llama-cpp-bin 10059.0.0__tar.gz → 10067.0.0__tar.gz

{llama_cpp_bin-10059.0.0/src/llama_cpp_bin.egg-info → llama_cpp_bin-10067.0.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: llama-cpp-bin
-Version: 10059.0.0
+Version: 10067.0.0
 Summary: llama.cpp server binary built from source
 Author-email: Vladyslav Tkachenko <vladlearns@gmail.com>
 License: AGPL-3.0-or-later

{llama_cpp_bin-10059.0.0 → llama_cpp_bin-10067.0.0}/llama.cpp/common/arg.cpp

@@ -361,6 +361,14 @@ common_models_handler common_models_handler_init(const common_params & params, l
                                         params.speculative.types.end(),
                                         COMMON_SPECULATIVE_TYPE_DRAFT_MTP) != params.speculative.types.end();
 
+    const bool spec_type_draft_dflash = std::find(params.speculative.types.begin(),
+                                           params.speculative.types.end(),
+                                           COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH) != params.speculative.types.end();
+
+    const bool spec_type_draft_eagle3 = std::find(params.speculative.types.begin(),
+                                           params.speculative.types.end(),
+                                           COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3) != params.speculative.types.end();
+
     // only download mmproj if the current example is using it
     bool use_mmproj = false;
     for (const auto & ex : mmproj_examples) {
@@ -373,6 +381,8 @@ common_models_handler common_models_handler_init(const common_params & params, l
     opts.bearer_token    = params.hf_token;
     opts.offline         = params.offline;
     opts.download_mtp    = spec_type_draft_mtp;
+    opts.download_eagle3 = spec_type_draft_eagle3;
+    opts.download_dflash = spec_type_draft_dflash;
     opts.download_mmproj = use_mmproj && !params.no_mmproj
                         && params.mmproj.path.empty() && params.mmproj.url.empty();
 
@@ -546,6 +556,26 @@ void common_models_handler_apply(common_models_handler & handler, common_params
             }
         });
     }
+    if (!plan.dflash.local_path.empty() && !had_spec_url) {
+        tasks.emplace_back(plan.dflash, opts, [&]() {
+            // only fall back to the discovered DFlash sidecar when no draft was explicitly provided
+            if (params.speculative.draft.mparams.empty()) {
+                params.speculative.draft.mparams.path = hf_cache::finalize_file(plan.dflash);
+            } else {
+                hf_cache::finalize_file(plan.dflash);
+            }
+        });
+    }
+    if (!plan.eagle3.local_path.empty() && !had_spec_url) {
+        tasks.emplace_back(plan.eagle3, opts, [&]() {
+            // only fall back to the discovered Eagle3 sidecar when no draft was explicitly provided
+            if (params.speculative.draft.mparams.empty()) {
+                params.speculative.draft.mparams.path = hf_cache::finalize_file(plan.eagle3);
+            } else {
+                hf_cache::finalize_file(plan.eagle3);
+            }
+        });
+    }
     if (!plan.preset.local_path.empty()) {
         tasks.emplace_back(plan.preset, opts, [&]() {
             // if HF repo is a preset repo, we simply run server in router mode with the preset.ini file
@@ -2815,6 +2845,20 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.speculative.types.push_back(COMMON_SPECULATIVE_TYPE_DRAFT_MTP);
         }
     ).set_examples({LLAMA_EXAMPLE_DOWNLOAD}));
+    add_opt(common_arg(
+        {"--dflash"},
+        "also download the DFlash sidecar, if available (default: unused)",
+        [](common_params & params) {
+            params.speculative.types.push_back(COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH);
+        }
+    ).set_examples({LLAMA_EXAMPLE_DOWNLOAD}));
+    add_opt(common_arg(
+        {"--eagle3"},
+        "also download the Eagle3 sidecar, if available (default: unused)",
+        [](common_params & params) {
+            params.speculative.types.push_back(COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3);
+        }
+    ).set_examples({LLAMA_EXAMPLE_DOWNLOAD}));
     add_opt(common_arg(
         {"--context-file"}, "FNAME",
         "file to load context from (use comma-separated values to specify multiple files)",

{llama_cpp_bin-10059.0.0 → llama_cpp_bin-10067.0.0}/llama.cpp/common/download.cpp

@@ -620,6 +620,16 @@ static hf_cache::hf_file find_best_mtp(const hf_cache::hf_files & files,
     return find_best_sibling(files, model, "mtp-");
 }
 
