MiniCPM-V-4.7-35B-A3B-DSpark EXL3 (8.0 bpw)

EXL3 quantization of the openbmb DSpark block-parallel speculative-decoding drafter for MiniCPM-V-4.7-35B-A3B. Pair with the target pack MiniCPM-V-4.7-35B-A3B-EXL3-4.15bpw-hq (same org); the drafter shares the target's embedding and lm_head, which the target pack provides.

• 5-layer GDN-free GQA backbone, block size 15 (14 draft tokens), markov head rank 256, confidence head enabled; tap_shift: 1 pinned in config.json (measured better than 0).
• Quantizer: exllamav3 1.6.1, LDLQ compensation, default calibration corpus.
• Measured vs target (RTX 3090, greedy, 8-workload mix): 2.05 accepted drafts/round with the STS Theta scheduler.

In this repository

• CONVERT.md - full conversion record; calibration.md - what was calibrated
• SWEEP.tsv - drafter bitrate sweep (acceptance vs the same quantised target)
• assets/ - STS scheduler calibration (temperatures, RTX 3090 round-time curve, refit tool)
• UPLOAD_MANIFEST.json - sha256 manifest

Usage

python eval/spec_decode.py -m <target-pack> -dm <this-pack> -ndt 14

Optional confidence-scheduled verification: calibrate with EXL3_DSPARK_CONF_DUMP, fit per-position temperatures, then run with EXL3_DSPARK_STS=<temps.json> and EXL3_DSPARK_SPS=<verify-batch->round-ms table>. Requires exllamav3 >= 1.6.1.