Are some zip files missing for tensorflow-windows-wheel/2.4.0/py38/CPU+GPU/cuda111cudnn8sse2/ ?

[ALIANS-PRODUCTIONS]

For tensorflow-windows-wheel/2.4.0/py38/CPU+GPU/cuda111cudnn8sse2/ there are only 3 7z files.

When i download and extract with 7zip on windows 10, i get a file tensorflow-2.4.0-cp38-cp38-win_amd64.7z 338kb.
However this file itself cannot be further extracted to geth the WHL file.

I am wondering if we are missing some more of the split 7z files ? FYI the AVX2 version of the directory has 5 7z files in total

If not, maybe i am completely misunderstanding how to install this
fyi : pip install tensorflow-2.4.0-cp38-cp38-win_amd64.7z
ERROR: Could not find a version that satisfies the requirement tensorflow-2.4.0-cp38-cp38-win_amd64.7z
ERROR: No matching distribution found for tensorflow-2.4.0-cp38-cp38-win_amd64.7z

[fo40225]

You need to click the file link on the repo file browser and press the "Download" button.

[ALIANS-PRODUCTIONS]

Thank you so much for your reply and my apologies for being silly :-) That worked!

If by any chance you are able to build TF 2.4.1 against Cuda 11.2 and latest CUDNN for windows 10, i would happily donate to your page for your time. I am researching ML models for timeseries analysis and right now 3090 owners are stuck with tf builds against cuda 11.1 that handicaps the card. I have a blank SSD card staring at me right now and i dread the thought of having to install Ubuntu and dual boot my pc just to get Nvidias tf cuda 11.2 wheels...

[fo40225]

Do you have encountered any issues when using 2.4.0/py38/CPU+GPU/cuda111cudnn8sse2 on 3090?

[ALIANS-PRODUCTIONS]

i have a cluster of machines, running different gpus:
the 2.4.0/py38/CPU+GPU/cuda111cudnn8sse2 will be to upgrade an old dell T5500 which has two xeon hex core 5570s non avx and a rtx 2080ti. Currently on Cuda 10.2
Anacoda doesnt have non avx builds beyond tf 2.3 and i cant figure out how to get newer versions of Cudnn working via Anaconda. Tensorflow just crashes if i use pip to install on that machine (since they dont have non avx builds)

So the wheel from your repo will help a lot in getting better performance on the 2080ti due to (hopefully) higher cuda and cudnn libs.

The 3090 is on a machine with avx2, am able to run tf 2.4.1 using pip install, cuda 11.0 and cudnn. But have had many strange situations where training on the 2080ti was running significantly faster. And my hunch is that the ampere cards really need Cuda 11.2 + cudnn 8.1.1 to get maximum performance.

Having all machines on tf 2.4 will make it easier to run the same code without version modifications (for things like mixed precision) for each machine...

[fo40225]

The CUDA 11.0 only support the A100.

cuda 11.1 and cudnn 8.0.5 should be fine with 3090.

[wwangyuxia]

You need to click the file link on the repo file browser and press the "Download" button.

I follow your steps to download the file. get"tensorflow-1.15.4+nv-cp38-cp38-win_amd64.7z.001". but I can't zip it,there is a error: "The compressed file is corrupted". How can I get the ".whl" file?
Look forward to your reply!

[fo40225]

#52 (comment)