+static hf_cache::hf_file find_best_eagle3(const hf_cache::hf_files & files,
+                                          const std::string        & model) {
+    return find_best_sibling(files, model, "eagle3-");
+}
+
+static hf_cache::hf_file find_best_dflash(const hf_cache::hf_files & files,
+                                          const std::string        & model) {
+    return find_best_sibling(files, model, "dflash-");
+}
+
 static bool gguf_filename_is_model(const std::string & filepath) {
     if (!string_ends_with(filepath, ".gguf")) {
         return false;
@@ -632,7 +642,9 @@ static bool gguf_filename_is_model(const std::string & filepath) {
 
     return filename.find("mmproj")  == std::string::npos &&
            filename.find("imatrix") == std::string::npos &&
-           filename.find("mtp-")    == std::string::npos;
+           filename.find("mtp-")    == std::string::npos &&
+           filename.find("eagle3-") == std::string::npos &&
+           filename.find("dflash-") == std::string::npos;
 }
 
 static hf_cache::hf_file find_best_model(const hf_cache::hf_files & files,
@@ -740,6 +752,12 @@ common_download_hf_plan common_download_get_hf_plan(const common_params_model &
     if (opts.download_mtp) {
         plan.mtp = find_best_mtp(all, primary.path);
     }
+    if (opts.download_dflash) {
+        plan.dflash = find_best_dflash(all, primary.path);
+    }
+    if (opts.download_eagle3) {
+        plan.eagle3 = find_best_eagle3(all, primary.path);
+    }
 
     return plan;
 }
@@ -911,8 +929,10 @@ std::vector<common_cached_model_info> common_list_cached_models() {
     for (const auto & f : files) {
         auto split = get_gguf_split_info(f.path);
         if (split.index != 1 || split.tag.empty() ||
-            split.prefix.find("mmproj") != std::string::npos ||
-            split.prefix.find("mtp-")   != std::string::npos) {
+            split.prefix.find("mmproj")  != std::string::npos ||
+            split.prefix.find("mtp-")    != std::string::npos ||
+            split.prefix.find("eagle3-") != std::string::npos ||
+            split.prefix.find("dflash-") != std::string::npos) {
             continue;
         }
         if (seen.insert(f.repo_id + ":" + split.tag).second) {

{llama_cpp_bin-10059.0.0 → llama_cpp_bin-10067.0.0}/llama.cpp/common/download.h

@@ -55,8 +55,10 @@ struct common_download_opts {
     std::string bearer_token;
     common_header_list headers;
     bool offline = false;
-    bool download_mmproj = false;
-    bool download_mtp = false;
+    bool download_mmproj  = false;
+    bool download_mtp     = false;
+    bool download_eagle3  = false;
+    bool download_dflash  = false;
     common_download_callback * callback = nullptr;
 };
 
@@ -106,6 +108,8 @@ struct common_download_hf_plan {
     hf_cache::hf_files model_files;
     hf_cache::hf_file mmproj;
     hf_cache::hf_file mtp;
+    hf_cache::hf_file eagle3;
+    hf_cache::hf_file dflash;
     hf_cache::hf_file preset; // if set, only this file is downloaded
 };
 common_download_hf_plan common_download_get_hf_plan(const common_params_model & model, const common_download_opts & opts);

{llama_cpp_bin-10059.0.0 → llama_cpp_bin-10067.0.0}/llama.cpp/ggml/CMakeLists.txt

@@ -4,7 +4,7 @@ project("ggml" C CXX ASM)
 
 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
-set(GGML_VERSION_MINOR 16)
+set(GGML_VERSION_MINOR 17)
 set(GGML_VERSION_PATCH 0)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
 

{llama_cpp_bin-10059.0.0 → llama_cpp_bin-10067.0.0}/llama.cpp/ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -876,7 +876,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_gemm_moe_q8_1_dp4a_q5k = nullptr;   // generic dp4a MoE GEMM (MOE_QT=5, q5_K), opt-in
     cl_kernel kernel_moe_expand_scale_q5_K = nullptr;    // q5_K 6-bit s[] -> uniform scale[16]/min[8]
     cl_kernel kernel_gemv_moe_q5_k_f32_ns, kernel_gemm_moe_q5_k_f32_ns;
-    cl_kernel kernel_gemv_moe_q6_k_f32_ns, kernel_gemm_moe_q6_k_f32_ns;
+    cl_kernel kernel_gemv_moe_q6_k_f32_ns, kernel_gemm_moe_q6_k_f32_ns, kernel_gemm_moe_q6_k_f32_ns_bin;
     cl_kernel kernel_gemm_moe_q6_k_q8_1_dp4a = nullptr;    // dp4a (int8) q6_K MoE prefill GEMM
     cl_kernel kernel_gemv_moe_mxfp4_f32, kernel_gemm_moe_mxfp4_f32;
     cl_kernel kernel_gemv_moe_mxfp4_f32_ns, kernel_gemm_moe_mxfp4_f32_ns, kernel_gemm_moe_mxfp4_f32_ns_bin;
@@ -4310,6 +4310,24 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) {
         GGML_LOG_CONT(".");
     }
 
