GLM-5.2-SIQ-Fruit-pilot

GLM-5.2-SIQ-Fruit-pilot — Clémentine

The original micro Fruit experiment: a 413M-parameter, about 30M-active GLM-5.2 architectural mimic trained from scratch in roughly 13 minutes on one RTX 5090, then exported through the SIQ/Trellis serving path.

Clémentine exists to make the entire train → quantize → load → generate → MTP loop cheap enough to debug. It is a toy language model and CI fixture, not a general assistant.

Runtime requirement: this SIQ checkpoint needs a compatible b12x/SparkInfer + vLLM build. Stock vLLM and Transformers do not implement its exl3-trellis expert tensors.

Architecture

• GlmMoeDsaForCausalLM: 3 dense + 3 MoE decoder layers + 1 MTP layer.
• Hidden size 512; 8 attention heads.
• 256 routed experts, top-8 routing, plus one shared expert.
• Production MLA dimensions retained: KV LoRA 512, QK no-RoPE/RoPE 192/64, value head 256.
• DSA indexer shape 32×128. Its weights are random-init/untrained; at contexts no longer than index_topk, the tested path is numerically dense.
• GLM tokenizer with 154,880 tokens.

The trained logical MoE intermediate size is 128. The corrected serving export stores it as 256 by appending exact zero rows/columns; this preserves the model function while satisfying the MTP Trellis kernel's CTA tile requirement.

Training data

15.6M tokens drawn from TinyStories, the GLM-5.2 REAP recall calibration set, and repeated SPDX license text. The model reliably continues TinyStories-style openings and has memorized fragments of common license templates. That is the extent of the intended capability.

The original BF16 source state remains in this repository as training/fruit_pilot.pt, SHA-256 7c7d07f0d7c2944faf8701de6e174fcac7ba48a854de87b194ef9e8e33eede49.

Corrected SIQ artifact

• Ordinary MoE layers: 96 K4 + 160 K3 experts.
• MTP layer: uniform K3.
• Non-expert tensors: BF16.
• Tensor payload: 595,150,336 bytes (0.554 GiB).
• RoPE theta: 10,000 in both supported configuration locations.
• Export conversion: 28 half-split-to-interleaved RoPE projection permutations plus the MTP eh_proj input-half swap.
• MANIFEST.sha256 covers the 17 serving files. Tools, logs, the card, and the training checkpoint are outside that serving-artifact manifest; the source checkpoint hash is pinned above.

Measured validation of the corrected export

RTX 5090; custom gilded-gnosis r25, fp8_ds_mla, eager mode:

MTP is slower on this tiny model because orchestration overhead dominates; its acceptance rate validates the draft path, not a speedup claim.

Observed greedy continuations include:

Once upon a time → , there was a little girl named Lily...

MIT License\n\nCopyright → (c) [year] [fullname]...

Release correction

Revision bfddc762ed7d75eac7b8706395eddff4fcb42220 copied legacy trainer weights into serving tensors without the required RoPE channel permutation or MTP eh_proj half-swap, retained theta 8,000,000 instead of the trained 10,000, and left the logical 128-wide MTP experts incompatible with the current Trellis kernel.

The 2026-08-07 rebuild authenticates the original source checkpoint, applies both convention conversions, exact-zero-pads the served MoE intermediate size to 256, re-encodes the expert tensors, regenerates the manifest, and passes the validation table above. Consumers pinned to the old revision should update.

Reproduction

The repository includes the original trainer/exporter snapshots and logs. The maintained, fail-closed exporter and parity/MTP harnesses live at github.com/malaiwah/proxy-fruit. For the larger production-shape fixture, use GLM-5.2-SIQ-Fruit.

Limitations

• 15.6M training tokens are insufficient for broad language capability.
• The indexer was not trained; this release does not establish long-context sparse-attention quality.
• License-template continuation is memorization behavior, not legal advice.
• Do not deploy this model.