Or merge the bottom 1/8 (or whatever) experts together and (optionally) do some minimal training with all other weights frozen. Would need to modify the MoE routers slightly to map old -> new expert indices so you don't…
The general rule of thumb when assessing MoE <-> Dense model intelligence is SQRT(Total_Params*Active_Params). For Deepseek, you end up with ~158B params. The economics of batch inferencing a ~158B model at scale are…
There's still a lot of opportunity for software optimizations here. Trouble is that really only two classes of systems get optimizations for Deepseek, namely 1 small GPU + a lot of RAM (ktransformers) and the system…
I've been using PyCharm for the debugger (and everything else) and VSCode + RooCode + Local LLM lately. I've heard decent things about the Windsurf extension in PyCharm, but not being able to use a local LLM is an…
MoE inference wouldn't be terrible. That being said, there's not a good MoE model in the 70-160B range as far as I'm aware.
If you want to split tensorwise yes. Layerwise splits could go over Ethernet. I would be interested to see how feasible hybrid approaches would be, e.g. connect each pair up directly via ConnectX and then connect the…
About $12k when Project Digits comes out.
Just to add onto this point, you expect different experts to be activated for every token, so not having all of the weights in fast memory can still be quite slow as you need to load/unload memory every token.
It's not really possible to say what's "best" because the criteria is super subjective. I personally like the Spline family, and I default to Spline36 for both upscaling and downscaling in ffmpeg. Most people can't tell…
It would not be that slow as it is an MoE model with 37b activated parameters. Still, 8x3090 gives you ~2.25 bits per weight, which is not a healthy quantization. Doing bifurcation to get up to 16x3090 would be…
It will be slower for a 70b model since Deepseek is an MoE that only activates 37b at a time. That's what makes CPU inference remotely feasible here.
An actual hardcore technical AI "psychology" program would actually be really cool. Could be a good onboarding for prompt engineering (if it still exists in 5 years).
I definitely agree with you in the interim regarding junior developers. However, I do think we will eventually have the AI coding equivalent of CICD built into perhaps our IDE. Basically, when an AI generated some code…
Check out their project digits announcement, 128GB unified memory with infiniband capabilities for $3k. For more of the fast VRAM you would be in Quadro territory.
I suspect the big AI companies try to adversarially train that out as it could be used to "jailbreak" their AI. I wonder though, what would be considered a meaningful punishment/reward to an AI agent? More/less training…
I wonder if you would want to use an earlier layer as opposed to the penultimate layer, I would imagine that the LLM uses that layer to "prepare" for the final dimensionality reduction to clean the signal such that it…
I kinda wish you could just take a course on a specific distribution. Like, here's the Poisson class where you learn all of its interesting properties and apply it to e.g. queuing problems.
If you are comfortable with purchasing used hardware, used 3090 are great value, they can be had for roughly a third of the price of a new 4090. How many GPUs you need is completely dependent on the size of your team,…
I would probably refer to category 1 as "Open Architecture". I wouldn't want to give anyone the false impression that category 1 is comparable in the slightest to Open Weights, which is vastly more useful.
You could always split one of the experts up across multiple GPUs. I tend to agree with your sentiment, I think researchers in this space tend to not optimize that well for inference deployment scenarios. To be fair,…
Generally speaking this works well, pending your definition of node and the interconnect between them. If by node you mean GPU, and you have multiple of them on the same system (interconnect is PCIE, doesn't need to be…
You would need to fit the 389B parameters in VRAM to have a speed that is usable. Different experts are activated on a per token basis, so you would need to load/unload a large chunk of the 52B active parameters every…
Perhaps I'm being charitable but I read OP's comment in the light of what you described with context length. Batching, context length, and attention implementation vary these numbers wildly. I can fit a 6bit quant…
CogVideoX seems to be the best offline model so far
Does anyone know of a CoT dataset somewhere for finetuning? I would think exposing it to that type of modality during a finetune/lora would help.
