vulkan: prefer host-visible memory buffers on UMA devices (#22930)

* implement UMA host-visible memory

* update based on 0cc4m's suggestion

common/speculative.cpp

@@ -140,6 +140,8 @@ struct common_speculative_impl {
     size_t n_gen_tokens = 0; // number of tokens generated by this implementation.
     size_t n_acc_tokens = 0; // number of tokens accepted by the target model.
 
+    std::vector<size_t> n_acc_tokens_per_pos; // number of tokens accepted per draft position.
+
     // TODO: track performance of most recent calls
     const bool gen_perf = true; // whether to generate performance stats.
 
@@ -2059,6 +2061,15 @@ void common_speculative_accept(common_speculative * spec, llama_seq_id seq_id, u
 
     {
         common_time_meas tm(impl->t_accept_us, !impl->gen_perf);
+
+        if (impl->n_acc_tokens_per_pos.size() < n_accepted) {
+            impl->n_acc_tokens_per_pos.resize(n_accepted, 0);
+        }
+
+        for (size_t i = 0; i < n_accepted; ++i) {
+            impl->n_acc_tokens_per_pos[i]++;
+        }
+
         if (n_accepted > 0) {
             impl->n_acc_drafts++;
             impl->n_acc_tokens += n_accepted;
@@ -2093,13 +2104,31 @@ void common_speculative_print_stats(const common_speculative * spec) {
             str_perf = "";
         }
 
-        LOG_INF("statistics %16s: #calls(b,g,a) = %4zu %6zu %6zu, #gen drafts = %6zu, #acc drafts = %5zu, #gen tokens = %6zu, #acc tokens = %5zu%s\n",
+        std::string str_stats;
+        if (impl->n_call_accept > 0) {
+            const double mean =
+                1.0 + (double) impl->n_acc_tokens / (double) impl->n_call_accept;
+            std::ostringstream tmp;
+            tmp << std::fixed << std::setprecision(3);
+            for (size_t i = 0; i < impl->n_acc_tokens_per_pos.size(); ++i) {
+                if (i > 0) {
+                    tmp << ", ";
+                }
+                tmp << (double) impl->n_acc_tokens_per_pos[i] / (double) impl->n_call_accept;
+            }
+            std::ostringstream oss;
+            oss << std::fixed << std::setprecision(2) << mean;
+            str_stats = ", #mean acc len = " + oss.str() + ", #acc rate/pos = (" + tmp.str() + ")";
+        }
+
+        LOG_INF("statistics %16s: #calls(b,g,a) = %4zu %6zu %6zu, #gen drafts = %6zu, #acc drafts = %5zu, #gen tokens = %6zu, #acc tokens = %5zu%s%s\n",
                 common_speculative_type_to_str(impl->type).c_str(),
                 impl->n_call_begin, impl->n_call_draft, impl->n_call_accept,
                 impl->n_gen_drafts,
                 impl->n_acc_drafts,
                 impl->n_gen_tokens,
                 impl->n_acc_tokens,
+                str_stats.c_str(),
                 str_perf.c_str());
     }
 }

