mirror of
https://github.com/huggingface/text-generation-inference.git
synced 2025-04-21 14:52:20 +00:00
abd58ff82c
# What does this PR do? Reworked the loading logic. Idea is to use cleaner loading code: - Remove need for `no_init_weights` - Remove all weird `bnb_linear` and `load_weights` and `post_load_weights`. New code layout: - New class `Weights` in charge of handling loading the weights from multiple files into appropiate tensors (potentially sharded) - TP layers now are "shells", they contain the code to know what kind of sharding we need + eventual `all_reduce`. They do not inherit from linear, but they contain some kind of Linear instead - the contained linear can be either FastLinear, BnbLinear or GPTq Linear next. - All modeling code is explictly made for sharding, process group is just no-ops for non sharded code (removes a lot of test cases)  --------- Co-authored-by: Ubuntu <ubuntu@ip-172-31-41-161.taildb5d.ts.net> Co-authored-by: Ubuntu <ubuntu@ip-172-31-41-161.ec2.internal> Co-authored-by: OlivierDehaene <olivier@huggingface.co> Co-authored-by: OlivierDehaene <23298448+OlivierDehaene@users.noreply.github.com>
67 lines
1.9 KiB
Python
67 lines
1.9 KiB
Python
import torch
|
|
import torch.distributed
|
|
|
|
from opentelemetry import trace
|
|
from transformers import AutoTokenizer
|
|
from typing import Optional
|
|
|
|
from text_generation_server.models import FlashCausalLM
|
|
from text_generation_server.models.custom_modeling.flash_rw_modeling import (
|
|
RWConfig,
|
|
FlashRWForCausalLM,
|
|
)
|
|
from text_generation_server.utils import (
|
|
initialize_torch_distributed,
|
|
weight_files,
|
|
Weights,
|
|
)
|
|
|
|
tracer = trace.get_tracer(__name__)
|
|
|
|
|
|
class FlashRWSharded(FlashCausalLM):
|
|
def __init__(
|
|
self,
|
|
model_id: str,
|
|
revision: Optional[str] = None,
|
|
quantize: Optional[str] = None,
|
|
trust_remote_code: bool = False,
|
|
):
|
|
self.process_group, rank, world_size = initialize_torch_distributed()
|
|
if torch.cuda.is_available():
|
|
device = torch.device(f"cuda:{rank}")
|
|
dtype = torch.float16
|
|
else:
|
|
raise NotImplementedError("FlashRW 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,
|
|
)
|
|
|
|
config = RWConfig.from_pretrained(
|
|
model_id, revision=revision, trust_remote_code=trust_remote_code
|
|
)
|
|
|
|
torch.distributed.barrier(group=self.process_group)
|
|
filenames = weight_files(model_id, revision=revision, extension=".safetensors")
|
|
weights = Weights(filenames, device, dtype, process_group=self.process_group)
|
|
|
|
config.quantize = quantize
|
|
|
|
model = FlashRWForCausalLM(config, weights)
|
|
|
|
torch.distributed.barrier(group=self.process_group)
|
|
super(FlashCausalLM, self).__init__(
|
|
model=model.to(device),
|
|
tokenizer=tokenizer,
|
|
requires_padding=False,
|
|
dtype=dtype,
|
|
device=device,
|
|
rank=rank,
|
|
world_size=world_size,
|
|
)
|