CUDA: fix FA FP16 accumulator overflow for Granite (#18614)

* Support Youtu-VL Model
---------

Co-authored-by: Xuan-Son Nguyen <son@huggingface.co>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

convert_hf_to_gguf.py

@@ -7212,6 +7212,7 @@ class DeepseekModel(TextModel):
     "DeepseekV3ForCausalLM",
     "KimiVLForConditionalGeneration",
     "YoutuForCausalLM",
+    "YoutuVLForConditionalGeneration"
 )
 class DeepseekV2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.DEEPSEEK2
@@ -10674,8 +10675,8 @@ class JanusProVisionModel(MmprojModel):
         return []
 
 
-@ModelBase.register("YOUTUVLForConditionalGeneration", "YOUTUVLForCausalLM")
-class YOUTUVLVisionModel(MmprojModel):
+@ModelBase.register("YoutuVLForConditionalGeneration")
+class YoutuVLVisionModel(MmprojModel):
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
         assert self.hparams_vision is not None

ggml/src/ggml-cuda/common.cuh

@@ -1065,6 +1065,7 @@ struct ggml_cuda_graph {
     int number_consecutive_updates = 0;
     bool cuda_graphs_enabled = false;
     std::vector<ggml_graph_node_properties> ggml_graph_properties;
+    std::vector<ggml_graph_node_properties> extraneous_srcs_properties;
 #endif
 };
 

ggml/src/ggml-cuda/fattn-common.cuh

@@ -11,10 +11,12 @@
 #define SOFTMAX_FTZ_THRESHOLD -20.0f                   // Softmax exp. of values smaller than this are flushed to zero to avoid NaNs.
 
 // log(2) = 0.6931, by adding this to the KQ maximum used for the softmax the numerical range representable
-//     by the VKQ accumulators is effectively being shifted up by a factor of 8.
+//     by the VKQ accumulators is effectively being shifted up by a factor of 2.
 // This reduces issues with numerical overflow but also causes larger values to be flushed to zero.
 // However, as the output from FlashAttention will usually be used as an input for a matrix multiplication this should be negligible.
-#define FATTN_KQ_MAX_OFFSET 0.6931f
+// Still, the value range should be shifted as much as necessary but as little as possible.
+// The macro on the following line shifts it by a factor of 2**3=8, as was needed to fix https://github.com/ggml-org/llama.cpp/issues/18606 .
+#define FATTN_KQ_MAX_OFFSET (3.0f*0.6931f)
 
 typedef void (* fattn_kernel_t)(
         const char * __restrict__ Q,

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -2973,15 +2973,16 @@ static bool is_cuda_graph_update_required(ggml_backend_cuda_context * cuda_ctx,
     }
 
     // Check if the graph size has changed
-    if (cuda_ctx->cuda_graph->ggml_graph_properties.size() != (size_t)cgraph->n_nodes) {
+    if (cuda_ctx->cuda_graph->ggml_graph_properties.size() != (size_t)cgraph->n_nodes + cgraph->n_leafs) {
         cuda_graph_update_required = true;
-        cuda_ctx->cuda_graph->ggml_graph_properties.resize(cgraph->n_nodes);
+        cuda_ctx->cuda_graph->ggml_graph_properties.resize(cgraph->n_nodes + cgraph->n_leafs);
     }
 
     // Loop over nodes in GGML graph to determine if CUDA graph update is required
     // and store properties to allow this comparison for the next token
     for (int i = 0; i < cgraph->n_nodes; i++) {
         bool has_matching_properties = true;
+
         if (!cuda_graph_update_required) {
             has_matching_properties = ggml_graph_node_has_matching_properties(cgraph->nodes[i], &cuda_ctx->cuda_graph->ggml_graph_properties[i]);
         }
@@ -2991,6 +2992,17 @@ static bool is_cuda_graph_update_required(ggml_backend_cuda_context * cuda_ctx,
         set_ggml_graph_node_properties(cgraph->nodes[i], &cuda_ctx->cuda_graph->ggml_graph_properties[i]);
     }
 
