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

242 lines
9.3 KiB
Python
Raw Normal View History

2023-01-31 10:53:56 -07:00
import torch
import torch.distributed
from typing import List, Optional
2023-01-31 10:53:56 -07:00
from accelerate import init_empty_weights
from safetensors import safe_open
from transformers import (
AutoTokenizer,
AutoModelForCausalLM,
AutoConfig,
)
from transformers.models.gpt_neox.parallel_layers import (
TensorParallelColumnLinear,
TensorParallelEmbedding,
TensorParallelRowLinear,
)
2023-03-07 10:52:22 -07:00
from text_generation_server.models import CausalLM
from text_generation_server.utils import (
2023-01-31 10:53:56 -07:00
initialize_torch_distributed,
weight_files,
)
HAS_BITS_AND_BYTES = True
try:
import bitsandbytes as bnb
from bitsandbytes.nn import Int8Params
except Exception as e:
HAS_BITS_AND_BYTES = False
class GPTNeoxSharded(CausalLM):
2023-01-31 10:53:56 -07:00
def __init__(
feat(server): GPTQ quantization (step1) (#277) Changes only the type from `bool` to `Option<Enum>` pretty much everywhere. - Use `Optional[str]` in Python (easier to manage than importing type everywhere). Except for the cli to get proper validation - Updated all models to handle gracefully new values. (Error out if unknown value, or gptq since not implemented). <!-- 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 -->
2023-05-12 06:46:41 -06:00
self,
model_id: str,
revision: Optional[str] = None,
quantize: Optional[str] = None,
2023-01-31 10:53:56 -07:00
):
self.process_group, rank, world_size = initialize_torch_distributed()
2023-01-31 10:53:56 -07:00
if torch.cuda.is_available():
device = torch.device(f"cuda:{rank}")
2023-05-10 07:51:10 -06:00
dtype = torch.float16
2023-01-31 10:53:56 -07:00
else:
device = torch.device("cpu")
dtype = torch.float32
tokenizer = AutoTokenizer.from_pretrained(
model_id, revision=revision, padding_side="left", truncation_side="left"
2023-01-31 10:53:56 -07:00
)
tokenizer.pad_token = tokenizer.eos_token
config = AutoConfig.from_pretrained(
model_id, revision=revision, tp_parallel=True
2023-01-31 10:53:56 -07:00
)
torch.distributed.barrier(group=self.process_group)
filenames = weight_files(model_id, revision=revision, extension=".safetensors")
2023-01-31 10:53:56 -07:00
with init_empty_weights():
model = AutoModelForCausalLM.from_config(config)
torch.distributed.barrier(group=self.process_group)
self.load_weights(
model,
filenames,
quantize=quantize,
device=device,
dtype=dtype,
rank=rank,
world_size=world_size,
2023-01-31 10:53:56 -07:00
)
self.model = model.eval()
2023-01-31 10:53:56 -07:00
torch.distributed.barrier(group=self.process_group)
super(CausalLM, self).__init__(
tokenizer=tokenizer,
requires_padding=True,
dtype=dtype,
2023-01-31 10:53:56 -07:00
device=device,
rank=rank,
world_size=world_size,
2023-01-31 10:53:56 -07:00
)
@staticmethod
def load_weights(
model,
filenames: List[str],
feat(server): GPTQ quantization (step1) (#277) Changes only the type from `bool` to `Option<Enum>` pretty much everywhere. - Use `Optional[str]` in Python (easier to manage than importing type everywhere). Except for the cli to get proper validation - Updated all models to handle gracefully new values. (Error out if unknown value, or gptq since not implemented). <!-- 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 -->
2023-05-12 06:46:41 -06:00
quantize: Optional[str],
2023-01-31 10:53:56 -07:00
device: torch.device,
dtype: torch.dtype,
2023-01-31 10:53:56 -07:00
rank: int,
world_size: int,
):
parameters = dict(model.named_parameters())
for file in filenames:
with safe_open(
2023-05-15 15:36:30 -06:00
file, framework="pt", device=str(device) if quantize is None else "cpu"
2023-01-31 10:53:56 -07:00
) as f:
for name in f.keys():
module_name, param_name = name.rsplit(".", 1)
module = model.get_submodule(module_name)
current_parameter_tensor = parameters.get(name, None)
slice_ = f.get_slice(name)
if isinstance(module, TensorParallelColumnLinear):
size = slice_.get_shape()[0]
block_size = size // world_size
start = rank * block_size
stop = (rank + 1) * block_size
tensor = slice_[start:stop]
elif isinstance(module, TensorParallelRowLinear):
if param_name == "weight":
size = slice_.get_shape()[1]
block_size = size // world_size
start = rank * block_size
stop = (rank + 1) * block_size
tensor = slice_[:, start:stop]
else:
tensor = slice_[:]
