Merge branch 'upstream' into concedo_experimental

# Conflicts:
#	.github/workflows/build-riscv-native.yml
#	.github/workflows/build.yml
#	ci/README.md
#	ci/run.sh
#	ggml/src/ggml-opencl/CMakeLists.txt
#	ggml/src/ggml-opencl/ggml-opencl.cpp
#	ggml/src/ggml-opencl/kernels/cvt.cl
#	tests/test-backend-ops.cpp

Author: LostRuins

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -1201,6 +1201,14 @@ struct vk_staging_memcpy {
     size_t n;
 };
 
+struct vk_staging_memset {
+    vk_staging_memset(void * _dst, uint32_t _val, size_t _n) : dst(_dst), val(_val), n(_n) {}
+
+    void * dst;
+    uint32_t val;
+    size_t n;
+};
+
 struct vk_context_struct {
     vk_submission * s;
     std::vector<vk_sequence> seqs;
@@ -1209,6 +1217,7 @@ struct vk_context_struct {
 
     std::vector<vk_staging_memcpy> in_memcpys;
     std::vector<vk_staging_memcpy> out_memcpys;
+    std::vector<vk_staging_memset> memsets;
 
     vk_command_pool * p {};
 };
@@ -1600,7 +1609,9 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin
     }
 
     vk::ComputePipelineCreateInfo compute_pipeline_create_info(
-        vk::PipelineCreateFlags{},
+        device->pipeline_executable_properties_support ?
+            vk::PipelineCreateFlagBits::eCaptureStatisticsKHR :
+            vk::PipelineCreateFlags{},
         pipeline_shader_create_info,
         pipeline->layout);
 
@@ -3396,7 +3407,6 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_ ## name [0], #name "_f32", name ## _f32_len, name ## _f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);  \
     ggml_vk_create_pipeline(device, device->pipeline_ ## name [1], #name "_f16", name ## _f16_len, name ## _f16_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
 
-    CREATE_UNARY(exp)
     CREATE_UNARY(gelu)
     CREATE_UNARY(gelu_erf)
     CREATE_UNARY(gelu_quick)
@@ -3408,6 +3418,17 @@ static void ggml_vk_load_shaders(vk_device& device) {
     CREATE_UNARY(hardswish)
 #undef CREATE_UNARY
 
+#define CREATE_UNARY_RTE(name)  \
+    if (device->float_controls_rte_fp16) {  \
+        ggml_vk_create_pipeline(device, device->pipeline_ ## name [0], #name "_f32_rte", name ## _f32_rte_len, name ## _f32_rte_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);   \
+        ggml_vk_create_pipeline(device, device->pipeline_ ## name [1], #name "_f16_rte", name ## _f16_rte_len, name ## _f16_rte_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);   \
+    } else {    \
+        ggml_vk_create_pipeline(device, device->pipeline_ ## name [0], #name "_f32", name ## _f32_len, name ## _f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);   \
+        ggml_vk_create_pipeline(device, device->pipeline_ ## name [1], #name "_f16", name ## _f16_len, name ## _f16_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);   \
+    }
+    CREATE_UNARY_RTE(exp)
+#undef CREATE_UNARY_RTE
+
 #define CREATE_GLU(name)  \
     if (device->float_controls_rte_fp16) {  \
         ggml_vk_create_pipeline(device, device->pipeline_ ## name [0], #name "_f32_rte", name ## _f32_rte_len, name ## _f32_rte_data, "main", 3, sizeof(vk_op_glu_push_constants), {512, 1, 1}, {}, 1, true);   \
@@ -5224,6 +5245,14 @@ static void deferred_memcpy(void * dst, const void * src, size_t size, std::vect
     }
 }
 
