Display total VRAM from all GPUs and whether Infinity Fabric Link/Bridge is present in "Your machine" area

[chafey]

What problem would this solve?

In a multi-gpu setup, the app currently displays just one of the GPUs and its VRAM. It would be nice for it to display the # of GPUs and total VRAM in some way. In addition to this, it would be nice if it displays whether or not infinity fabric link/bridge is active as there is no way to determine this from Mac OS X and sometimes the bridge card is not function.

Proposed behavior

For a 2x Vega II Duo setup with infinity fabric link:
Today: AMD Radeon Vega II Duo 32GB VRAM
New: 2x Vega II Duo 32GB (128GB VRAM IFB)

For a 2x Vega II Duo setup no infinity fabric link:
Today: AMD Radeon Vega II Duo 32GB VRAM
New: 2x Vega II Duo 32GB (128GB VRAM)

Alternatives considered

No response

Additional context

No response

Before submitting

• I searched the existing issues and did not find the same request.

[engeldlgado]

Thanks for asking for this, and specially for writing the example line out. That is what made it obvious that showing a total on its own was not enough: the card had to keep saying which card is which.

It is in 0.85.7, which is building right now. Update when the app offers it, or grab the DMG from https://github.com/engeldlgado/toshllm/releases once it shows up there.

What you should see with more than one card:

• one line per GPU model, 4 × AMD Radeon Pro Vega II Duo · 32 GB each, and then the machine total.
• a line for the link. Metal puts the cards a fabric joins into a peer group, so two Duos with a working bridge come out as one group of four, and two Duos with nothing joining them come out as two groups of two. That line also carries the memory behind each link, which is the memory a split model moves without going through the host.
• a Details button that opens the full breakdown, per model and per link.
• in the GPUs card, the linked ones share a coloured bar and a badge, the unlinked ones have none, and every row now carries its device index so it matches the split picker and the log.

About your question of whether the bridge is really working, that is the part I want you to check, because it is the reason the line exists: macOS tells you nowhere, and the app was already offering an Infinity Fabric option that silently falls back when there is no bridge. If your 2x Vega II Duo show as one group of four, the bridge is alive. If they show as two groups of two, each Duo is talking to itself only and the bridge is doing nothing.

One thing I should be straight about: this machine has a single GPU, so I could only test the multi-card layout with simulated data, using a device listing another user posted from a nine-GPU Mac Pro. The logic has tests and the single-card view is checked, but nobody has seen the real thing yet. So when you open the app, a screenshot of the Home panel would help a lot: I want to see the two cards on real hardware, and whether the numbers and the grouping read right to you.

Thanks again for the report, and for the example line.

[chafey]

Awesome, here is a screen shot from my 2x Vega II Duo System with IFB:

It is possible my IFB is not working as I have had no way to see that Mac OS X detects or any applications to use it. Here are my settings which as you can see has an option to enable IFBL which is confusing since it seemed like you said the option wouldn't be available if there IFB was not detected.

Note that I also have a 2x w6800x Duo system with IFB so can test on that too. I also have coding skills so can debug/fix it if needed

[chafey]

Also I just noticed it says 4x AMD Radeon Pro Vega II Duo. Shouldn't that be 2x AMD Radeon Pro Vega II Duo or 4x AMD Radeon Pro Vega II?

[engeldlgado]

I'll look into what you mentioned. Also, when you enable the toggle, does it show which items have IF Link enabled on the main panel? Could you please share that information with me?

[chafey]

When I hit the toggle, the screen stays the same (says Infinity Fabric: no link between cards). I have the single card IFB as well - will plug those in and report back. PS - Tensor parallelism errors on my machine when using all GPUs, but by layers works. Will create another bug with log file for that issue

…

On Sat, Aug 22, 2026 at 2:09 PM Engelbert Vizcaya ***@***.***> wrote: *engeldlgado* left a comment (engeldlgado/toshllm#67) <#67 (comment)> I'll look into what you mentioned. Also, when you enable the toggle, does it show which items have IF Link enabled on the main panel? Could you please share that information with me? — Reply to this email directly, view it on GitHub <#67?email_source=notifications&email_token=AAJVXWXSK4SLJKGKQ4FTI2D5LHVWLA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMZYGIYTANRYGQ22M4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KYZTPN52GK4S7MNWGSY3L#issuecomment-5382106845>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAJVXWS6NDVZ6QH5J7F6HOL5LHVWLAVCNFSNUABGKJSXA33TNF2G64TZHMYTENRZGA4TSOJWG45US43TOVSTWNJSGIZDCMRWGU2DHILWAI> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/AAJVXWQRBMS2BYGKAYFYGHD5LHVWLA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMZYGIYTANRYGQ22M4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KUZTPN52GK4S7NFXXG> and Android <https://github.com/notifications/mobile/android/AAJVXWRVTOGGRAMIGG3ROH35LHVWLA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMZYGIYTANRYGQ22M4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2K4ZTPN52GK4S7MFXGI4TPNFSA>. Download it today! You are receiving this because you authored the thread.Message ID: ***@***.***>

[chafey]

It seems to detect it properly when using separate IFB for each card:

[engeldlgado]

Yes, please share all the logs you can get from your card combo so I can visualize this better. Thanks for testing

[chafey]

Tensor parallelism did not error out when using just two gpus. It doesn't seem to run any faster with Qwen3-14B-Q4_K_M which I believe is a dense model

[engeldlgado]

Yes, it's something I know for now. I'm working on implementing it correctly, since for now I've been able to get it working, but it's still far below its real capacity, I think. But I've been doing this blindly, without having any cards with IF Link physically; I've been doing everything with the tester logs that were provided to me. I hope to resolve this soon so that it works at its full capacity. For now, what increases is the prompt processing, which greatly increases its speed, but the generation does not yet.

[chafey]

Here is a screen shot from 2x w6800x Duos

[engeldlgado]

Both parts are fixed in 0.85.8, building now.

The counting was exactly as you said it. Metal hands the app one device per GPU and a Duo shows up twice under the same name, so two cards read as four. It now separates the two things: 2 × AMD Radeon Pro Vega II Duo · 4 GPUs · 32 GB each. The machine total was already right, it was only the model line that doubled the cards.

The other one is the toggle you found confusing, and you were right to say so. It was offered even with nothing bridged. It is now disabled when Metal reports no peer group, with a line explaining that a bridge would show up as one.

Now the part I think matters most to you, because your screenshots answered your own question better than I could:

• Your W6800X Duo machine shows one group of four, 128 GB. That bridge works.
• Your Vega II Duo machine with a bridge on each card shows Fabric A and Fabric B, 64 GB each. Each Duo is talking to itself, which is what you would expect from that layout.
• The Vega II Duo machine in the first configuration you tried showed no group at all, not even the internal one each Duo has in the other screenshot. That is the interesting one: if the cross card bridge were doing anything, those four would come out as one group like the W6800X does. So on that machine, in that layout, it is not working.

That line exists precisely because macOS tells you nowhere, and it took real hardware to see it. Thanks for the screenshots from both machines, they were worth more than the report itself.

[chafey]

I have been unable to get the machine to boot with the 2x Vega II Duo and single IFB connecting them. I am pretty sure I had this working before, but there are people online saying it is in invalid configuration and does not work so I am not sure if its that or my IFB connector is broken.