eaddario/Llama-xLAM-2-8b-fc-r-GGUF

Experimental layer-wise quantization of Salesforce/Llama-xLAM-2-8b-fc-r

Using LLaMA C++ release b5180 for quantization.

Original model: Salesforce/Llama-xLAM-2-8b-fc-r

From the original model creators:

Large Action Models (LAMs) are advanced language models designed to enhance decision-making by translating user intentions into executable actions. As the brains of AI agents, LAMs autonomously plan and execute tasks to achieve specific goals, making them invaluable for automating workflows across diverse domains. This model release is for research purposes only.

The new xLAM-2 series, built on our most advanced data synthesis, processing, and training pipelines, marks a significant leap in multi-turn conversation and tool usage. Trained using our novel APIGen-MT framework, which generates high-quality training data through simulated agent-human interactions. Our models achieve state-of-the-art performance on BFCL and τ-bench benchmarks, outperforming frontier models like GPT-4o and Claude 3.5. Notably, even our smaller models demonstrate superior capabilities in multi-turn scenarios while maintaining exceptional consistency across trials.

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'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 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	Naive	Repo	Shrinkage
Llama-xLAM-2-8b-fc-r-IQ3_M	3.78	3.69	2.4%
Llama-xLAM-2-8b-fc-r-IQ3_S	3.68	3.43	6.8%
Llama-xLAM-2-8b-fc-r-IQ4_NL	4.71	4.39	6.2%
Llama-xLAM-2-8b-fc-r-Q3_K_L	4.32	3.76	13.0%
Llama-xLAM-2-8b-fc-r-Q3_K_M	4.02	3.56	11.4%
Llama-xLAM-2-8b-fc-r-Q3_K_S	3.66	3.31	9.6%
Llama-xLAM-2-8b-fc-r-Q4_K_M	4.92	4.41	10.4%
Llama-xLAM-2-8b-fc-r-Q4_K_S	4.69	4.28	8.7%
Llama-xLAM-2-8b-fc-r-Q5_K_M	5.73	5.38	6.1%
Llama-xLAM-2-8b-fc-r-Q5_K_S	5.60	5.24	6.4%
Llama-xLAM-2-8b-fc-r-Q6_K	6.60	6.57	0.5%
Llama-xLAM-2-8b-fc-r-Q8_0	8.54	7.73	9.5%