Or merge the bottom 1/8 (or whatever) experts together and (optionally) do some minimal training with all other weights frozen. Would need to modify the MoE routers slightly to map old -> new expert indices so you don't…
The general rule of thumb when assessing MoE <-> Dense model intelligence is SQRT(Total_Params*Active_Params). For Deepseek, you end up with ~158B params. The economics of batch inferencing a ~158B model at scale are…
There's still a lot of opportunity for software optimizations here. Trouble is that really only two classes of systems get optimizations for Deepseek, namely 1 small GPU + a lot of RAM (ktransformers) and the system…
I've been using PyCharm for the debugger (and everything else) and VSCode + RooCode + Local LLM lately. I've heard decent things about the Windsurf extension in PyCharm, but not being able to use a local LLM is an…
MoE inference wouldn't be terrible. That being said, there's not a good MoE model in the 70-160B range as far as I'm aware.
If you want to split tensorwise yes. Layerwise splits could go over Ethernet. I would be interested to see how feasible hybrid approaches would be, e.g. connect each pair up directly via ConnectX and then connect the…
About $12k when Project Digits comes out.
Just to add onto this point, you expect different experts to be activated for every token, so not having all of the weights in fast memory can still be quite slow as you need to load/unload memory every token.
It's not really possible to say what's "best" because the criteria is super subjective. I personally like the Spline family, and I default to Spline36 for both upscaling and downscaling in ffmpeg. Most people can't tell…
It would not be that slow as it is an MoE model with 37b activated parameters. Still, 8x3090 gives you ~2.25 bits per weight, which is not a healthy quantization. Doing bifurcation to get up to 16x3090 would be…
It will be slower for a 70b model since Deepseek is an MoE that only activates 37b at a time. That's what makes CPU inference remotely feasible here.
An actual hardcore technical AI "psychology" program would actually be really cool. Could be a good onboarding for prompt engineering (if it still exists in 5 years).
I definitely agree with you in the interim regarding junior developers. However, I do think we will eventually have the AI coding equivalent of CICD built into perhaps our IDE. Basically, when an AI generated some code…
Check out their project digits announcement, 128GB unified memory with infiniband capabilities for $3k. For more of the fast VRAM you would be in Quadro territory.
I suspect the big AI companies try to adversarially train that out as it could be used to "jailbreak" their AI. I wonder though, what would be considered a meaningful punishment/reward to an AI agent? More/less training…
I wonder if you would want to use an earlier layer as opposed to the penultimate layer, I would imagine that the LLM uses that layer to "prepare" for the final dimensionality reduction to clean the signal such that it…
I kinda wish you could just take a course on a specific distribution. Like, here's the Poisson class where you learn all of its interesting properties and apply it to e.g. queuing problems.
If you are comfortable with purchasing used hardware, used 3090 are great value, they can be had for roughly a third of the price of a new 4090. How many GPUs you need is completely dependent on the size of your team,…
I would probably refer to category 1 as "Open Architecture". I wouldn't want to give anyone the false impression that category 1 is comparable in the slightest to Open Weights, which is vastly more useful.
You could always split one of the experts up across multiple GPUs. I tend to agree with your sentiment, I think researchers in this space tend to not optimize that well for inference deployment scenarios. To be fair,…
Generally speaking this works well, pending your definition of node and the interconnect between them. If by node you mean GPU, and you have multiple of them on the same system (interconnect is PCIE, doesn't need to be…
You would need to fit the 389B parameters in VRAM to have a speed that is usable. Different experts are activated on a per token basis, so you would need to load/unload a large chunk of the 52B active parameters every…
Perhaps I'm being charitable but I read OP's comment in the light of what you described with context length. Batching, context length, and attention implementation vary these numbers wildly. I can fit a 6bit quant…
CogVideoX seems to be the best offline model so far
Does anyone know of a CoT dataset somewhere for finetuning? I would think exposing it to that type of modality during a finetune/lora would help.