Update: no longer reproducible on 7.0.0-34-generic. I re-ran the minimal reproducer (load a 28GB checkpoint through comfy.sd.load_checkpoint_guess_config, the exact step that hung at param.copy_(input_param)) on three kernels:
kernel step1 step2 (hang point) step3 step4 result 7.0.0-28 0.0s HANG (>300s) - - hung 6.17.0-35 0.0s 7.4s 5.6s 0.0s OK 7.0.0-34 0.0s 6.3s 4.9s 0.0s OK (warm cache) On 7.0.0-34 the previously hanging step completes in the same time as on the working 6.17 kernel. Tested twice, both runs clean. Likely root cause of the fix: "drm/amdgpu: drop retry loop in amdgpu_hmm_range_get_pages", which is absent in 7.0.0-28 and present from 7.0.0-30 onwards (changelog entry under LP: #2158267, "Performance regression causes SDXL inference slowdown (~42x)"). That patch touches amdgpu_hmm_range_get_pages, i.e. the HMM/SVM page migration path reported here, and matches the boot-time "svm_range_restore_work ... hogged CPU" warnings. Upstream this is a Linux 7.0.13 stable fix. Environment note: the original report used ROCm 7.15; that environment can no longer run the current ComfyUI (missing comfy_aimdo), so the re-test used the current environment (torch 2.14.0+rocm10.1.0a20260822, hip 7.16.26332) on the same hardware (Ryzen AI Max+ 395 / Radeon 8060S, 512MB VRAM + 120GB GTT). This looks like a duplicate / severe manifestation of LP: #2158267. I'm happy to have it closed as fixed (or marked duplicate). -- You received this bug notification because you are a member of Ubuntu Bugs, which is subscribed to Ubuntu. https://bugs.launchpad.net/bugs/2161667 Title: Kernel 7.0 KFD deadlock loading models via ROCm on Strix Halo To manage notifications about this bug go to: https://bugs.launchpad.net/ubuntu/+source/linux-hwe-7.0/+bug/2161667/+subscriptions -- ubuntu-bugs mailing list [email protected] https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs
