MikeRoz/MiniMax-M2.5-exl3

exllamav3 quantizations of MiniMaxAI/MiniMax-M2.5. Quantized using commit 89b841d of the dev branch.

Note that tensor parallelism is not currently supported for this architecture, so multi-GPU setups will have a harder time fitting this model than they would otherwise (you'll get more context out of 1x96 GB GPU than 4x24 GB GPUs).

Quant	Size	KLD	PPL	GPU Requirement Hint
2.00 bpw h6	61.054 GiB	0.42365	9.31452	3x24 GB w/ 49152 FP16 context
2.10 bpw h6 (optimized)	57.292 GiB	0.36355	9.20850	3x24GB w/ 40960 FP16 context
2.50 bpw h6 (optimized)	67.838 GiB	0.30152	8.88802	4x24GB w/ 90112 FP16 context
3.00 bpw h6	81.613 GiB	0.17263	8.58626	4x24GB w/ 16384 FP16 context
3.06 bpw h6 (optimized)	82.656 GiB	0.15648	8.66856	4x24GB w/ 12288 FP16 context
3.50 bpw h6 (optimized)	94.328 GiB	0.12513	8.58743	5x24 GB w/ 49152 FP16 context
4.00 bpw h6	108.087 GiB	0.07882	8.45404	6x24GB w/ 49152 FP16 context
5.00 bpw h6	134.561 GiB	-	-	5x24GB + 1x32GB w/ 24576 FP16 context (will not load for me with 6x24GB)

K/L-D and PPL graphs

Measurements for creating optimized quants

measurement.json - 2.0bpw_H6 vs 3.0bpw_H6
measurement.json - 3.0bpw_H6 vs 4.0bpw_H6
measurement.json - 4.0bpw_H6 vs 5.0bpw_H6

How to use these quants

The documentation for exllamav3 is your best bet here, as wall as that of TabbyAPI or Text Generation Web UI (oobabooga). In short:

• You need to have sufficient VRAM to fit the model and your context cache. I give some pointers above that may be helpful.
• At this point, your GPUs need to be nVidia. AMD/ROCm, Intel, and offloading to system RAM are not currently supported.
• You will need a software package capable of loading exllamav3 models. I'm still somewhat partial to oobabooga, but TabbyAPI is another popular option. Follow the documenation for your choice in order to get yourself set up.

How to create a quant

The documentation for exllamav3 is again the authoritative source. But for a short primer, click below to continue.

Expand for more details

Quantization happens a layer at a time, so you don't need nearly as much VRAM to quant as you do to load the whole model.

Not all architectures are supported by exllamav3. Check the documentation to ensure the model you want to quantize is supported.

To create a quant, you'll need to:

• Download your source model
• git clone exllamav3
• Set up a Python environment with all requirements from requirements.txt
• Run convert.py:

python convert.py -w [path/to/work_area] -i [path/to/source_model] -o [path/to/output_model] -b [bitrate] -hb [head bitrate]

Where:

• path/to/work_area is a folder where the script can save intermediate checkpoints as it works. If the process crashes, you can pass the --resume flag to pick up from where it left off.
• path/to/source_model folder containing the source model you downloaded
• path/to/output_model destination folder for your completed quant (will be created if it does not exist)
• bitrate The average number of bits to use for each weight. Needs to be a float (pass 4.0 if you want just 4 even).
• head bitrate Number of bits to use for attention head weights. 6 is usually most useful here. 8 is generally considered overkill, but may be useful in some situations.

How to create optimized quants

It's possible to produce quants that are better for a given size than the ones you get by performing a quant directly to a given target bitrate. The process involves comparing two quants, measuring which modules are more affected by the quantization process, and selecting those modules first when targeting some in-between bitrate.

Expand for more details

exllamav3 includes a measurement script util/measure.py that will compare two exllamav3 models module by module against the original model. The goal is to see which modules are the most affected by the decrease in precision involved in going from a larger quant to a smaller quant.