+    for (int i = 0; i < cgraph->n_leafs; i++) {
+        bool has_matching_properties = true;
+        if (!cuda_graph_update_required) {
+            has_matching_properties = ggml_graph_node_has_matching_properties(cgraph->leafs[i], &cuda_ctx->cuda_graph->ggml_graph_properties[cgraph->n_nodes + i]);
+        }
+        if (!has_matching_properties) {
+            cuda_graph_update_required = true;
+        }
+        set_ggml_graph_node_properties(cgraph->leafs[i], &cuda_ctx->cuda_graph->ggml_graph_properties[cgraph->n_nodes + i]);
+    }
+
     return cuda_graph_update_required;
 }
 

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

@@ -2898,39 +2898,41 @@ static void ggml_vk_load_shaders(vk_device& device) {
         const uint32_t tk_m = device->coopmat_support ? device->coopmat_k : 1;
         const uint32_t tk_s = device->coopmat_support ? device->coopmat_k : 1;
 
-        l_warptile = { 128, 128, 128, 16, subgroup_size_8 * 2, 64, 2, tm_l, tn_l, tk_l, subgroup_size_8 };
-        m_warptile = { 128,  64,  64, 16, subgroup_size_8, 32, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
-        s_warptile = { subgroup_size_16, 32, 32, 16, 32, 32, 2, tm_s, tn_s, tk_s, subgroup_size_8 };
+        const uint32_t s_warptile_wm = device->subgroup_size == 8 ? 8 : 32;
 
-        l_warptile_mmq = { 128, 128, 128, 32, subgroup_size_8 * 2, 64, 2, tm_l, tn_l, tk_l, subgroup_size_8 };
-        m_warptile_mmq = { 128,  64,  64, 32, subgroup_size_8, 32, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
-        s_warptile_mmq = { subgroup_size_32, 32, 32, 32, 32, 32, 2, tm_s, tn_s, tk_s, subgroup_size_8 };
+        l_warptile = { 128,             128, 128, 16, subgroup_size_8 * 2, 64, 2, tm_l, tn_l, tk_l, subgroup_size_8 };
+        m_warptile = { 128,              64,  64, 16, subgroup_size_8,     32, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
+        s_warptile = { subgroup_size_32, 32,  32, 16, s_warptile_wm,       32, 2, tm_s, tn_s, tk_s, subgroup_size_8 };
+
+        l_warptile_mmq = { 128,             128, 128, 32, subgroup_size_8 * 2, 64, 2, tm_l, tn_l, tk_l, subgroup_size_8 };
+        m_warptile_mmq = { 128,              64,  64, 32, subgroup_size_8,     32, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
+        s_warptile_mmq = { subgroup_size_32, 32,  32, 32, s_warptile_wm,       32, 2, tm_s, tn_s, tk_s, subgroup_size_8 };
 
         // Integer MMQ has a smaller shared memory profile, but heavier register use
-        l_warptile_mmq_int = { 128, 128, 128, 32, subgroup_size_8 * 2, 64, 2, 4, 4, 1, subgroup_size_8 };
-        m_warptile_mmq_int = { 128,  64,  64, 32, subgroup_size_8,     32, 2, 2, 2, 1, subgroup_size_8 };
-        s_warptile_mmq_int = { subgroup_size_32, 32, 32, 32, 32,       32, 2, 2, 1, 1, subgroup_size_8 };
+        l_warptile_mmq_int = { 128,             128, 128, 32, subgroup_size_8 * 2, 64, 2, 4, 4, 1, subgroup_size_8 };
+        m_warptile_mmq_int = { 128,              64,  64, 32, subgroup_size_8,     32, 2, 2, 2, 1, subgroup_size_8 };
+        s_warptile_mmq_int = { subgroup_size_32, 32,  32, 32, s_warptile_wm,       32, 2, 2, 1, 1, subgroup_size_8 };
 
