Model : Add support for Kimi-K2

[gabriellarson]

I used the same set_vocab approach as the HunYuanMoE, and attempted to accurately represent the kimi_tokenization.py regex in unicode.cpp .

[anikifoss]

Thanks for the patch, I'm excited to try this model! Running convert_hf_to_gguf based on your branch, will report back with the results.

[gabriellarson]

Uploaded Q2_K and BF16 GGUF to HuggingFace (thanks @danielhanchen for the initial BF16 conversion!) https://huggingface.co/gabriellarson/Kimi-K2-Instruct-GGUF

[danielhanchen]

Nice work @gabriellarson ! I was also trying to get a PR going - master...unslothai:llama.cpp:master - but I was primarily stuck on the K2's special regex handling - will try yours out to see if the regex works!

[thad0ctor]

has anyone been able to get quantization to work, I got Q3 to successfully quantize but have failed with Q2, its really a pain with this model because the GGUF ended up being 2 TB when I converted it and I only have 384GB of RAM (only) and Q3 comes it at about 440GB so loading it is a nightmare

[ubergarm]

has anyone been able to get quantization to work, I got Q3 to successfully quantize but have failed with Q2, its really a pain with this model because the GGUF ended up being 2 TB when I converted it and I only have 384GB of RAM (only) and Q3 comes it at about 440GB so loading it is a nightmare

i believe @gabriellarson made a Q2_K if you scroll down to the bottom of their huggingface repo files and that is a link to a discussion where some folks are sharing information. going to try this PR shortly and with luck release an imatrix file (or two eventually...) 🤞

[gabriellarson]

@CISC I think this is ready for review now

[thad0ctor]

has anyone been able to get quantization to work, I got Q3 to successfully quantize but have failed with Q2, its really a pain with this model because the GGUF ended up being 2 TB when I converted it and I only have 384GB of RAM (only) and Q3 comes it at about 440GB so loading it is a nightmare

i believe @gabriellarson made a Q2_K if you scroll down to the bottom of their huggingface repo files and that is a link to a discussion where some folks are sharing information. going to try this PR shortly and with luck release an imatrix file (or two eventually...) 🤞

Thank you! The Q2 I made outputs complete gibberish - not sure if my own tweaks to the SRC caused this issue or if its inherent to the commit as a whole right now. I saw the recent commits so I will start over from a clean slate. I need to test some more to see if FA or KV quant was the issue.

Gibberish aside, I was able to get about 13-15 t/s on 4x RTX 5090s which is about par on what I get with Deepseek.

[csabakecskemeti]

I've just started converting -> testing (based on my own dequantized BF16).
Will report back (will take some time)

[anikifoss]

I haven't tested any edge-cases involving complex character handling in the patch, but it works for English:

• convert_hf_to_gguf worked
• quantized model loads with llama.cpp compiled from this branch and the output looks good (able to one-shot the spinning hexagon with 20 balls)

[ubergarm]

Thanks @anikifoss for the confirmation. I too have been able to fp8 to bf16 cast, then run convert_hf_to_gguf to get bf16 GGUF and use that to quantize a "pure" q8_0 which successfully inferenced with llama-server in a few short chats.

I have my methodology details and screenshots on the hf repo discussion and updating as I go along.

Running into issue now with imatrix dropping a lot of experts due to only 99.74% partial data, might need to look at #9400 (comment) to get a better imatrix here on mainline.

[RodriMora]

Gibberish aside, I was able to get about 13-15 t/s on 4x RTX 5090s which is about par on what I get with Deepseek.

@thad0ctor How many layers did you have to offload to vram and at what context size did you get that speed?

[usrlocalben]

I tested briefly with the provided Q2_K quant and observed a lot of repetition in the output. Whole paragraphs repeating verbatim 3-4 times under different but similar headings. Bottom Line, In Summary, etc. (temp=0.3 and temp=0.6). I'm building Q8 and will try again.

[danielhanchen]

@gabriellarson I can confirm the new regex seems to work well based on tokenization ID matches!

@ubergarm Yes you're correct on experts being zeros - I think I also found this to be the case.

