# Known CUDA Regressions Affecting RTX 2080 Ti (sm_75)

## Issue #24514: 25% Performance Drop b9301 → b9305

- **Status:** Open (unconfirmed)
- **Affects:** All CUDA users with MTP spec decoding
- **Symptoms:** Generation speed drops ~25%, prompt eval collapses in extreme cases (584→0.4 t/s)
- **First bad build:** b9305
- **Impact on our fleet:** Gemma-12b went from 117.8 to 96.2 t/s (~18% drop)
- **Workaround:** None — upstream bug, not configurable

## Issue #24670: draft-mtp Not Activating on Turing (sm_75)

- **Status:** Open
- **Affects:** All sm_75 GPUs (GTX 1650 SUPER, RTX 2080 Ti) with hybrid SSM+attention models
- **Symptoms:** draft-mtp initializes but never activates; generation speed matches no-spec baseline
- **Root cause:** Draft MTP context has no recurrent state (n_rs_seq=0) while main context has n_rs_seq=2
- **Workaround:** None — upstream bug, sm_75 specific

## Issue #24066: Vulkan Performance Drop (not CUDA)

- Not applicable to our CUDA setup
- Included for reference: Vulkan users on AMD saw similar symptoms