eaddario/OLMo-2-1124-7B-Instruct-GGUF

Experimental layer-wise quantization of allenai/OLMo-2-1124-7B-Instruct

Using LLaMA C++ release b5220 for quantization.

Original model: allenai/OLMo-2-1124-7B-Instruct

From the original model creators:

OLMo 2 7B Instruct November 2024 is post-trained variant of the OLMo-2 7B November 2024 model, which has undergone supervised finetuning on an OLMo-specific variant of the Tülu 3 dataset and further DPO training on this dataset, and finally RLVR training using this data. Tülu 3 is designed for state-of-the-art performance on a diversity of tasks in addition to chat, such as MATH, GSM8K, and IFEval. Check out the OLMo 2 paper or Tülu 3 paper for more details!

OLMo is a series of Open Language Models designed to enable the science of language models. These models are trained on the Dolma dataset. We are releasing all code, checkpoints, logs (coming soon), and associated training details. The core models released in this batch include the following:

Stage	OLMo 2 7B	OLMo 2 13B
Base Model	allenai/OLMo2-7B-1124	allenai/OLMo-2-13B-1124
SFT	allenai/OLMo-2-1124-7B-SFT	allenai/OLMo-2-1124-13B-SFT
DPO	allenai/OLMo-2-1124-7B-DPO	allenai/OLMo-2-1124-13B-DPO
Final Models (RLVR)	allenai/OLMo-2-1124-7B-Instruct	allenai/OLMo-2-1124-13B-Instruct
Reward Model (RM)	allenai/OLMo-2-1124-7B-RM	allenai/OLMo-2-1124-13B-RM

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 and llama-quantize to identify 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.

As of version b5125 llama-quantize can now 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).

The modified version of llama-imatrix generates useful statistics to guide the tensor selection process, --show-statistics will display:

• Σ(Bias): the sum of all activations over the tensor (i.e. the Importance Scores)
• Min & Max: minimum and maximum activation values
• μ & σ: activations' mean and standard deviation
• % Active: proportion of elements whose average activation exceeds a very small threshold (1e-6). Helpful to determine how alive/dormant the tensor is during inference
• N: number of activations in the tensor
• Entropy: entropy of the 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 individial 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(Bias) 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 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 the modified version of llama-imatrix
5. Select an appropiate quant level for each tensor and quantize the model using 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	Repo	Shrinkage
OLMo-2-1124-7B-Instruct-IQ3_M	3.78	3.69	2.4%
OLMo-2-1124-7B-Instruct-IQ3_S	3.68	3.43	6.8%
OLMo-2-1124-7B-Instruct-IQ4_NL	4.71	4.39	6.2%
OLMo-2-1124-7B-Instruct-Q3_K_L	4.32	3.76	13.0%
OLMo-2-1124-7B-Instruct-Q3_K_M	4.02	3.56	11.4%
OLMo-2-1124-7B-Instruct-Q3_K_S	3.66	3.31	9.6%
OLMo-2-1124-7B-Instruct-Q4_K_M	4.92	4.41	10.4%
OLMo-2-1124-7B-Instruct-Q4_K_S	4.69	4.28	8.7%
OLMo-2-1124-7B-Instruct-Q5_K_M	5.73	5.38	6.1%
OLMo-2-1124-7B-Instruct-Q5_K_S	5.60	5.24	6.4%
OLMo-2-1124-7B-Instruct-Q6_K	6.60	6.57	0.5%
OLMo-2-1124-7B-Instruct-Q8_0	8.54	7.73	9.5%

