keys GLM-5.3 DSpark drafter (ft2: realigned to the 2.75 bpw Abliterated EXL3 target)

A DSpark speculative-decoding drafter for full GLM-5.3, fine-tuned from RedHatAI/GLM-5.3-speculator.dspark to the token distribution of drowzeys/keys-GLM-5.3-EXL3-2.75BPW-Abliterated as served at TP=4 on four DGX Spark (GB10). Same architecture and file layout as the RedHat drafter (block 8, taps after layers 1, 19, 38, 57, 74); drop-in for TF_GLM53_DSPARK.

Training

  • Data: 800 on-policy captures (thinking on, temperature 1.0) of the Abliterated target serving its own replies: 1.33 M labelled tokens (top-32 target logits + the drafter's taps); a fixed 2 % holdout (23 files by name hash).
  • Recipe: LR 1e-5, loss 0.1 CE + 0.9 TV (RedHat's mix), 4,000 steps, held-out early stop.
  • Held-out acceptance length: 3.712 (RedHat stock) -> 3.770. More data (target-generated or public text) and a 6-layer up-scale did not raise it further (tested 2026-10-06/07).

Measured (TP=4, 4x DGX Spark, thinking ON, temperature 1.0, same session, 2026-10-07)

DSpark drafting alone (--mtp-drafts 0), confidence policy 0.3 (TF_GLM53_DSPARK_POLICY=confidence, TF_GLM53_DSPARK_CONFIDENCE=0.3), against the built-in MTP head (k = 2) on the same weights:

tok/s DSpark ft2 MTP change
prose short / 32K / ~120K 38.2 / 33.8 / 31.8 36.8 / 32.2 / 29.2 +4 / +5 / +9 %
code short / 32K / ~120K 45.8 / 39.0 / 37.4 39.9 / 34.8 / 31.5 +15 / +12 / +19 %
prefill 32K / 128K 1,219 / 1,093 1,185 / 1,045 +3 / +5 %

Needle at 120K: found. 1M context: a 1M boot found the needle in an 830K-token prompt (prefill 2,351 s, 13 GB minimum free memory per node). The confidence policy matters: with the default cost policy (calibrated at context 0) the same drafter drafts too deep at long context (31.8 prose at 32K). Recipe and image: drowzeys/keys-GLM-5.3-EXL3-2.75BPW-MixedK-Ablit-MLX-Serve-4-DGXSparks.

License and credits

GLM-5.3 License (see LICENSE, Copyright (c) 2026 Z.AI), inherited from the target and from the RedHat drafter. Base drafter: RedHatAI (speculators DSpark). Target: Z.AI GLM-5.3. Serving stack and its third-party licences: see the recipe's NOTICE.

Downloads last month
711
Safetensors
Model size
1B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for drowzeys/keys-GLM-5.3-speculator.dspark-ft2

Base model

zai-org/GLM-5.3
Finetuned
(1)
this model