mirror of
https://github.com/huggingface/text-generation-inference.git
synced 2025-04-20 22:32:07 +00:00
c35f39cf83
# Add AWQ quantization inference support
Fixes
https://github.com/huggingface/text-generation-inference/issues/781
This PR (partially) adds support for AWQ quantization for inference.
More information on AWQ [here](https://arxiv.org/abs/2306.00978). In
general, AWQ is faster and more accurate than GPTQ, which is currently
supported by TGI.
This PR installs 4-bit GEMM custom CUDA kernels released by AWQ authors
(in `requirements.txt`, just one line change).
Quick way to test this PR would be bring up TGI as follows:
```
text-generation-server download-weights abhinavkulkarni/codellama-CodeLlama-7b-Python-hf-w4-g128-awq
text-generation-launcher \
--huggingface-hub-cache ~/.cache/huggingface/hub/ \
--model-id abhinavkulkarni/codellama-CodeLlama-7b-Python-hf-w4-g128-awq \
--trust-remote-code --port 8080 \
--max-input-length 2048 --max-total-tokens 4096 --max-batch-prefill-tokens 4096 \
--quantize awq
```
Please note:
* This PR was tested with FlashAttention v2 and vLLM.
* This PR adds support for AWQ inference, not quantizing the models.
That needs to be done outside of TGI, instructions
[here](f084f40bd9).
* This PR only adds support for `FlashLlama` models for now.
* Multi-GPU setup has not been tested.
* No integration tests have been added so far, will add later if
maintainers are interested in this change.
* This PR can be tested on any of the models released
[here](https://huggingface.co/abhinavkulkarni?sort_models=downloads#models).
Please refer to the linked issue for benchmarks for
[abhinavkulkarni/meta-llama-Llama-2-7b-chat-hf-w4-g128-awq](https://huggingface.co/abhinavkulkarni/meta-llama-Llama-2-7b-chat-hf-w4-g128-awq)
vs
[TheBloke/Llama-2-7b-Chat-GPTQ](https://huggingface.co/TheBloke/Llama-2-7b-Chat-GPTQ).
Please note, AWQ has released faster (and in case of Llama, fused)
kernels for 4-bit GEMM, currently at the top of the `main` branch at
https://github.com/mit-han-lab/llm-awq, but this PR uses an older commit
that has been tested to work. We can switch to latest commit later on.
## Who can review?
@OlivierDehaene OR @Narsil
---------
Co-authored-by: Abhinav Kulkarni <abhinav@concentric.ai>
82 lines
2.6 KiB
Python
82 lines
2.6 KiB
Python
import torch
|
|
import torch.distributed
|
|
|
|
from opentelemetry import trace
|
|
from transformers import AutoConfig, AutoTokenizer
|
|
from transformers.models.llama import LlamaTokenizer
|
|
from typing import Optional
|
|
|
|
from text_generation_server.models import FlashCausalLM
|
|
from text_generation_server.models.custom_modeling.flash_llama_modeling import (
|
|
FlashLlamaForCausalLM,
|
|
LlamaConfig,
|
|
)
|
|
from text_generation_server.utils import (
|
|
initialize_torch_distributed,
|
|
weight_files,
|
|
Weights,
|
|
)
|
|
|
|
tracer = trace.get_tracer(__name__)
|
|
|
|
|
|
class FlashLlama(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("FlashLlama is only available on GPU")
|
|
|
|
try:
|
|
tokenizer = LlamaTokenizer.from_pretrained(
|
|
model_id,
|
|
revision=revision,
|
|
padding_side="left",
|
|
truncation_side="left",
|
|
trust_remote_code=trust_remote_code,
|
|
)
|
|
except Exception:
|
|
tokenizer = AutoTokenizer.from_pretrained(
|
|
model_id,
|
|
revision=revision,
|
|
padding_side="left",
|
|
truncation_side="left",
|
|
trust_remote_code=trust_remote_code,
|
|
)
|
|
|
|
config = LlamaConfig.from_pretrained(
|
|
model_id, revision=revision, trust_remote_code=trust_remote_code
|
|
)
|
|
config.quantize = quantize
|
|
|
|
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)
|
|
if config.quantize in ["gptq", "awq"]:
|
|
weights._set_gptq_params(model_id)
|
|
|
|
model = FlashLlamaForCausalLM(config, weights)
|
|
|
|
torch.distributed.barrier(group=self.process_group)
|
|
super(FlashLlama, self).__init__(
|
|
model=model,
|
|
tokenizer=tokenizer,
|
|
num_layers=len(model.model.layers),
|
|
num_kv_heads=model.model.num_key_value_heads,
|
|
head_size=model.model.head_size,
|
|
dtype=dtype,
|
|
device=device,
|
|
rank=rank,
|
|
world_size=world_size,
|
|
)
|