Perplexity and KL Divergence scores

Model	μPPL	𝜌PPL	μKLD	RMS Δp
OLMo-2-1124-7B-Instruct-IQ3_M	9.201710 ±0.071255	96.92%	0.153149 ±0.000797	11.390 ±0.060
OLMo-2-1124-7B-Instruct-IQ3_S	9.306264 ±0.071084	95.94%	0.197699 ±0.000965	12.938 ±0.062
OLMo-2-1124-7B-Instruct-IQ4_NL	8.680650 ±0.065689	98.37%	0.076583 ±0.000454	8.111 ±0.049
OLMo-2-1124-7B-Instruct-Q3_K_L	9.252820 ±0.070404	95.74%	0.204708 ±0.001020	13.139 ±0.063
OLMo-2-1124-7B-Instruct-Q3_K_M	9.242884 ±0.069850	95.38%	0.220640 ±0.001086	13.659 ±0.065
OLMo-2-1124-7B-Instruct-Q3_K_S	9.651383 ±0.073494	94.01%	0.287772 ±0.001362	15.534 ±0.069
OLMo-2-1124-7B-Instruct-Q4_K_M	8.683512 ±0.065748	98.46%	0.071862 ±0.000424	7.858 ±0.048
OLMo-2-1124-7B-Instruct-Q4_K_M-bartowski	7.951677 ±0.058104	97.11%	0.144072 ±0.001123	10.504 ±0.065
OLMo-2-1124-7B-Instruct-Q4_K_S	8.665009 ±0.065466	98.38%	0.076159 ±0.000445	8.104 ±0.049
OLMo-2-1124-7B-Instruct-Q5_K_M	8.475671 ±0.064030	99.41%	0.025628 ±0.000174	4.820 ±0.037
OLMo-2-1124-7B-Instruct-Q5_K_S	8.494382 ±0.064237	99.39%	0.026960 ±0.000180	4.932 ±0.038
OLMo-2-1124-7B-Instruct-Q6_K	8.425234 ±0.063616	99.67%	0.013181 ±0.000105	3.474 ±0.032
OLMo-2-1124-7B-Instruct-Q8_0	8.416597 ±0.063592	99.74%	0.009659 ±0.000089	2.993 ±0.031
OLMo-2-1124-7B-Instruct-F16	8.368713 ±0.062985	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
OLMo-2-1124-7B-Instruct-IQ3_M	64.9333 ±1.7436	82.53	40.9333 ±1.7967	34.8000 ±1.7405	72.5333 ±1.6309	59.15
OLMo-2-1124-7B-Instruct-IQ3_S	65.2000 ±1.7405	82.66	41.3333 ±1.7993	33.6000 ±1.7259	71.7333 ±1.6453	58.91
OLMo-2-1124-7B-Instruct-IQ4_NL	67.0667 ±1.7172	83.33	41.7333 ±1.8018	37.3333 ±1.7674	74.4000 ±1.5947	60.77
OLMo-2-1124-7B-Instruct-Q3_K_L	64.5333 ±1.7481	81.47	40.9333 ±1.7967	33.3333 ±1.7225	72.5333 ±1.6309	58.56
OLMo-2-1124-7B-Instruct-Q3_K_M	63.4667 ±1.7594	81.86	41.3333 ±1.7993	33.6000 ±1.7259	73.2000 ±1.6184	58.69
OLMo-2-1124-7B-Instruct-Q3_K_S	64.9333 ±1.7436	81.60	40.4000 ±1.7930	33.2000 ±1.7207	71.8667 ±1.6430	58.40
OLMo-2-1124-7B-Instruct-Q4_K_M	66.5333 ±1.7242	83.87	42.0000 ±1.8034	36.9333 ±1.7635	71.4667 ±1.6500	60.16
OLMo-2-1124-7B-Instruct-Q4_K_M-bartowski	65.8667 ±1.7325	82.40	42.1333 ±1.8042	34.0000 ±1.7309	74.2667 ±1.5974	59.73
OLMo-2-1124-7B-Instruct-Q4_K_S	66.2667 ±1.7276	83.87	42.5333 ±1.8065	36.6667 ±1.7608	71.3333 ±1.6523	60.13
OLMo-2-1124-7B-Instruct-Q5_K_M	67.4667 ±1.7119	83.33	42.0000 ±1.8034	37.6000 ±1.7699	74.4000 ±1.5947	60.96
OLMo-2-1124-7B-Instruct-Q5_K_S	67.3333 ±1.7137	83.47	42.0000 ±1.8034	37.2000 ±1.7661	74.8000 ±1.5864	60.96
OLMo-2-1124-7B-Instruct-Q6_K	67.0667 ±1.7172	83.33	42.2667 ±1.8050	37.4667 ±1.7686	74.4000 ±1.5947	60.91
OLMo-2-1124-7B-Instruct-Q8_0	66.6667 ±1.7225	83.20	42.4000 ±1.8057	37.7333 ±1.7711	73.8667 ±1.6054	60.77
OLMo-2-1124-7B-Instruct-F16	67.3333 ±1.7137	83.20	41.8667 ±1.8026	37.8667 ±1.7724	72.6667 ±1.6284	60.59

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
OLMo-2-1124-7B-Instruct-Q4_K_M	3.73 GiB	7.30 B	Metal,BLAS	6	pp512	331.23 ± 0.55
OLMo-2-1124-7B-Instruct-Q4_K_M	3.73 GiB	7.30 B	Metal,BLAS	6	tg128	29.25 ± 0.19
OLMo-2-1124-7B-Instruct-Q4_K_M	3.73 GiB	7.30 B	Metal,BLAS	6	pp1024+tg1024	44.26 ± 0.13
OLMo-2-1124-7B-Instruct-Q4_K_M-bartowski	4.16 GiB	7.30 B	Metal,BLAS	6	pp512	345.11 ± 0.95
OLMo-2-1124-7B-Instruct-Q4_K_M-bartowski	4.16 GiB	7.30 B	Metal,BLAS	6	tg128	27.54 ± 0.15
OLMo-2-1124-7B-Instruct-Q4_K_M-bartowski	4.16 GiB	7.30 B	Metal,BLAS	6	pp1024+tg1024	42.76 ± 0.18

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 on Huggingface, and a really big Thank You! to Georgi Gerganov for his amazing work with llama.cpp and the ggml/gguf libraries.