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

@@ -911,8 +911,8 @@ struct vk_device_struct {
     vk_pipeline pipeline_pool2d_f32;
     vk_pipeline pipeline_rwkv_wkv6_f32;
     vk_pipeline pipeline_rwkv_wkv7_f32;
-    // [size_idx][kda] where size_idx: 0=d32, 1=d64, 2=d128
-    vk_pipeline pipeline_gated_delta_net[3][2];
+    // [size_idx][kda] where size_idx: 0=d16, 1=d32, 2=d64, 3=d128
+    vk_pipeline pipeline_gated_delta_net[4][2];
     vk_pipeline pipeline_ssm_scan_f32_d128;
     vk_pipeline pipeline_ssm_scan_f32_d256;
     vk_pipeline pipeline_ssm_conv_f32;
@@ -3080,8 +3080,10 @@ static vk_buffer ggml_vk_create_buffer_device(vk_device& device, size_t size) {
             buf = ggml_vk_create_buffer(device, size, {vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent,
                                                        vk::MemoryPropertyFlagBits::eDeviceLocal});
         } else if (device->uma) {
-            // Fall back to host memory type
-            buf = ggml_vk_create_buffer(device, size, {vk::MemoryPropertyFlagBits::eDeviceLocal,
+            // On UMA, prefer host-visible memory so direct tensor borrowing works.
+            // If unavailable, fall back to device-local memory.
+            buf = ggml_vk_create_buffer(device, size, {vk::MemoryPropertyFlagBits::eDeviceLocal | vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent,
+                                                       vk::MemoryPropertyFlagBits::eDeviceLocal,
                                                        vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent});
         } else if (device->disable_host_visible_vidmem) {
             if (device->allow_sysmem_fallback) {
@@ -5231,14 +5233,14 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
     ggml_vk_create_pipeline(device, device->pipeline_rwkv_wkv7_f32, "rwkv_wkv7_f32", rwkv_wkv7_f32_len, rwkv_wkv7_f32_data, "main", 8, sizeof(vk_op_rwkv_wkv7_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
 
     {
-        const uint32_t gdn_sizes[] = {32, 64, 128};
+        const uint32_t gdn_sizes[] = {16, 32, 64, 128};
         const char * gdn_names[][2] = {
+            {"gated_delta_net_f32_d16",     "gated_delta_net_f32_d16_kda"},
             {"gated_delta_net_f32_d32",     "gated_delta_net_f32_d32_kda"},
             {"gated_delta_net_f32_d64",     "gated_delta_net_f32_d64_kda"},
             {"gated_delta_net_f32_d128",    "gated_delta_net_f32_d128_kda"},
         };
-        const bool use_subgroup_reduce = device->subgroup_arithmetic;
-        for (uint32_t si = 0; si < 3; si++) {
+        for (uint32_t si = 0; si < 4; si++) {
             const uint32_t S_V = gdn_sizes[si];
             GGML_ASSERT(is_pow2(S_V));
 
@@ -5252,10 +5254,29 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
                 lanes_per_column = std::min(S_V, device->subgroup_size);
             }
 
-            const bool need_clustered_shader = lanes_per_column != 1 && (lanes_per_column < device->subgroup_size);
+            // gated_delta_net.comp relies on S_V % COLS_PER_WG == 0 and
+            // S_V % LANES_PER_COLUMN == 0 to avoid bounds checks.
+            while (lanes_per_column > 1u) {
+                const bool valid_lanes = (device->subgroup_size % lanes_per_column) == 0 &&
+                                         (S_V % lanes_per_column) == 0;
+                const uint32_t cols_per_wg = valid_lanes ? device->subgroup_size / lanes_per_column : 0;
+                if (valid_lanes && cols_per_wg > 0 && (S_V % cols_per_wg) == 0) {
+                    break;
+                }
+                lanes_per_column >>= 1u;
+            }
+
+            GGML_ASSERT((device->subgroup_size % lanes_per_column) == 0);
+            GGML_ASSERT((S_V % lanes_per_column) == 0);
+            GGML_ASSERT((S_V % (device->subgroup_size / lanes_per_column)) == 0);
+
+            const bool need_partial_subgroup_reduce = lanes_per_column != 1u && lanes_per_column < device->subgroup_size;
+            const bool use_clustered_reduce = device->subgroup_arithmetic && device->subgroup_clustered && need_partial_subgroup_reduce;
+            const bool use_subgroup_reduce = device->subgroup_arithmetic && !need_partial_subgroup_reduce;
+            const bool use_subgroup_ops = use_clustered_reduce || use_subgroup_reduce;
             size_t gdn_len;
             const void * gdn_data;
-            if (use_subgroup_reduce && need_clustered_shader) {
+            if (use_clustered_reduce) {
                 gdn_len = gated_delta_net_f32_len;
                 gdn_data = (const void *)gated_delta_net_f32_data;
             } else if (use_subgroup_reduce) {
@@ -5272,7 +5293,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
             for (uint32_t kda = 0; kda < 2; kda++) {
                 ggml_vk_create_pipeline(device, device->pipeline_gated_delta_net[si][kda],
                     gdn_names[si][kda], gdn_len, gdn_data, "main", 7, sizeof(vk_op_gated_delta_net_push_constants),
-                    wg_denoms, {S_V, kda, device->subgroup_size, lanes_per_column}, 1, true, use_subgroup_reduce, device->subgroup_size);
+                    wg_denoms, {S_V, kda, device->subgroup_size, lanes_per_column}, 1, true, use_subgroup_ops, device->subgroup_size);
             }
         }
     }
@@ -10746,9 +10767,10 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             const uint32_t kda = (dst->src[3]->ne[0] == (int64_t)S_v) ? 1 : 0;
             uint32_t si;
             switch (S_v) {
-                case 32:  si = 0; break;
-                case 64:  si = 1; break;
-                case 128: si = 2; break;
+                case 16:  si = 0; break;
+                case 32:  si = 1; break;
+                case 64:  si = 2; break;
+                case 128: si = 3; break;
                 default: return nullptr;
             }
             return ctx->device->pipeline_gated_delta_net[si][kda];
@@ -17193,7 +17215,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_GATED_DELTA_NET:
             {
                 const uint32_t S_v = op->src[2]->ne[0];
-                if (S_v != 32 && S_v != 64 && S_v != 128) {
+                if (S_v != 16 && S_v != 32 && S_v != 64 && S_v != 128) {
                     return false;
                 }
                 for (int i = 0; i < 6; i++) {

