eaddario/Hammer2.1-7b-GGUF

Experimental layer-wise quantization of MadeAgents/Hammer2.1-7b

Using LLaMA C++ release b5170 for quantization.

Original model: MadeAgents/Hammer2.1-7b

From the original model creators:

Hammer refers to a series of lightweight Large Action Models. Currently, we are releasing Hammer 2.1 models (0.5B, 1.5B, 3B, and 7B) with strong function calling capability. These models are based on the Qwen 2.5 coder series and utilize function masking techniques and other advanced technologies. Hammer 2.1 series bring significant enhancements, while still maintaining the basic functionality of Hammer 2.0's Single-Turn interaction and further strengthening other capabilities.

The Hammer 2.1 models, fine-tuned from the Qwen 2.5 coder series, inherit Hammer 2.0's advantages and are enhanced as follows:

• Multi-Step Function Calling: The assistant can perform multiple internal function calls to handle a single user request, actively planning and gathering information to fulfill complex tasks.
• Multi-Turn Function Calling: Enables continuous and context-aware interactions over multiple exchanges, with each turn potentially containing multiple steps, for a more natural conversation experience.
• Enhanced Irrelevant Information Inspection: Better at identifying when provided functions are irrelevant to a user query, by providing a non-function call response.

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'd normally use models produced by Unsloth (Daniel and Michael Han do some really advanced level stuff!) and Bartowski (see credits below), but they don't provide GGUF versions of this model, so 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 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	Naive	Repo	Shrinkage
Hammer2.1-7b-IQ3_M	3.57	3.48	2.5%
Hammer2.1-7b-IQ3_S	3.50	3.25	7.1%
Hammer2.1-7b-IQ4_NL	4.44	4.17	6.1%
Hammer2.1-7b-Q3_K_L	4.09	3.54	13.4%
Hammer2.1-7b-Q3_K_M	3.81	3.37	11.5%
Hammer2.1-7b-Q3_K_S	3.49	3.14	10.0%
Hammer2.1-7b-Q4_K_M	4.68	4.18	10.7%
Hammer2.1-7b-Q4_K_S	4.46	4.06	9.0%
Hammer2.1-7b-Q5_K_M	5.44	5.09	6.4%
Hammer2.1-7b-Q5_K_S	5.31	4.96	6.6%
Hammer2.1-7b-Q6_K	6.25	6.23	0.3%
Hammer2.1-7b-Q8_0	8.10	7.27	10.2%

