apresence/internlm2_5-7b-chat-GGUF_with-tool-fix

Fixed Tool Calls

This is a version of InternLM 2.5 7B Chat in GGUF format with a fix for tool call tokens missing from the model output.

Because the tool tokens ('<|plugin|>', '<|interpreter|>', '<|action_end|>', '<|action_start|>') were flagged as special tokens in the original model, llama.cpp was excluding them from the output. This was fixed by clearing the special flag on the tokens and then re-converting the model to GGUF.

A clone of the original transformers repo with this fix applied is available here.

Be sure you are using llama.cpp release b3368 or newer. This is the version that was used for quantization and the version I tested with.

BE ADVISED:

• The original transformers model uses dynamic RoPE scaling by default, with an option to use linear. In my testing, the model perplexity is bad with anything other than dynamic
• Llama.cpp uses linear by default, and has no dynamic option, thus tool calls are inconsistent and perplexity quickly degrades as context length increases
• Opened Issue 8405 to address
• Temporary work-around: Use --rope-scaling none

Please visit the original model card for licensing details, example code, etc.

Credits

• InternLM on HF for the original model posted here.
• InternLM on GitHub for the detailed chat format info.
• bartowski on HF for the README.md template and dataset used to make the imatrix quants.

Prompt format

<|im_start|>system
{system_prompt}<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant

Prompt format for tool calls (example)

<|im_start|>system
You are InternLM2-Chat, a harmless AI assistant<|im_end|>
<|im_start|>system name=<|plugin|>
[
    {
        "name": "get_current_weather",
        "description": "Get the current weather in a given location",
        "parameters": {
            "type": "object",
            "properties": {
                "location": {
                    "type": "string",
                    "description": "The city and state, e.g. San Francisco, CA",
                },
                "unit": {"type": "string"},
            },
            "required": ["location"],
        },
    }
]
<|im_end|>
<|im_start|>user
I want to know today's weather in Shanghai<|im_end|>
<|im_start|>assistant
Sure, I will search for the weather of Shanghai.<|action_start|><|plugin|>
{"name": "get_current_weather", "parameters": {"location": "Shanghai"}}<|action_end|><|im_end|>
<|im_start|>environment name=<|plugin|>
{"temperature": 22}<|im_end|>
<|im_start|>assistant
The weather in Shanghai is 22 celsius<|im_end|>

Testing tool calls with llama-cli

Command:

./llama-cli --predict 512 --gpu-layers 32 --temp 0.8 --top-p 0.8 --top-k 50 -r '<|im_end|>\n' -if --multiline-input --model internlm2_5-7b-chat-Q4_K_M.gguf

Copy & paste the following:

<|im_start|>system
You are a helpful assistant.<|im_end|>
<|im_start|>system name=<|plugin|>
[{"name": "generate_image", "description": "Generates an image based on the given text prompt", "parameters": {"type": "object", "properties": {"prompt": {"type": "string", "description": "The text prompt used to guide image generation"}}, "required": ["prompt"]}}]<|im_end|>
<|im_start|>user
Draw a picture of a kitten.<|im_end|>
<|im_start|>assistant
\

The model should respond along these lines:

I will call an image generation api to generate image<|action_start|><|plugin|>
{"name": "generate_image", "parameters": {"prompt": "A cute and playful kitten, sitting on a soft cushion, looking up with big curious eyes, bright and vibrant colors, cartoon style, high resolution, and a touch of whimsy."}}<|action_end|>

Simulate the tool return value by copying & pasting this:

<|im_start|>environment name=<|plugin|>
{"image_url": "http://127.0.0.1/data/123253252345.png"}<|im_end|>
<|im_start|>assistant
\

The model should respond along these lines:

Here is a link to the image: [link](http://127.0.0.1/data/123253252345.png).

Direct download links

Pick the quant(s) you want to download from the below list. Only one is required.

Notes:

• The sweet spot seems to be at Q4_K_M.
• In my testing, the perplexity is so bad at any quant lower than Q3_K_L that tool calls do not work.
• Experimental quants were made with --output-tensor-type f16 --token-embedding-type f16 per ZeroWw's suggestion, please provide any feedback on quality differences you spot.
• The imatrix quants were made with one of bartowski's datasets located here.

