eaddario/DeepSeek-R1-Distill-Llama-8B-GGUF

Experimental layer-wise quantization of deepseek-ai/DeepSeek-R1-Distill-Llama-8B

Using LLaMA C++ release b5120 for quantization.

Original model: deepseek-ai/DeepSeek-R1-Distill-Llama-8B

From the original model creators:

DeepSeek-R1-Zero, a model trained via large-scale reinforcement learning (RL) without supervised fine-tuning (SFT) as a preliminary step, demonstrated remarkable performance on reasoning. With RL, DeepSeek-R1-Zero naturally emerged with numerous powerful and interesting reasoning behaviors. However, DeepSeek-R1-Zero encounters challenges such as endless repetition, poor readability, and language mixing. To address these issues and further enhance reasoning performance, we introduce DeepSeek-R1, which incorporates cold-start data before RL. DeepSeek-R1 achieves performance comparable to OpenAI-o1 across math, code, and reasoning tasks. To support the research community, we have open-sourced DeepSeek-R1-Zero, DeepSeek-R1, and six dense models distilled from DeepSeek-R1 based on Llama and Qwen. DeepSeek-R1-Distill-Qwen-32B outperforms OpenAI-o1-mini across various benchmarks, achieving new state-of-the-art results for dense models.

NOTE: Before running DeepSeek-R1 series models locally, we kindly recommend reviewing the Usage Recommendation section.

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 custom versions of llama-imatrix and llama-quantize to identify the influential tensors, and quantize the most important layers to higher bit precision and the less important to lower bits. 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.

There’re two pull requests (imatrix & quantize) to merge these changes back into the core llama.cpp project. This may or may not ever happen so, until then, the modified versions 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).

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 modeled, and it helps to counterbalance the negative effects of quantization and pruning.

The process to generate these models is roughly as follows:

1. Convert the the original model's tensors to GGUF F16*
2. Estimate the Perplexity score for the F16 model (baseline) using the wikitext-2-raw-v1 dataset, and save the logits
3. Generate an imatrix from selected calibration datasets
4. Determine tensor and layer Importance Score contribution using a modified version of llama-imatrix
5. Select an appropiate quant level for each tensor using a modified version of llama-quantize
6. Calculate Perplexity, KL Divergence, ARC (Easy+Challenge), HellaSwag, MMLU, Truthful QA and WinoGrande scores for each quantized model
7. Keep versions with the best scores
8. 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
DeepSeek-R1-Distill-Llama-8B-IQ3_M	3.78	N/A	3.69	2.5%
DeepSeek-R1-Distill-Llama-8B-IQ3_S	N/A	N/A	3.43	N/A
DeepSeek-R1-Distill-Llama-8B-IQ4_NL	4.68	N/A	4.39	6.1%
DeepSeek-R1-Distill-Llama-8B-Q3_K_L	4.32	N/A	3.76	13.0%
DeepSeek-R1-Distill-Llama-8B-Q3_K_M	4.02	4.02	3.56	11.3%
DeepSeek-R1-Distill-Llama-8B-Q3_K_S	3.66	N/A	3.31	9.7%
DeepSeek-R1-Distill-Llama-8B-Q4_K_M	4.92	4.92	4.41	10.5%
DeepSeek-R1-Distill-Llama-8B-Q4_K_S	4.69	N/A	4.28	8.8%
DeepSeek-R1-Distill-Llama-8B-Q5_K_M	5.73	5.73	5.38	6.2%
DeepSeek-R1-Distill-Llama-8B-Q5_K_S	5.60	N/A	5.24	6.4%
DeepSeek-R1-Distill-Llama-8B-Q6_K	6.60	6.60	6.57	0.5%
DeepSeek-R1-Distill-Llama-8B-Q8_0	8.54	8.54	7.73	9.4%

