Frequently Asked Questions

[Titus-von-Koeller]

Please submit questions here that will likely affect a lot of other users and that haven't been sufficiently covered in the documentation.

[TezRomacH]

How to implement/add a new optimizer? Because I would really love to have AdaBelief8bit 😄

[Titus-von-Koeller]

Good question! I'll check-in with Tim when I get the chance and try to get back to you about it soon. We have a lot going on right now, so can't make promises, but put it on my list. Thanks for bringing this up and being interested to contribute 🤗

[Titus-von-Koeller]

Tim gave the following feedback on this question:

If you lock at the kernels.cu file, there is a swtich statement for Adam/Lion etc. You can use that switch statement to define a new optimizer. You can search for an old PR on the lion optimizer that was implemented by another contributor. They can follow the same implementation details to integrate their optimizer

[Titus-von-Koeller]

Would you be interested to work on a PR? We could help you out as you go along.

[TezRomacH]

Unfortunately I cannot commit on that, because I know very little about writing on cuda :(

[TezRomacH]

But I'll try my best

[Titus-von-Koeller]

In case you're interested in that rabbit hole, there's a nice Discord community for that with a Github repo (find the Discord invite link at the top of that page). The lectures are on Youtube, I specifically recommend lecture 3 as an intro. On of the newer lectures is even on "optimizing optimizers" I see now :)

[GM-git-dotcom]

Hi, I was wondering if bnb.nn.Embedding defined here quantizes Embedding layers to 4bits or 8bits? We are currently working on a PR at HF PEFT and this information would be very useful. Thank you!
https://github.com/TimDettmers/bitsandbytes/blob/cfd6ac75eff48e8c06b03cd8e721302a713c77a8/bitsandbytes/nn/modules.py#L133

[GM-git-dotcom]

Update: I have determined that it does not quantize and I made an issue for it. Thank you anyway!

[Titus-von-Koeller]

Yes, the embedding is not quantized usually as the errors would be too large and its not very efficient to quantize it. So lots of pains with little gain.

[solomonmanuelraj]

Hi Team,

when i am running the above qlora code for owl-vit model (google/owlvit-base-patch32) with below 4 bits bnbconfig , the fine tuning is taking place without any error.

bnb_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_use_double_quant=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_compute_dtype=torch.bfloat16
)

once i change the config with below information

bnb_config = BitsAndBytesConfig(
load_in_8bit=True
)

i receive the following error trace.

#########################################################################################
RuntimeError Traceback (most recent call last)
Cell In[25], line 1
----> 1 trainer.train()

File ~/miniconda3/envs/testenv/lib/python3.10/site-packages/transformers/trainer.py:1537, in Trainer.train(self, resume_from_checkpoint, trial, ignore_keys_for_eval, **kwargs)
1535 hf_hub_utils.enable_progress_bars()
1536 else:
-> 1537 return inner_training_loop(
1538 args=args,
1539 resume_from_checkpoint=resume_from_checkpoint,
1540 trial=trial,
1541 ignore_keys_for_eval=ignore_keys_for_eval,
1542 )

File ~/miniconda3/envs/testenv/lib/python3.10/site-packages/transformers/trainer.py:1854, in Trainer._inner_training_loop(self, batch_size, args, resume_from_checkpoint, trial, ignore_keys_for_eval)
1851 self.control = self.callback_handler.on_step_begin(args, self.state, self.control)
1853 with self.accelerator.accumulate(model):
-> 1854 tr_loss_step = self.training_step(model, inputs)
1856 if (
1857 args.logging_nan_inf_filter
1858 and not is_torch_tpu_available()
1859 and (torch.isnan(tr_loss_step) or torch.isinf(tr_loss_step))
1860 ):
1861 # if loss is nan or inf simply add the average of previous logged losses
1862 tr_loss += tr_loss / (1 + self.state.global_step - self._globalstep_last_logged)

File ~/miniconda3/envs/testenv/lib/python3.10/site-packages/transformers/trainer.py:2744, in Trainer.training_step(self, model, inputs)
2742 scaled_loss.backward()
2743 else:
-> 2744 self.accelerator.backward(loss)
2746 return loss.detach() / self.args.gradient_accumulation_steps

