Seem to be having a huge slowdown since the recent updates. Had to roll back to version 2.7.0 to gain my speeds again with Minimax H3. Gemini wrote this issue report. I can confirm that rolling back to 2.7.0 and disabling all the MMGP stuff fixed my issue and I'm back to normal speeds. Disabling MMGP on version 3.1 didn't work, was still 3x slower. Gemini says it's a bug with MMGP on 16gb cards.
Performance Regression on 16GB GPUs (RTX 5080)
Issue: Updating to v3.0 / v3.1 causes generation times on the MiniMax H3 model to drop significantly (3x slower), going from ~1 hour per generation to over 4 hours (step times ballooning from ~3 mins to 6+ mins).
Hardware:
• GPU: Nvidia RTX 5080 (16GB VRAM)
• RAM: 64GB System RAM
Details:
Disabling MMGP or switching backend profiles back to raw PyTorch in v3.1 does not resolve the slowdown. Rolling the codebase back entirely bypasses the issue and restores normal generation tracking speeds, indicating that the new MMGP memory planning scheduler / weight-shuttling logic is over-allocating or aggressively paging out VRAM on 16GB cards, severely degrading rendering performance.
Seem to be having a huge slowdown since the recent updates. Had to roll back to version 2.7.0 to gain my speeds again with Minimax H3. Gemini wrote this issue report. I can confirm that rolling back to 2.7.0 and disabling all the MMGP stuff fixed my issue and I'm back to normal speeds. Disabling MMGP on version 3.1 didn't work, was still 3x slower. Gemini says it's a bug with MMGP on 16gb cards.
Performance Regression on 16GB GPUs (RTX 5080)
Issue: Updating to v3.0 / v3.1 causes generation times on the MiniMax H3 model to drop significantly (3x slower), going from ~1 hour per generation to over 4 hours (step times ballooning from ~3 mins to 6+ mins).
Hardware:
• GPU: Nvidia RTX 5080 (16GB VRAM)
• RAM: 64GB System RAM
Details:
Disabling MMGP or switching backend profiles back to raw PyTorch in v3.1 does not resolve the slowdown. Rolling the codebase back entirely bypasses the issue and restores normal generation tracking speeds, indicating that the new MMGP memory planning scheduler / weight-shuttling logic is over-allocating or aggressively paging out VRAM on 16GB cards, severely degrading rendering performance.