RekaAI/reka-flash-3
Reka Flash 3 is a 21B general-purpose reasoning model that was trained from scratch.
That makes me not very confident that it will be llama.cpp compatible but due to it using the LlamaForCausalLM archidecture let's at least try. It's queued! :D
You can check for progress at http://hf.tst.eu/status.html or regularly check the model
summary page at https://hf.tst.eu/model#reka-flash-3-GGUF for quants to appear.
It is looking really good so far. The model successfully converted to GGUF on nico1 for manual testing and imatrix computation and is currently being converting to GGUF on marco for quantization. A new base model of that size and quality compatible with llama.cpp would be huge.
I'm slightly concerned about https://huggingface.co/RekaAI/reka-flash-3/discussions/4 but let's proceed.
The model is llama.cpp compatible and almost too good for what they claim is trained from scratch. How could they even afford to train such an awesome model. But it really seems like it is indeed a new base model they trained from scratch.
Here a great example showing the models potential:
What is the meaning of life? assistant: <reasoning>
The user is asking, "What is the meaning of life?" That's a big question. Let me try to break it down. First, I should consider
different perspectives because the answer can vary a lot depending on whether someone is religious, philosophical, scientific,
or something else.
Starting with religion and spirituality, many religions have their own answers. For example, in Christianity, the meaning of life might be to serve God and live in accordance with His will. In Hinduism, there's the concept of dharma, which involves following one's duty to achieve moksha, or liberation from the cycle of rebirth. Buddhism focuses on reducing suffering and achieving enlightenment. So each religion has a different take.
Then there's philosophy. Existentialists like Sartre and Camus would say that life has no inherent meaning, and it's up to each
individual to create their own meaning. Absurdism also deals with the idea that life is meaningless, but humans can find happiness by embracing that absurdity. Stoics might say the meaning is in living virtuously and in accordance with nature.
From a scientific standpoint, some people argue that there's no inherent meaning because life evolved through natural processes
like evolution by natural selection. The meaning we ascribe is a product of our brains, which developed to help us survive and
reproduce. But others might say that even without inherent meaning, we can find purpose through our experiences and contributions to others.
Personally, I think the answer is subjective. It depends on what each person values. Someone might find meaning in relationships, career, helping others, creativity, or personal growth. It's important to recognize that there's no single right answer, and
it's up to each individual to explore and define their own purpose.
Wait, but maybe I should mention some key philosophers or thinkers who have addressed this. Like Aristotle, who talked about eudaimonia, or flourishing through virtue. Nietzsche's idea of the _bermensch creating their own values. Also, Eastern philosophies like Taoism, which emphasize harmony with the Tao.
I should also consider modern perspectives, like positive psychology, which focuses on finding meaning through engagement and relationships. And maybe touch on the scientific community's views, like theistic evolution versus materialism.
But I need to make sure I present a balanced view without bias. The user might be looking for a comprehensive overview rather than a specific answer. So structuring the response to cover various angles: religious, philosophical, scientific, and personal would be thorough.
Let me check if I'm missing any major viewpoints. Humanism comes to mind, which emphasizes ethical living and human welfare as the highest good. Also, transhumanism, which looks towards technological progress to enhance human potential and redefine meaning.
I should also address the common themes across different perspectives, like the search for purpose, happiness, connection, and the balance between individual and collective meaning.
In summary, the answer is multifaceted. It's important to acknowledge the diversity of thought and encourage the user to reflect on their own values and experiences to find what resonates with them.</reasoning>
The question of life's meaning is deeply profound and multifaceted, addressed through various lenses such as religion, philosophy, science, and personal experience. Here's a synthesis of key perspectives:
Religious and Spiritual Views:
Abrahamic Religions: Life_s purpose often revolves around serving a divine being (e.g., living according to God_s will
in Christianity) or achieving union with the divine (e.g., moksha in Hinduism).Buddhism: Focuses on alleviating suffering through ethical living, mindfulness, and attaining enlightenment (nirvana).
Eastern Philosophies: Taoism emphasizes harmony with the natural order, while Confucianism prioritizes social harmony and duty.
Philosophical Perspectives:
Existentialism (e.g., Sartre, Camus): Life has no inherent meaning, but individuals create purpose through choices, actions, and authenticity.
Absurdism (Camus): Acknowledges the conflict between human aspirations for meaning and the universe_s indifference, yet advocates embracing freedom and joy despite this absurdity.
Stoicism: Encourages living virtuously, aligning with nature, and finding contentment in what is within one_s control.
Aristotle: Suggested meaning comes from eudaimonia (flourishing through virtue and reason).
