Update README.md

README.md

@@ -20,24 +20,13 @@ Using <a href="https://github.com/ggml-org/llama.cpp/">llama.cpp</a> release <a
 
 Original model: https://huggingface.co/MiniMaxAI/MiniMax-M3
 
-All quants made using imatrix option with dataset from [here](https://gist.github.com/bartowski1182/82ae9b520227f57d79ba04add13d0d0d)
+**Model details:**
+- Parameter count: 427B
+- Input support: text, image (with mmproj file) - [details](#multimodal)
+- MTP: no
+- imatrix: yes - [details](#imatrix)
 
-Run them in your choice of tools:
-
-- [llama.cpp](https://github.com/ggml-org/llama.cpp)
-- [ramalama](https://github.com/containers/ramalama)
-- [LM Studio](https://lmstudio.ai/)
-- [koboldcpp](https://github.com/LostRuins/koboldcpp)
-- [Jan AI](https://www.jan.ai/)
-- [Text Generation Web UI](https://github.com/oobabooga/text-generation-webui)
-- [LoLLMs](https://github.com/ParisNeo/lollms)
-- [Atomic Chat](https://atomic.chat/)
-
-Note: if it's a newly supported model, you may need to wait for an update from the developers.
-
-**Important**:
-
-Make sure you use llama.cpp release [b10165](https://github.com/ggml-org/llama.cpp/releases/tag/b10165) or higher for fixed chat template parsing, PR for reference: [#26210](https://github.com/ggml-org/llama.cpp/pull/26210)
+[How to run](#how-to-run)
 
 ## Prompt format
 
@@ -56,7 +45,9 @@ Current thinking mode: adaptive. You are encouraged to think for complex decisio
 ]~b]ai
 ```
 
-## Download a file (not the whole branch) from below:
+**Don't know which to choose?** Grab [Q4_K_M](https://huggingface.co/bartowski/MiniMax-M3-GGUF/tree/main/MiniMax-M3-Q4_K_M) (261.28GB) - usually a good mix of size and performance. Download instructions available [here](#downloading-using-the-hugging-face-cli)
+
+## Available files:
 
 | Filename | Quant type | File Size | Split | Description |
 | -------- | ---------- | --------- | ----- | ----------- |
@@ -67,8 +58,8 @@ Current thinking mode: adaptive. You are encouraged to think for complex decisio
 | [MiniMax-M3-Q4_1.gguf](https://huggingface.co/bartowski/MiniMax-M3-GGUF/tree/main/MiniMax-M3-Q4_1) | Q4_1 | 268.89GB | true | Legacy format, similar performance to Q4_K_S but with improved tokens/watt on Apple silicon. |
 | [MiniMax-M3-Q4_K_M.gguf](https://huggingface.co/bartowski/MiniMax-M3-GGUF/tree/main/MiniMax-M3-Q4_K_M) | Q4_K_M | 261.28GB | true | Good quality, default size for most use cases, *recommended*. |
 | [MiniMax-M3-Q4_K_S.gguf](https://huggingface.co/bartowski/MiniMax-M3-GGUF/tree/main/MiniMax-M3-Q4_K_S) | Q4_K_S | 251.36GB | true | Slightly lower quality with more space savings, *recommended*. |
-| [MiniMax-M3-Q4_0.gguf](https://huggingface.co/bartowski/MiniMax-M3-GGUF/tree/main/MiniMax-M3-Q4_0) | Q4_0 | 243.64GB | true | Legacy format, offers online repacking for ARM and AVX CPU inference. |
-| [MiniMax-M3-IQ4_NL.gguf](https://huggingface.co/bartowski/MiniMax-M3-GGUF/tree/main/MiniMax-M3-IQ4_NL) | IQ4_NL | 242.75GB | true | Similar to IQ4_XS, but slightly larger. Offers online repacking for ARM CPU inference. |
+| [MiniMax-M3-Q4_0.gguf](https://huggingface.co/bartowski/MiniMax-M3-GGUF/tree/main/MiniMax-M3-Q4_0) | Q4_0 | 243.64GB | true | Legacy format, kept for compatibility with older tools. |
+| [MiniMax-M3-IQ4_NL.gguf](https://huggingface.co/bartowski/MiniMax-M3-GGUF/tree/main/MiniMax-M3-IQ4_NL) | IQ4_NL | 242.75GB | true | Similar to IQ4_XS, but slightly larger. |
 | [MiniMax-M3-IQ4_XS.gguf](https://huggingface.co/bartowski/MiniMax-M3-GGUF/tree/main/MiniMax-M3-IQ4_XS) | IQ4_XS | 229.69GB | true | Decent quality, smaller than Q4_K_S with similar performance, *recommended*. |
 | [MiniMax-M3-Q3_K_XL.gguf](https://huggingface.co/bartowski/MiniMax-M3-GGUF/tree/main/MiniMax-M3-Q3_K_XL) | Q3_K_XL | 206.05GB | true | Uses Q8_0 for embed and output weights. Lower quality but usable, good for low RAM availability. |
 | [MiniMax-M3-IQ3_M.gguf](https://huggingface.co/bartowski/MiniMax-M3-GGUF/tree/main/MiniMax-M3-IQ3_M) | IQ3_M | 205.52GB | true | Medium-low quality, new method with decent performance comparable to Q3_K_M. |
@@ -86,88 +77,66 @@ Current thinking mode: adaptive. You are encouraged to think for complex decisio
 | [MiniMax-M3-IQ1_M.gguf](https://huggingface.co/bartowski/MiniMax-M3-GGUF/tree/main/MiniMax-M3-IQ1_M) | IQ1_M | 100.74GB | true | Extremely low quality, *not* recommended. |
 | [MiniMax-M3-IQ1_S.gguf](https://huggingface.co/bartowski/MiniMax-M3-GGUF/tree/main/MiniMax-M3-IQ1_S) | IQ1_S | 90.53GB | true | Extremely low quality, *not* recommended. |
 
-## Embed/output weights
-
-Some of these quants (Q3_K_XL, Q4_K_L etc) are the standard quantization method with the embeddings and output weights quantized to Q8_0 instead of what they would normally default to.
-
-## Downloading using huggingface-cli
+## Downloading using the Hugging Face CLI
 
 <details>
   <summary>Click to view download instructions</summary>
 
-First, make sure you have huggingface-cli installed:
+First, make sure you have the Hugging Face CLI installed:
 
 ```
 pip install -U "huggingface_hub[cli]"
 ```
 
-Then, you can target the specific file you want:
+The files marked `true` in the Split column above are stored as multiple parts in a folder. To download all the parts to a local folder, run:
 
 ```
-huggingface-cli download bartowski/MiniMax-M3-GGUF --include "MiniMax-M3-Q4_K_M.gguf" --local-dir ./
+hf download bartowski/MiniMax-M3-GGUF --include "MiniMax-M3-Q8_0/*" --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:
+You can either specify a new local-dir (MiniMax-M3-Q8_0) or download them all in place (./)
+
+</details>
+
+## How to run
+
+These quants run with [llama.cpp](https://github.com/ggml-org/llama.cpp) - installable in one line via [llama.app](https://llama.app/):
 
 ```
-huggingface-cli download bartowski/MiniMax-M3-GGUF --include "MiniMax-M3-Q8_0/*" --local-dir ./
+curl -LsSf https://llama.app/install.sh | sh
+llama-server -hf bartowski/MiniMax-M3-GGUF:Q4_K_M
 ```
 
-You can either specify a new local-dir (MiniMax-M3-Q8_0) or download them all in place (./)
+llama-server includes a built-in chat web UI, served at http://localhost:8080 by default.
 
-</details>
+These quants were made with llama.cpp release b10141 - if this model's architecture is newly supported, you'll need that release or newer to run them.
 
