Eval bug: crash when pooling_type == LLAMA_POOLING_TYPE_MEAN

[ivan-tkatchev]

Name and Version

Revision 9b169a4

Operating systems

Linux

GGML backends

CPU

Hardware

ARM Ampere

Models

Qwen2.5-14B-Instruct-1M-Q5_K_M

Problem description & steps to reproduce

Setting pooling_type = LLAMA_POOLING_TYPE_MEAN and calling llama_init_from_model() causes this crash:

/build/source/ggml/src/ggml.c:2738: GGML_ASSERT(ggml_can_mul_mat(a, b)) failed

Setting to LLAMA_POOLING_TYPE_LAST and changing nothing else works correctly.

First Bad Commit

No response

Relevant log output

/build/source/ggml/src/ggml.c:2738: GGML_ASSERT(ggml_can_mul_mat(a, b)) failed

[ivan-tkatchev]

Possibly a duplicate of #12517