mirror of
https://github.com/huggingface/text-generation-inference.git
synced 2025-04-21 14:52:20 +00:00
8a5f564942
# What does this PR do? During the safetensor conversion, duplicate weights are removed. However, which of the duplicates gets removed, differs per checkpoint. In some, like `h2oai/h2ogpt-oig-oasst1-falcon-40b`, the weight `transformer.word_embeddings.weightSafetensor` gets removed. In others, `lm_head.weight` gets removed. Long story long, we need to support both. Originally,f018143mapped `lm_head` to `word_embeddings`. Thenac736fdswitched this around. This commit merges them and allows for both. ## Before submitting - [x] This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case). - [x] 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? @Narsil, you wrote both commits I referenced in this PR. I think you'll understand this change :)
81 lines
2.5 KiB
Python
81 lines
2.5 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,
|
|
dtype: Optional[torch.dtype] = 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 if dtype is None else dtype
|
|
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,
|
|
aliases={
|
|
"lm_head.weight": ["transformer.word_embeddings.weight"],
|
|
"transformer.word_embeddings.weight": ["lm_head.weight"],
|
|
},
|
|
)
|
|
|
|
config.quantize = quantize
|
|
if config.quantize == "gptq":
|
|
weights._set_gptq_params(model_id)
|
|
|
|
model = FlashRWForCausalLM(config, weights)
|
|
|
|
torch.distributed.barrier(group=self.process_group)
|
|
super(FlashRWSharded, self).__init__(
|
|
model=model.to(device),
|
|
tokenizer=tokenizer,
|
|
num_layers=len(model.transformer.h),
|
|
num_kv_heads=model.transformer.cache_size,
|
|
head_size=model.transformer.head_size,
|
|
dtype=dtype,
|
|
device=device,
|
|
rank=rank,
|
|
world_size=world_size,
|
|
)
|