m8than/GLM-5.2-DSparkFlash-Writer-v2

GLM-5.2 DSparkFlash Writer (v2)

Experimental GLM-5.2 speculative draft: the unmodified RedHat DSpark base backbone plus a conditional Markov selector retrained on creator-chat captures and native prompts.

This is not the v1 release. v1 used a creative-tuned backbone. v2 does not: its transformer blocks are the donor's, so the name "Writer" describes the selector's training data, not a fine-tuned writing backbone. This is not an official Red Hat release.

It is not a standalone language model and does not include the GLM-5.2 target. It requires the Furnace runtime; unmodified upstream SGLang or vLLM are not validated for this head.

Use with Furnace

Use the same launch as the v1 release, pointing the draft at this repository:

SGLANG_RAGGED_VERIFY_MODE=static python -m sglang.launch_server \
  ... \
  --speculative-algorithm DSPARK \
  --speculative-draft-model-path m8than/GLM-5.2-DSparkFlash-Writer \
  --speculative-draft-model-quantization unquant \
  --speculative-num-draft-tokens 9 \
  --speculative-draft-window-size 2048 \
  --speculative-draft-attention-backend triton \
  --speculative-draft-kv-cache-dtype bfloat16 \
  --superfast

The ... stands for the target and server flags. Keep SGLANG_RAGGED_VERIFY_MODE=static. --superfast is lossy, and its approximate prefix-cache reuse can make outputs depend on cache history. This release does not establish lossless serving.

What was trained

1. Backbone: RedHatAI/GLM-5.2-speculator.dspark at revision cc714308fc4ad68b667a6a71bbf2a344c2ef903b, unchanged (Apache-2.0).
2. Selector: only the markov_head.* tensors (hidden-state projection) were trained, with AdamW (weight decay 0.01, batch 128 anchors). The learning rate was swept over 3e-5, 1e-4 and 3e-4 for up to 2,048 steps; validation selected 1e-4 at step 1,984.
3. Data: supervised labels are greedy decodes from GLM-5.2 E2M1 in Furnace, one 8-token suffix per decode step. Only the assistant reply is supervised; the system prompt, history and tool context are context. Sources: 58 native prompt jobs and 1,363 synthetic creator-chat conversations generated for this project (no real user data). No conversations, outputs or feature caches are in this repository.

Evaluation: offline, not live acceptance

Fixed-anchor replay on the held-out split: 36,972 anchors (35,994 creator, 978 native). The metric is mean accepted length, which counts the anchor.

Head	Overall	Creator	Native
Donor head, no selector	1.826	–	–
Neutral top-16 selector, untrained	1.913	1.794	2.346
This release	1.965	1.835	2.313
Oracle upper bound	2.586

The gain over the untrained selector is small, +0.04 on creator anchors. On native prompts the trained selector is slightly worse, −0.03. The v1 card's figures use different prompts and a different metric and are not comparable to these.

No live server acceptance, throughput or speedup has been measured for this release. Unlike v1, it has not been through the live Furnace smoke check.

Limitations

Scoped to GLM-5.2 on the Furnace ROCm gfx942 setup. Creator-chat traffic is synthetic. Other hardware, GLM-5.3 and general quality gains are not established. Experimental; not a replacement for the default model.