]> git.djapps.eu Git - pkg/ggml/sources/ggml/commit
ggml : fix SpaceMit IME array out-of-bounds in task assignment (llama/16629)
authormuggle-stack <redacted>
Fri, 17 Oct 2025 10:01:23 +0000 (18:01 +0800)
committerGeorgi Gerganov <redacted>
Tue, 21 Oct 2025 15:14:33 +0000 (18:14 +0300)
commit40f00112b25ef1feb2b5a7ed6360ae5cc4999899
tree555ed4906322942c384e6b9416fdcfc9cf8880d6
parent8c30a31813f0f36607cc33351a4fdf4933861e31
ggml : fix SpaceMit IME array out-of-bounds in task assignment (llama/16629)

Fix incorrect task-to-batch index calculation in the quantization phase.

The bug caused out-of-bounds access to qnbitgemm_args array when
compute_idx exceeded per_gemm_block_count_m, leading to invalid
pointer dereferences and SIGBUS errors.

Correctly map tasks to batches by dividing compute_idx by
per_gemm_block_count_m instead of block_size_m.

Example:
  batch_feature=1, gemm_m=30, block_size_m=4
  per_gemm_block_count_m = 8, task_count = 8

  Old: gemm_idx = 4/4 = 1 (out of bounds  New: gemm_idx = 4/8 = 0 (correct)

Tested on SpaceMit K1 RISC-V64 with qwen2.5:0.5b model.

Co-authored-by: muggle <redacted>
src/ggml-cpu/spacemit/ime.cpp