nvComp: Unknown output data type

[imbalu007]

The following snippet fails in encoding step:

from nvidia import nvcomp
import torch

codec = nvcomp.Codec(algorithm="ANS", data_type="f16")
data_gpu = torch.randn(4096*4096, dtype=torch.float16).cuda()
nvarr_d = nvcomp.as_array(data_gpu)
com_arr = codec.encode(nvarr_d)

Fails in encode step with error:
ValueError: Unknown output data type

nvcomp version: 4.1.0
nvcomp cuda version: 12050
OS: Ubuntu

[wjd6910502]

it looks linke not support data_type params if you delet this params it works。
i met the same problem is decode “Traceback (most recent call last):
File "/code/sglang_v6/test_nbf16_new.py", line 40, in
decoded_array = codec1.decode(encoded_array, data_type='f4')
ValueError: Unknown output data type
”

[imbalu007]

Without data_type, the input is treated as int8 (default). Not sure if that will impact compression rate or throughput for fp16 arrays

[baliika]

Thank you @imbalu007 for raising this point. We had a typo in our documentation, and with the new patch release of nvCOMP (4.1.1) we also updated this.

2-byte floats should be referred to as <f2, i.e., 2-byte little-endian float using the Python array-protocol type strings. Note, that indicating endianness is also required. With this change, your example changes to:

from nvidia import nvcomp
import torch

codec = nvcomp.Codec(algorithm="ANS", data_type="<f2")
data_gpu = torch.randn(4096*4096, dtype=torch.float16).cuda()
nvarr_d = nvcomp.as_array(data_gpu)
com_arr = codec.encode(nvarr_d)

We also created a table where we match nvCOMP types to array-protocol type strings for more clarity while using the Python API:
https://docs.nvidia.com/cuda/nvcomp/py_api.html#data-type-association

Please let us know if you have other questions!