SIGN IN SIGN UP

fix: seq_rm in recurrent memory should fall through on rollback failure

The llama_memory_recurrent::seq_rm function was returning false when
a partial rollback failed, causing GGML_ABORT in common_context_seq_rm.
This crashed the server when streaming + tools + conversation history
with reasoning were used together.

The partial rollback is an optimization for recurrent models (Mamba/RWKV).
When it fails, the code should fall through to the regular cell clearing
loop which handles this case correctly.

This fixes the crash on Qwen3.6-35B with streaming + tools + history.
F
fewtarius committed
b83d23022d1252c1c005e95e44db2cdc9fd154f8
Parent: b9ed083