Experimental layer-wise + pruned (layers 4 and 5) quantization of mistralai/Devstral-Small-2505

Using LLaMA C++ release b5890 for quantization.

Original model: mistralai/Devstral-Small-2505

From the original model creators:

Devstral Small 1.0

Devstral is an agentic LLM for software engineering tasks built under a collaboration between Mistral AI and All Hands AI 🙌. Devstral excels at using tools to explore codebases, editing multiple files and power software engineering agents. The model achieves remarkable performance on SWE-bench which positionates it as the #1 open source model on this benchmark.

It is finetuned from Mistral-Small-3.1, therefore it has a long context window of up to 128k tokens. As a coding agent, Devstral is text-only and before fine-tuning from Mistral-Small-3.1 the vision encoder was removed.

For enterprises requiring specialized capabilities (increased context, domain-specific knowledge, etc.), we will release commercial models beyond what Mistral AI contributes to the community.

Learn more about Devstral in our blog post.

PLEASE READ THIS BEFORE USING THESE EXPERIMENTAL VERSIONS!

An area of personal interest is finding ways to optimize the inference performance of LLMs when deployed in resource-constrained environments like commodity hardware, desktops, laptops, mobiles, edge devices, etc. There are many approaches to accomplish this, including architecture simplification and knowledge distillation, but my focus has been primarily on quantization and pruning.

The method used to produce these experimental versions is covered in Squeezing Tensor Bits: the quest for smaller LLMs, but at a high level it involves using a custom version of llama-imatrix to identify influential tensors, quantize the most important layers to higher bit precision and the less important to lower bits, and remove (prune) one or more layers. This process was partly inspired by Dumitru's et al Layer-Wise Quantization: A Pragmatic and Effective Method for Quantizing LLMs Beyond Integer Bit-Levels, and Xin Men's et al ShortGPT: Layers in Large Language Models are More Redundant Than You Expect

As of version b5125, llama-quantize can perform tensor-wide quantization (TWQ), whereby user-defined tensors are quantized at a specific level, or perform layer-wise quantization (LWQ) by selecting different quantization types per tensor/layer. For example, --tensor-type attn_v=q6_k will quantize all Attention Value tensors at q6_k (TWQ), and --tensor-type "\.([0-9]|1[01257]|31)\.attn_k=q4_k" will quantize Attention Key tensors on layers 0 to 9, 10, 11, 12, 15, 17 and 31 at q4_k, leaving the remaining layers at their default value (LWQ).

As of version b5740, llama-quantize can also prune models during quantisation by providing a comma-separated list in the --prune-layers command line option. The pruning operation will renumber remaining layers to avoid gaps in the sequence, update the relevant model metadata and, if an imatrix is available, it will use the correct importance score vector. This option can be used alongside --tensor-type to perform tensor/layer-wise quantization on selected tensor types, whilst at the same time pruning others. For example:

llama-quantize --tensor-type attn=q6_k --prune-layers 3,7,11 --imatrix imatrix.dat model-f32.gguf model-q4_k_m.gguf q4_k_m

An enhanced version of llama-imatrix generates useful statistics to guide the tensor and layer selection process. --show-statistics will display:

Σ(Act²): the sum of all squared activations over the tensor (i.e. the Importance Scores)
Min & Max: minimum and maximum squared activation values
μ & σ: activations' mean and standard deviation
% Active: proportion of elements whose average squared activation exceeds a very small threshold (1e-5). Helpful to determine how alive/dormant the tensor is during inference
N: number of squared activations in the tensor
Entropy: entropy of the squared activation distribution, in bits (standard Shannon entropy measurement)
E (norm): Normalized entropy.
ZD Score: z-score distribution as described in 3.1 Layer Importance Scores in the Layer-Wise Quantization paper
CosSim: cosine similarity between same type tensors with respect to the previous layer (i.e. blk.7.attn_k and blk.6.attn_k)

Please note that statistics are calculated for each individual tensor and should be used to compare between tensors of the same type only. For example, assuming that attn_k in layer 10 has a higher influence during inference than attn_k in layer 7 because its Σ(Act²) is larger makes sense, whilst concluding the same between attn_k and ffn_down does not.

There’s a pull request to merge these changes back into the core llama.cpp project. This may or may not ever happen so, until then, the modified version will be available on GitHub.

