mirror of
https://github.com/games-on-whales/wolf.git
synced 2026-10-15 21:12:08 +00:00
[GH-ISSUE #417] H.264 failing on blackwell, works on ampere #271
Labels
No labels
🧠 1 - 🤖 999
AMD
bug
bug
cannot reproduce
documentation
duplicate
easy
enhancement
enhancement
enhancement
fixed?
fixed?
good first issue
hard
help wanted
Intel
invalid
LXC
medium
moonlight-protocol
Multi GPU
nixos
not a bug
nvidia
pull-request
question
wontfix
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
games-on-whales/wolf#271
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Originally created by @j-dobr on GitHub (May 31, 2026).
Original GitHub issue: https://github.com/games-on-whales/wolf/issues/417
I have a machine with 2x RTX 3060, RTX 5070ti and RTX PRO 4000. One of Moonlight clients can use only H.264, but it is failing to negotiate working encoder when using either of Blackwell cards. Using Ampere (RTX 3060), H.264 works (but 3060 does not support AV1 for other clients).
When using RTX 3060
At this moment, I'm in Wolf UI user selection on client.
When using RTX 5070ti / RTX PRO 4000 (same result)
At this moment, client exits Wolf UI claiming no video stream.
Setup
Setup is same in both cases, config is default. GPU switching is done through docker-compose.yml visible devices.
ghcr.io/games-on-whales/wolf:stable sha256:7ade85d69af9eabdc80ebc8787f45d0e10998c5128712c2c72de58199971737f
Nvidia drivers are passed to container manual way (thorough volume), but it's rebuilt on each run.
@ABeltramo commented on GitHub (May 31, 2026):
Have you tried not using the
NVIDIA_*env variables and just set the encoder node for Wolf accordingly?I feel that the issue is just that you set
renderD128for Wolf which will then pass that device to the app container but that's not whatrenderD128is on the host and it results in a mismatch. Just setWOLF_RENDER_NODEas specified on the docs https://games-on-whales.github.io/wolf/stable/user/configuration.html#_multiple_gpu@j-dobr commented on GitHub (May 31, 2026):
Same result unfortunately:
(As far as host goes, I have a nice table of ID, UUID and renderD for each GPU making sure it always matches... truth is I have no idea what happens in containers but given GPU utilization when it works, it should match all the way down)
@j-dobr commented on GitHub (May 31, 2026):
And also without those
NVIDIA_*variables, it detects AV1 encoder even though renderD points to RTX 3060 and subsequently fails (as it uses blackwell for H.264, not RTX 3060).@JBailes commented on GitHub (May 31, 2026):
To be clear, these are all on the same box?
@j-dobr commented on GitHub (May 31, 2026):
Yes.
Edit: also my table for reference:
@ABeltramo commented on GitHub (May 31, 2026):
Can you provide a full run with the env variable
WOLF_LOG_LEVEL=DEBUG? And please post the full yaml that you are using to run Wolf@j-dobr commented on GitHub (May 31, 2026):
Failing state - RTX 5070ti
docker-compose.yml
Output
Edit: Got to go, will be back in 1 week for further debugging if needed.
@ABeltramo commented on GitHub (May 31, 2026):
Can you remove the env variables
?
@j-dobr commented on GitHub (Jun 7, 2026):
Same result. h.264 not working while h.265 within same run works. Below is result from purely h.264 run.
docker-compose.yml
Output
@JBailes commented on GitHub (Jun 8, 2026):
I have reproduced this locally and am looking into a fix.
@JBailes commented on GitHub (Jun 8, 2026):
This issue is specific to Blackwell and is due to a Blackwell-specific change in CUDA.
@JBailes commented on GitHub (Jun 8, 2026):
Oof. @ABeltramo Blackwell cudaupload does not accept DMABuf — only system, GLMemory, or CUDAMemory. The compositor emits CUDAMemory BGRA directly (the broken Blackwell interop). cudaupload on that is a passthrough, so the bad memory reaches cudaconvertscale and fails.
I'm working on a fix for blackwell and non-blackwell architectures.