+static void deferred_memset(void * dst, uint32_t val, size_t size, std::vector<vk_staging_memset>* memsets = nullptr) {
+    if (memsets == nullptr) {
+        memset(dst, val, size);
+    } else {
+        memsets->emplace_back(dst, val, size);
+    }
+}
+
 static void ggml_vk_ensure_sync_staging_buffer(vk_device& device, size_t size) {
     if (device->sync_staging == nullptr || device->sync_staging->size < size) {
         VK_LOG_MEMORY("ggml_vk_ensure_sync_staging_buffer(" << size << ")");
@@ -5419,6 +5448,10 @@ static void ggml_vk_buffer_write_2d(vk_buffer& dst, size_t offset, const void *
             memcpy(cpy.dst, cpy.src, cpy.n);
         }
 
+        for (auto& mset : subctx->memsets) {
+            memset(mset.dst, mset.val, mset.n);
+        }
+
         ggml_vk_submit(subctx, dst->device->fence);
         VK_CHECK(dst->device->device.waitForFences({ dst->device->fence }, true, UINT64_MAX), "vk_buffer_write_2d waitForFences");
         dst->device->device.resetFences({ dst->device->fence });
@@ -5558,12 +5591,25 @@ static void ggml_vk_buffer_copy(vk_buffer& dst, size_t dst_offset, vk_buffer& sr
 static void ggml_vk_buffer_memset_async(vk_context& ctx, vk_buffer& dst, size_t offset, uint32_t c, size_t size) {
     VK_LOG_DEBUG("ggml_vk_buffer_memset_async(" << offset << ", " << c << ", " << size << ")");
 
+    if (dst->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible &&
+        dst->device->uma) {
+        deferred_memset((uint8_t*)dst->ptr + offset, c, size, &ctx->memsets);
+        return;
+    }
+
+    // Fall back to GPU fillBuffer for non-UMA or non-host-visible buffers
     ctx->s->buffer.fillBuffer(dst->buffer, offset, size, c);
 }
 
 static void ggml_vk_buffer_memset(vk_buffer& dst, size_t offset, uint32_t c, size_t size) {
     VK_LOG_DEBUG("ggml_vk_buffer_memset(" << offset << ", " << c << ", " << size << ")");
 
+    if (dst->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible &&
+        dst->device->uma) {
+        memset((uint8_t*)dst->ptr + offset, c, size);
+        return;
+    }
+
     std::lock_guard<std::recursive_mutex> guard(dst->device->mutex);
     vk_context subctx = ggml_vk_create_temporary_context(dst->device->transfer_queue.cmd_pool);
     ggml_vk_ctx_begin(dst->device, subctx);
@@ -11198,6 +11244,10 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph *
             memcpy(cpy.dst, cpy.src, cpy.n);
         }
 
+        for (auto& mset : subctx->memsets) {
+            memset(mset.dst, mset.val, mset.n);
+        }
+
         if (almost_ready && !ctx->almost_ready_fence_pending && !use_fence) {
             ggml_vk_submit(subctx, ctx->almost_ready_fence);
             ctx->almost_ready_fence_pending = true;
@@ -11220,6 +11270,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph *
         }
         subctx->in_memcpys.clear();
         subctx->out_memcpys.clear();
+        subctx->memsets.clear();
     }
 
     return true;

ggml/src/ggml-vulkan/vulkan-shaders/exp.comp

@@ -1,5 +1,6 @@
 #version 450
 
+#include "rte.comp"
 #include "generic_head.comp"
 #include "types.comp"
 

ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp

@@ -31,10 +31,22 @@
 #include "types.comp"
 
 #ifndef LOAD_VEC_A
-#define LOAD_VEC_A 2
+#define LOAD_VEC_A 1
 #endif
 #ifndef LOAD_VEC_B
-#define LOAD_VEC_B 2
+#define LOAD_VEC_B 1
+#endif
+
+// Load 2 values at once without affecting index calculations through LOAD_VEC
+#if (defined(DATA_A_F32) || defined(DATA_A_F16) || defined(DATA_A_BF16)) && !defined(ALIGNED)
+#define LOAD_VEC_BATCH_A 2
+#else
+#define LOAD_VEC_BATCH_A 1
+#endif
+#if !defined(ALIGNED)
+#define LOAD_VEC_BATCH_B 2
+#else
+#define LOAD_VEC_BATCH_B 1
 #endif
 
