Single 3090 with Q4 Qwen 27B, context dropped from 137k to 14k with MTP enabled. Is it normal?

In three linesUser reports dramatic context size drop (137k → 14k) when enabling MTP (Multi-Token Prediction) with speculative decoding on Qwen 27B Q4 in llama.cpp. Asks if this behavior is expected.

Summary generated by Claude — human-verified