# XXX: Hack for Rowlinear to add the bias only once.
if rank != 0:
tensor = torch.zeros_like(tensor)
elif isinstance(module, TensorParallelEmbedding):
size = slice_.get_shape()[0]
block_size = size // world_size
start = rank * block_size
stop = (rank + 1) * block_size
tensor = slice_[start:stop]
elif name == "embed_out.weight" and model.gpt_neox.tp_embeddings:
2023-01-31 10:53:56 -07:00
size = slice_.get_shape()[0]
block_size = size // world_size
start = rank * block_size
stop = (rank + 1) * block_size
tensor = slice_[start:stop]
else:
try:
tensor = slice_[:]
except:
tensor = f.get_tensor(name)
if (
current_parameter_tensor is not None
and current_parameter_tensor.shape != tensor.shape
):
raise ValueError(
f"Name {name} -- Current {current_parameter_tensor.shape} and got {tensor.shape}"
)
tensor = tensor.contiguous().to(dtype)
2023-01-31 10:53:56 -07:00
feat(server): GPTQ quantization (step1) (#277) Changes only the type from `bool` to `Option<Enum>` pretty much everywhere. - Use `Optional[str]` in Python (easier to manage than importing type everywhere). Except for the cli to get proper validation - Updated all models to handle gracefully new values. (Error out if unknown value, or gptq since not implemented). <!-- 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 -->
2023-05-12 06:46:41 -06:00
if quantize == "bitsandbytes":
2023-01-31 10:53:56 -07:00
if not HAS_BITS_AND_BYTES:
raise ImportError(
"bitsandbytes is not available on your machine either because it is not installed "
"or you don't have a GPU.\n"
"You can install it with `pip install bitsandbytes`."
)
if (
type(module)
in [TensorParallelRowLinear, TensorParallelColumnLinear]
and param_name == "weight"
2023-01-31 10:53:56 -07:00
):
tensor = Int8Params(
tensor,
has_fp16_weights=False,
requires_grad=False,
).to(device)
state = bnb.MatmulLtState()
state.threshold = 6.0
state.has_fp16_weights = False
state.memory_efficient_backward = False
state.use_pool = True
state.CB = tensor.CB
state.SCB = tensor.SCB
tensor.CB = None
tensor.SCB = None
def replace_linear(state):
def linear(input, weight, bias):
out = bnb.matmul(
input,
weight,
state=state,
threshold=state.threshold,
bias=bias,
)
if state.CB 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 state.CB
weight.data = state.CxB
return out
return linear
module.linear = replace_linear(state)
feat(server): GPTQ quantization (step1) (#277) Changes only the type from `bool` to `Option<Enum>` pretty much everywhere. - Use `Optional[str]` in Python (easier to manage than importing type everywhere). Except for the cli to get proper validation - Updated all models to handle gracefully new values. (Error out if unknown value, or gptq since not implemented). <!-- 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 -->
2023-05-12 06:46:41 -06:00
elif quantize == "gptq":
raise NotImplementedError(
"`gptq` is not implemented for now"
)
elif quantize is None:
2023-01-31 10:53:56 -07:00
tensor = tensor.to(device)
feat(server): GPTQ quantization (step1) (#277) Changes only the type from `bool` to `Option<Enum>` pretty much everywhere. - Use `Optional[str]` in Python (easier to manage than importing type everywhere). Except for the cli to get proper validation - Updated all models to handle gracefully new values. (Error out if unknown value, or gptq since not implemented). <!-- 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 -->
2023-05-12 06:46:41 -06:00
else:
raise ValueError(f"Unexpected quantize `{quantize}`")
2023-01-31 10:53:56 -07:00
if current_parameter_tensor is not None:
module._parameters[param_name] = tensor
else:
module._buffers[param_name] = tensor
def forward(
self, input_ids, attention_mask, position_ids, past_key_values: Optional = None
):
if self.model.gpt_neox.tp_embeddings:
outputs = self.model.forward(
input_ids=input_ids,
attention_mask=attention_mask,
position_ids=position_ids,
past_key_values=past_key_values,
use_cache=True,
)
# Logits are sharded, so we need to gather them
logits = [torch.empty_like(outputs.logits) for _ in range(self.world_size)]
torch.distributed.all_gather(
logits, outputs.logits, group=self.process_group
)
logits = torch.cat(logits, dim=2)
return logits, outputs.past_key_values
# While the model itself is sharded, the embeddings might not as they might not be dividable by num-shard
else:
return super(GPTNeoxSharded, self).forward(
input_ids, attention_mask, position_ids, past_key_values
)