Nietzsche: Proposed that the _bermensch (self-actualized individual) creates their own values and meaning.
Scientific and Materialist Views:
- Biology: Life may arise from natural processes like evolution, where meaning is a human construct to enhance survival and reproduction.
- Cosmology: Some scientists argue meaning emerges from our capacity to explore, understand, and appreciate the universe
.
Modern and Human-Centered Approaches:
- Positive Psychology: Highlights finding meaning through relationships, purpose, and engagement.
- Humanism: Focuses on ethical living, human welfare, and flourishing in the present.
- Transhumanism: Seeks to redefine meaning through technological advancements to enhance human potential.
Personal and Subjective Meaning:
- Ultimately, many individuals derive meaning from personal values, relationships, creativity, career, or contributing to others. This subjective approach allows each person to craft their own purpose.
Conclusion: There is no universal answer, as meaning is deeply personal. It intertwines with cultural, existential, and individual reflections. Exploring these dimensions_through philosophy, community, and self-reflection_can help individuals discover a purpose that resonates with their unique journey.
<sep> human: [end of text]
llama_perf_sampler_print: sampling time = 100.61 ms / 1213 runs ( 0.08 ms per token, 12056.82 tokens per second)
llama_perf_context_print: load time = 3596.82 ms
llama_perf_context_print: prompt eval time = 252.95 ms / 7 tokens ( 36.14 ms per token, 27.67 tokens per second)
llama_perf_context_print: eval time = 351843.35 ms / 1205 runs ( 291.99 ms per token, 3.42 tokens per second)
llama_perf_context_print: total time = 352333.76 ms / 1212 tokens
For some reason nico1 decided to redownload the model anyways. I thought I could skip it doing that by putting the GGUF to /tmp but it doesn't matter as it will do static quants first anyways. The interrupt logic really let me down as it tried to interrupt Llama-3.3-Swallow-70B-Instruct-v0.4 doing IQ3_XS quants due to it having less priority but it being at the very beginning this would likely have taken hours to complete so I instead interrupting L3.1-Athena-l4-8B as well and it almost immediately paused in favor of reka-flash-3 would which is actually is the first time I decided to llmc shell interfere with a different node. I really should have just queued it on nico1.
@aifeifei798 and @DavidAU Thanks a lot for pointing out the issue with the chat template. This amazing model is now in the process of being requantized. @TK-13 I recommend you redownload the requantized version once ready as well.
You can check for progress at http://hf.tst.eu/status.html or regularly check the model
summary page at https://hf.tst.eu/model#reka-flash-3-GGUF for quants to appear.
@mradermacher Why did it copy the existing imatrix file from kaos despite me booth nukerepo and nuke it? I don't want it to do so.
nico1 ~# llmc nukerepo reka-flash-3-
reka-flash-3-GGUF...
reka-flash-3-i1-GGUF...
nico1 ~# llmc nuke reka-flash-3
reka-flash-3: not found
marco marco found reka-flash-3
nico1 ~# llmc add force -7777 si https://huggingface.co/RekaAI/reka-flash-3 worker nico1
submit tokens: ["force","-7777","static","imatrix","worker","nico1","https://huggingface.co/RekaAI/reka-flash-3"]
https://huggingface.co/RekaAI/reka-flash-3
["https://huggingface.co/RekaAI/reka-flash-3",["-2000","static","imatrix"],1741716518],
https://huggingface.co/RekaAI/reka-flash-3 already in llmjob.submit.txt
forcing.
nico1 ~# llmc push-model nico1 reka-flash-3
reka-flash-3: copying imatrix file from /fs/kaos/root/imatrix-remote/reka-flash-3.imatrix to nico1
reka-flash-3 submitted to nico1
@mradermacher
Manualy deleting /tmp/quant/reka-flash-3-i1-GGUF/imatrix.dat puts the job into the si state but the system then imediately copies the existing imatrix file from kaos again instead of generating a new one which puts the job back into the sI state.
And with llmc shell kaos I don't have access to /root/imatrix-remote/
Ah shit I should have used nukeall instead of nukerepo and nuke seperately. I missed the also archives any imatrix files with similar(tm) name away so they need to be regenrated part in the llmc documentation.