For testing and comparison I use models produced by Unsloth (Daniel and Michael Han do some really advanced level stuff!) and Bartowski (see credits below) but if they don't provide versions of the required model, all tests and comparisons are done against naive quantizations obtained by simply running llama-quantize with no further optimization.

All experimental versions were generated using an appropriate imatrix created from calibration datasets available at eaddario/imatrix-calibration. At its core, an Importance Matrix (imatrix) is a table or, more broadly, a structured representation that scores the relative importance of different features or parameters in a machine learning model. It essentially quantifies the "impact" each feature has on a specific outcome, prediction, or relationship being modelled, and it helps to counterbalance the negative effects of quantization and pruning.

The process to generate these models is roughly as follows:

Convert the original model's tensors to GGUF F16*
Estimate the Perplexity score for the F16 model (baseline) using the wikitext-2-raw-v1 dataset, and save the logits
Generate an imatrix from selected calibration datasets
Determine tensor and layer Importance Score contribution using the enhanced version of llama-imatrix
Select an appropriate quant level for each tensor and quantize/prune the model using llama-quantize. In this model's case, layers 4 and 5 have been pruned
Calculate Perplexity, KL Divergence, ARC (Easy+Challenge), HellaSwag, MMLU, Truthful QA and WinoGrande scores for each quantized model
Keep versions with the best scores
Repeat until all desired quants are created. I find that quantizations below Q3/IQ3 are not fit for my purposes and therefore do not usually generate them, but happy to provide other quants on request.

*BF16 would be preferred, but Apple's GPUs don't support it yet, and therefore any operations are executed in the CPU, making it unacceptably slow. This is expected to change in the near term but until then, if you are using Apple kit avoid using any models tagged BF16

Models

Sizes (in GB)

Model	Bartowski	Unsloth	Repo	Shrinkage
Devstral-Small-2505-pruned-IQ3_M	10.7	N/A	10.3	3.7%
Devstral-Small-2505-pruned-IQ3_S	9.9	N/A	9.6	3.1%
Devstral-Small-2505-pruned-IQ4_NL	13.5	13.5	12.5	7.4%
Devstral-Small-2505-pruned-Q3_K_L	12.4	N/A	10.9	12.1%
Devstral-Small-2505-pruned-Q3_K_M	11.5	11.5	10.0	13.0%
Devstral-Small-2505-pruned-Q3_K_S	10.4	10.4	9.0	13.5%
Devstral-Small-2505-pruned-Q4_K_M	14.3	14.3	12.5	12.6%
Devstral-Small-2505-pruned-Q4_K_S	13.5	13.5	11.6	14.1%
Devstral-Small-2505-pruned-Q5_K_M	16.8	16.8	15.3	8.9%
Devstral-Small-2505-pruned-Q5_K_S	16.3	16.3	14.5	11.0%
Devstral-Small-2505-pruned-Q6_K	19.3	19.3	18.9	2.1%
Devstral-Small-2505-pruned-Q8_0	25.1	25.1	20.3	19.1%