+    // gemm_moe_q6_k_f32_ns_bin
+    {
+        size_t bin_size = 0;
+        backend_ctx->kernel_gemm_moe_q6_k_f32_ns_bin = nullptr;
+
+        if (use_adreno_bin_kernels(backend_ctx)) {
+            const char * kernel_bin = (const char *)backend_ctx->get_adreno_bin_kernel("gemm_moe_q6_k_f32_ns_ila", &bin_size);
+            if (kernel_bin && bin_size > 0) {
+                cl_program prog =
+                    build_program_from_binary(backend_ctx->context, backend_ctx->device, kernel_bin, CL_moe_compile_opts, bin_size);
+
+                CL_CHECK((backend_ctx->kernel_gemm_moe_q6_k_f32_ns_bin = clCreateKernel(prog, "kernel_gemm_moe_q6_k_f32_ns_ila", &err), err));
+                CL_CHECK(clReleaseProgram(prog));
+                GGML_LOG_CONT(".");
+            }
+        }
+    }
+
     // gemm_moe_q6_k_q8_1_dp4a (dp4a q6_K MoE prefill GEMM)
     if (backend_ctx->has_integer_dot) {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -6012,7 +6030,8 @@ static void transpose_2d(
     cl_kernel kernel,
     cl_mem src, cl_mem dst, size_t size,
     cl_int stride, cl_int rows,
-    bool blocking = true
+    bool blocking = true,
+    bool auto_local = false // let driver pick local size for non-uniform workgroups
 ) {
     static ggml_cl_buffer buf;
 
@@ -6038,7 +6057,7 @@ static void transpose_2d(
     size_t local_size[3] = {64, 1, 1};
     size_t global_size[3] = {(size_t)stride, (size_t)rows, 1};;
     CL_CHECK(clEnqueueNDRangeKernel(backend_ctx->queue, kernel, 3, NULL,
-        global_size, local_size, 0, NULL, NULL));
+        global_size, auto_local ? NULL : local_size, 0, NULL, NULL));
 
     if (blocking) {
         CL_CHECK(clEnqueueCopyBuffer(backend_ctx->queue, trans, dst, 0, 0, size, 0, NULL, &evt));
@@ -6055,10 +6074,11 @@ static void transpose_2d_as_8b(
     ggml_backend_opencl_context * backend_ctx,
     cl_mem src, cl_mem dst, size_t size,
     cl_int stride, cl_int rows,
-    bool blocking = true
+    bool blocking = true,
+    bool auto_local = false
 ) {
     transpose_2d(backend_ctx, backend_ctx->kernel_transpose_8_buf,
-        src, dst, size, stride, rows, blocking);
+        src, dst, size, stride, rows, blocking, auto_local);
 }
 
 static void transpose_2d_as_16b(
@@ -9054,6 +9074,9 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
             transpose_2d_as_16b(backend_ctx, extra->q, extra->q, size_q, K/4, M);
             transpose_2d_as_16b(backend_ctx, extra->d, extra->d, size_d, K/256, M);
             transpose_2d_as_16b(backend_ctx, extra->dm, extra->dm, size_dm, K/256, M);
+
+            // Transpose s as uchar
+            transpose_2d_as_8b(backend_ctx, extra->s, extra->s, size_s, K/256*12, M, true, true);
         }
 #endif // GGML_OPENCL_USE_ADRENO_KERNELS
         return;
@@ -10222,23 +10245,27 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,
             size_t size_q  = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*ggml_blck_size(tensor->type)/2;
             size_t size_d  = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*sizeof(ggml_fp16_t);
             size_t size_dm = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*sizeof(ggml_fp16_t);
+            size_t size_s  = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*12;
 
             static ggml_cl_buffer buf_trans_q;
             static ggml_cl_buffer buf_trans_d;
             static ggml_cl_buffer buf_trans_dm;
+            static ggml_cl_buffer buf_trans_s;
 
             buf_trans_q.allocate(backend_ctx->context, size_q);
             buf_trans_d.allocate(backend_ctx->context, size_d);
             buf_trans_dm.allocate(backend_ctx->context, size_dm);
+            buf_trans_s.allocate(backend_ctx->context, size_s);
 
