I built my 2x3090 with parts from eBay… MB (x299 Giga), i9 CPU, 64G RAM and 2 3090s… I did spring for a new, heavy duty PS and case with big fans.
All in, I spent about $2k.
System runs 70B models like Llama-2-70B-Orca-200k just fine at 11 T/s…
I feel like there’s not a ton of downside - I think the 3090s will be valuable for a while yet, and that’s over half the value of the system.
Having the hardware right here means I can have thing running all the time - when I read about a new model, I can download, play, etc. in minutes. Spinning up a runpod feels frustratingly slow to me. I went that route for a while, but found that the friction involved meant I tried fewer things. Having a system that might be slower, but is always available just works for my way of working.
So no “regerts” here.
I’m not quite understanding something - when I run the example colab code linked, it actually loads a different model (model_name = “codellama/CodeLlama-34b-hf”) which I can load fine.
When I search for THIS model (defog/sqlcoder-34b-hf) and find it on HF, I tried to load it locally on my 2x3090 system and I get out of memory - same error with 8-bit and 4-bit set per the Notebook.
So… which is the correct model? The one mentioned here in the post or the one in the Colab demo?
Assuming the HF repository for defog/sqlcoder-34b-alpha is the correct one… I have run into the following problem:
Using the colab notebook, but changing to the correct model mentioned in this post…
Download of model is fine, loads to GPUs and shows about 1/2 of my memory is used… (with the 4-bit option set)…
However, when I run the model.generate I get out-of-memory error…
Code Cell:
inputs = tokenizer(prompt, return_tensors=“pt”).to(“cuda”)
generated_ids = model.generate(
**inputs,
num_return_sequences=1,
eos_token_id=eos_token_id,
pad_token_id=eos_token_id,
max_new_tokens=400,
do_sample=False,
num_beams=1,
)
OutOfMemoryError: CUDA out of memory. Tried to allocate 14.41 GiB. GPU 0 has a total capacty of 23.69 GiB of which 6.11 GiB is free. Process 3797 has 306.20 MiB memory in use. Including non-PyTorch memory, this process has 17.26 GiB memory in use. Of the allocated memory 16.02 GiB is allocated by PyTorch, and 958.46 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
From the model card.
"Hardware Requirements
SQLCoder-34B has been tested on a 4xA10 GPU with float16
weights. You can also load an 8-bit and 4-bit quantized version of the model on consumer GPUs with 20GB or more of memory – like RTX 4090, RTX 3090, and Apple M2 Pro, M2 Max, or M2 Ultra Chips with 20GB or more of memory."
Any thoughts appreciated.