Perplexity and KL Divergence scores

Model	μPPL	𝜌PPL	μKLD	RMS Δp
Hammer2.1-7b-IQ3_M	9.827809 ±0.071236	98.97%	0.059970 ±0.000226	6.574 ±0.032
Hammer2.1-7b-IQ3_S	9.942560 ±0.071243	98.64%	0.079935 ±0.000289	7.445 ±0.038
Hammer2.1-7b-IQ4_NL	9.585928 ±0.068060	99.51%	0.026627 ±0.000108	4.346 ±0.026
Hammer2.1-7b-Q3_K_L	10.010842 ±0.071669	98.55%	0.080230 ±0.000317	7.556 ±0.040
Hammer2.1-7b-Q3_K_M	10.109660 ±0.072549	98.44%	0.086444 ±0.000342	7.845 ±0.041
Hammer2.1-7b-Q3_K_S	10.224307 ±0.073520	98.16%	0.101951 ±0.000397	8.487 ±0.044
Hammer2.1-7b-Q4_K_M	9.513981 ±0.067589	99.56%	0.023792 ±0.000107	4.057 ±0.026
Hammer2.1-7b-Q4_K_M (naive)	9.469517 ±0.067164	99.70%	0.016770 ±0.000081	3.364 ±0.022
Hammer2.1-7b-Q4_K_S	9.529467 ±0.067693	99.53%	0.025429 ±0.000116	4.191 ±0.026
Hammer2.1-7b-Q5_K_M	9.411472 ±0.066924	99.88%	0.005881 ±0.000035	2.028 ±0.015
Hammer2.1-7b-Q5_K_S	9.410802 ±0.066883	99.88%	0.006252 ±0.000036	2.085 ±0.015
Hammer2.1-7b-Q6_K	9.384561 ±0.066615	99.96%	0.001779 ±0.000018	1.101 ±0.012
Hammer2.1-7b-Q8_0	9.379380 ±0.066571	99.98%	0.000692 ±0.000012	0.706 ±0.010
Hammer2.1-7b-F16	9.366577 ±0.066397	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
Hammer2.1-7b-IQ3_M	56.5333 ±1.8113	71.46	38.0000 ±1.7736	29.8667 ±1.6723	70.6667 ±1.6636	53.31
Hammer2.1-7b-IQ3_S	55.8667 ±1.8143	70.26	38.6667 ±1.7794	30.9333 ±1.6889	70.4000 ±1.6680	53.23
Hammer2.1-7b-IQ4_NL	57.3333 ±1.8072	71.20	38.5333 ±1.7783	30.1333 ±1.6766	69.8667 ±1.6766	53.41
Hammer2.1-7b-Q3_K_L	57.2000 ±1.8079	69.86	36.6667 ±1.7608	29.6000 ±1.6680	70.9333 ±1.6591	52.85
Hammer2.1-7b-Q3_K_M	55.7333 ±1.8149	69.87	37.2000 ±1.7661	29.8667 ±1.6723	70.4000 ±1.6680	52.61
Hammer2.1-7b-Q3_K_S	56.5333 ±1.8113	68.80	37.3333 ±1.7674	29.0667 ±1.6591	69.3333 ±1.6849	52.21
Hammer2.1-7b-Q4_K_M	57.8667 ±1.8042	71.20	38.0000 ±1.7736	29.8667 ±1.6723	69.8667 ±1.6766	53.36
Hammer2.1-7b-Q4_K_M (naive)	57.8313 ±1.8080	73.47	35.0667 ±1.7436	33.9683 ±2.6727	70.8000 ±1.6614	54.23
Hammer2.1-7b-Q4_K_S	57.8667 ±1.8042	71.07	37.8667 ±1.7724	29.6000 ±1.6680	69.6000 ±1.6807	53.20
Hammer2.1-7b-Q5_K_M	58.0000 ±1.8034	71.33	38.6667 ±1.7794	30.2667 ±1.6787	70.6667 ±1.6636	53.79
Hammer2.1-7b-Q5_K_S	58.4000 ±1.8010	71.47	38.8000 ±1.7805	30.5333 ±1.6828	70.6667 ±1.6636	53.97
Hammer2.1-7b-Q6_K	58.8000 ±1.7984	72.00	38.9333 ±1.7816	29.7333 ±1.6702	70.4000 ±1.6680	53.97
Hammer2.1-7b-Q8_0	58.5333 ±1.8002	72.00	38.9333 ±1.7816	30.1333 ±1.6766	70.2667 ±1.6702	53.97
Hammer2.1-7b-F16	58.9333 ±1.7976	72.13	39.0667 ±1.7827	30.4000 ±1.6807	70.5333 ±1.6658	54.21

Tokens per Second - Benchmarks

Scores generated using llama-bench. Naive Q4_K_M quantization included for comparison.

model	size	params	backend	threads	test	t/s
Hammer2.1-7b-Q4_K_M	3.89 GiB	7.61 B	Metal,BLAS	6	pp512	336.11 ± 0.60
Hammer2.1-7b-Q4_K_M	3.89 GiB	7.61 B	Metal,BLAS	6	tg128	29.32 ± 0.14
Hammer2.1-7b-Q4_K_M	3.89 GiB	7.61 B	Metal,BLAS	6	pp1024+tg1024	48.00 ± 0.16
Hammer2.1-7b-Q4_K_M (naive)	4.35 GiB	7.61 B	Metal,BLAS	6	pp512	355.08 ± 0.19
Hammer2.1-7b-Q4_K_M (naive)	4.35 GiB	7.61 B	Metal,BLAS	6	tg128	28.21 ± 0.00
Hammer2.1-7b-Q4_K_M (naive)	4.35 GiB	7.61 B	Metal,BLAS	6	pp1024+tg1024	45.92 ± 1.18

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