         // K-quants use even more registers, mitigate by setting WMITER to 1
-        l_warptile_mmq_int_k = { 128, 128, 128, 32, subgroup_size_8 * 2, 64, 1, 4, 4, 1, subgroup_size_8 };
-        m_warptile_mmq_int_k = { 128,  64,  64, 32, subgroup_size_8,     32, 1, 2, 2, 1, subgroup_size_8 };
-        s_warptile_mmq_int_k = { subgroup_size_32, 32, 32, 32, 32,       32, 1, 2, 1, 1, subgroup_size_8 };
+        l_warptile_mmq_int_k = { 128,               128, 128, 32, subgroup_size_8 * 2, 64, 1, 4, 4, 1, subgroup_size_8 };
+        m_warptile_mmq_int_k = { 128,                64,  64, 32, subgroup_size_8,     32, 1, 2, 2, 1, subgroup_size_8 };
+        s_warptile_mmq_int_k = { subgroup_size_32,   32,  32, 32, s_warptile_wm,       32, 1, 2, 1, 1, subgroup_size_8 };
 
-        l_warptile_id = { 128, 128, 128, 16, mul_mat_subgroup_size_16 * 2, 64, 2, tm_l, tn_l, tk_l, mul_mat_subgroup_size_16 };
-        m_warptile_id = { 128,  64,  64, 16, mul_mat_subgroup_size_16, 32, 2, tm_m, tn_m, tk_m, mul_mat_subgroup_size_16 };
-        s_warptile_id = { mul_mat_subgroup_size_16, 32, 32, 16, 32, 32, 2, tm_s, tn_s, tk_s, mul_mat_subgroup_size_16 };
+        l_warptile_id = { 128,                      128, 128, 16, mul_mat_subgroup_size_16 * 2, 64, 2, tm_l, tn_l, tk_l, mul_mat_subgroup_size_16 };
+        m_warptile_id = { 128,                       64,  64, 16, mul_mat_subgroup_size_16,     32, 2, tm_m, tn_m, tk_m, mul_mat_subgroup_size_16 };
+        s_warptile_id = { mul_mat_subgroup_size_16,  32,  32, 16, s_warptile_wm,                32, 2, tm_s, tn_s, tk_s, mul_mat_subgroup_size_16 };
 
-        l_warptile_mmqid = { 128, 128, 128, 32, mul_mat_subgroup_size_8 * 2, 64, 2, tm_l, tn_l, tk_l, mul_mat_subgroup_size_8 };
-        m_warptile_mmqid = { 128,  64,  64, 32, mul_mat_subgroup_size_8, 32, 2, tm_m, tn_m, tk_m, mul_mat_subgroup_size_8 };
-        s_warptile_mmqid = { mul_mat_subgroup_size_32, 32, 32, 32, 32, 32, 2, tm_s, tn_s, tk_s, mul_mat_subgroup_size_8 };
+        l_warptile_mmqid = { 128,                       128, 128, 32, mul_mat_subgroup_size_8 * 2, 64, 2, tm_l, tn_l, tk_l, mul_mat_subgroup_size_8 };
+        m_warptile_mmqid = { 128,                        64,  64, 32, mul_mat_subgroup_size_8,     32, 2, tm_m, tn_m, tk_m, mul_mat_subgroup_size_8 };
+        s_warptile_mmqid = { mul_mat_subgroup_size_32,   32,  32, 32, s_warptile_wm,               32, 2, tm_s, tn_s, tk_s, mul_mat_subgroup_size_8 };
 
-        l_warptile_mmqid_int = { 128, 128, 128, 32, mul_mat_subgroup_size_8 * 2, 64, 2, 4, 4, 1, mul_mat_subgroup_size_8 };
-        m_warptile_mmqid_int = { 128,  64,  64, 32, mul_mat_subgroup_size_8,     32, 2, 2, 2, 1, mul_mat_subgroup_size_8 };
-        s_warptile_mmqid_int = { mul_mat_subgroup_size_32, 32, 32, 32, 32,       32, 2, 2, 1, 1, mul_mat_subgroup_size_8 };
+        l_warptile_mmqid_int = { 128,                       128, 128, 32, mul_mat_subgroup_size_8 * 2, 64, 2, 4, 4, 1, mul_mat_subgroup_size_8 };
+        m_warptile_mmqid_int = { 128,                        64,  64, 32, mul_mat_subgroup_size_8,     32, 2, 2, 2, 1, mul_mat_subgroup_size_8 };
+        s_warptile_mmqid_int = { mul_mat_subgroup_size_32,   32,  32, 32, s_warptile_wm,               32, 2, 2, 1, 1, mul_mat_subgroup_size_8 };
 
