llamap: per-tensor GGUF precision maps from measured output KL (beats Unsloth Dynamic 2.0 at equal size on Llama 3.1 8B / Qwen3-8B)

[RedOxer]

Hi all. This is my first public project in the llama.cpp ecosystem. My background is post-training quantization for embedded CNNs, so I may have missed conventions or prior work here: corrections and critiques are very welcome.

Below, I compare against two references:

• Unsloth: Unsloth's published Dynamic 2.0 GGUFs (UD-Qn_K_XL);
• target-bpw: EAddario's --target-bpw / --target-size mode from PR #15550, which picks per-tensor types under a budget from a weight-based error estimate.

What it does. For each block matrix and each format (Q2_K…Q8_0, IQ), llamap estimates the KL divergence at the model output when that matrix alone is quantized. It uses a Fisher approximation: one backward pass, 128 samples. Candidates are quantized in memory through ggml's ggml_quantize_chunk with the imatrix, byte-identical to llama-quantize. A multiple-choice knapsack picks one format per matrix under a byte budget, and the result is a --tensor-type-file for stock llama-quantize. About 15 min for the first budget on an 8B, then about 1 min per extra budget.

Validation. Ground truth from the degrade-one protocol (everything Q8_0, one tensor in Q2_K, ΔKL vs BF16), Spearman correlation:

Cost estimate	Qwen3-1.7B	SmolLM2-1.7B
target-bpw's own per-tensor error (from its log)	0.45	0.44
llamap, exact output KL	0.93	0.95
llamap, Fisher (default)	0.93	0.95

Fisher vs exact on Llama 3.1 8B: 0.998 over 2,464 tensor × format pairs. Measured map KL ≈ sum of costs (1.01 at Q3/Q4, 1.24 at Q2).

Results, at the exact sizes of Unsloth's files (same token_embd/output, sizes within 0.2 MB; target-bpw and llamap use the same imatrix). KL vs BF16 averaged over 6–7 texts disjoint from calibration: wikitext, a book, code, chat, chat with template, Qwen3 thinking traces.

Model	Budget	vs Unsloth	vs target-bpw
Llama 3.1 8B	Q2 / Q3 / Q4 / Q5	−37 / −26 / −17 / −13 %	−8 / −4 / −4 / −14 %
Qwen3-8B	Q2 / Q3 / Q4 / Q5	−35 / −39 / −19 / −22 %	−2 / −9 / −10 / −25 %

Against Unsloth, every text is better at every budget. Against target-bpw, the gap is small at Q2–Q4 (within noise on some texts) and large at Q5. The maps show why: llamap puts attn_v in Q8_0 at almost every depth. CUDA speed is unchanged (pp512 within 1 %, tg128 at most −4 %).

Limits. Dense llama and qwen3 only (no MoE yet). It needs the HF weights, PyTorch and a 32 GB GPU for an 8B. It measures KL fidelity, not task scores. Max KL and |ΔPPL| improve in only about 2/3 of cases.

Repo, with commits, file revisions, eval texts and every command: https://github.com/RedOxer/Llamap

Most welcome:

• results on other models;
• critiques of the evaluation;
• whether an output-KL cost could be useful inside target-bpw.