Bits per Weight, Perplexity and KL Divergence scores

Model	BPW	μPPL	𝜌PPL	μKLD	RMS Δp
Devstral-Small-2505-pruned-IQ3_M	3.6735	25.296399 ±0.207824	68.77%	1.589584 ±0.004716	40.684 ±0.088
Devstral-Small-2505-pruned-IQ3_S	3.4052	33.072181 ±0.266788	66.25%	1.858256 ±0.004862	43.468 ±0.087
Devstral-Small-2505-pruned-IQ4_NL	4.4435	25.870345 ±0.201328	67.87%	1.610465 ±0.004621	42.340 ±0.088
Devstral-Small-2505-pruned-Q3_K_L	3.8705	25.391481 ±0.200241	68.62%	1.590939 ±0.004603	41.557 ±0.088
Devstral-Small-2505-pruned-Q3_K_M	3.5467	25.452478 ±0.201209	68.69%	1.593512 ±0.004603	41.653 ±0.088
Devstral-Small-2505-pruned-Q3_K_S	3.2060	25.605200 ±0.202837	69.46%	1.599039 ±0.004488	41.245 ±0.088
Devstral-Small-2505-pruned-Q4_K_M	4.4492	21.875174 ±0.172411	70.04%	1.440707 ±0.004475	39.909 ±0.088
Devstral-Small-2505-Q4_K_M-bartowski	4.8620	5.151812 ±0.029360	99.47%	0.020939 ±0.000193	4.633 ±0.046
Devstral-Small-2505-Q4_K_M-unsloth	4.8620	5.153395 ±0.029344	99.47%	0.021015 ±0.000194	4.649 ±0.047
Devstral-Small-2505-pruned-Q4_K_S	4.1390	22.721916 ±0.178813	69.70%	1.478272 ±0.004493	40.367 ±0.088
Devstral-Small-2505-pruned-Q5_K_M	5.4555	20.789833 ±0.163143	70.86%	1.391234 ±0.004397	39.303 ±0.088
Devstral-Small-2505-pruned-Q5_K_S	5.1466	21.078977 ±0.165583	70.53%	1.404281 ±0.004435	39.453 ±0.088
Devstral-Small-2505-pruned-Q6_K	6.7205	20.781395 ±0.162964	70.85%	1.390876 ±0.004395	39.291 ±0.087
Devstral-Small-2505-pruned-Q8_0	7.2324	20.620931 ±0.161512	71.05%	1.383438 ±0.004372	39.180 ±0.087
Devstral-Small-2505-pruned-F16	16.0003	5.050590 ±0.028422	100%	N/A	N/A

ARC, HellaSwag, MMLU, Truthful QA and WinoGrande scores

Scores generated using llama-perplexity with 750 tasks per test, and a context size of 768 tokens.

For the test data used in the generation of these scores, follow the appropiate links: HellaSwag, ARC, MMLU, Truthful QA and WinoGrande

Model	ARC	HellaSwag	MMLU	Truthful QA	WinoGrande	Avg Score
Devstral-Small-2505-pruned-IQ3_M	62.4000 ±1.7699	77.87	38.9333 ±1.7816	31.8667 ±1.7026	70.9333 ±1.6591	56.40
Devstral-Small-2505-pruned-IQ3_S	61.7333 ±1.7759	77.73	40.6667 ±1.7948	31.3333 ±1.6949	71.8667 ±1.6430	56.67
Devstral-Small-2505-pruned-IQ4_NL	60.5333 ±1.7860	78.80	39.7333 ±1.7880	36.0000 ±1.7539	72.2667 ±1.6358	57.47
Devstral-Small-2505-pruned-Q3_K_L	58.6667 ±1.7993	77.33	42.4000 ±1.8057	35.2000 ±1.7451	72.6667 ±1.6284	57.25
Devstral-Small-2505-pruned-Q3_K_M	60.0000 ±1.7900	77.07	43.4667 ±1.8113	35.4667 ±1.7481	72.8000 ±1.6260	57.76
Devstral-Small-2505-pruned-Q3_K_S	55.3333 ±1.8165	77.20	40.5333 ±1.7939	34.1333 ±1.7325	69.7333 ±1.6787	55.39
Devstral-Small-2505-pruned-Q4_K_M	59.2000 ±1.7958	78.00	40.2667 ±1.7920	34.5333 ±1.7374	71.0667 ±1.6569	56.61
Devstral-Small-2505-Q4_K_M-bartowski	66.9333 ±1.7190	82.80	43.2000 ±1.8100	35.8667 ±1.7525	79.8667 ±1.4652	61.73
Devstral-Small-2505-Q4_K_M-unsloth	66.9333 ±1.7190	82.80	43.6000 ±1.8119	36.4000 ±1.7581	79.8667 ±1.4652	61.92
Devstral-Small-2505-pruned-Q4_K_S	60.9333 ±1.7827	77.87	40.4000 ±1.7930	35.0667 ±1.7436	72.2667 ±1.6358	57.31
Devstral-Small-2505-pruned-Q5_K_M	59.6000 ±1.7930	77.60	41.2000 ±1.7984	35.0667 ±1.7436	73.3333 ±1.6158	57.36
Devstral-Small-2505-pruned-Q5_K_S	59.8667 ±1.7910	77.47	41.3333 ±1.7993	34.2667 ±1.7342	72.9333 ±1.6235	57.17
Devstral-Small-2505-pruned-Q6_K	61.0667 ±1.7816	77.73	41.2000 ±1.7984	34.9333 ±1.7420	72.1333 ±1.6382	57.41
Devstral-Small-2505-pruned-Q8_0	60.6667 ±1.7849	77.33	41.3333 ±1.7993	34.6667 ±1.7389	72.4000 ±1.6334	57.28
Devstral-Small-2505-pruned-F16	66.8000 ±1.7207	83.47	43.7333 ±1.8126	36.0000 ±1.7539	79.7333 ±1.4688	61.95

