Cross platform model

[ghost]

Hi! I have a model trained on 64 bit Linux(Intel) machine.

But when I trying to use this model on arm64 with 32bit logic/pointers I’m always getting an exception while bin model loading.

It’s happened because train machine use few 64bit data types(like ‘size_t’, ‘long’) which used in model files, but arm64_32 has different data types. For example, size_t is a 4 bit value instead 8 bit on training machine.

So, when I trying to read this model on arm( 32bit data types), model load class use wrong offset(in ‘template get’) - 4 bit instead 8 bit in model and fails.

Is there a solution to this problem? Thanks!

[emjotde]

Sorry, ARM is not an architecture we ever tested on or support. I am actually amazed you got Marian to compile there.

[hieuhoang]

out of interest, what machine are you using? Apple M1? Surface Go? AWS server?

[ghost]

Sorry, ARM is not an architecture we ever tested on or support. I am actually amazed you got Marian to compile there.

Yes, I understand. But i think the same problem will be on any 32 bit machine.

[frankseide]

Is this a bin model or an NPY? I think the NPY file Should he readable across all platforms, since NPY is a standard.

If the NPY file can be loaded, Maybe you can try to use the ARM build to convert it to BIN?

Get Outlook for iOShttps://aka.ms/o0ukef

---

From: Marcin Junczys-Dowmunt notifications@github.com Sent: Tuesday, January 5, 2021 11:56:48 PM To: marian-nmt/marian marian@noreply.github.com Cc: Subscribed subscribed@noreply.github.com Subject: Re: [marian-nmt/marian] Cross platform model (#352)

Sorry, ARM is not an architecture we ever tested on or support. I am actually amazed you got Marian to compile there.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHubhttps://github.com/marian-nmt/marian/issues/352#issuecomment-755145834, or unsubscribehttps://github.com/notifications/unsubscribe-auth/AD7GDFX5DN67DUHIHCRX4B3SYQJUBANCNFSM4VXECC3Q.

[ghost]

out of interest, what machine are you using? Apple M1? Surface Go? AWS server?

Apple Watches Arm64_32(ARMv8-s)

[ghost]

Is this a bin model or an NPY? I think the NPY file Should he readable across all platforms, since NPY is a standard.

If the NPY file can be loaded, Maybe you can try to use the ARM build to convert it to BIN?

Get Outlook for iOShttps://aka.ms/o0ukef

---

From: Marcin Junczys-Dowmunt notifications@github.com Sent: Tuesday, January 5, 2021 11:56:48 PM To: marian-nmt/marian marian@noreply.github.com Cc: Subscribed subscribed@noreply.github.com Subject: Re: [marian-nmt/marian] Cross platform model (#352)

Sorry, ARM is not an architecture we ever tested on or support. I am actually amazed you got Marian to compile there.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHubhttps://github.com/marian-nmt/marian/issues/352#issuecomment-755145834, or unsubscribehttps://github.com/notifications/unsubscribe-auth/AD7GDFX5DN67DUHIHCRX4B3SYQJUBANCNFSM4VXECC3Q.

I use BIN model. Thanks! I'll try NPY.

[emjotde]

Where does the *.bin file come from?

[seidefrank]

I think the BIN model is more or less a direct memory dump, that’s kind of their point :) . I would consider BIN models as inherently non portable. If NPY fails as well, then that’s something worth debugging.

[emjotde]

Yes, as Frank said, the *.bin models are meant to be memory-mapped on 64-bit machines (apart from the small header).

And careful: it's .npz not .npy.

I think *.npy won't do anything, hopefully it will complain :)

[seidefrank]

Indeed NPZ, not NPY :) been away from Marian for too long already!!

[emjotde]

I guess a .bin file created from a .npz file on an ARM machine would have a chance to work then?

[seidefrank]

It should, or if not, it should be an easy fix if the NPZ works.

[kpu]

We've been compiling Marian for WASM in https://github.com/browsermt/marian-dev/pull/6 . WASM also has a 32-bit size_t. This https://github.com/marian-nmt/marian-dev/pull/779 should fix one problem for you. For what it's worth, an npz trained on an amd64 machine just worked on WASM.

[frankseide]

BTW you shouldn't need to retrain the model for this. You just need to convert the BIN model to NPZ. Not sure if marian-convert has a function to go that way, but it should be easy to hack if not.

[ghost]

Thanks to all! NPZ models works. Also needs to undefine FASTOPT feature.

[emjotde]

marian-convert can work both ways, based on the suffix.