Enntity/GLM-5.3-Flash-DFlash2-enntity

GLM-5.3 Flash DFlash2 drafter, Enntity: trained from scratch on 2x DGX Spark

A DFlash2 speculative-decoding drafter for GLM-5.3 Flash, trained from scratch (fresh random initialization, no third-party drafter weights) on two DGX Sparks. DFlash2 architecture and tensor layout (5 layers, 1.17B trained parameters, 81 BF16 tensors, format atlas-dflash2-bf16), so SparkGLM (Atlas) serves it as a drop-in DFlash2 drafter. Results are measured against the reference drafter: the best public DFlash2 drafter for GLM-5.3-Flash, incoai/GLM-5.3-Flash-DFlash2. The embedding and lm_head are the target's and are not included.

This repository was Enntity/GLM-5.3-Flash-DFlash2-enntity-v80; the old name redirects here. Each release is a git tag, so pin a tag or commit rather than main.

version	tag	released	model.safetensors sha256
v81 (current, main)	v81	2026-11-05	d14e1edb080711d46aa507651bfd0e92c50c2a4f1082e138662f6f0274003cf6
v80	v80 (commit 8416cdfb)	2026-11-01	61190f81c66cbd0f6ae6a27c5d516d60af1e050f002e41fccbe653b475bf4d72

hf download Enntity/GLM-5.3-Flash-DFlash2-enntity --revision v81 --local-dir GLM-5.3-Flash-DFlash2-enntity-v81

Atlas loads the directory with --draft-model. Training lineage and hashes are in PROVENANCE.json (v81's parent is v80's fp32 master).

Progress

Live speed (Atlas, 2026-11-05)

SparkGLM on two GB10 nodes (NVFP4 target, greedy, one request at a time, adaptive verify width as shipped). Every arm ran on the same server and settings and differed only in --draft-model. Public probes: v81 and v80 two visits each (tok/step identical between an arm's visits), the reference one visit. Held-out agentic: one visit per arm. Engine decode tok/s, with accepted tokens per verify step in brackets.

category	v81	v80	v81 vs v80	reference	v81 / reference
prose, 12 prompts x 256 tokens	40.9 (2.12)	40.1 (2.06)	+1.9%	46.8 (2.50)	87%
code, 12 prompts x 256 tokens	63.0 (4.11)	61.5 (3.99)	+2.4%	69.2 (4.89)	91%
structured output, 4 prompts	79.6 (6.01)	76.9 (5.74)	+3.4%	87.3 (6.98)	91%
RigMark prose	36.0 (1.86)	35.8 (1.83)	+0.6%	39.0 (2.04)	92%
RigMark code	63.3 (4.27)	61.8 (4.12)	+2.4%	66.9 (4.73)	95%
RigMark structured	84.2 (6.73)	79.8 (6.25)	+5.6%	91.0 (7.64)	93%
held-out coding-agent turns, set A (16)	51.1 (3.42)	52.0 (3.46)	-1.6%	42.0 (2.70)	122%
held-out coding-agent turns, set B (24)	48.8 (3.17)	47.6 (3.10)	+2.4%	42.3 (2.68)	115%

Run-to-run noise is about ±2%. All 32 public exact-match texts were identical across every visit and drafter; on the held-out agentic sets 39 of 40 matched, and the one difference is the engine's forced end-of-thinking at a token budget, whose position follows verify-step boundaries.

Offline (held-out, teacher-forced expected accepted length per 8-token block)

group	v80	v81
agentic (96 rows)	3.714	3.763
template	3.240	3.282
structured	5.672	5.816
prose	2.510	2.569
code	4.211	4.360

Training (v81)

• Start: v80's fp32 master.
• Data: 76,427 captured rows (75,957 train, 470 held-out), 119M stored tokens. Sources: Enntity-authored template, prose, code and structured-output coverage prompts with GLM-5.3-Flash teacher completions; Enntity-internal chat and coding-agent sessions; GLM-5.3-Flash coding-agent sessions on Enntity-authored tasks. Hidden states captured from the served target.
• Objective: selector_prefix with nla unary loss and expected-accepted-length weight 1.0.
• Schedule: one-epoch cosine plan (17,000 steps, lr 2.5e-5, floor 0.2x, 50 warmup), stopped at step 13,200 once held-out gains flattened; 4,096 anchors per step, about 315M supervised tokens.
• Two-node plain averaging (DiLoCo, 200 inner steps, outer lr 1, no momentum). Published as drafter network global v81.

License and provenance

Weights: Apache-2.0. Trained from a fresh initialization (lineage root enntity-scratch-init, no warm start from any third-party drafter). Contains none of the reference drafter's weights, configs or text. Target model GLM-5.3-Flash is MIT. Training code: https://github.com/Enntity/glm-flash-drafter (Apache-2.0).