ExternalVendorCode/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-05-28 21:14:11 +00:00
Code Issues Actions 4 Packages Projects Releases Wiki Activity

vulkan: Allow up to 4096 elements for mul_mat_id row_ids (llama/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 Georgi Gerganov
parent e8a7f1b7bb
commit e46df4850f
4 changed files with 5 additions and 5 deletions
Show Stats 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];

2
ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
View File

@ -103,7 +103,7 @@ shared FLOAT_TYPE buf_a[BM * SHMEM_STRIDE];
shared FLOAT_TYPE buf_b[BN * SHMEM_STRIDE];
#ifdef MUL_MAT_ID
shared u16vec2 row_ids[3072];
shared u16vec2 row_ids[4096];
#endif // MUL_MAT_ID
#define NUM_WARPS (BLOCK_SIZE / WARP)

2
ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp
View File

@ -92,7 +92,7 @@ layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
#ifdef MUL_MAT_ID
layout (binding = 3) readonly buffer IDS {int data_ids[];};
shared u16vec4 row_ids[3072];
shared u16vec4 row_ids[4096];
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufB {
B_TYPE b[];

2
ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp
View File

@ -101,7 +101,7 @@ shared FLOAT_TYPE_VEC2 buf_b_ds[BN];
#define LOAD_VEC_B 4
#ifdef MUL_MAT_ID
shared u16vec2 row_ids[3072];
shared u16vec2 row_ids[4096];
#endif // MUL_MAT_ID
#define NUM_WARPS (BLOCK_SIZE / WARP)