUE-FALCON retrieval baseline scripts
│   ├── prompts/                       # multilingual trace smoke suite (7 langs × 7 domains)
│   ├── reports/                       # 37 common-axes report artifacts
│   ├── traces/                        # common-axes trace datasets (batch/smoke/synth gitignored)
│   ├── phase3_dsa_unblock/            # rejected DSA-patch forensic artifacts
│   └── reap37/                        # REAP37 MLX 4-bit experimental track
├── mixed-precision-quantization/
│   └── scripts/                       # build_llamacpp.sh, quant_glm52_mixed.sh, verify_glm52_mixed.py
├── layer-level-structured-pruning/
│   ├── scripts/                       # prune_gguf.py, prune_layers.py, analyze_bi_scores.py
│   ├── reports/                       # 12 BI-plan artifacts + prune inventory
│   └── traces/                        # 327 trace files (BI calibration + forensic 3-way + baselines)
└── vendor/                            # git submodules
    ├── gguf2mlx/                      # converter fork (see above)
    └── llama.cpp/                     # patched llama.cpp fork (see above)

Canonical docs (root)

• AGENTS.md — contract: how findings get tracked, what counts as a durable finding, where operational binaries live.
• LOCAL_SETUP.md — submodule init + env vars (LLAMA_CPP_DIR, MODEL_DIR, REAP37_MODEL_DIR) + how to set them locally, plus a JANGTQ_K (MLX) quantize + run section covering the alternative MLX/vMLX stack.
• PLAN.md — GLM-5.2 glm-dsa converter feature plan (the converter implementation itself lives at vendor/gguf2mlx, see submodules).
• GLM52_SESSION_MEMORY.md — 3,751-line canonical findings narrative (Phase 1 → Phase 8 + post-mortem).
• GLM52_TRACE_PLAN.md — tracing methodology + user stories / acceptance criteria.
• REAP37_EXPERIMENTS.md — REAP37 MLX 4-bit track.

Install

git submodule update --init --recursive      # first-time only
uv pip install -e .                           # registers glm52_kitchen + vendored gguf2mlx
pytest -q                                     # 71 tests pass + 1 skipped
python -c "from gguf2mlx import convert"      # converter available via submodule

The kitchen pyproject.toml declares the gguf2mlx submodule as a uv path dependency, so uv pip install -e . resolves the converter automatically.

Cross-references to outside

• gguf2mlx converter fork: https://github.com/Deviad/gguf2mlx/tree/feature/update_for_glm5.2_cooking (tracked as submodule at vendor/gguf2mlx)
• llama.cpp patched fork: https://github.com/Deviad/llama.cpp/tree/feature/patch_used_to_create_mixed_quantization_of_glm5.2 (tracked as submodule at vendor/llama.cpp)
• Both forks descend from the originals at https://github.com/barrontang/gguf2mlx and https://github.com/ggml-org/llama.cpp respectively.