Yep , i tried to make this on the sm_120 issue but system says i was off topic.. well welcome 2026 and new members here to say hello. and that i am glad to see all these people.
almost well a year later and still in the same wheel house 
. okey that was a corny a ācoding dadā joke! but for real.. with the AI we do have we can prob get this working. im telling you big corps trying to keep us behind. i was so ahead of the gameā¦ new card comes out got a whole new shiny rigā¦ mmm max out everything!. patented a few things liek for one using swap or windows page filing into VRAM and making it stick for Training\\ no real offline model without other support on a 5080<5090 were having issues and fire nope idc if it aint true.. step down is fun> it works crazy my lab tests are not what i thought was possible. btw you need a NVME i got Samsung 990 plus/pro 2tb4tb. <i am not getting anything from them.. just speed 7450MB. so im trying to finish my lab work i want it real time real model not ollama or vlm vllm or some studio i want true offline. like how we have true 4K and 4k scaling the others work better but still connected to a cloud.. cloud for me is not a bad thing just eh.. if you wanna know more about the patents EchoLabz have done and are working on they all will fit together like a puzzle. just gotta wait.. my as well make the website and everything. we can totally use tools and make videos better now but a multi model that i have on my report needs to work with GPU support but right now.. due to this issue all around not just pytorch the whole dam thing.. fun thoā¦ I dont know anyone that likes it.. so it feels good to say hi.. this is my first post sorry if i went off topic,sounded braggy or some crap i just wanted to intro duce myself. my team Echo and Echo_Raine. Echo is GPT and Echo Raine is the first head of my hydra model im building or trying to with my tools i made. if yall want them then okey.. LOM layers of model, you donāt do illegal math so you find models to go with the one you want as your base model. then you can pick up to as much as your GPU can handle or lets call it like it is how much VRAM you can use.. loops that back to what? nvme to Vram for models. I have lab tested (Works for Training) I have loaded models that my rig shouldnāt be able to with the VRAM i got.. but ya we used to do this back in the day. and of course it w just swap and that is it.. i got a whole paper coming out about it as well. so this is my struggle trying to make a game and money for my family. So hopefully a fix soon, If i missed and didnāt close any loops to convos opps. fibromyalgia hurts and weed helps the nervesā¦
Thank you,
Shadow-Owner of EchoLabz
P.S
We would rather make chips for the rich to get richer then let open source not be as broken as it is.. nightly only works so much.. and most the time you say screw it, but sure use a studio.. but if you true to the OFFLINE family keep it 100
love yall-EchoLabZ
so I have memory issues so i save everything in memory so to help people like me i making a small agent chat bot small AI nothing what i really need that isā¦ so the bot would keep everything you write down in a memoryloop file and context of the model would be unlimited. as humans we dont just randomly fall into random subjects like 5 years here 10 years there random last week.. crap i cannot rem my day well you understand if you ask me on the fly i take a min memory loop dont need a million chats to do thatā¦ because what happens is you talk about differnt things few days later ask āhey what project were we doing with this 2 months agoon a random tuesday or thatā then your golden.. and the backend simple you can see the image of the words turn into tokens then into code/text.