mirror of
https://github.com/huggingface/text-generation-inference.git
synced 2025-04-21 23:12:07 +00:00
fd89d9dfae
# What does this PR do? <!-- Congratulations! You've made it this far! You're not quite done yet though. Once merged, your PR is going to appear in the release notes with the title you set, so make sure it's a great title that fully reflects the extent of your awesome contribution. Then, please replace this with a description of the change and which issue is fixed (if applicable). Please also include relevant motivation and context. List any dependencies (if any) that are required for this change. Once you're done, someone will review your PR shortly (see the section "Who can review?" below to tag some potential reviewers). They may suggest changes to make the code even better. If no one reviewed your PR after a week has passed, don't hesitate to post a new comment @-mentioning the same persons---sometimes notifications get lost. --> <!-- Remove if not applicable --> Fixes # (issue) ## Before submitting - [ ] This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case). - [ ] Did you read the [contributor guideline](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md#start-contributing-pull-requests), Pull Request section? - [ ] Was this discussed/approved via a Github issue or the [forum](https://discuss.huggingface.co/)? Please add a link to it if that's the case. - [ ] Did you make sure to update the documentation with your changes? Here are the [documentation guidelines](https://github.com/huggingface/transformers/tree/main/docs), and [here are tips on formatting docstrings](https://github.com/huggingface/transformers/tree/main/docs#writing-source-documentation). - [ ] Did you write any new necessary tests? ## Who can review? Anyone in the community is free to review the PR once the tests have passed. Feel free to tag members/contributors who may be interested in your PR. <!-- Your PR will be replied to more quickly if you can figure out the right person to tag with @ @OlivierDehaene OR @Narsil -->
107 lines
3.5 KiB
Python
107 lines
3.5 KiB
Python
import torch
|
|
from loguru import logger
|
|
from functools import lru_cache
|
|
import bitsandbytes as bnb
|
|
from bitsandbytes.nn import Int8Params, Params4bit
|
|
|
|
|
|
@lru_cache(1)
|
|
def warn_deprecate_bnb():
|
|
logger.warning(
|
|
"Bitsandbytes 8bit is deprecated, using `eetq` is a drop-in replacement, and has much better performnce"
|
|
)
|
|
|
|
|
|
class Linear8bitLt(torch.nn.Module):
|
|
def __init__(
|
|
self,
|
|
weight,
|
|
bias,
|
|
has_fp16_weights=True,
|
|
memory_efficient_backward=False,
|
|
threshold=0.0,
|
|
index=None,
|
|
):
|
|
super().__init__()
|
|
assert (
|
|
not memory_efficient_backward
|
|
), "memory_efficient_backward is no longer required and the argument is deprecated in 0.37.0 and will be removed in 0.39.0"
|
|
self.state = bnb.MatmulLtState()
|
|
self.index = index
|
|
|
|
# Necessary for stacked layers
|
|
self.state.threshold = threshold
|
|
self.state.has_fp16_weights = has_fp16_weights
|
|
self.state.memory_efficient_backward = memory_efficient_backward
|
|
if threshold > 0.0 and not has_fp16_weights:
|
|
self.state.use_pool = True
|
|
|
|
self.weight = Int8Params(
|
|
weight.data,
|
|
has_fp16_weights=has_fp16_weights,
|
|
requires_grad=has_fp16_weights,
|
|
)
|
|
self.weight.cuda(weight.device)
|
|
self.bias = bias
|
|
|
|
def init_8bit_state(self):
|
|
self.state.CB = self.weight.CB
|
|
self.state.SCB = self.weight.SCB
|
|
self.weight.CB = None
|
|
self.weight.SCB = None
|
|
|
|
def forward(self, x: torch.Tensor):
|
|
self.state.is_training = self.training
|
|
if self.weight.CB is not None:
|
|
self.init_8bit_state()
|
|
|
|
# weights are cast automatically as Int8Params, but the bias has to be cast manually
|
|
if self.bias is not None and self.bias.dtype != x.dtype:
|
|
self.bias.data = self.bias.data.to(x.dtype)
|
|
|
|
out = bnb.matmul(x, self.weight, bias=self.bias, state=self.state)
|
|
|
|
if not self.state.has_fp16_weights:
|
|
if self.state.CB is not None and self.state.CxB is not None:
|
|
# we converted 8-bit row major to turing/ampere format in the first inference pass
|
|
# we no longer need the row-major weight
|
|
del self.state.CB
|
|
self.weight.data = self.state.CxB
|
|
return out
|
|
|
|
|
|
class Linear4bit(nn.Module):
|
|
def __init__(self, weight, bias, quant_type):
|
|
super().__init__()
|
|
self.weight = Params4bit(
|
|
weight.data,
|
|
requires_grad=False,
|
|
compress_statistics=True,
|
|
quant_type=quant_type,
|
|
)
|
|
self.compute_dtype = None
|
|
self.weight.cuda(weight.device)
|
|
self.bias = bias
|
|
|
|
def forward(self, x: torch.Tensor):
|
|
# weights are cast automatically as Int8Params, but the bias has to be cast manually
|
|
if self.bias is not None and self.bias.dtype != x.dtype:
|
|
self.bias.data = self.bias.data.to(x.dtype)
|
|
|
|
if getattr(self.weight, "quant_state", None) is None:
|
|
print(
|
|
"FP4 quantization state not initialized. Please call .cuda() or .to(device) on the LinearFP4 layer first."
|
|
)
|
|
inp_dtype = x.dtype
|
|
if self.compute_dtype is not None:
|
|
x = x.to(self.compute_dtype)
|
|
|
|
bias = None if self.bias is None else self.bias.to(self.compute_dtype)
|
|
out = bnb.matmul_4bit(
|
|
x, self.weight.t(), bias=bias, quant_state=self.weight.quant_state
|
|
)
|
|
|
|
out = out.to(inp_dtype)
|
|
|
|
return out
|