-            // Transpose q, d, dm back
+            // Transpose q, d, dm, s back
             transpose_2d_as_16b(backend_ctx, extra->q,  buf_trans_q.buffer,  size_q,  M, K/4);
             transpose_2d_as_16b(backend_ctx, extra->d,  buf_trans_d.buffer,  size_d,  M, K/256);
             transpose_2d_as_16b(backend_ctx, extra->dm, buf_trans_dm.buffer, size_dm, M, K/256);
+            transpose_2d_as_8b (backend_ctx, extra->s,  buf_trans_s.buffer,  size_s,  M, K/256*12, true, true);
 
             cl_kernel kernel = backend_ctx->kernel_restore_block_q4_K_noshuffle;
             CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &buf_trans_q.buffer));
-            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->s));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &buf_trans_s.buffer));
             CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &buf_trans_d.buffer));
             CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &buf_trans_dm.buffer));
             CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &data_device));
@@ -22282,8 +22309,8 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
                         bool use_moe_dp4a = q5kmdp4a_on
                             && backend_ctx->kernel_gemm_moe_q8_1_dp4a_q5k != nullptr
                             && extra0_q5_K->scale != nullptr;
-                        // bin kernel takes precedence
-                        use_moe_dp4a = use_moe_dp4a && backend_ctx->kernel_gemm_moe_q4_k_f32_ns_bin == nullptr;
+                        // dot prod has to be available
+                        use_moe_dp4a = backend_ctx->has_integer_dot && use_moe_dp4a;
 
                         if (use_moe_dp4a) {
                             const size_t tok_slots = (size_t)max_post_router_tile * n_tile_size;
@@ -22501,6 +22528,9 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
 
                 } else { // for gemm
                     kernel = backend_ctx->kernel_gemm_moe_q6_k_f32_ns;
+                    if (backend_ctx->kernel_gemm_moe_q6_k_f32_ns_bin) {
+                        kernel = backend_ctx->kernel_gemm_moe_q6_k_f32_ns_bin;
+                    }
 
                     // Reorder router if called from test-backend-ops or when new router is generated.
                     // Otherwise reuse the reordered result from previous mul_mat_id call.
@@ -22521,6 +22551,8 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
                                                             || backend_ctx->adreno_gen == ADRENO_GPU_GEN::X1E);
                     // dot prod has to be available
                     use_moe_dp4a = backend_ctx->has_integer_dot && use_moe_dp4a;
+                    // bin kernel takes precedence
+                    use_moe_dp4a = use_moe_dp4a && backend_ctx->kernel_gemm_moe_q6_k_f32_ns_bin == nullptr;
 
                     cl_buffer_region region;
                     region.origin = 0;
@@ -22557,6 +22589,11 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
                         CL_CHECK(status);
                         cl_image_format image_format_buf_src1 = {CL_RGBA, CL_FLOAT};
                         cl_image_desc image_desc_buf_src1 = {CL_MEM_OBJECT_IMAGE1D_BUFFER, static_cast<size_t>(ne00 * max_post_router_tile * n_tile_size / 4), 0,0,0,0,0,0,0, {buf_src1_reordered}};
+                        if (backend_ctx->kernel_gemm_moe_q6_k_f32_ns_bin) {
+                            // bin kernel uses slightly different image format
+                            image_format_buf_src1 = {CL_R, CL_FLOAT};
+                            image_desc_buf_src1.image_width = static_cast<size_t>(ne00 * max_post_router_tile * n_tile_size);
+                        }
                         image_src1_reordered = clCreateImage(backend_ctx->context, CL_MEM_READ_ONLY, &image_format_buf_src1, &image_desc_buf_src1, NULL, &status);
                         CL_CHECK(status);
 

{llama_cpp_bin-10059.0.0 → llama_cpp_bin-10067.0.0}/llama.cpp/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_q8_1_dp4a.cl

