Commit 6f5c9ad21 for llama.cpp

commit 6f5c9ad21746f69bbb7f3cb295a90e53fe316963
Author: Ruben Ortlam <rortlam@redhat.com>
Date:   Fri Nov 20 18:11:59 2026 +0100

    vulkan: add support for COL2IM_1D misalign bytes and split workgroups (#32673)

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan-push-constants.h b/ggml/src/ggml-vulkan/ggml-vulkan-push-constants.h
index 0a04c1d56..7a48c08dc 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan-push-constants.h
+++ b/ggml/src/ggml-vulkan/ggml-vulkan-push-constants.h
@@ -695,6 +695,7 @@ struct vk_op_col2im_1d_push_constants {
     uint32_t K;
     int32_t  stride;
     int32_t  p0;
+    uint32_t src_offset;
 };

 struct vk_op_conv_transpose_1d_push_constants {
@@ -1032,6 +1033,15 @@ template <> inline void init_pushconst_tensor_offsets(ggml_backend_vk_context *
     GGML_UNUSED(src3);
 }

+template <> inline void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_col2im_1d_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst) {
+    p.src_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
+    GGML_ASSERT(get_misalign_bytes(ctx, dst) == 0);
+
+    GGML_UNUSED(src1);
+    GGML_UNUSED(src2);
+    GGML_UNUSED(src3);
+}
+
 template <> inline void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_dsv4_hc_comb_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst) {
     p.m_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
     p.s_offset = get_misalign_bytes(ctx, src1) / ggml_type_size(src1->type);
diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index a9d60d6af..cc0071033 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -10610,6 +10610,12 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     case GGML_OP_COL2IM_1D:
         {
             elements = { uint32_t(dst->ne[0]), uint32_t(dst->ne[1]), 1 };
+            const uint32_t groups_x = CEIL_DIV(elements[0], pipeline->wg_denoms[0]);
+            const uint32_t max_groups_x = ctx->device->properties.limits.maxComputeWorkGroupCount[0];
+            if (groups_x > max_groups_x) {
+                elements[2] = CEIL_DIV(groups_x, max_groups_x);
+                elements[0] = CEIL_DIV(groups_x, elements[2]) * pipeline->wg_denoms[0];
+            }
         } break;
     case GGML_OP_POOL_1D:
         {
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/col2im_1d.comp b/ggml/src/ggml-vulkan/vulkan-shaders/col2im_1d.comp
index a23de380f..ba5054485 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/col2im_1d.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/col2im_1d.comp
@@ -15,10 +15,12 @@ layout (push_constant) uniform parameter {
     uint32_t K;
     int32_t  stride;
     int32_t  p0;
+    uint32_t src_offset;
 } p;

 // Load A_TYPE to float
 float load_col(uint32_t idx) {
+    idx += p.src_offset;
 #if defined(DATA_A_BF16)
     return bf16_to_fp32(uint32_t(data_a[idx]));
 #else
@@ -36,7 +38,7 @@ void store_dst(uint32_t idx, float v) {
 }

 void main() {
-    const uint32_t t_out = gl_GlobalInvocationID.x;
+    const uint32_t t_out = gl_GlobalInvocationID.x + gl_GlobalInvocationID.z * gl_NumWorkGroups.x * gl_WorkGroupSize.x;
     const uint32_t oc    = gl_GlobalInvocationID.y;
     if (t_out >= p.T_out || oc >= p.OC) return;