-        l_warptile_mmqid_int_k = { 128, 128, 128, 32, mul_mat_subgroup_size_16 * 2, 64, 1, 4, 4, 1, mul_mat_subgroup_size_16 };
-        m_warptile_mmqid_int_k = { 128,  64,  64, 32, mul_mat_subgroup_size_16,     32, 1, 2, 2, 1, mul_mat_subgroup_size_16 };
-        s_warptile_mmqid_int_k = { mul_mat_subgroup_size_32, 32, 32, 32, 32,       32, 1, 2, 1, 1, mul_mat_subgroup_size_16 };
+        l_warptile_mmqid_int_k = { 128,                     128, 128, 32, mul_mat_subgroup_size_16 * 2, 64, 1, 4, 4, 1, mul_mat_subgroup_size_16 };
+        m_warptile_mmqid_int_k = { 128,                      64,  64, 32, mul_mat_subgroup_size_16,     32, 1, 2, 2, 1, mul_mat_subgroup_size_16 };
+        s_warptile_mmqid_int_k = { mul_mat_subgroup_size_32, 32,  32, 32, s_warptile_wm,                32, 1, 2, 1, 1, mul_mat_subgroup_size_16 };
 
         // chip specific tuning
         if ((device->architecture == AMD_GCN) && (device->driver_id != vk::DriverId::eAmdProprietary)) {
@@ -6773,7 +6775,12 @@ static void ggml_vk_quantize_q8_1(ggml_backend_vk_context * ctx, vk_context& sub
 
     vk_pipeline pipeline = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1);
 
-    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, std::array<uint32_t, 1>{ne}, { ne, 1, 1 });
+    const uint32_t num_blocks = CEIL_DIV(ne, pipeline->wg_denoms[0]);
+    // clamp the number of elements to the max workgroup count. The shader will iterate over the total number of blocks.
+    const uint64_t max_elements = std::min<uint64_t>(uint64_t{ctx->device->properties.limits.maxComputeWorkGroupCount[0]} * pipeline->wg_denoms[0], std::numeric_limits<uint32_t>::max());
+    const uint32_t elements = std::min(ne, static_cast<uint32_t>(max_elements));
+
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, std::array<uint32_t, 2>{ ne, num_blocks }, { elements, 1, 1 });
     ggml_vk_sync_buffers(ctx, subctx);
 }
 

ggml/src/ggml-vulkan/vulkan-shaders/quantize_q8_1.comp

@@ -15,6 +15,7 @@
 layout (push_constant) uniform parameter
 {
     uint ne;
+    uint num_blocks;
 } p;
 
 #include "types.glsl"
@@ -33,8 +34,7 @@ layout (binding = 1) writeonly buffer D {block_q8_1_x4 data_b[];};
 shared float shmem[GROUP_SIZE];
 #endif
 
-void quantize() {
-    const uint wgid = gl_WorkGroupID.x;
+void quantize(const uint wgid) {
     const uint tid = INVOCATION_ID;
 
     // Each thread handles a vec4, so 8 threads handle a block
@@ -45,11 +45,7 @@ void quantize() {
     const uint ib = wgid * blocks_per_group + block_in_wg;
     const uint iqs = tid % 8;
 
-#ifndef QBLOCK_X4
-    if (ib >= gl_NumWorkGroups.x * blocks_per_group) {
-        return;
-    }
-#else
+#ifdef QBLOCK_X4
     const uint ibx4_outer = ib / 4;
     const uint ibx4_inner = ib % 4;
 
