For the LLaMa-7B fp16 model with a maximum length of 2048, the server requires approximately 1030MB of GPU memory to store kv_cache for each concurrent session created. This means that even an A100 80G can only serve a limited number of users.
To reduce runtime GPU memory usage, we have implemented PTQ quantization for kv cache, using the following formula:
zp = (min+max) / 2
scale = (max-min) / 255
quant: q = round( (f-zp) / scale)
dequant: f = q * scale + zp
Convert the Hugging Face model format to the TurboMind inference format to create a workspace directory.
python3 -m lmdeploy.serve.turbomind.deploy internlm-chat-7b /path/to/internlm-chat-7b
If you already have a workspace directory, skip this step.
Get the quantization parameters by these two steps:
# get minmax
python3 -m lmdeploy.lite.apis.calibrate \
--model $HF_MODEL \
--calib_dataset 'c4' \ # Support c4, ptb, wikitext2, pileval
--calib_samples 128 \ # Number of samples in the calibration set, if the memory is not enough, it can be adjusted appropriately
--calib_seqlen 2048 \ # Length of a single text, if the memory is not enough, you can adjust it appropriately
--work_dir $WORK_DIR \ # Directory for saving quantized statistical parameters and quantized weights in Pytorch format
# get quant parameters
python3 -m lmdeploy.lite.apis.kv_qparams \
--work_dir $WORK_DIR \ # Directory of the last output
--turbomind_dir workspace/triton_models/weights/ \ # Directory to save the quantization parameters
--kv_sym False \ # Symmetric or asymmetric quantization, default is False
--num_tp 1 \ # Number of GPUs used for Tensor parallelization, keep it consistent with deploy.py
kv_qparams
will generate fp32 scaling factors in the weights
directory. The file format is a binary produced by numpy.tofile
.
You can also first set turbomind_dir
to a private directory, then copy the scaling factors into workspace/triton_models/weights/
.
Modify workspace/triton_models/weights/config.ini
:
- Set use_context_fmha to 0, which means turning off flashattention
- Set quant_policy to 4. This means enabling kv_cache int8
This is because there are two versions of flashattention, v1 and v2, and kv_cache int8 has also previously realized the symmetric version.
Considering there are four combinations of kernels needed to be implemented, premature optimization when the algorithm is uncertain can be disastrous for software.
Test the chat performance.
python3 -m lmdeploy.turbomind.chat ./workspace
The test object is the internlm-chat-7b model. Testing method:
- Use
deploy.py
to convert the model, modify the maximum concurrency in theworkspace
configuration; adjust the number of requests inllama_config.ini
. - Compile and run
bin/llama_triton_example
to obtain the GPU memory situation of the fp16 version under different batch_size. - Enable quantization, re-run
bin/llama_triton_example
to obtain the GPU memory situation of the int8 version under different batch_size.
Below shows the comparison of GPU memory between the two versions:
batch_size | fp16 memory(MiB) | int8 memory(MiB) | diff(MiB) |
---|---|---|---|
8 | 22337 | 18241 | -4096 |
16 | 30593 | 22369 | -8224 |
32 | 47073 | 30625 | -16448 |
48 | 63553 | 38881 | -24672 |
Compared to directly quantizing Weight (such as GPTQ-for-LLaMa), we have done a comparative estimation of memory growth in the 7B model for both methods, with some data from llama.cpp.
As can be seen, the fp16 version requires 1030MB of GPU memory for each concurrency, so quantizing kv_cache can significantly reduce the rate of increase of runtime memory.
The test object is the internlm-chat-7b command model.
Below is the result of PTQ quantization of kCacheKVInt8
method with only 128 randomly selected data from the c4 dataset. The accuracy was tested using opencompass before and after quantization.
task | dataset | metric | int8 | fp16 | diff |
---|---|---|---|---|---|
Language | winogrande | accuracy | 60.77 | 61.48 | -0.71 |
Knowledge | nq | score | 2.69 | 2.60 | +0.09 |
Reasoning | gsm8k | accuracy | 33.28 | 34.72 | -1.44 |
Reasoning | bbh | naive_average | 20.12 | 20.51 | -0.39 |
Understanding | openbookqa_fact | accuracy | 82.40 | 82.20 | +0.20 |
Understanding | eprstmt-dev | accuracy | 90.62 | 88.75 | +1.87 |
Safety | crows_pairs | accuracy | 32.56 | 31.43 | +1.13 |
Note that both kCacheKVInt8
and WeightInt4
methods can be enabled at the same time.