This used to work but no matter how I start ROCm I always get:
ROCm preparing model tensor mappings: 80.45 GiBds4: ROCm model arena alloc failed for tensor-span:128 (1792.00 MiB chunk): out of memory
With auto VRAM BIOS configuration I get that earlier, around 58 or something ... with --ssd-streaming it works but it's noticeably slower than it used to be before ... is Strix Halo these days not a q2 (reduced) target anymore with its VRAM?
Thank you!
This used to work but no matter how I start ROCm I always get:
With auto VRAM BIOS configuration I get that earlier, around 58 or something ... with
--ssd-streamingit works but it's noticeably slower than it used to be before ... is Strix Halo these days not a q2 (reduced) target anymore with its VRAM?Thank you!