-## ARM/AVX information
+They also work in: [LM Studio](https://lmstudio.ai/) · [koboldcpp](https://github.com/LostRuins/koboldcpp) · [ramalama](https://github.com/containers/ramalama) · [Jan AI](https://www.jan.ai/) · [Text Generation Web UI](https://github.com/oobabooga/text-generation-webui) · [LoLLMs](https://github.com/ParisNeo/lollms) · [Atomic Chat](https://atomic.chat/)
 
-Previously, you would download Q4_0_4_4/4_8/8_8, and these would have their weights interleaved in memory in order to improve performance on ARM and AVX machines by loading up more data in one pass.
+## Multimodal
 
-Now, however, there is something called "online repacking" for weights. details in [this PR](https://github.com/ggml-org/llama.cpp/pull/9921). If you use Q4_0 and your hardware would benefit from repacking weights, it will do it automatically on the fly.
+This model supports multimodal input. Alongside the quants, this repo includes the multimodal projector files [mmproj-MiniMax-M3-f16.gguf](https://huggingface.co/bartowski/MiniMax-M3-GGUF/blob/main/mmproj-MiniMax-M3-f16.gguf) and [mmproj-MiniMax-M3-bf16.gguf](https://huggingface.co/bartowski/MiniMax-M3-GGUF/blob/main/mmproj-MiniMax-M3-bf16.gguf), which pair with any quant above.
 
-As of llama.cpp build [b4282](https://github.com/ggml-org/llama.cpp/releases/tag/b4282) you will not be able to run the Q4_0_X_X files and will instead need to use Q4_0.
+llama.cpp downloads the mmproj automatically when using `-hf` as shown above; if you're loading files manually, pass it with `--mmproj`.
 
-Additionally, if you want to get slightly better quality, you can use IQ4_NL thanks to [this PR](https://github.com/ggml-org/llama.cpp/pull/10541) which will also repack the weights for ARM, though only the 4_4 for now. The loading time may be slower but it will result in an overall speed increase.
+## imatrix
 
-<details>
-  <summary>Click to view Q4_0_X_X information (deprecated)</summary>
+All quants made using imatrix option with dataset from [here](https://gist.github.com/bartowski1182/82ae9b520227f57d79ba04add13d0d0d). The imatrix is available here: [MiniMax-M3-imatrix.gguf](https://huggingface.co/bartowski/MiniMax-M3-GGUF/blob/main/MiniMax-M3-imatrix.gguf).
 
-I'm keeping this section to show the potential theoretical uplift in performance from using the Q4_0 with online repacking.
+## Embed/output weights
 