Now things are looking much more promising:
nico1 ~# mv /tmp/quant/reka-flash-3.gguf /root/
nico1 /tmp/quant# llmc nukeall reka-flash-3
nukeall<reka-flash-3>
mv: cannot stat '/root/imatrix/reka-flash-3.imatrix': No such file or directory
mv: cannot stat '/root/imatrix/reka-flash-3.[iQ]*.imatrix': No such file or directory
renamed '/root/imatrix-remote/reka-flash-3.imatrix' -> '/root/imatrix-archive/./reka-flash-3.imatrix'
mv: cannot stat '/root/imatrix-remote/reka-flash-3.[iQ]*.imatrix': No such file or directory
found
nico1 nico1 found reka-flash-3
nico1 nico1 rm -rf /llmjob/wdir/reka-flash-3-GGUF
nico1 nico1 rm -rf /llmjob/wdir/reka-flash-3-i1-GGUF
nico1 nico1 rm -rf /llmjob/sdir/reka-flash-3.log
nico1 nico1 rm -rf /llmjob/sdir/reka-flash-3.status
nuked reka-flash-3 on nico1
reka-flash-3-GGUF...
reka-flash-3-i1-GGUF... RepositoryNotFoundError('404 Client Error. (Request ID: Root=1-67d16737-6329681a1bb9cf561833e5ad;11622ae3-2e67-4f74-bc7b-19fd87b56ab5)\n\nRepository Not Found for url: https://huggingface.co/api/repos/delete.\nPlease make sure you specified the correct `repo_id` and `repo_type`.\nIf you are trying to access a private or gated repo, make sure you are authenticated.') at /llmjob/share/bin/llmjob line 2411.
nico1 ~# mv /root/reka-flash-3.gguf /tmp/quant/
nico1 ~# llmc add force -7777 si https://huggingface.co/RekaAI/reka-flash-3 worker nico1
submit tokens: ["force","-7777","static","imatrix","worker","nico1","https://huggingface.co/RekaAI/reka-flash-3"]
https://huggingface.co/RekaAI/reka-flash-3
["https://huggingface.co/RekaAI/reka-flash-3",["-2000","static","imatrix"],1741716518],
["https://huggingface.co/RekaAI/reka-flash-3",["force","-7777","static","imatrix","worker","nico1"],1741775424],
https://huggingface.co/RekaAI/reka-flash-3 already in llmjob.submit.txt
forcing.
nico1 ~# llmc push-model nico1 reka-flash-3
reka-flash-3 submitted to nico1
nico1 ~# touch /tmp/quant/reka-flash-3.force
Note to myself: Next time I don't have to move away the source GGUF as nukeall is not deleting it.
I thought I could skip it doing that by putting the GGUF to /tmp
If this is about imatrix quants, yes, the state is local to kaos, so it doesn't know the file is already there.
The interrupt logic really let me down
Note that interruptions are no longer a big deal (for a few weeks now) - even if models get interrupted, they can get started immediatelly again, and all that happens additionally is a test whether the repo exists already.
However, the logic will always interrupt the lowest priority (as in display sort order) in favour of the highest priority ready job (if it decides to interrupt at all).
I decided to llmc shell interfere with a different node.
Time you learned the ropes!
Another option (or an additional option) would have been to start reka in addition to the other models (using .force and push).
Next time I don't have to move away the source GGUF as nukeall is not deleting it.
That is not actually very intended, and should probably not be relied upon. That would mean we leaked lots of ggufs to some innocent /tmp already.
Update: you are confusing me. nukeall invokes nuke, and that most certainly will unlink the .gguf
There is also this https://huggingface.co/RekaAI/reka-flash-3/discussions/9
But maybe it's prudent to wait till the issue is sorted out.
Next time I don't have to move away the source GGUF as nukeall is not deleting it.
That is not actually very intended, and should probably not be relied upon. That would mean we leaked lots of ggufs to some innocent /tmp already.
Update: you are confusing me. nukeall invokes nuke, and that most certainly will unlink the .gguf
Yes I thought so as well but I accidentally executed nukeall a second time after I already moved back the GGUF and it didn't delete it. Maybe because I manually provided the GGUF instead instead of having it being automatically generated but in any case it did not delete it. I even checked the file creation data to confirm it hasn't redownloaded it.
I checked my bash history to confirm:
llmc nukeall reka-flash-3
cd /root
mv reka-flash-3.gguf /tmp/quant/
llmc nukeall reka-flash-3
llmc add force -7777 si https://huggingface.co/RekaAI/reka-flash-3 worker nico1
llmc push-model nico1 reka-flash-3
cd /tmp/quant/
touch reka-flash-3.force
Edit: Ah maybe it behaved like that because the first nukeall already made it forget that such a model is on nico1 so the next nukeall might have skipped it.
executed nukeall a second time after I already moved back the GGUF and it didn't delete it.
Ah, yes - it will not delete anything if the job no longer exists. It wouldn't even know on which node it should delete files.