Filename	Quant type	File Size	Description
internlm2_5-7b-chat-f32.gguf	f32	29GB	Full quality, no quantization. Primarily useful as a source for creating quants.
internlm2_5-7b-chat-Q8_0_L.gguf	Q8_0_L	8.93GB	Experimental, uses f16 for embed and output weights. Please provide any feedback of differences. Extremely high quality, generally unneeded but max available quant.
internlm2_5-7b-chat-Q8_0.gguf	Q8_0	8.22GB	Extremely high quality, generally unneeded but max available quant.
internlm2_5-7b-chat-Q6_K_L.gguf	Q6_K_L	7.24GB	Experimental, uses f16 for embed and output weights. Please provide any feedback of differences. Very high quality, near perfect, recommended.
internlm2_5-7b-chat-Q6_K.gguf	Q6_K	6.35GB	Very high quality, near perfect, recommended.
internlm2_5-7b-chat-Q5_K_L.gguf	Q5_K_L	6.45GB	Experimental, uses f16 for embed and output weights. Please provide any feedback of differences. High quality, recommended.
internlm2_5-7b-chat-Q5_K_M.gguf	Q5_K_M	5.50GB	High quality, recommended.
internlm2_5-7b-chat-Q5_K_S.gguf	Q5_K_S	5.37GB	High quality, recommended.
internlm2_5-7b-chat-Q4_K_L.gguf	Q4_K_L	5.70GB	Experimental, uses f16 for embed and output weights. Please provide any feedback of differences. Good quality, uses about 4.83 bits per weight, recommended.
internlm2_5-7b-chat-Q4_K_M.gguf	Q4_K_M	4.71GB	Good quality, uses about 4.83 bits per weight, recommended.
internlm2_5-7b-chat-Q4_K_S.gguf	Q4_K_S	4.48GB	Slightly lower quality with more space savings, recommended.
internlm2_5-7b-chat-IQ4_XS.gguf	IQ4_XS	4.24GB	Decent quality, smaller than Q4_K_S with similar performance, recommended.
internlm2_5-7b-chat-Q3_K_XL.gguf	Q3_K_XL	5.17GB	Experimental, uses f16 for embed and output weights. Please provide any feedback of differences. Lower quality but usable, good for low RAM availability.
internlm2_5-7b-chat-Q3_K_L.gguf	Q3_K_L	4.13GB	Lower quality but usable, good for low RAM availability.
internlm2_5-7b-chat-Q3_K_M.gguf	Q3_K_M	3.83GB	Even lower quality.
internlm2_5-7b-chat-IQ3_M.gguf	IQ3_M	3.59GB	Medium-low quality, new method with decent performance comparable to Q3_K_M.
internlm2_5-7b-chat-Q3_K_S.gguf	Q3_K_S	3.47GB	Low quality, not recommended.
internlm2_5-7b-chat-IQ3_XS.gguf	IQ3_XS	3.33GB	Lower quality, new method with decent performance, slightly better than Q3_K_S.
internlm2_5-7b-chat-IQ3_XXS.gguf	IQ3_XXS	3.10GB	Lower quality, new method with decent performance, comparable to Q3 quants.
internlm2_5-7b-chat-Q2_K.gguf	Q2_K	3.00GB	Very low quality but surprisingly usable.
internlm2_5-7b-chat-IQ2_M.gguf	IQ2_M	2.77GB	Very low quality, uses SOTA techniques to also be surprisingly usable.
internlm2_5-7b-chat-IQ2_S.gguf	IQ2_S	2.58GB	Very low quality, uses SOTA techniques to be usable.
internlm2_5-7b-chat-IQ2_XS.gguf	IQ2_XS	2.45GB	Very low quality, uses SOTA techniques to be usable.

Downloading using huggingface-cli

First, make sure you have hugginface-cli installed:

pip install -U "huggingface_hub[cli]"

Then, you can target the specific file you want:

huggingface-cli download apresence/internlm2_5-7b-chat-GGUF_with-tool-fix --include "internlm2_5-7b-chat-Q4_K_M.gguf" --local-dir ./

If the model is bigger than 50GB, it will have been split into multiple files. In order to download them all to a local folder, run:

huggingface-cli download apresence/internlm2_5-7b-chat-GGUF_with-tool-fix --include "internlm2_5-7b-chat-Q8_0.gguf/*" --local-dir internlm2_5-7b-chat-Q8_0

You can either specify a new local-dir (internlm2_5-7b-chat-Q8_0) or download them all in place (./)

Which file should I choose?

A great write up with charts showing various performances is provided by Artefact2 here

The first thing to figure out is how big a model you can run. To do this, you'll need to figure out how much RAM and/or VRAM you have.

If you want your model running as FAST as possible, you'll want to fit the whole thing on your GPU's VRAM. Aim for a quant with a file size 1-2GB smaller than your GPU's total VRAM.

If you want the absolute maximum quality, add both your system RAM and your GPU's VRAM together, then similarly grab a quant with a file size 1-2GB Smaller than that total.

Next, you'll need to decide if you want to use an 'I-quant' or a 'K-quant'.

If you don't want to think too much, grab one of the K-quants. These are in format 'QX_K_X', like Q5_K_M.

If you want to get more into the weeds, you can check out this extremely useful feature chart:

llama.cpp feature matrix

But basically, if you're aiming for below Q4, and you're running cuBLAS (Nvidia) or rocBLAS (AMD), you should look towards the I-quants. These are in format IQX_X, like IQ3_M. These are newer and offer better performance for their size.

These I-quants can also be used on CPU and Apple Metal, but will be slower than their K-quant equivalent, so speed vs performance is a tradeoff you'll have to decide.

The I-quants are not compatible with Vulcan, which is also AMD, so if you have an AMD card double check if you're using the rocBLAS build or the Vulcan build. At the time of writing this, LM Studio has a preview with ROCm support, and other inference engines have specific builds for ROCm.

Want to support my work? Visit my ko-fi page here: https://ko-fi.com/apresence