@@ -37,15 +37,17 @@ static inline float e8m0_to_fp32(uchar x) {
 // One token's dp4a dot (8 uints = 32 K elems) + mxfp4 block-scale epilogue.
 // blk_scale already carries the 0.5 factor (== 0.5 * 2^e).
 #define MOE_MXFP4_DP4A_T(t) do {                                     \
+        uint4 a0 = vload4(0, &sh_qa[t][0]);                          \
+        uint4 a1 = vload4(0, &sh_qa[t][4]);                          \
         int raw = 0;                                                 \
-        raw = dot_acc_sat_4x8packed_ss_int(qw[0], sh_qa[t][0], raw); \
-        raw = dot_acc_sat_4x8packed_ss_int(qw[1], sh_qa[t][1], raw); \
-        raw = dot_acc_sat_4x8packed_ss_int(qw[2], sh_qa[t][2], raw); \
-        raw = dot_acc_sat_4x8packed_ss_int(qw[3], sh_qa[t][3], raw); \
-        raw = dot_acc_sat_4x8packed_ss_int(qw[4], sh_qa[t][4], raw); \
-        raw = dot_acc_sat_4x8packed_ss_int(qw[5], sh_qa[t][5], raw); \
-        raw = dot_acc_sat_4x8packed_ss_int(qw[6], sh_qa[t][6], raw); \
-        raw = dot_acc_sat_4x8packed_ss_int(qw[7], sh_qa[t][7], raw); \
+        raw = dot_acc_sat_4x8packed_ss_int(qw[0], a0.s0, raw);       \
+        raw = dot_acc_sat_4x8packed_ss_int(qw[1], a0.s1, raw);       \
+        raw = dot_acc_sat_4x8packed_ss_int(qw[2], a0.s2, raw);       \
+        raw = dot_acc_sat_4x8packed_ss_int(qw[3], a0.s3, raw);       \
+        raw = dot_acc_sat_4x8packed_ss_int(qw[4], a1.s0, raw);       \
+        raw = dot_acc_sat_4x8packed_ss_int(qw[5], a1.s1, raw);       \
+        raw = dot_acc_sat_4x8packed_ss_int(qw[6], a1.s2, raw);       \
+        raw = dot_acc_sat_4x8packed_ss_int(qw[7], a1.s3, raw);       \
         acc[t] += blk_scale * (float)sh_d[t] * (float)raw;           \
     } while (0)
 
@@ -133,11 +135,13 @@ kernel void kernel_gemm_moe_mxfp4_q8_1_dp4a(
         qw[6] = mxfp4_pack((ushort)(r3));        qw[7] = mxfp4_pack((ushort)(r3 >> 16));
 
         // cooperatively stage the n_real-token x 32-K int8 activations
-        const uint stage_lim = (uint)n_real * 8;
-        for (uint idx = lid; idx < stage_lim; idx += 64) {
-            const uint t = idx >> 3;
-            const uint u = idx & 7;
-            sh_qa[t][u] = src1_qa[(col + t) * ne00_u + (step >> 2) + u];
+        // Stage each token's 8 activation uints as two 128-bit uint4 loads/stores.
+        const uint vlim = (uint)n_real * 2;
+        for (uint idx = lid; idx < vlim; idx += 64) {
+            const uint t = idx >> 1;
+            const uint h = (idx & 1) << 2;   // 0 or 4
+            uint4 v = vload4(0, &src1_qa[(col + t) * ne00_u + (step >> 2) + h]);
+            vstore4(v, 0, &sh_qa[t][h]);
         }
         if (lid < (uint)n_real) {
             sh_d[lid] = src1_da[(col + lid) * num_blocks + sub];

{llama_cpp_bin-10059.0.0 → llama_cpp_bin-10067.0.0}/llama.cpp/ggml/src/ggml-opencl/kernels/gemm_moe_q4_0_q8_1_dp4a.cl

@@ -17,15 +17,17 @@
 
 // One token's dp4a dot (8 uints = 32 K elems) + q4_0 scale/zero-point epilogue.
 #define MOE_Q40_DP4A_T(t) do {                                       \
+        uint4 a0 = vload4(0, &sh_qa[t][0]);                          \
+        uint4 a1 = vload4(0, &sh_qa[t][4]);                          \
         int raw = 0;                                                 \
-        raw = dot_acc_sat_4x8packed_ss_int(qw[0], sh_qa[t][0], raw); \
-        raw = dot_acc_sat_4x8packed_ss_int(qw[1], sh_qa[t][1], raw); \
-        raw = dot_acc_sat_4x8packed_ss_int(qw[2], sh_qa[t][2], raw); \
-        raw = dot_acc_sat_4x8packed_ss_int(qw[3], sh_qa[t][3], raw); \
-        raw = dot_acc_sat_4x8packed_ss_int(qw[4], sh_qa[t][4], raw); \
-        raw = dot_acc_sat_4x8packed_ss_int(qw[5], sh_qa[t][5], raw); \
-        raw = dot_acc_sat_4x8packed_ss_int(qw[6], sh_qa[t][6], raw); \
-        raw = dot_acc_sat_4x8packed_ss_int(qw[7], sh_qa[t][7], raw); \
+        raw = dot_acc_sat_4x8packed_ss_int(qw[0], a0.s0, raw);       \
+        raw = dot_acc_sat_4x8packed_ss_int(qw[1], a0.s1, raw);       \
+        raw = dot_acc_sat_4x8packed_ss_int(qw[2], a0.s2, raw);       \
+        raw = dot_acc_sat_4x8packed_ss_int(qw[3], a0.s3, raw);       \
+        raw = dot_acc_sat_4x8packed_ss_int(qw[4], a1.s0, raw);       \
+        raw = dot_acc_sat_4x8packed_ss_int(qw[5], a1.s1, raw);       \
+        raw = dot_acc_sat_4x8packed_ss_int(qw[6], a1.s2, raw);       \
+        raw = dot_acc_sat_4x8packed_ss_int(qw[7], a1.s3, raw);       \
         acc[t] += d_val * ((float)sh_d[t] * (float)raw - 8.0f * (float)sh_s[t]); \
     } while (0)
 
