m8than/GLM-5.2-DFlash2-B12-Writer

GLM-5.2 DFlash2 B12 Furnace — Conversations v3

Experimental 2.29B-parameter BF16 DFlash2 draft, continued for one additional epoch on a frozen 966-conversation snapshot. The selected checkpoint is step 712, with learning rate 3e-5. This is not a standalone language model: use it with GLM-5.2 and the compatible Furnace runtime.

The paired held-out conversation test improved draft acceptance from 8.32% to 9.77%, but the speed results were mixed. This checkpoint was not promoted over the previous conversation-trained draft on the originating server. It is published separately so the older public model is preserved.

Compatibility and launch

Architecture: DFlash2DraftModel, six layers, hidden size 6,144, 64 query heads, eight KV heads, head dimension 96; target layers [1, 16, 31, 45, 60, 75]. The draft uses full attention, not a sliding-window architecture. Its block size is 12: one target anchor plus eleven proposed draft positions. The target verifies those proposals; the block size is not an acceptance guarantee.

Add these flags to a configured Furnace GLM-5.2 launch:

--speculative-algorithm DFLASH \
--speculative-draft-model-path m8than/GLM-5.2-DFlash2-B12-Furnace-Conversations-v3 \
--speculative-draft-model-quantization unquant \
--speculative-dflash-block-size 12 \
--speculative-draft-attention-backend triton \
--speculative-draft-kv-cache-dtype bfloat16

Use the target model's tokenizer, embedding and output head. No target weights are included here. The separately published quantized target is GLM-5.2-E2M1-Furnace. See Furnace's START_COMMANDS.md for the TP8 MI325X launch recipe. Arbitrary upstream SGLang versions and GLM-5.3 are not validated with this checkpoint.

Training and data provenance

• Initialization: the previous local two-epoch conversation-trained draft, weights SHA256 955e12ab35e9341296159a3f144ed25a89deeebd2448f4490d958b7b68d2eb84. This is a continuation, not training from scratch or directly from the donor.
• Frozen snapshot: 966 conversations; 852 train, 62 validation, 52 test.
• Within the 8,192-token training limit: 5,693 training examples with 36,096,263 sequence tokens and 521,112 supervised assistant tokens; 397 validation examples with 35,225 supervised assistant tokens. Over-limit examples were excluded, not truncated.
• Replies were existing zai-org/GLM-5.2 outputs from Featherless, re-tokenized locally with one appended GLM stop token. They were not freshly decoded by the local Furnace teacher for this refresh.
• Teacher hidden states were captured locally using unchanged, full-attention Furnace GLM-5.2-FP8 at revision f33c6dc501ee5a2c7e35155653b1b1abbc320951, not the lossy E2M1 superfast target.
• Full conversation context conditions the draft, but only the final assistant reply contributes to loss. System/user/history tokens are masked.
• Hard labels plus teacher hidden states; not soft-logprob KL distillation.
• One epoch, 712 steps, global batch eight on eight MI325X GPUs. Two 32-step validation-only LR trials compared 3e-5 and 1e-4; the selected 3e-5 full run restarted from the same original initializer. No held-out test tuning.
• Checkpoint selection used cached validation prefix score: 2.4678 before, 2.5444 at step 356, 2.5548 at step 712. These are not live acceptance rates.

The original backbone donor was SubconsciousDev/glm-5.2-fp8-dflash-v2 at b45e4c46ed2902771d47c2cd09ee17870bb8be24. Earlier conversion, pilot, long-run and conversation adaptations precede this refresh; the current assistant-only data description must not be attributed to the entire lineage. No conversations, prompts, responses, per-conversation IDs, or teacher caches are published in this repository.

Evaluation

Matched 43 prompts from 12 held-out conversation families, covering 12 languages; frozen histories, temperature zero, top-p one, up to 256 generated tokens, concurrency four. Target: E2M1 GLM-5.2 with Furnace's lossy superfast prefill, TP8 on eight MI325X GPUs, BF16 KV cache. Acceptance is accepted draft tokens divided by proposed draft tokens.

Draft	Accepted / proposed	Acceptance	Output tokens / verification
Previous conversation-trained draft	2,421 / 29,084	8.3242%	1.9119
This refresh	2,649 / 27,126	9.7655%	2.0673

This is not a head-to-head comparison with the older pre-conversation 17.08% public release: that result used a different test set and setup. All 43 greedy output sequences differed between the two drafts in this test; these are matched prompts, not fixed-token replay. The small result does not establish a task-quality gain, general language ranking, or lossless serving.

Separate single-stream speed tests used 256 output tokens, cold cache, one excluded warmup and three measured trials per context. Values below are medians; metrics.json and the graph also report observed min/max, not CIs.

Input tokens	Before decode tok/s	New decode tok/s	New effective prefill tok/s
1,024	258.6	211.2	9,785.5
8,192	174.2	147.7	19,428.5
32,768	169.1	175.8	20,502.5

Effective prefill is input tokens divided by time to first token, not an isolated kernel measurement. No consistent speed win or 250 tok/s decode achievement is claimed. No new broad task-quality benchmark was run for this draft, and no clean-machine deployment was retested for publication.

License and limitations

The target declares MIT, but the inherited donor draft's redistribution terms remain undeclared/unresolved. This release does not assign a new license to inherited donor weights or infer their license from the target. Review the donor terms before redistribution or broader use. The model is experimental; acceptance, long-context quality and speed depend strongly on workload/runtime.