#28860·llama.cpp

Eval 问题: 当启用 ngram-mod 时, SYCL 需要极大的临时存储空间分配 (2GB+)。

作者: jintakhan创建于 2026年9月13日更新于 2026年9月17日
标签bug-unconfirmed
日志
5.40.399.350 D que  yield_to_que: done yielding to queue
5.40.399.358 D que  yield_to_que: yielding to queue
5.40.399.392 D que  process_new_: processing task, id = 37
5.40.399.393 D srv  process_sing: decoding, decline task, id_task = 37
5.40.399.393 D que  process_new_: task declined, id = 37
5.40.399.419 D que  yield_to_que: done yielding to queue
5.40.400.137 D res          send: sending result for task id = 8
5.40.400.144 D res          send: task id = 8 pushed to result queue
5.40.400.151 D slot process_toke: id  1 | task 8 | n_gen = 15, n_remaining = -1, next token:  4243 'cd'
5.40.400.152 D que    start_loop: waiting for new tasks
5.40.400.152 D que    start_loop: processing new tasks
5.40.400.153 D que  process_new_: processing task, id = 37
5.40.400.153 D que    start_loop: update slots
5.40.400.153 D srv  update_slots: posting NEXT_RESPONSE
5.40.400.155 D que          post: new task, id = 38, front = 0
5.40.400.162 D slot get_n_draft_: id  1 | task 8 | max possible draft: 96515
5.40.400.180 D que  yield_to_que: yielding to queue
5.40.400.201 D spec common_specu: called impl ngram-mod, hist size = 34555, call_count = 15, gen = 59
5.40.400.239 D que  process_new_: processing task, id = 38
5.40.400.240 D srv  process_sing: decoding, decline task, id_task = 38
5.40.400.240 D que  process_new_: task declined, id = 38
5.40.400.268 D que  yield_to_que: done yielding to queue
5.40.400.390 D srv    operator(): http: streamed chunk: data:
…

内容来源: ggml-org/llama.cpp