tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-28 12:25:03 +00:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: Allow up to 4096 elements for mul_mat_id row_ids (#13326)

Browse Source 
This assert fired running Qwen_Qwen3-30B-A3B-Q2_K.gguf:

GGML_ASSERT(nei0 * nei1 <= 3072);

The tensor is 8 x 512. Increase this array size to accommodate.
This commit is contained in:
Jeff Bolz
2025-05-09 02:23:41 -05:00
committed by GitHub
parent d9c4accaff
commit 02115dcd9a
4 changed files with 5 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View File

@ -1632,7 +1632,7 @@ static bool ggml_vk_matmul_shmem_support(const vk_device& device, const std::vec
const uint32_t warps = warptile[0] / warptile[10];
const uint32_t load_bufs = (warptile[1] + warptile[2]) * (warptile[3] + bank_conflict_offset) * type_size;
const uint32_t mmid_row_ids = mul_mat_id ? 3072 * sizeof(uint32_t) : 0;
const uint32_t mmid_row_ids = mul_mat_id ? 4096 * sizeof(uint32_t) : 0;
const uint32_t coopmat_stage = device->coopmat_support ? warptile[7] * warptile[8] / warps * sizeof(float) : 0;
const uint32_t total_size = load_bufs + mmid_row_ids + coopmat_stage + lut_size;
@ -5260,7 +5260,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
const uint64_t nei0 = ids->ne[0];
const uint64_t nei1 = ids->ne[1];
GGML_ASSERT(nei0 * nei1 <= 3072);
GGML_ASSERT(nei0 * nei1 <= 4096);
const uint32_t nbi1 = ids->nb[1];
const uint32_t nbi2 = ids->nb[2];