Perplexity and KL Divergence scores

Model	μPPL	𝜌PPL	μKLD	RMS Δp
Llama-xLAM-2-8b-fc-r-IQ3_M	8.471225 ±0.059374	98.14%	0.096730 ±0.000436	9.339 ±0.048
Llama-xLAM-2-8b-fc-r-IQ3_S	8.675839 ±0.060418	97.37%	0.137925 ±0.000554	11.245 ±0.051
Llama-xLAM-2-8b-fc-r-IQ4_NL	8.337503 ±0.060156	99.09%	0.047455 ±0.000243	6.280 ±0.039
Llama-xLAM-2-8b-fc-r-Q3_K_L	8.894129 ±0.063417	97.22%	0.136754 ±0.000659	11.276 ±0.057
Llama-xLAM-2-8b-fc-r-Q3_K_M	8.991141 ±0.063906	96.89%	0.152094 ±0.000706	11.870 ±0.058
Llama-xLAM-2-8b-fc-r-Q3_K_S	9.352260 ±0.066573	95.91%	0.198689 ±0.000870	13.526 ±0.061
Llama-xLAM-2-8b-fc-r-Q4_K_M	8.230419 ±0.058263	99.18%	0.041808 ±0.000219	5.988 ±0.037
Llama-xLAM-2-8b-fc-r-Q4_K_M (naive)	8.072811 ±0.057158	99.60%	0.019868 ±0.000110	4.044 ±0.029
Llama-xLAM-2-8b-fc-r-Q4_K_S	8.239495 ±0.058176	99.10%	0.045691 ±0.000240	6.262 ±0.039
Llama-xLAM-2-8b-fc-r-Q5_K_M	8.062572 ±0.057549	99.77%	0.011576 ±0.000073	3.136 ±0.025
Llama-xLAM-2-8b-fc-r-Q5_K_S	8.057947 ±0.057474	99.75%	0.012330 ±0.000075	3.210 ±0.026
Llama-xLAM-2-8b-fc-r-Q6_K	7.983587 ±0.056711	99.91%	0.004239 ±0.000034	1.912 ±0.018
Llama-xLAM-2-8b-fc-r-Q8_0	7.982215 ±0.056796	99.94%	0.002365 ±0.000026	1.449 ±0.019
Llama-xLAM-2-8b-fc-r-F16	7.968796 ±0.056714	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
Llama-xLAM-2-8b-fc-r-IQ3_M	64.6667 ±1.7466	76.67	38.5333 ±1.7783	29.6000 ±1.6680	74.5333 ±1.5919	56.80
Llama-xLAM-2-8b-fc-r-IQ3_S	60.8000 ±1.7838	72.40	38.0000 ±1.7736	30.9333 ±1.6889	72.5333 ±1.6309	54.93
Llama-xLAM-2-8b-fc-r-IQ4_NL	66.0000 ±1.7309	77.73	39.0667 ±1.7827	30.8000 ±1.6869	73.7333 ±1.6080	57.47
Llama-xLAM-2-8b-fc-r-Q3_K_L	65.0667 ±1.7420	76.67	38.6667 ±1.7794	29.6000 ±1.6680	71.6000 ±1.6477	56.32
Llama-xLAM-2-8b-fc-r-Q3_K_M	64.4000 ±1.7496	76.93	37.4667 ±1.7686	30.0000 ±1.6744	71.8667 ±1.6430	56.13
Llama-xLAM-2-8b-fc-r-Q3_K_S	61.6000 ±1.7771	75.73	38.4000 ±1.7771	30.2667 ±1.6787	72.4000 ±1.6334	55.68
Llama-xLAM-2-8b-fc-r-Q4_K_M	65.6000 ±1.7358	77.33	39.0667 ±1.7827	30.4000 ±1.6807	73.8667 ±1.6054	57.25
Llama-xLAM-2-8b-fc-r-Q4_K_M (naive)	66.2667 ±1.7276	77.47	39.6000 ±1.7870	31.2000 ±1.6929	72.5333 ±1.6309	57.41
Llama-xLAM-2-8b-fc-r-Q4_K_S	65.7333 ±1.7342	77.07	39.3333 ±1.7849	30.6667 ±1.6849	73.6000 ±1.6106	57.28
Llama-xLAM-2-8b-fc-r-Q5_K_M	66.4000 ±1.7259	78.80	39.8667 ±1.7890	32.0000 ±1.7045	74.0000 ±1.6027	58.21
Llama-xLAM-2-8b-fc-r-Q5_K_S	65.7333 ±1.7342	78.93	39.4667 ±1.7860	32.0000 ±1.7045	74.1333 ±1.6001	58.05
Llama-xLAM-2-8b-fc-r-Q6_K	65.4667 ±1.7374	78.80	39.3333 ±1.7849	32.2667 ±1.7082	73.6000 ±1.6106	57.89
Llama-xLAM-2-8b-fc-r-Q8_0	66.0000 ±1.7309	78.93	39.2000 ±1.7838	32.0000 ±1.7045	74.4000 ±1.5947	58.11
Llama-xLAM-2-8b-fc-r-F16	65.7333 ±1.7342	78.40	40.0000 ±1.7900	31.3333 ±1.6949	74.1333 ±1.6001	57.92

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
Llama-xLAM-2-8b-fc-r-Q4_K_M	4.10 GiB	8.03 B	Metal,BLAS	6	pp512	314.18 ± 0.08
Llama-xLAM-2-8b-fc-r-Q4_K_M	4.10 GiB	8.03 B	Metal,BLAS	6	tg128	28.09 ± 0.03
Llama-xLAM-2-8b-fc-r-Q4_K_M	4.10 GiB	8.03 B	Metal,BLAS	6	pp1024+tg1024	44.57 ± 0.09
Llama-xLAM-2-8b-fc-r-Q4_K_M (naive)	4.58 GiB	8.03 B	Metal,BLAS	6	pp512	327.86 ± 0.57
Llama-xLAM-2-8b-fc-r-Q4_K_M (naive)	4.58 GiB	8.03 B	Metal,BLAS	6	tg128	26.15 ± 0.10
Llama-xLAM-2-8b-fc-r-Q4_K_M (naive)	4.58 GiB	8.03 B	Metal,BLAS	6	pp1024+tg1024	42.97 ± 0.10

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.