@@ -112,11 +114,13 @@ kernel void kernel_gemm_moe_q4_0_q8_1_dp4a(
         qw[6] = EXP4(r3);        qw[7] = EXP4(r3 >> 16);
 
         // cooperatively stage the n_real-token x 32-K int8 activations
-        const uint stage_lim = (uint)n_real * 8;
-        for (uint idx = lid; idx < stage_lim; idx += 64) {
-            const uint t = idx >> 3;
-            const uint u = idx & 7;
-            sh_qa[t][u] = src1_qa[(col + t) * ne00_u + (step >> 2) + u];
+        // Stage each token's 8 activation uints as two 128-bit uint4 loads/stores.
+        const uint vlim = (uint)n_real * 2;
+        for (uint idx = lid; idx < vlim; idx += 64) {
+            const uint t = idx >> 1;
+            const uint h = (idx & 1) << 2;   // 0 or 4
+            uint4 v = vload4(0, &src1_qa[(col + t) * ne00_u + (step >> 2) + h]);
+            vstore4(v, 0, &sh_qa[t][h]);
         }
         if (lid < (uint)n_real) {
             sh_d[lid] = src1_da[(col + lid) * num_blocks + sub];

{llama_cpp_bin-10059.0.0 → llama_cpp_bin-10067.0.0}/llama.cpp/ggml/src/ggml-opencl/kernels/gemm_moe_q4_k_q8_1_dp4a.cl

@@ -37,16 +37,21 @@ inline void get_scale_min_k4(
                   (((uint)((u) & 0xF000u)) << 12) )
 
 // One token's dp4a dot (8 uints = 32 K elems) + q4_K scale/min epilogue into acc[t].
+// The 8 activation uints are read as two 128-bit uint4 loads staged to private (Adreno
+// wants 128-bit local reads, and a __local operand fed straight to the dp4a builtin is
+// slower and can miscompile).
 #define MOE_Q4K_DP4A_T(t) do {                                       \
+        uint4 a0 = vload4(0, &sh_qa[t][0]);                          \
+        uint4 a1 = vload4(0, &sh_qa[t][4]);                          \
         int raw = 0;                                                 \
-        raw = dot_acc_sat_4x8packed_ss_int(qw[0], sh_qa[t][0], raw); \
-        raw = dot_acc_sat_4x8packed_ss_int(qw[1], sh_qa[t][1], raw); \
-        raw = dot_acc_sat_4x8packed_ss_int(qw[2], sh_qa[t][2], raw); \
-        raw = dot_acc_sat_4x8packed_ss_int(qw[3], sh_qa[t][3], raw); \
-        raw = dot_acc_sat_4x8packed_ss_int(qw[4], sh_qa[t][4], raw); \
-        raw = dot_acc_sat_4x8packed_ss_int(qw[5], sh_qa[t][5], raw); \
-        raw = dot_acc_sat_4x8packed_ss_int(qw[6], sh_qa[t][6], raw); \
-        raw = dot_acc_sat_4x8packed_ss_int(qw[7], sh_qa[t][7], raw); \
+        raw = dot_acc_sat_4x8packed_ss_int(qw[0], a0.s0, raw);       \
+        raw = dot_acc_sat_4x8packed_ss_int(qw[1], a0.s1, raw);       \
+        raw = dot_acc_sat_4x8packed_ss_int(qw[2], a0.s2, raw);       \
+        raw = dot_acc_sat_4x8packed_ss_int(qw[3], a0.s3, raw);       \
+        raw = dot_acc_sat_4x8packed_ss_int(qw[4], a1.s0, raw);       \
+        raw = dot_acc_sat_4x8packed_ss_int(qw[5], a1.s1, raw);       \
+        raw = dot_acc_sat_4x8packed_ss_int(qw[6], a1.s2, raw);       \
+        raw = dot_acc_sat_4x8packed_ss_int(qw[7], a1.s3, raw);       \
         acc[t] += scale * (float)sh_d[t] * (float)raw - minv * (float)sh_s[t]; \
     } while (0)
 
