hf_text-generation-inference/server/text_generation_server/models/flash_neox.py

72 lines
2.3 KiB
Python
Raw Permalink Normal View History

2023-03-24 07:02:14 -06:00
import torch
import torch.distributed
from opentelemetry import trace
2023-04-03 11:06:42 -06:00
from transformers import AutoTokenizer, AutoConfig
from typing import Optional
2023-03-24 07:02:14 -06:00
2023-04-03 11:06:42 -06:00
from text_generation_server.models import FlashCausalLM
from text_generation_server.models.custom_modeling.flash_neox_modeling import (
2023-03-24 07:02:14 -06:00
FlashGPTNeoXForCausalLM,
)
from text_generation_server.utils import (
initialize_torch_distributed,
weight_files,
Weights,
2023-03-24 07:02:14 -06:00
)
tracer = trace.get_tracer(__name__)
class FlashNeoXSharded(FlashCausalLM):
2023-03-24 07:02:14 -06:00
def __init__(
2023-05-15 15:36:30 -06:00
self,
model_id: str,
revision: Optional[str] = None,
quantize: Optional[str] = None,
dtype: Optional[torch.dtype] = None,
trust_remote_code: bool = False,
2023-03-24 07:02:14 -06:00
):
self.process_group, rank, world_size = initialize_torch_distributed()
2023-03-24 07:02:14 -06:00
if torch.cuda.is_available():
device = torch.device(f"cuda:{rank}")
dtype = torch.float16 if dtype is None else dtype
2023-03-24 07:02:14 -06:00
else:
raise NotImplementedError("FlashNeoX is only available on GPU")
tokenizer = AutoTokenizer.from_pretrained(
model_id,
revision=revision,
padding_side="left",
truncation_side="left",
trust_remote_code=trust_remote_code,
2023-03-24 07:02:14 -06:00
)
config = AutoConfig.from_pretrained(
model_id, revision=revision, trust_remote_code=trust_remote_code
2023-03-24 07:02:14 -06:00
)
config.quantize = quantize
2023-03-24 07:02:14 -06:00
torch.distributed.barrier(group=self.process_group)
filenames = weight_files(model_id, revision=revision, extension=".safetensors")
weights = Weights(
filenames, device=device, dtype=dtype, process_group=self.process_group
)
feat(server): Using `quantize_config.json` instead of GPTQ_BITS env variables. (#671) - Current PR is not great because we're side stepping the `Weights.__init__` but Weights shouldn't requires anything related to the config or the model_id as it aims to be a simple Wrapper over multi file loading. - Ideal solution would be to use something like Rust enum ``` enum Quantize{ Bitandbytes(Bitsandbytes), GPTQ(bits: usize, groupsize: usize) ``` And passing that around during load. Unfortunately we don't have access to this, so for now, side-stepping seems easier. - Re-enabling groupsize<0 with exllama (confirmed it works.) Helps #601 In next steps we should make sure our quantization script uses that format and make it standard. # 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 -->
2023-07-25 05:00:27 -06:00
if config.quantize == "gptq":
weights._set_gptq_params(model_id)
2023-03-24 07:02:14 -06:00
model = FlashGPTNeoXForCausalLM(config, weights)
2023-03-24 07:02:14 -06:00
torch.distributed.barrier(group=self.process_group)
super(FlashNeoXSharded, self).__init__(
model=model.to(device),
2023-03-24 07:02:14 -06:00
tokenizer=tokenizer,
num_layers=len(model.gpt_neox.layers),
num_kv_heads=model.gpt_neox.num_heads,
head_size=model.gpt_neox.head_size,
dtype=dtype,
2023-03-24 07:02:14 -06:00
device=device,
rank=rank,
world_size=world_size,
2023-03-24 07:02:14 -06:00
)