Fix Vulkan DeepSeek-Coder-V2-Lite MoE support

This commit is contained in:
0cc4m 2024-07-21 10:58:05 +02:00
parent 6274b3f835
commit 3252afb323
2 changed files with 2 additions and 4 deletions

View file

@ -3451,7 +3451,7 @@ 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 <= 2048);
GGML_ASSERT(nei0 * nei1 <= 3072);
const uint32_t nbi1 = ids->nb[1];
const uint32_t nbi2 = ids->nb[2];
@ -3463,8 +3463,6 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context *
const uint64_t n_as = ne02;
GGML_ASSERT(n_as <= 8);
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra;
ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra;
ggml_tensor_extra_gpu * extra_src1 = (ggml_tensor_extra_gpu *) src1->extra;

View file

@ -71,7 +71,7 @@ shared FLOAT_TYPE buf_a[BM * (BK+1)];
shared FLOAT_TYPE buf_b[BN * (BK+1)];
#ifdef MUL_MAT_ID
shared u16vec2 row_ids[2048];
shared u16vec2 row_ids[3072];
#endif
void main() {