happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

vulkan: intel mmv fix attempt

This commit is contained in:
0cc4m 2025-11-23 10:13:19 +01:00
parent 54d83bbe85
commit 722f9defe9
1 changed files with 7 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View File

@ -5618,7 +5618,7 @@ static vk_subbuffer ggml_vk_tensor_subbuffer(
const ggml_backend_vk_context * ctx, const ggml_tensor * tensor, bool allow_misalign = false) { const ggml_backend_vk_context * ctx, const ggml_tensor * tensor, bool allow_misalign = false) {
vk_buffer buffer = nullptr; vk_buffer buffer = nullptr;
size_t offset = 0; uint64_t offset = 0;
if (ctx->device->uma) { if (ctx->device->uma) {
ggml_vk_host_get(ctx->device, tensor->data, buffer, offset); ggml_vk_host_get(ctx->device, tensor->data, buffer, offset);
} }
@ -5629,9 +5629,9 @@ static vk_subbuffer ggml_vk_tensor_subbuffer(
} }
GGML_ASSERT(buffer != nullptr); GGML_ASSERT(buffer != nullptr);
size_t size = ggml_nbytes(tensor); uint64_t size = ggml_nbytes(tensor);
size_t misalign_bytes = offset & (ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1); uint64_t misalign_bytes = offset & (ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
// The shader must support misaligned offsets when indexing into the buffer // The shader must support misaligned offsets when indexing into the buffer
GGML_ASSERT(allow_misalign || misalign_bytes == 0); GGML_ASSERT(allow_misalign || misalign_bytes == 0);
offset &= ~misalign_bytes; offset &= ~misalign_bytes;
@ -6895,13 +6895,13 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
vk_subbuffer d_X, d_Y; vk_subbuffer d_X, d_Y;
if (qx_needs_dequant) { if (qx_needs_dequant) {
d_X = { ctx->prealloc_x, 0, ctx->prealloc_x->size }; d_X = { ctx->prealloc_x, 0, x_sz };
} else { } else {
d_X = d_Qx; d_X = d_Qx;
GGML_ASSERT(qx_sz == x_sz); GGML_ASSERT(qx_sz == x_sz);
} }
if (qy_needs_dequant || quantize_y) { if (qy_needs_dequant || quantize_y) {
d_Y = { ctx->prealloc_y, 0, ctx->prealloc_y->size }; d_Y = { ctx->prealloc_y, 0, y_sz };
} else { } else {
d_Y = d_Qy; d_Y = d_Qy;
} }
@ -7602,12 +7602,12 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
vk_subbuffer d_X, d_Y; vk_subbuffer d_X, d_Y;
if (qx_needs_dequant) { if (qx_needs_dequant) {
d_X = { ctx->prealloc_x, 0, ctx->prealloc_x->size }; d_X = { ctx->prealloc_x, 0, x_sz };
} else { } else {
d_X = d_Qx; d_X = d_Qx;
} }
if (qy_needs_dequant) { if (qy_needs_dequant) {
d_Y = { ctx->prealloc_y, 0, ctx->prealloc_y->size }; d_Y = { ctx->prealloc_y, 0, y_sz };
} else { } else {
d_Y = d_Qy; d_Y = d_Qy;
} }