ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages 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
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 5 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

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)