rtx-5090-benchmarks / reports /quant-tax.md
witcheer's picture
quant-tax report (t090)
4dcc135 verified
|
Raw
History Blame Contribute Delete
5.5 kB

The quant tax on Qwen3.6-27B: near-zero down to Q4, one real step at Q3

Rig: one RTX 5090 32GB (sm_120) · llama.cpp CUDA build (canary-matched sm_120, quantize sha 0a50d990) · Qwen/Qwen3.6-27B base at commit 6a9e13b · plain K-quants, no imatrix (stated, not implied) · greedy (temp 0), thinking off · every rung served and scored on the same harness, same day (t090 ladder, 2026-08-03).

This is the ladder the NVFP4 report kept pointing at: same subject, one pin, quality measured across the GGUF K-quant rungs so "does quantization cost accuracy" gets a number. Five rungs, Q8_0 down to Q3_K_M, against a common harness. The BF16 source (54.7GB) is the provenance root, not a serving target on a 32GB card.

The finding

Down to Q4_K_M the tax is inside the noise: composite quality moves 92.3 to 92.0 from Q8 to Q4 (−0.4 pt) while single-stream decode goes 52.8 to 80.4 tok/s (+52%) and the file shrinks 29.0 to 16.8 GB. Q3_K_M is where the ladder first bends: another +12% decode (90.2 tok/s) for a −1.4 pt quality step, the first drop larger than the between-run noise.

Two operator takeaways from the grid:

  • Q6_K strictly dominates Q8_0 here. Identical scores on all three suites (208/240 MMLU, 244/250 GSM8K, 152/164 HumanEval), 6 GB smaller, 21% faster decode. On this subject there is no reason to serve Q8.
  • Q4_K_M is the sweet spot. 80 tok/s, 16.8 GB (half a 5090's VRAM, room for a second model or long context), and quality within 0.4 pt of Q8.

The numbers

Same models, same harness, one day. Decode is single-stream tg; VRAM is the served peak; quality suites are pass@1, greedy, think-off.

rung size decode tok/s VRAM MMLU (/240) GSM8K (/250) HumanEval (/164) composite
Q8_0 29.0 GB 52.8 27.6 GiB 86.67 97.60 92.68 92.32
Q6_K 22.4 GB 63.7 21.7 GiB 86.67 97.60 92.68 92.32
Q5_K_M 19.5 GB 71.8 19.2 GiB 85.83 97.60 93.29 92.24
Q4_K_M 16.8 GB 80.4 16.8 GiB 84.58 98.00 93.29 91.96
Q3_K_M 13.5 GB 90.2 13.8 GiB 83.75 96.40 91.46 90.54

Composite = mean of the three suites. Prefill tok/s is compute-bound and effectively flat across rungs (1360–1580, run-to-run noise dominates the quant effect); the clean, monotonic signal is decode, which is memory-bound and tracks file size directly.

Reading the ladder per suite

The aggregate hides the per-suite shape, and the per-suite shape is where the honest caveat lives:

  • MMLU is the only suite that declines smoothly with bits: 86.67, 86.67, 85.83, 84.58, 83.75, a gentle monotonic slope. Knowledge recall is where the low-bit rounding shows up first and most steadily.
  • GSM8K and HumanEval are flat until Q3. GSM8K sits at 244/250 through Q8 to Q5, blips to 245 at Q4, drops to 241 at Q3. HumanEval holds 152 to 153/164 through Q4, drops to 150 at Q3. These are one- and two-question moves on samples of 250 and 164, inside the ±1 to 1.5 pt single-flip noise floor until the Q3 step clears it.

So "the tax is near-zero down to Q4" is a claim about aggregate and about GSM8K/HumanEval; MMLU is already paying a small, steady toll the whole way down. Per-suite is a stronger statement than the composite, and only MMLU discriminates the top of the ladder.

What's not here (and why)

  • No perplexity column. At the ladder's -c 4096 context the corpus is only 34 chunks, and llama-perplexity returned a non-physical ordering: Q3_K_M lowest (8.03), Q4_K_M highest (9.51), a spread the stated ±0.11 error bars come nowhere near explaining, and one that contradicts the monotonic MMLU from the same GGUF files. Re-running the two extremes at the standard -c 512 stride (274 chunks, same binary and corpus) collapses it: Q8_0 lands at 7.05 ±0.07 and Q3_K_M at 7.02 ±0.07, tied inside the error bars. The scramble was sampling noise from 34 chunks, not a fault in the weights or the perplexity path. And the corrected result agrees with the rest of this report: at honest sampling, wikitext2 perplexity barely separates these quants, so the accuracy suites (MMLU above all) are the discriminator. A full 5-rung -c 512 column is a cheap follow-up, unlikely to add signal the accuracy grid does not already carry.
  • Three suites, not five. This ladder ran MMLU/GSM8K/HumanEval, not the full leaderboard q_avg (which adds ARC-C and HellaSwag). The composite here is a 3-suite mean and is not comparable to a leaderboard q_avg row.
  • Plain K-quants, no imatrix. An importance matrix would lift the low rungs (especially Q3) and is the obvious next rung of the study. These numbers are the no-imatrix floor.
  • One subject. Qwen3.6-27B only. The shape (flat to Q4, knee at Q3) is this model's; for other models it is a hypothesis, not established.

Reproduce

# on capsule (GPU), Donald down first; run_ladder restores it via the chain wrapper
cd ~/benchmark-rig
bash scripts/quant_tax/run_ladder.sh          # serves each rung, runs speed + 3 suites + ppl
python3 scripts/quant_tax/grade.py results/quant_tax/*.gens.json   # offline grading

# chart (on the Mac, matplotlib on system python3)
python3 scripts/chart_quant_tax.py            # -> reports/chart_quant_tax_5090.png

Provenance: ~/t090/gguf/PROVENANCE.json (base commit, quantize build, imatrix state). Ladder log: ~/t090/logs/ladder-run-2026-08-03.log.