-<details>
-  <summary>Click to view benchmarks on an AVX2 system (EPYC7702)</summary>
-
-| model                          |       size |     params | backend    | threads |          test |                  t/s |  % (vs Q4_0)  |
-| ------------------------------ | ---------: | ---------: | ---------- | ------: | ------------: | -------------------: |-------------: |
-| qwen2 3B Q4_0                  |   1.70 GiB |     3.09 B | CPU        |      64 |         pp512 |        204.03 ± 1.03 |          100% |
-| qwen2 3B Q4_0                  |   1.70 GiB |     3.09 B | CPU        |      64 |        pp1024 |        282.92 ± 0.19 |          100% |
-| qwen2 3B Q4_0                  |   1.70 GiB |     3.09 B | CPU        |      64 |        pp2048 |        259.49 ± 0.44 |          100% |
-| qwen2 3B Q4_0                  |   1.70 GiB |     3.09 B | CPU        |      64 |         tg128 |         39.12 ± 0.27 |          100% |
-| qwen2 3B Q4_0                  |   1.70 GiB |     3.09 B | CPU        |      64 |         tg256 |         39.31 ± 0.69 |          100% |
-| qwen2 3B Q4_0                  |   1.70 GiB |     3.09 B | CPU        |      64 |         tg512 |         40.52 ± 0.03 |          100% |
-| qwen2 3B Q4_K_M                |   1.79 GiB |     3.09 B | CPU        |      64 |         pp512 |        301.02 ± 1.74 |          147% |
-| qwen2 3B Q4_K_M                |   1.79 GiB |     3.09 B | CPU        |      64 |        pp1024 |        287.23 ± 0.20 |          101% |
-| qwen2 3B Q4_K_M                |   1.79 GiB |     3.09 B | CPU        |      64 |        pp2048 |        262.77 ± 1.81 |          101% |
-| qwen2 3B Q4_K_M                |   1.79 GiB |     3.09 B | CPU        |      64 |         tg128 |         18.80 ± 0.99 |           48% |
-| qwen2 3B Q4_K_M                |   1.79 GiB |     3.09 B | CPU        |      64 |         tg256 |         24.46 ± 3.04 |           83% |
-| qwen2 3B Q4_K_M                |   1.79 GiB |     3.09 B | CPU        |      64 |         tg512 |         36.32 ± 3.59 |           90% |
-| qwen2 3B Q4_0_8_8              |   1.69 GiB |     3.09 B | CPU        |      64 |         pp512 |        271.71 ± 3.53 |          133% |
-| qwen2 3B Q4_0_8_8              |   1.69 GiB |     3.09 B | CPU        |      64 |        pp1024 |       279.86 ± 45.63 |          100% |
-| qwen2 3B Q4_0_8_8              |   1.69 GiB |     3.09 B | CPU        |      64 |        pp2048 |        320.77 ± 5.00 |          124% |
-| qwen2 3B Q4_0_8_8              |   1.69 GiB |     3.09 B | CPU        |      64 |         tg128 |         43.51 ± 0.05 |          111% |
-| qwen2 3B Q4_0_8_8              |   1.69 GiB |     3.09 B | CPU        |      64 |         tg256 |         43.35 ± 0.09 |          110% |
-| qwen2 3B Q4_0_8_8              |   1.69 GiB |     3.09 B | CPU        |      64 |         tg512 |         42.60 ± 0.31 |          105% |
-
-Q4_0_8_8 offers a nice bump to prompt processing and a small bump to text generation
+Some of these quants (Q3_K_XL, Q4_K_L etc) are the standard quantization method with the embeddings and output weights quantized to Q8_0 instead of what they would normally default to.
 
-</details>
+## ARM/AVX information
 
-</details>
+llama.cpp automatically "repacks" weights into an interleaved layout at load time for faster inference on ARM and AVX machines - details in [this PR](https://github.com/ggml-org/llama.cpp/pull/9921). This once required downloading special Q4_0_4_4/4_8/8_8 files; those are long gone. Online repacking now covers Q4_0, IQ4_NL, and most K-quants, so no special quant choice is needed for CPU inference.
 
 ## Which file should I choose?
 
 <details>
   <summary>Click here for details</summary>
 
-A great write up with charts showing various performances is provided by Artefact2 [here](https://gist.github.com/Artefact2/b5f810600771265fc1e39442288e8ec9)
+An older (early 2024) but still useful write-up with charts comparing quant performances is provided by Artefact2 [here](https://gist.github.com/Artefact2/b5f810600771265fc1e39442288e8ec9)
 
 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.
 
@@ -175,6 +144,8 @@ If you want your model running as FAST as possible, you'll want to fit the whole
 
 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.
 
+Hugging Face can also do this math for you: add your hardware in your [Local Apps settings](https://huggingface.co/settings/local-apps) and the model page will show which files fit.
+
 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.
@@ -196,3 +167,4 @@ Thank you kalomaze and Dampf for assistance in creating the imatrix calibration
 Thank you ZeroWw for the inspiration to experiment with embed/output.
 
 Want to support my work? Visit my ko-fi page here: https://ko-fi.com/bartowski
+