tools/llama-bench/llama-bench.cpp

@@ -323,6 +323,7 @@ struct cmd_params {
     std::vector<std::string>         hf_repo;
     std::vector<std::string>         hf_file;
     std::string                      hf_token;
+    bool                             offline;
     std::vector<int>                 n_prompt;
     std::vector<int>                 n_gen;
     std::vector<std::pair<int, int>> n_pg;
@@ -367,6 +368,7 @@ static const cmd_params cmd_params_defaults = {
     /* hf_repo              */ {},
     /* hf_file              */ {},
     /* hf_token             */ "",
+    /* offline              */ false,
     /* n_prompt             */ { 512 },
     /* n_gen                */ { 128 },
     /* n_pg                 */ {},
@@ -437,6 +439,8 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("                                              (default: unused)\n");
     printf("  -hft, --hf-token <token>                    Hugging Face access token\n");
     printf("                                              (default: value from HF_TOKEN environment variable)\n");
+    printf("  --offline                                   Offline mode: forces use of cache, prevents network access\n");
+    printf("                                              (default: disabled)\n");
     printf("  -p, --n-prompt <n>                          (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
     printf("  -n, --n-gen <n>                             (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
     printf("  -pg <pp,tg>                                 (default: %s)\n", join(transform_to_str(cmd_params_defaults.n_pg, pair_str), ",").c_str());
@@ -558,6 +562,8 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                     break;
                 }
                 params.hf_token = argv[i];
+            } else if (arg == "--offline") {
+                params.offline = true;
             } else if (arg == "-p" || arg == "--n-prompt") {
                 if (++i >= argc) {
                     invalid_param = true;
@@ -1040,6 +1046,7 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
 
             common_download_opts opts;
             opts.bearer_token = params.hf_token;
+            opts.offline         = params.offline;
             auto download_result = common_download_model(model, opts);
             if (download_result.model_path.empty()) {
                 fprintf(stderr, "error: failed to download model from HuggingFace\n");