 #if !defined(TO_FLOAT_TYPE)
@@ -236,13 +248,13 @@ void main() {
     const uint warp_r = warp_i % (BM / WM);
     const uint warp_c = warp_i / (BM / WM);
 
-    const uint loadr_a = gl_LocalInvocationID.x % (BK / LOAD_VEC_A);
-    const uint loadc_a = gl_LocalInvocationID.x / (BK / LOAD_VEC_A);
-    const uint loadr_b = gl_LocalInvocationID.x % (BK / LOAD_VEC_B);
-    const uint loadc_b = gl_LocalInvocationID.x / (BK / LOAD_VEC_B);
+    const uint loadr_a = gl_LocalInvocationID.x % (BK / LOAD_VEC_A / LOAD_VEC_BATCH_A);
+    const uint loadc_a = gl_LocalInvocationID.x / (BK / LOAD_VEC_A / LOAD_VEC_BATCH_A);
+    const uint loadr_b = gl_LocalInvocationID.x % (BK / LOAD_VEC_B / LOAD_VEC_BATCH_B);
+    const uint loadc_b = gl_LocalInvocationID.x / (BK / LOAD_VEC_B / LOAD_VEC_BATCH_B);
 
-    const uint loadstride_a = gl_WorkGroupSize.x * LOAD_VEC_A / BK;
-    const uint loadstride_b = gl_WorkGroupSize.x * LOAD_VEC_B / BK;
+    const uint loadstride_a = gl_WorkGroupSize.x * LOAD_VEC_A * LOAD_VEC_BATCH_A / BK;
+    const uint loadstride_b = gl_WorkGroupSize.x * LOAD_VEC_B * LOAD_VEC_BATCH_B / BK;
 
 #ifdef MUL_MAT_ID
 #ifdef MUL_MAT_ID_USE_SUBGROUPS

ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.comp

@@ -14,8 +14,8 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
             FLOAT_TYPE_VEC4 aa = FLOAT_TYPE_VEC4(data_a[idx]);
             buf_a[buf_idx    ] = aa.xy;
             buf_a[buf_idx + 1] = aa.zw;
-#else // LOAD_VEC_A == 2
-            const uint idx = pos_a * 2 + col * p.stride_a + row * 2;
+#else // LOAD_VEC_BATCH_A == 2
+            const uint idx = pos_a + col * p.stride_a + row * 2;
             const uint buf_idx = col * SHMEM_STRIDE + row;
             if (idx_m < p.M && block + row * 2 + 1 < end_k) {
                 buf_a[buf_idx] = FLOAT_TYPE_VEC2(data_a[idx],
@@ -33,8 +33,8 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
             FLOAT_TYPE_VEC4 aa = FLOAT_TYPE_VEC4(TO_FLOAT_TYPE(data_a[idx]));
             buf_a[buf_idx    ] = aa.xy;
             buf_a[buf_idx + 1] = aa.zw;
-#else // LOAD_VEC_A == 2
-            const uint idx = pos_a * 2 + col * p.stride_a + row * 2;
+#else // LOAD_VEC_BATCH_A == 2
+            const uint idx = pos_a + col * p.stride_a + row * 2;
             const uint buf_idx = col * SHMEM_STRIDE + row;
             if (idx_m < p.M && block + row * 2 + 1 < end_k) {
                 buf_a[buf_idx] = FLOAT_TYPE_VEC2(TO_FLOAT_TYPE(data_a[idx]),
@@ -500,8 +500,8 @@ void load_b_to_shmem(const uint pos_b, const uint row, const uint col, const uin
 #endif
             buf_b[buf_idx + 0] = bb.xy;
             buf_b[buf_idx + 1] = bb.zw;
-#else // LOAD_VEC_B == 2
-            const uint idx = pos_b * 2 + col * p.stride_b + row * 2;
+#else // LOAD_VEC_BATCH_B == 2
+            const uint idx = pos_b + col * p.stride_b + row * 2;
             const uint buf_idx = col * SHMEM_STRIDE + row;
             if (idx_n < p.N && block + row * 2 + 1 < end_k) {
                 buf_b[buf_idx] = FLOAT_TYPE_VEC2(TO_FLOAT_TYPE(data_b[idx]),
@@ -536,17 +536,17 @@ void load_b_to_shmem(const uint pos_b, const uint row, const uint col, const uin
 #endif
             buf_b[buf_idx + 0] = bb.xy;
             buf_b[buf_idx + 1] = bb.zw;
-#else // LOAD_VEC_B == 2
+#else // LOAD_VEC_BATCH_B == 2
             const uint row_i = ic * BN + col;
             const uint buf_idx = col * SHMEM_STRIDE + row;
             if (row_i < _ne1 && block + row * 2 + 1 < end_k) {
                 const u16vec2 row_idx = row_ids[col];
-                const uint idx = pos_b * 2 + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + row * 2;
+                const uint idx = pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + row * 2;
                 buf_b[buf_idx] = FLOAT_TYPE_VEC2(TO_FLOAT_TYPE(data_b[idx]),
                                                  TO_FLOAT_TYPE(data_b[idx + 1]));
             } else if (row_i < _ne1 && block + row * 2 < end_k) {
                 const u16vec2 row_idx = row_ids[col];
-                const uint idx = pos_b * 2 + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + row * 2;
+                const uint idx = pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + row * 2;
                 buf_b[buf_idx] = FLOAT_TYPE_VEC2(TO_FLOAT_TYPE(data_b[idx]), 0.0f);
             } else {
                 buf_b[buf_idx] = FLOAT_TYPE_VEC2(0.0f);

ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -468,7 +468,7 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
 
         std::string data_a_key = "DATA_A_" + to_uppercase(tname);
         // For unaligned, load one at a time for f32/f16, or two at a time for quants
-        std::string load_vec_a_unaligned = coopmat2 ? "1" : (tname == "f32" || tname == "f16" || tname == "bf16") ? "2" : load_vec_quant;
+        std::string load_vec_a_unaligned = (coopmat2 || tname == "f32" || tname == "f16" || tname == "bf16") ? "1" : load_vec_quant;
         // For aligned matmul loads
         std::string load_vec_a = (coopmat2 || tname == "f32" || tname == "f16" || tname == "bf16") ? load_vec : load_vec_quant;
 
@@ -718,8 +718,11 @@ void process_shaders() {
 
     string_to_spv("upscale_f32", "upscale.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
 
-    string_to_spv("exp_f16",        "exp.comp",         {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
-    string_to_spv("exp_f32",        "exp.comp",         {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
+    for (auto rte : {false, true}) {
+        std::string suffix = rte ? "_rte" : "";
+        string_to_spv("exp_f16" + suffix,        "exp.comp",         {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"},   {"RTE16", rte ? "1" : "0"}});
+        string_to_spv("exp_f32" + suffix,        "exp.comp",         {{"A_TYPE", "float"},       {"D_TYPE", "float"}    ,   {"RTE16", rte ? "1" : "0"}});
+    }
     string_to_spv("gelu_f16",       "gelu.comp",        {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
     string_to_spv("gelu_f32",       "gelu.comp",        {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
     string_to_spv("gelu_erf_f16",   "gelu_erf.comp",    {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});

koboldcpp.py

@@ -64,7 +64,7 @@
 extra_images_max = 4
 
 # global vars
-KcppVersion = "1.99.1"
+KcppVersion = "1.99.2"
 showdebug = True
 kcpp_instance = None #global running instance
 global_memory = {"tunnel_url": "", "restart_target":"", "input_to_exit":False, "load_complete":False, "restart_override_config_target":""}

tools/main/main.cpp

@@ -179,7 +179,7 @@ int main(int argc, char ** argv) {
             return 1;
         }
 
-        // Start the non-batch threadpool in the paused state
+        // start the non-batch threadpool in the paused state
         tpp.paused = true;
     }