Tokens per Second - Benchmarks

Scores generated using llama-bench. Naive (llama-quantize with no optimization) Q4_K_M quantization included for comparison.

model	size	params	backend	threads	test	t/s
Devstral-Small-2505-pruned-Q4_K_M	11.63 GiB	22.46 B	Metal,BLAS	12	pp512	266.12 ±6.25
Devstral-Small-2505-pruned-Q4_K_M	11.63 GiB	22.46 B	Metal,BLAS	12	tg128	28.07 ±0.33
Devstral-Small-2505-pruned-Q4_K_M	11.63 GiB	22.46 B	Metal,BLAS	12	pp1024+tg1024	47.14 ±0.22
Devstral-Small-2505-Q4_K_M-bartowski	13.34 GiB	23.57 B	Metal,BLAS	12	pp512	254.77 ±13.57
Devstral-Small-2505-Q4_K_M-bartowski	13.34 GiB	23.57 B	Metal,BLAS	12	tg128	27.64 ±0.39
Devstral-Small-2505-Q4_K_M-bartowski	13.34 GiB	23.57 B	Metal,BLAS	12	pp1024+tg1024	45.62 ±0.16
Devstral-Small-2505-Q4_K_M-unsloth	13.34 GiB	23.57 B	Metal,BLAS	12	pp512	259.37 ±5.28
Devstral-Small-2505-Q4_K_M-unsloth	13.34 GiB	23.57 B	Metal,BLAS	12	tg128	28.43 ±0.23
Devstral-Small-2505-Q4_K_M-unsloth	13.34 GiB	23.57 B	Metal,BLAS	12	pp1024+tg1024	45.62 ±0.15

Metrics used

Perplexity: one of the key metrics used in NLP evaluation. It measures the quality of a language model by evaluating how well it predicts the next token given a particular sequence of words. A PPL of 1 indicates an exact match between predicted and actual, whereas values greater than one indicate a degree of "surprise" the generated token differs from the expected.

Kullback–Leibler (KL) Divergence: a statistical measure of how much a probability distribution differs from another. When quantizing models (or altering the original tensors in any way for that matter), the closest we can preserve the weights' probability distribution to the original model the better, thus the closest to 0 the better.

AI2 Reasoning Challenge (ARC): a benchmark to evaluate the ability of AI models to answer complex science questions that require logical reasoning beyond pattern matching.

HellaSwag: the Harder Endings, Longer contexts, and Low-shot Activities for Situations With Adversarial Generations (bit of a mouthful!) is a benchmark designed to test commonsense natural language inference. It requires the model to predict the most likely ending of a sentence.

MMLU: the Massive Multitask Language Understanding evaluates LLMs’ general knowledge and problem-solving abilities across 57 subjects, including elementary mathematics, US history, computer science, and law.

Truthful QA: evaluates how well LLMs generate truthful responses to questions. It identifies whether AI models can avoid generating false or misleading information, particularly in areas where human knowledge is prone to misconceptions.

Winogrande: based on the Winograd Schema Challenge, is a natural language understanding task requiring models to resolve ambiguities in sentences involving pronoun references.

Credits

LLaMa C++ has a large and vibrant community of contributors (~1,200 last time I checked) that actively maintain and extend its functionality, adding new models and architectures almost as fast as they appear (considering the breakneck speed at which the AI/ML field is advancing, this alone is a remarkable feat!), and whilst I'm grateful to each and everyone of them, I want to recognise three people in particular: Thank You! Colin Kealty for the many contributions and for being one of the best sources of high quality quantized models available on Hugging Face, and a really big Thank You! to Georgi Gerganov for his amazing work with llama.cpp and the ggml/gguf libraries, and Iwan Kawrakow for being one of the key authors behind the many quantisation algorithms and the imatrix functionality.