@@ -145,12 +150,14 @@ kernel void kernel_gemm_moe_q4_k_q8_1_dp4a(
         qw[4] = EXP4(r2);        qw[5] = EXP4(r2 >> 16);
         qw[6] = EXP4(r3);        qw[7] = EXP4(r3 >> 16);
 
-        // --- cooperatively stage the n_real-token x 32-K int8 activations to LDS ---
-        const uint stage_lim = (uint)n_real * 8;
-        for (uint idx = lid; idx < stage_lim; idx += 64) {
-            const uint t = idx >> 3;
-            const uint u = idx & 7;
-            sh_qa[t][u] = src1_qa[(col + t) * ne00_u + (step >> 2) + u];
+        // cooperatively stage the n_real-token x 32-K int8 activations to lm
+        // Stage each token's 8 activation uints as two 128-bit uint4 loads/stores.
+        const uint vlim = (uint)n_real * 2;
+        for (uint idx = lid; idx < vlim; idx += 64) {
+            const uint t = idx >> 1;
+            const uint h = (idx & 1) << 2;   // 0 or 4
+            uint4 v = vload4(0, &src1_qa[(col + t) * ne00_u + (step >> 2) + h]);
+            vstore4(v, 0, &sh_qa[t][h]);
         }
         if (lid < (uint)n_real) {
             sh_d[lid] = src1_da[(col + lid) * ne00_b + sub];

{llama_cpp_bin-10059.0.0 → llama_cpp_bin-10067.0.0}/llama.cpp/ggml/src/ggml-opencl/kernels/gemm_moe_q6_k_q8_1_dp4a.cl

@@ -41,8 +41,10 @@ inline int dp4a_q6(uint qw0, uint qw1, uint qw2, uint qw3,
 
 // One token's q6_K dp4a dot (two halves, per-16 scales) + epilogue into acc[t].
 #define MOE_Q6K_DP4A_T(t) do {                                                                            \
-        const int raw1 = dp4a_q6(qw[0], qw[1], qw[2], qw[3], sh_qa[t][0], sh_qa[t][1], sh_qa[t][2], sh_qa[t][3]); \
-        const int raw2 = dp4a_q6(qw[4], qw[5], qw[6], qw[7], sh_qa[t][4], sh_qa[t][5], sh_qa[t][6], sh_qa[t][7]); \
+        uint4 a0 = vload4(0, &sh_qa[t][0]);                                                               \
+        uint4 a1 = vload4(0, &sh_qa[t][4]);                                                               \
+        const int raw1 = dp4a_q6(qw[0], qw[1], qw[2], qw[3], a0.s0, a0.s1, a0.s2, a0.s3);                 \
+        const int raw2 = dp4a_q6(qw[4], qw[5], qw[6], qw[7], a1.s0, a1.s1, a1.s2, a1.s3);                 \
         const float a_d = (float)sh_d[t];                                                                 \
         acc[t] += scale0 * a_d * (float)raw1 + scale1 * a_d * (float)raw2;                                \
     } while (0)
@@ -144,11 +146,13 @@ kernel void kernel_gemm_moe_q6_k_q8_1_dp4a(
         qw[6] = SIGN6(EXP4(r3)       | EXP2((qh2 >> 16) & 0xFFu));
         qw[7] = SIGN6(EXP4(r3 >> 16) | EXP2((qh2 >> 24) & 0xFFu));
 
-        const uint stage_lim = (uint)n_real * 8;
-        for (uint idx = lid; idx < stage_lim; idx += 64) {
-            const uint t = idx >> 3;
-            const uint u = idx & 7;
-            sh_qa[t][u] = src1_qa[(col + t) * ne00_u + (step >> 2) + u];
+        // Stage each token's 8 activation uints as two 128-bit uint4 loads/stores.
+        const uint vlim = (uint)n_real * 2;
+        for (uint idx = lid; idx < vlim; idx += 64) {
+            const uint t = idx >> 1;
+            const uint h = (idx & 1) << 2;   // 0 or 4
+            uint4 v = vload4(0, &src1_qa[(col + t) * ne00_u + (step >> 2) + h]);
+            vstore4(v, 0, &sh_qa[t][h]);
         }
         if (lid < (uint)n_real) {
             sh_d[lid] = src1_da[(col + lid) * ne00_b + sub];