File ~/miniconda3/envs/testenv/lib/python3.10/site-packages/accelerate/accelerator.py:1907, in Accelerator.backward(self, loss, **kwargs)
1905 return
1906 elif self.scaler is not None:
-> 1907 self.scaler.scale(loss).backward(**kwargs)
1908 else:
1909 loss.backward(**kwargs)

File ~/miniconda3/envs/testenv/lib/python3.10/site-packages/torch/_tensor.py:492, in Tensor.backward(self, gradient, retain_graph, create_graph, inputs)
482 if has_torch_function_unary(self):
483 return handle_torch_function(
484 Tensor.backward,
485 (self,),
(...)
490 inputs=inputs,
491 )
--> 492 torch.autograd.backward(
493 self, gradient, retain_graph, create_graph, inputs=inputs
494 )

File ~/miniconda3/envs/testenv/lib/python3.10/site-packages/torch/autograd/init.py:251, in backward(tensors, grad_tensors, retain_graph, create_graph, grad_variables, inputs)
246 retain_graph = create_graph
248 # The reason we repeat the same comment below is that
249 # some Python versions print out the first line of a multi-line function
250 # calls in the traceback and some print out the last line
--> 251 Variable.execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
252 tensors,
253 grad_tensors,
254 retain_graph,
255 create_graph,
256 inputs,
257 allow_unreachable=True,
258 accumulate_grad=True,
259 )

File ~/miniconda3/envs/testenv/lib/python3.10/site-packages/torch/autograd/function.py:288, in BackwardCFunction.apply(self, *args)
282 raise RuntimeError(
283 "Implementing both 'backward' and 'vjp' for a custom "
284 "Function is not allowed. You should only implement one "
285 "of them."
286 )
287 user_fn = vjp_fn if vjp_fn is not Function.vjp else backward_fn
--> 288 return user_fn(self, *args)

File ~/miniconda3/envs/testenv/lib/python3.10/site-packages/bitsandbytes/autograd/functions.py:491, in MatMul8bitLt.backward(ctx, grad_output)
485 print("state.CxB",state.CxB)
486 print("State ",state)
488 CB = (
489 undo_layout(state.CxB, state.tile_indices)
490 .to(ctx.dtype_A)
--> 491 .mul(state.SCB.unsqueeze(1).mul(1.0 / 127.0))
492 )
493 grad_A = torch.matmul(grad_output, CB).view(ctx.grad_shape).to(ctx.dtype_A)
494 else:

RuntimeError: The size of tensor a (32) must match the size of tensor b (4) at non-singleton dimension 0

#######################################################################################
need your help.

thanks.

[solomonmanuelraj]

Hi Team,

any response for my above query. As i have mentioned earlier, i am receiving the error "RuntimeError: The size of tensor a (32) must match the size of tensor b (4) at non-singleton dimension 0" when i am trying to use 8 bit quantization and finetuning using lora for owl-vit base 32 model.

waiting for your response.

thanks

[sirkitree]

Few suggests for an FAQ:

• How to use this? Assume starting from a fresh system.
• Requirements for using this? Things like minimum hardware specs and basic software systems/packages.
• Links to data sets?
• Basic tutorial? Assume the above and provide directions.

[younesbelkada]

Hi @sirkitree
thanks for the feedback, we do have a documentation page: https://huggingface.co/docs/bitsandbytes/v0.43.0/en/index with good pointers on how to get started, some placeholders are still WIP and welcome to contributions. Do you have any feedback / see any way we could improve that?

[joseph-mueller]

I am trying to get bitsandbytes compiled for Windows 11 with CUDA. I am using the specified versions of the Microsoft compiler (17.9.6) and CUDA (12.3). The last step in the documentation (https://huggingface.co/docs/bitsandbytes/main/en/installation?source=Windows) says to run:
python -m build --wheel

However, there is no build.py in the repo. I have tried running the setup.py, but having various difficulties, but suspect the problem is I am doing something else wrong.

BTW, the build had several warnings, not clear if any of them are fatal.

Sorry if this too elementary for this forum...

[CHNRyan]

Hello! I want to know if bnb support deepspeed zero3_init? #1249