ProprietaryLegal/pli-local-llm-serving-results

PLI local LLM serving results

Measured speed and numerical-accuracy results for serving two large mixture-of-experts models, GLM-5.3-Flash (Q4_K_M GGUF) and DeepSeek-V4.1-Flash (MXFP4 GGUF), with llama.cpp on one server with four RTX 3090s and 503 GiB of RAM. Published by PLI (ProprietaryLegal.ai), which builds legal-workflow tools on locally hosted models.

The serving recipes, patches and method notes are in the companion GitHub repository Mermiges/pli-local-llm-serving. This dataset holds the same result tables in machine-readable form.

Hardware

AMD EPYC 7302P (16 cores, Zen 2, AVX2), 8 × 64 GB DDR4-3200 (503 GiB), 4 × RTX 3090 24 GB on PCIe Gen4 (three x16, one x8), no NVLink. Ubuntu 24.04, NVIDIA driver 580.126.20, CUDA 12.6.

Files

Config	File	Rows describe
glm53_speed	results/glm53-speed.csv	GLM-5.3-Flash prefill and decode tokens/s per configuration
glm53_numerical	results/glm53-numerical.csv	KL divergence, top-1 agreement and perplexity against a CPU-expert reference
deepseek41_speed	results/deepseek41-speed.csv	DeepSeek-V4.1-Flash prefill and decode tokens/s

results/results.md explains every table in prose.

Headline results

GLM-5.3-Flash Q4_K_M, upstream llama.cpp 3d65c90d0, 16 routed-expert layers on the GPUs (A16), ubatch 2048, 16 threads, F16 KV. Warm, cool-memory medians (n fresh processes), 400 greedy tokens, measured 2026-10-11:

Configuration	Prompt	Prefill tok/s	Decode tok/s	n
Baseline (ctx 65,536)	2,000	186.59	17.48	3
Baseline (ctx 65,536)	30,000	178.18	17.18	3
Baseline (ctx 66,560)	64,000	159.58	16.59	2
+ built-in MTP, draft n = 2	30,000	174.84	18.68	3
+ built-in MTP, draft n = 2	64,000	149.65	15.72	2
+ DFlash2 drafter, n_max 4 (ctx 32,768)	30,000	173.76	20.43	3
A14, ubatch 1024, ctx 292,096 (long-context)	30,000	99.06	16.54	1

The baseline placement passed the numerical gate against all-experts-on-CPU at the same ubatch: mean KLD 0.0014, same top-1 99.16%, perplexity 2.8612 vs 2.8605.

Correction (2026-10-11). The first version of this card gave 169.48 / 12.42 tok/s for the 30k baseline and +64% decode for DFlash2. That baseline was very likely slowed by hot memory: on this machine DDR4 above about 70 °C cost about 28% decode (12.66 vs 17.41 tok/s for the same 2k configuration at 73–75 °C vs 61–64 °C). Against the cool baseline, DFlash2 is about +19% at 30k. Superseded rows remain in glm53_speed with status superseded-thermal.

Column conventions

• page_cache: prewarmed = every model file was read to EOF inside the same lock hold before the server started; uncontrolled = measured before that rule, so the OS page cache state is unknown. Do not compare the two groups directly. On this machine the two models do not fit in RAM together, and page-cache eviction alone moved decode by 30–40%. Rows marked "cool" were also taken with the hottest DIMM at ≤ 64 °C at the start and ≤ 70 °C at the end.
• samples / statistic: median is over that many fresh server processes; single is one.
• Prefill and decode rates are the server's own prompt_per_second and predicted_per_second.
• In glm53_numerical, KLD and top-1 are over 8,190 scored positions (two 8,192-token chunks). Gate for ubatch ≥ 64: KLD ≤ 0.005, top-1 ≥ 98%, PPL within ±0.5%, reference at the same ubatch. Below ubatch 64 the gate is relative to the same build with the feature off (see results.md).

Limits

• All numbers are from one machine. Rows marked uncontrolled and single samples are indicative only.
• The EXL3 rows used a different prompt template and natural-EOS output lengths, so they are not a matched comparison with the llama.cpp rows.
• Perplexity is on one English text; it shows relative differences between configurations, not task quality.
• No legal evaluation data is included. Aggregate results of our private legal battery are in results.md; in short, GLM-5.3-Flash failed the drafting fidelity check (with and without MTP) while passing tool calling 6/6 and long-document questions 7/7.

Licence

MIT. The model weights referred to here are not included and keep their own licences (the DFlash2 drafter is CC BY-NC-ND 4.0).