tools/server/bench/bench.py

@@ -40,6 +40,7 @@ def main(args_in: list[str] | None = None) -> None:
                         required=True)
     parser.add_argument("--hf-repo", type=str, help="Hugging Face model repository", required=True)
     parser.add_argument("--hf-file", type=str, help="Hugging Face model file", required=True)
+    parser.add_argument("--offline", action="store_true", default=False, help="Offline mode: forces use of cache, prevents network access")
     parser.add_argument("-ngl", "--n-gpu-layers", type=int, help="layers to the GPU for computation", required=True)
     parser.add_argument("--ctx-size", type=int, help="Set the size of the prompt context", required=True)
     parser.add_argument("--parallel", type=int, help="Set the number of slots for process requests", required=True)
@@ -268,6 +269,8 @@ def start_server_background(args):
     ]
     server_args.extend(['--hf-repo', args.hf_repo])
     server_args.extend(['--hf-file', args.hf_file])
+    if args.offline:
+        server_args.append('--offline')
     server_args.extend(['--n-gpu-layers', args.n_gpu_layers])
     server_args.extend(['--ctx-size', args.ctx_size])
     server_args.extend(['--parallel', args.parallel])

tools/server/server-context.cpp

@@ -201,6 +201,8 @@ struct server_slot {
     // Speculative decoding stats
     int32_t n_draft_total = 0;      // Total draft tokens generated
     int32_t n_draft_accepted = 0;   // Draft tokens actually accepted
+    int32_t n_draft_verif_steps = 0; // Total draft token verification steps by the target model
+    std::vector<int32_t> n_accepted_per_pos; // Accepted tokens per draft position
 
     void reset() {
         SLT_DBG(*this, "%s", "\n");
@@ -227,6 +229,8 @@ struct server_slot {
         // clear speculative decoding stats
         n_draft_total = 0;
         n_draft_accepted = 0;
+        n_draft_verif_steps = 0;
+        n_accepted_per_pos.clear();
 
         task_prev = std::move(task);
         task.reset();
@@ -509,10 +513,22 @@ struct server_slot {
                 llama_perf_context(ctx_tgt).n_reused);
 
         if (n_draft_total > 0) {
-            const float draft_ratio = (float) n_draft_accepted / n_draft_total;
+            const float  draft_ratio  = (float) n_draft_accepted / n_draft_total;
+            const double mean_acc_len = n_draft_verif_steps > 0 ? 1.0 + (double) n_draft_accepted / (double) n_draft_verif_steps : 1.0;
+
+            std::string acceptance_rates_per_pos;
+            if (n_draft_verif_steps > 0) {
+                for (size_t i = 0; i < n_accepted_per_pos.size(); ++i) {
+                    if (i > 0) {
+                        acceptance_rates_per_pos += ", ";
+                    }
+                    acceptance_rates_per_pos += string_format("%.3f", (double) n_accepted_per_pos[i] / (double) n_draft_verif_steps);
+                }
+            }
+
             SLT_INF(*this,
-                    "draft acceptance = %0.5f (%5d accepted / %5d generated)\n",
-                    draft_ratio, n_draft_accepted, n_draft_total);
+                    "draft acceptance = %0.5f (%5d accepted / %5d generated), mean acceptance length = %5.2f, acceptance rate per position = (%s)\n",
+                    draft_ratio, n_draft_accepted, n_draft_total, mean_acc_len, acceptance_rates_per_pos.c_str());
         }
 
         common_speculative_print_stats(spec);
@@ -3543,6 +3559,14 @@ private:
 
                 // update how many tokens out of those tested were accepted
                 slot.n_draft_accepted += ids.size() - 1;
+                slot.n_draft_verif_steps += 1;
+
+                if (slot.n_accepted_per_pos.empty()) {
+                    slot.n_accepted_per_pos.resize(common_speculative_n_max(&params_base.speculative), 0);
+                }
+                for (size_t i = 0; i < ids.size() - 1 && i < slot.n_accepted_per_pos.size(); ++i) {
+                    slot.n_accepted_per_pos[i]++;
+                }
 
                 // add accepted tokens to the prompt
                 slot.prompt.tokens.keep_first(slot.prompt.n_tokens() - n_draft);