satgeze/glm-dsa-mtp-audit

GLM-5.2 (GLM_DSA) NextN/MTP for llama.cpp — validation & wiring audit

Reference evidence for the llama.cpp pull request adding NextN / multi-token-prediction (MTP) speculative decoding for the GLM_DSA architecture (GLM-5.2) as a --spec-type draft-mtp target.

This repository exists so reviewers can inspect the raw correctness and performance evidence directly, rather than taking summary numbers on faith. It contains: the operation-by-operation wiring audit against the vLLM reference, the benchmark scripts and their raw outputs, and the determinism control run.

• Hardware: dual Blackwell workstation — 2× RTX PRO 6000 Blackwell (96 GB each), CUDA 12.9, SM120a, -sm layer.
• Model: GLM-5.2 744B-A40B, IQ1_S (the quant that fits fully in VRAM on this rig).
• Benchmark format: am17an's mtp-bench.py (9 fixed prompts; the in-tree house format).

---

1. Decode throughput — MTP vs baseline

Aggregate decode 57.7 → ~79 tok/s ≈ 1.37× greedy, across all nine task types. Every task speeds up; the lowest-acceptance task still matches baseline.

2. Per-position draft acceptance

Greedy, ungated (n_max=3, p_min=0): acceptance 0.834 / 0.653 / 0.494 across the three draft positions, mean accepted length 2.98 tokens. Healthy at all depths, so the framework defaults are the correct configuration for this model — no per-arch p_min override is proposed.

3. Acceptance by task

Aggregate greedy-ungated acceptance 0.6621 (3037 accepted / 4587 drafted over the 9-prompt suite).

4. Acceptance across stacks (context)

Metrics are not identical across stacks — each bar carries its own definition. This is context, not a like-for-like ranking. The vLLM figure is that stack's own published GLM-5 speculator number (per-token mean acceptance at k); ours is aggregate draft_n_accepted / draft_n; the field number is from a real 2-hour agentic coding session.

---

Numbers table

Measurement	Condition	Draft acceptance	Notes
Baseline	no spec	—	57.7 tok/s decode
MTP house bench	stochastic sampling, n_max=3, seed 42	0.5619 aggregate	mtp-bench.py as-is
MTP greedy ungated	temp 0, top_k 1, p_min 0	0.6621 aggregate	per-position 0.834 / 0.653 / 0.494, mean len 2.98
MTP field	2h agentic session, 264K ctx, q4_0 KV	0.86207	real Claude Code workload, zero server failures

Determinism control (why greedy MTP-on ≠ MTP-off byte-for-byte)

Batched speculative verification evaluates N draft tokens' logits in one forward pass; its floating-point reduction order differs from sequential decode, flipping argmax at near-tie tokens. This is framework-level behavior, not specific to this port. The control/ directory holds an A/B on an already-merged draft-mtp architecture (hy_v3, greedy top_k=1 temp=0 seed=42, 5 prompts × 500 tokens): all five prompts diverge identically (first difference as early as byte 33), with the merged arch's own MTP genuinely engaged (acceptance 0.21–0.54). Quality-equivalent, high-acceptance, faster — never claimed byte-identical.

Wiring audit

wiring-audit.md — operation-by-operation comparison of the llama.cpp glm_dsa::graph_mtp graph against vLLM's GLM MTP reference (deepseek_mtp.py, which vLLM routes glm_moe_dsa through). 8 of 9 checklist items are exact matches (embedding, enorm/hnorm, concat order [embed, hidden], eh_proj, attention dims/ranks/scale, MoE gating, shared-head order, next-step hidden recycle). The one flagged divergence: the MTP layer runs dense MLA — matching llama.cpp's own GLM_DSA trunk — where vLLM uses the sparse DSA lightning indexer. This is a pre-existing whole-model design choice in llama.cpp, not an MTP wiring decision, and is disclosed in full.

Contents

charts/           the four figures above (PNG)
benchmarks/       mtp-bench.py, greedy-accept.py, and raw JSON + server-log outputs
control/          hy_v3 determinism A/B run (base vs MTP outputs + result)
wiring-audit.md   the full vLLM-reference wiring audit

Reproduce

llama-server -m GLM-5.2-IQ1_S.gguf -ngl 99 -sm layer -c 32768 -np 1 -fa on --jinja [--spec-type draft-mtp]
python3 benchmarks/mtp-bench.py --url http://127.0.0.1:8080 --out bench.json
python3 benchmarks/greedy-accept.py http://127.0.0.1:8080 greedy.json   # greedy, ungated

---

Contribution developed by @satindergrewal. Implementation assisted by an AI system (Anthropic Claude), reviewed and hardware-tested by the author. This repository is a validated reference for the corresponding llama.cpp PR.