@@ -123,5 +119,9 @@ void quantize() {
 }
 
 void main() {
-    quantize();
+    uint wgid = gl_WorkGroupID.x;
+    while (wgid < p.num_blocks) {
+        quantize(wgid);
+        wgid += gl_NumWorkGroups.x;
+    }
 }

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

@@ -101,6 +101,10 @@ void main() {
     const uint lane = gl_SubgroupInvocationID;
 
     float probs[experts_per_thread];
+    [[unroll]]
+    for (int i = 0; i < experts_per_thread; i++) {
+        probs[i] = -INFINITY;
+    }
 
     [[unroll]]
     for (uint i = 0; i < n_experts; i += WARP_SIZE) {
@@ -112,8 +116,9 @@ void main() {
         softmax_warp_inplace(probs, n_experts, lane, nexperts_use_push);
     } else if (gating_func == GATING_FUNC_SIGMOID) {
         [[unroll]]
-        for (int i = 0; i < experts_per_thread; i++) {
-            probs[i] = 1.f / (1.f + exp(-probs[i]));
+        for (uint i = 0; i < n_experts; i += WARP_SIZE) {
+            const uint expert = i + lane;
+            probs[i / WARP_SIZE] = (n_experts % WARP_SIZE == 0 || expert < n_experts) ? 1.f / (1.f + exp(-probs[i / WARP_SIZE])) : -INFINITY;
         }
     }
 
@@ -150,11 +155,11 @@ void main() {
         uint   max_expert = lane;
 
         [[unroll]]
-        for (int i = 1; i < experts_per_thread; i++) {
-            const uint expert = lane + i * WARP_SIZE;
-            if ((n_experts % WARP_SIZE == 0 || expert < n_experts) && selection_probs[i] > max_val_s) {
-                max_val    = probs[i];
-                max_val_s  = selection_probs[i];
+        for (uint i = WARP_SIZE; i < n_experts; i += WARP_SIZE) {
+            const uint expert = i + lane;
+            if ((n_experts % WARP_SIZE == 0 || expert < n_experts) && selection_probs[i / WARP_SIZE] > max_val_s) {
+                max_val    = probs[i / WARP_SIZE];
+                max_val_s  = selection_probs[i / WARP_SIZE];
                 max_expert = expert;
             }
         }

src/llama-graph.cpp

@@ -1326,6 +1326,10 @@ ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
 
     res->add_input(std::move(inp));
 
+    // make sure the produced embeddings are immediately materialized in the ggml graph
+    // ref: https://github.com/ggml-org/llama.cpp/pull/18599
+    ggml_build_forward_expand(gf, cur);
+
     return cur;
 }
 

tests/test-backend-ops.cpp

@@ -7625,6 +7625,10 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 64, 77, 77, {12,1}, {1,1}));
 
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q4_0, GGML_TYPE_F32, 576, 512, 576, {1,1}, {1,1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q4_0, GGML_TYPE_F32, 1, 2048, 8192, {1,  1}, {1, 1}));
+    for (ggml_type type_a : all_types) {
+        test_cases.emplace_back(new test_mul_mat(type_a, GGML_TYPE_F32, 1, 64, 256, {1,  1}, {1, 1}));
+    }
 
 #if 0
     // test the mat-mat path for Metal
@@ -8180,6 +8184,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
                     test_cases.emplace_back(new test_topk_moe({71, 22, 1, 1}, 8, with_norm, bias_probs, gate, scale_w));
                     test_cases.emplace_back(new test_topk_moe({128, 1, 1, 1}, 128, with_norm, bias_probs, gate, scale_w));
                     test_cases.emplace_back(new test_topk_moe({129, 1, 1, 1}, 128, with_norm, bias_probs, gate, scale_w));
+                    test_cases.emplace_back(new test_topk_moe({160, 4, 1, 1}, 160, with_norm, bias_probs, gate, scale_w));
                 }
             }
         }

tools/server/server-models.cpp

@@ -21,11 +21,13 @@
 
 #ifdef _WIN32
 #include <winsock2.h>
+#include <windows.h>
 #else
 #include <sys/socket.h>
 #include <netinet/in.h>
 #include <arpa/inet.h>
 #include <unistd.h>
+extern char **environ;
 #endif
 