The command is:

python util/measure.py -l [level] -d [device] -ms [max_sys_memory] -i [path/to/quant1] [path/to/quant2] -r [path/to/original_model] -o [path/to/measurement.json]

Where:

• level is an integer between 0 and 3 that determines the resolution of the measurement. 0 is fastest but least granular, 2 is default, 3 is most granular and slowest.
• device is the index of the CUDA device that will perform the work
• max_sys_memory is the amount of memory that can be used for state data to speed things up, in GiB
• path/to/quant1 and path/to/quant2 are the paths to the two quants to compare
• path/to/original_model is the path to the original model
• path/to/measurement.json is the path to the resulting json measurement file

The masurement fie I created above compared my 2.0bpw_H6 and my 3.0bpw_H6 quants.

You can then feed this measurement file, along with the two quants, to util/optimize.py to create optimized quants that draw modules from both quants where appropriate to get the best result for a given bitrate.

The command is:

python util/optimize.py -i [path/to/quant1] [path/to/quant2] -o [path/to/resulting_model] -m [path/to/measurement.json] -b [target_bitrate]

Where:

• path/to/quant1 and path/to/quant2 are paths to the two source models
• path/to/resulting_model is the output path
• target_bitrate is the target bitrate as a number a decimal point

You can use a measurement script from one pair of quants with another pair of quants of the same model. When I tried to use 2.0bpw and 4.0bpw quants to create a 2.25bpw quant, the size of the resulting model was larger than requested because of the substitution at 2.48 bpw, but it was still an improvement over a straight 2.48bpw quant. An explicitly-requested 2.48bpw quant drawing from the 2.0bpw and 3.0bpw quants proved to be even better (in terms of k/l divergence). Finally, I tried creating a 3.25bpw quant from 3.0bpw and 4.0bpw quants, still using my 2.0-vs-3.0 measurement file. This was not as successful as the optimized 2.25bpw quant, and may have benefitted from a 'correct' measurement file that matched the two actual sources.

How to measure Perplexity and KL Divergence

Expand for details

Measuring KL/D is a process that involves comparing the outputs of the quantized model to outputs of the original model. If the original model is too large for your hardware to load without quantization, you can run a script to generate logits which can then be passed into the comparison script, sparing you the need to load the whole source model.

First, you'll need to create a dataset spec file. I based mine on eval/spec/wiki2_llama3_large.json.

{
    "tokenize_fn": "transformers",
    "tokenizer_dir": "path/to/full_model",
    "dataset": "wiki2",
    "eval_stride": 512,
    "eval_len": 2048,
    "max_rows": 100
}

I passed this into eval/compare_q_logits.py as follows:

python eval/compare_q_logits.py -m [path/to/full_model] -o [path/to/output_logits.safetensors] -d [path/to/dataset_spec.json] -rpb [rows_per_batch] -dev [device_index]

Where:

• path/to/full_model is the path to the model
• path/to/output_logits.safetensors is the path to the output logits file
• path/to/dataset_spec.json is the path to the dataset spec file described above
• rows_per_batch - I would run out of memory without this parameter. I set it to 32768.
• device_index - optional CUDA device index

Next, you'll need a model spec file that describes all the quants you want in the graph. You'll need to be able to load any model you'd like compared. Here's a sample of the one I used for these quants:

[
    {
        "load_fn": "exllamav3",
        "fwd_fn": "exllamav3",
        "label": "EXL3 2.0bpw H6",
        "model_dir": "path/to/MiniMaxAI_MiniMax-M2.5-2.0bpw-h6-exl3"
    },
    {
        "load_fn": "exllamav3",
        "fwd_fn": "exllamav3",
        "label": "EXL3 2.1bpw H6 (optimized)",
        "model_dir": "path/to/MiniMaxAI_MiniMax-M2.5-2.1bpw-h6-exl3"
    }
]

This spec file can be passed in to the following command:

python eval/compare_q.py -d [path/to/dataset_spec.json] -m [path/to/model_spec.json] -lf [path/to/logits.safetensors] -p [-kld] -t [chart_title]

Where:

• path/to/dataset_spec.json is the path to the dataset spec file described above
• path/to/model_spec.json is the path to the model spec file described above
• path/to/logits.safetensors is the path to the full model's logits, created above
• -kld the script creates a perplexity chart by default, add this if you want K/L-d instead
• chart_title the chart title in the resulting plot

Results are cached, so if the process crashes after processing one or more models, you just need to restart the script until every model has been tested (don't use the argument that clears the cache). Also note that if you're running this via SSH like me, you may not see anything - the script uses plt.show(). I hacked in an extra arg and a plt.savefig() call instead.