I also made some 245GB, 281GB (IQ1_S) dynamic quants + Q2_K_XL, Q4_K_XL quants at https://huggingface.co/unsloth/Kimi-K2-Instruct-GGUF - it should work fine with this PR or using my fork https://github.com/unslothai/llama.cpp - guide to run them here: https://docs.unsloth.ai/basics/kimi-k2-how-to-run-locally#run-kimi-k2-tutorials

[thad0ctor]

Gibberish aside, I was able to get about 13-15 t/s on 4x RTX 5090s which is about par on what I get with Deepseek.

@thad0ctor How many layers did you have to offload to vram and at what context size did you get that speed?

I am pretty sure it was default context (2k) and 5 layers per card which left about 4-5 GB headroom on each card (Q_4 kv cache). I suspect that the llama.cpp splitting may be inefficiently mapping layers so I suspect that moe enhancements on ik_llama or -ot args may add some perfomance once this model gets flushed out a bit more

It's unfortunately the Kimi team didn't work with the community pre-release to get the ball rolling on compatability with common interface engines for those of us who aren't swimming in VRAM lol

[ubergarm]

@danielhanchen

Yes you're correct on experts being zeros - I think I also found this to be the case.

Okay thanks for confirming! I checked your hf repo but didn't see your Kimi-K2-Instruct-GGUF/imatrix_unsloth.dat there, perhaps you don't release them anymore it seems? Also given it is a .dat I assume it is missing data for a lot of exps.

I'm trying @compilade 's new imatrix.gguf which seems to go ahead and save even with partial data.

I also made some 245GB, 281GB (IQ1_S) dynamic quants + Q2_K_XL, Q4_K_XL quants

Oh hey we already discussed this, but looks like your scripts mistakenly named another quant TQ1_0 given it doesn't actually contain that quantization type and is conflating the rough BPW range with an actual ternary model only quantization type.

Its great there are a lot of options in the smaller size ranges these days, but just trying to keep the naming conventions accurate! Thanks and great job getting this beast of a model going!

[ubergarm]

@gabriellarson @CISC

I just got it working with this, my first attempt forgot to add_ass, but without it the model would give "empty replies". I've had some folks testing with good success now using this code which would go into llama-chat.cpp as seen in this similar PR: ikawrakow/ik_llama.cpp#612 tested with this model ubergarm/Kimi-K2-Instruct-GGUF IQ2_KL 345.687 GiB (2.892 BPW) Final estimate: PPL = 3.2741 +/- 0.01689 (upload complete in 30 minutes lol)

+    } else if (tmpl == LLM_CHAT_TEMPLATE_KIMI_K2) {
+        // moonshotai/Kimi-K2-Instruct
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << "<|im_system|>system<|im_middle|>" << message->content << "<|im_end|>";
+            } else if (role == "assistant") {
+                ss << "<|im_user|>user<|im_middle|>" << message->content << "<|im_end|>";
+            } else {
+                ss << "<|im_assistant|>assistant<|im_middle|>" << message->content << "<|im_end|>";
+            }
+        }
+        if (add_ass) {
+            ss << "<|im_assistant|>assistant<|im_middle|>";
+        }

[gabriellarson]

@ubergarm I'll add the if (add_ass){}.

i also include the tool role in mine, is the tool role not necessary?

[ubergarm]

@gabriellarson

i also include the tool role in mine, is the tool role not necessary?

I wasn't sure myself honestly, but yes yours does look correct to me given my understanding of the official template. Should be fine, but I don't have a quant to test it and honestly don't know how to use proper tool calling 😅