Perplexity and KL Divergence scores

Model	μPPL	𝜌PPL	μKLD	RMS Δp
DeepSeek-R1-Distill-Llama-8B-IQ3_M	16.574922 ±0.145677	94.24%	0.367217 ±0.001523	15.613 ±0.067
DeepSeek-R1-Distill-Llama-8B-IQ3_S	17.505471 ±0.156184	92.97%	0.465573 ±0.001748	17.507 ±0.069
DeepSeek-R1-Distill-Llama-8B-IQ4_NL	14.243482 ±0.113690	96.12%	0.240811 ±0.001175	13.331 ±0.064
DeepSeek-R1-Distill-Llama-8B-Q3_K_L	15.685298 ±0.134793	94.79%	0.321187 ±0.001430	14.828 ±0.069
DeepSeek-R1-Distill-Llama-8B-Q3_K_M	15.756085 ±0.134240	94.45%	0.341147 ±0.001464	15.306 ±0.069
DeepSeek-R1-Distill-Llama-8B-Q3_K_S	16.481412 ±0.138861	92.23%	0.469890 ±0.001833	17.842 ±0.072
DeepSeek-R1-Distill-Llama-8B-Q4_K_M	13.991103 ±0.118848	98.03%	0.121711 ±0.000730	9.041 ±0.056
DeepSeek-R1-Distill-Llama-8B-Q4_K_S	14.070792 ±0.119167	97.72%	0.139625 ±0.000813	9.709 ±0.058
DeepSeek-R1-Distill-Llama-8B-Q5_K_M	13.259667 ±0.110576	98.94%	0.062326 ±0.000539	6.358 ±0.053
DeepSeek-R1-Distill-Llama-8B-Q5_K_S	13.254089 ±0.110505	98.85%	0.068798 ±0.000581	6.644 ±0.053
DeepSeek-R1-Distill-Llama-8B-Q6_K	13.223880 ±0.110295	99.19%	0.047895 ±0.000556	5.514 ±0.059
DeepSeek-R1-Distill-Llama-8B-Q8_0	13.204903 ±0.110236	99.26%	0.043927 ±0.000549	5.242 ±0.061
DeepSeek-R1-Distill-Llama-8B-F16	13.052024 ±0.108483	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
DeepSeek-R1-Distill-Llama-8B-IQ3_M	49.0667 ±1.8266	70.80	34.6667 ±1.7389	30.4000 ±1.6807	66.1333 ±1.7292	50.21
DeepSeek-R1-Distill-Llama-8B-IQ3_S	48.0000 ±1.8255	70.67	32.1333 ±1.7063	28.5333 ±1.6500	65.0667 ±1.7420	48.88
DeepSeek-R1-Distill-Llama-8B-IQ4_NL	51.6000 ±1.8260	74.53	33.7333 ±1.7276	34.8000 ±1.7405	67.8667 ±1.7063	52.51
DeepSeek-R1-Distill-Llama-8B-Q3_K_L	51.8667 ±1.8257	70.80	34.8000 ±1.7405	31.2000 ±1.6929	67.7333 ±1.7082	51.28
DeepSeek-R1-Distill-Llama-8B-Q3_K_M	52.6667 ±1.8244	70.80	35.0667 ±1.7436	30.5333 ±1.6828	67.7333 ±1.7082	51.36
DeepSeek-R1-Distill-Llama-8B-Q3_K_S	51.4667 ±1.8262	71.87	32.8000 ±1.7155	31.4667 ±1.6968	67.7333 ±1.7082	51.07
DeepSeek-R1-Distill-Llama-8B-Q4_K_M	49.8667 ±1.8270	72.40	37.6000 ±1.7699	31.4667 ±1.6968	68.5333 ±1.6968	51.97
DeepSeek-R1-Distill-Llama-8B-Q4_K_M-bartowski	50.3347 ±1.8306	74.40	34.8000 ±1.7405	37.1069 ±2.7133	69.4667 ±1.6828	53.22
DeepSeek-R1-Distill-Llama-8B-Q4_K_M-unsloth	52.4766 ±1.8284	73.20	33.2000 ±1.7207	36.0000 ±2.6667	68.4000 ±1.6988	52.66
DeepSeek-R1-Distill-Llama-8B-Q4_K_S	50.2667 ±1.8269	72.00	36.0000 ±1.7539	31.4667 ±1.6968	67.2000 ±1.7155	51.39
DeepSeek-R1-Distill-Llama-8B-Q5_K_M	51.4667 ±1.8262	73.73	36.2667 ±1.7567	32.2667 ±1.7082	68.4000 ±1.6988	52.43
DeepSeek-R1-Distill-Llama-8B-Q5_K_S	51.4667 ±1.8262	73.73	35.4667 ±1.7481	32.4000 ±1.7100	67.6000 ±1.7100	52.13
DeepSeek-R1-Distill-Llama-8B-Q6_K	50.4000 ±1.8269	74.13	36.1333 ±1.7553	31.3333 ±1.6949	67.6000 ±1.7100	51.92
DeepSeek-R1-Distill-Llama-8B-Q8_0	50.1333 ±1.8270	74.13	37.0667 ±1.7648	32.1333 ±1.7063	67.7333 ±1.7082	52.24
DeepSeek-R1-Distill-Llama-8B-F16	50.0000 ±1.8270	74.40	36.6667 ±1.7608	32.0000 ±1.7045	67.7333 ±1.7082	52.16

Tokens per Second - Benchmarks

Scores generated using llama-bench. Q4_K_M quantizations from Bartowski and Unsloth included for comparison.

model	size	params	backend	threads	test	t/s
DeepSeek-R1-Distill-Llama-8B-Q4_K_M	4.10 GiB	8.03 B	Metal,BLAS	6	pp512	310.89 ± 2.20
DeepSeek-R1-Distill-Llama-8B-Q4_K_M	4.10 GiB	8.03 B	Metal,BLAS	6	tg128	27.69 ± 0.26
DeepSeek-R1-Distill-Llama-8B-Q4_K_M	4.10 GiB	8.03 B	Metal,BLAS	6	pp1024+tg1024	44.43 ± 0.24
DeepSeek-R1-Distill-Llama-8B-Q4_K_M-bartowski	4.58 GiB	8.03 B	Metal,BLAS	6	pp512	329.03 ± 0.11
DeepSeek-R1-Distill-Llama-8B-Q4_K_M-bartowski	4.58 GiB	8.03 B	Metal,BLAS	6	tg128	25.79 ± 0.92
DeepSeek-R1-Distill-Llama-8B-Q4_K_M-bartowski	4.58 GiB	8.03 B	Metal,BLAS	6	pp1024+tg1024	42.35 ± 0.93
DeepSeek-R1-Distill-Llama-8B-Q4_K_M-unsloth	4.58 GiB	8.03 B	Metal,BLAS	6	pp512	328.93 ± 0.15
DeepSeek-R1-Distill-Llama-8B-Q4_K_M-unsloth	4.58 GiB	8.03 B	Metal,BLAS	6	tg128	26.43 ± 0.01
DeepSeek-R1-Distill-Llama-8B-Q4_K_M-unsloth	4.58 GiB	8.03 B	Metal,BLAS	6	pp1024+tg1024	42.38 ± 0.97

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

A big Thank You! to Colin Kealty for the many contributions and for being one of the best sources of high quality quantized models available in Hugginface, and a really big Thank You! to Georgi Gerganov for his amazing work with llama.cpp and the ggml/gguf libraries.