 #if defined(__APPLE__) && defined(__MACH__)
@@ -99,6 +101,49 @@ static void unset_reserved_args(common_preset & preset, bool unset_model_args) {
     }
 }
 
+#ifdef _WIN32
+static std::string wide_to_utf8(const wchar_t * ws) {
+    if (!ws || !*ws) {
+        return {};
+    }
+
+    const int len = static_cast<int>(std::wcslen(ws));
+    const int bytes = WideCharToMultiByte(CP_UTF8, 0, ws, len, nullptr, 0, nullptr, nullptr);
+    if (bytes == 0) {
+        return {};
+    }
+
+    std::string utf8(bytes, '\0');
+    WideCharToMultiByte(CP_UTF8, 0, ws, len, utf8.data(), bytes, nullptr, nullptr);
+
+    return utf8;
+}
+#endif
+
+static std::vector<std::string> get_environment() {
+    std::vector<std::string> env;
+
+#ifdef _WIN32
+    LPWCH env_block = GetEnvironmentStringsW();
+    if (!env_block) {
+        return env;
+    }
+    for (LPWCH e = env_block; *e; e += wcslen(e) + 1) {
+        env.emplace_back(wide_to_utf8(e));
+    }
+    FreeEnvironmentStringsW(env_block);
+#else
+    if (environ == nullptr) {
+        return env;
+    }
+    for (char ** e = environ; *e != nullptr; e++) {
+        env.emplace_back(*e);
+    }
+#endif
+
+    return env;
+}
+
 void server_model_meta::update_args(common_preset_context & ctx_preset, std::string bin_path) {
     // update params
     unset_reserved_args(preset, false);
@@ -117,14 +162,11 @@ void server_model_meta::update_args(common_preset_context & ctx_preset, std::str
 server_models::server_models(
         const common_params & params,
         int argc,
-        char ** argv,
-        char ** envp)
+        char ** argv)
             : ctx_preset(LLAMA_EXAMPLE_SERVER),
               base_params(params),
+              base_env(get_environment()),
               base_preset(ctx_preset.load_from_args(argc, argv)) {
-    for (char ** env = envp; *env != nullptr; env++) {
-        base_env.push_back(std::string(*env));
-    }
     // clean up base preset
     unset_reserved_args(base_preset, true);
     // set binary path

tools/server/server-models.h

@@ -105,7 +105,7 @@ private:
     void add_model(server_model_meta && meta);
 
 public:
-    server_models(const common_params & params, int argc, char ** argv, char ** envp);
+    server_models(const common_params & params, int argc, char ** argv);
 
     void load_models();
 
@@ -147,8 +147,8 @@ struct server_models_routes {
     common_params params;
     json webui_settings = json::object();
     server_models models;
-    server_models_routes(const common_params & params, int argc, char ** argv, char ** envp)
-            : params(params), models(params, argc, argv, envp) {
+    server_models_routes(const common_params & params, int argc, char ** argv)
+            : params(params), models(params, argc, argv) {
         if (!this->params.webui_config_json.empty()) {
             try {
                 webui_settings = json::parse(this->params.webui_config_json);

tools/server/server.cpp

@@ -66,7 +66,7 @@ static server_http_context::handler_t ex_wrapper(server_http_context::handler_t
     };
 }
 
-int main(int argc, char ** argv, char ** envp) {
+int main(int argc, char ** argv) {
     // own arguments required by this example
     common_params params;
 
@@ -126,7 +126,7 @@ int main(int argc, char ** argv, char ** envp) {
     if (is_router_server) {
         // setup server instances manager
         try {
-            models_routes.emplace(params, argc, argv, envp);
+            models_routes.emplace(params, argc, argv);
         } catch (const std::exception & e) {
             LOG_ERR("%s: failed to initialize router models: %s\n", __func__, e.what());
             return 1;