👈 chat template decoder script

output

$ python chat_template_tester.py moonshotai/Kimi-K2-Instruct
>> chat template <<
<|im_system|>system<|im_middle|>example system prompt<|im_end|><|im_user|>user<|im_middle|>example user turn 1<|im_end|><|im_assistant|>assistant<|im_middle|>example assistant turn 1<|im_end|><|im_user|>user<|im_middle|>example user turn 2<|im_end|><|im_assistant|>assistant<|im_middle|>example assistant turn 2<|im_end|><|im_system|>tool<|im_middle|>## Return of \nsome kind of tool call maybe<|im_end|><|im_assistant|>assistant<|im_middle|>

python script

$ cat chat_template_tester.py
# uv pip install transformers jinja2
# (and sometimes also sentencepiece torch statsmodels, looking at you ERNIE4.5)
from transformers import AutoTokenizer
import argparse

parser = argparse.ArgumentParser()
parser.add_argument("model", help="Name of Hugging Face LLM repo (org/model format)")
args = parser.parse_args()

tokenizer = AutoTokenizer.from_pretrained(args.model, trust_remote_code = True)

chat = [
    {"role": "system", "content": "example system prompt"},
    {"role": "user", "content": "example user turn 1"},
    {"role": "assistant", "content": "example assistant turn 1"},
    {"role": "user", "content": "example user turn 2"},
    {"role": "assistant", "content": "example assistant turn 2"},
    {"role": "tool", "content": "some kind of tool call maybe"},
]

print(">> chat template <<")
print(tokenizer.apply_chat_template(chat, add_generation_prompt=True, tokenize=False))
print(">> end of chat template <<")

[jukofyork]

I've managed to create a draft model, but unsure yet if I can actually train it:

https://huggingface.co/jukofyork/Kimi-K2-Instruct-DRAFT-0.6B-UNTRAINED

https://huggingface.co/jukofyork/Kimi-K2-Instruct-DRAFT-0.6B-UNTRAINED-GGUF

Even the untrained draft gives some improvements for highly "draftable" refactoring prompts:

prompt eval time =   55366.84 ms /  1832 tokens (   30.22 ms per token,    33.09 tokens per second)
       eval time =  241439.49 ms /  1618 tokens (  149.22 ms per token,     6.70 tokens per second)
      total time =  296806.34 ms /  3450 tokens

prompt eval time =   55209.16 ms /  1832 tokens (   30.14 ms per token,    33.18 tokens per second)
       eval time =  169682.33 ms /  1524 tokens (  111.34 ms per token,     8.98 tokens per second)
      total time =  224891.50 ms /  3356 tokens
draft acceptance rate = 0.59985 (  814 accepted /  1357 generated)

I'll try to see if I can fine-tune it, but it all depends on if I can get transformers to load it and/or figure out a similar TikToken <--> SentencePiece hack...

[jukofyork]

I managed to get them to train:

https://huggingface.co/jukofyork/Kimi-K2-Instruct-DRAFT-0.6B-v2.0
https://huggingface.co/jukofyork/Kimi-K2-Instruct-DRAFT-0.6B-v2.0-GGUF

[ngxson]

I made a small fix to the chat template, would appreciate if anyone can test it: #14852

[jukofyork]

There's a good chance we can improve Kimi K2 a little by limiting to 32k context and changing the YaRN scaling parameter, eg:

--override-kv deepseek2.context_length=int:32768 --override-kv deepseek2.rope.scaling.factor=float:8.0 --ctx_size 32768

The tech report seems to suggest that there was no actual long-context training beyond 32k:

unlike DeepSeek V3 which used the YaRN parameters during training:

It doesn't look like it works that great above 32k anyway:

so it's probably a good idea to avoid using the 128k YaRN parameters unless you really have to, and only then use the minimum you actually need >32k, eg:

--override-kv deepseek2.context_length=int:65536 --override-kv deepseek2.rope.scaling.factor=float:16.0 --ctx_size 65536

and so on.

[jukofyork]

I raised a question here: MoonshotAI/Kimi-K2#55

[jukofyork]

I raised a question here: MoonshotAI/Kimi-K2#55

They replied:

Starting from 4k -> 32k stage, we set yarn_scale_factor=32, original_max_position_embeddings=4096 and rope_theta=50000, i.e. we use 128k settings even in 32k stages.

This is the same as what deepseek-v3 did, so it's probably not a good idea to alter the YaRN scaling parameter.

[danielhanchen]

@jukofyork So during "mid-training" they essentially did long context extension it seems?

@ngxson Nice work :)