{llama_cpp_bin-10059.0.0 → llama_cpp_bin-10067.0.0}/llama.cpp/ggml/src/ggml-opencl/kernels/gemm_moe_q8_1_dp4a.cl

@@ -102,8 +102,10 @@ inline int dp4a4(uint w0,uint w1,uint w2,uint w3,uint a0,uint a1,uint a2,uint a3
 
 // One token's two-half dp4a + uniform scale/min epilogue into acc[t].
 #define MOE_DP4A_T(t) do {                                                                  \
-        const int raw1 = dp4a4(qw[0],qw[1],qw[2],qw[3], sh_qa[t][0],sh_qa[t][1],sh_qa[t][2],sh_qa[t][3]); \
-        const int raw2 = dp4a4(qw[4],qw[5],qw[6],qw[7], sh_qa[t][4],sh_qa[t][5],sh_qa[t][6],sh_qa[t][7]); \
+        uint4 a0 = vload4(0, &sh_qa[t][0]);                                                 \
+        uint4 a1 = vload4(0, &sh_qa[t][4]);                                                 \
+        const int raw1 = dp4a4(qw[0],qw[1],qw[2],qw[3], a0.s0,a0.s1,a0.s2,a0.s3);           \
+        const int raw2 = dp4a4(qw[4],qw[5],qw[6],qw[7], a1.s0,a1.s1,a1.s2,a1.s3);           \
         const float a_d = (float)sh_d[t];                                                   \
         acc[t] += sc0*a_d*(float)raw1 + sc1*a_d*(float)raw2 - mn*(float)sh_s[t];             \
     } while (0)
@@ -178,10 +180,13 @@ kernel void kernel_gemm_moe_q8_1_dp4a(
 
         LOAD_QW(step, sub)
 
-        const uint stage_lim = (uint)n_real * 8;
-        for (uint idx = lid; idx < stage_lim; idx += 64) {
-            const uint t = idx >> 3, u = idx & 7;
-            sh_qa[t][u] = src1_qa[(col + t) * ne00_u + (step >> 2) + u];
+        // Stage each token's 8 activation uints as two 128-bit uint4 loads/stores.
+        const uint vlim = (uint)n_real * 2;
+        for (uint idx = lid; idx < vlim; idx += 64) {
+            const uint t = idx >> 1;
+            const uint h = (idx & 1) << 2;   // 0 or 4
+            uint4 v = vload4(0, &src1_qa[(col + t) * ne00_u + (step >> 2) + h]);
+            vstore4(v, 0, &sh_qa[t][h]);
         }
         if (lid < (uint)n_real) {
             sh_d[lid] = src1_da[(col + lid) * ne00_b + sub];

{llama_cpp_bin-10059.0.0 → llama_cpp_bin-10067.0.0}/llama.cpp/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_k_f32.cl

@@ -8,9 +8,11 @@
 #define QK_K         256
 #define K_SCALE_SIZE 12
 
+// scales are transposed: consecutive codes of a row are `stride` apart
 inline void get_scale_min_k4(
     int j,
     global const uchar * q,
+    int stride,
     uchar * d,
     uchar * m,
     uchar mask_d6,
@@ -18,11 +20,11 @@ inline void get_scale_min_k4(
     uchar mask_hi2
 ) {
     if (j < 4) {
-        *d = q[j]   & mask_d6;
-        *m = q[j+4] & mask_d6;
+        *d = q[j*stride]     & mask_d6;
+        *m = q[(j+4)*stride] & mask_d6;
     } else {
-        *d = (q[j+4] & mask_d4) | ((q[j-4] & mask_hi2) >> 2);
-        *m = ((q[j+4] >> 4) & mask_d4) | ((q[j]   & mask_hi2) >> 2);
+        *d = (q[(j+4)*stride] & mask_d4) | ((q[(j-4)*stride] & mask_hi2) >> 2);
+        *m = ((q[(j+4)*stride] >> 4) & mask_d4) | ((q[j*stride] & mask_hi2) >> 2);
     }
 }
 
@@ -55,7 +57,6 @@ kernel void kernel_gemm_noshuffle_q4_k_f32(
     half8 B;
     half4 dequantized_weights;
 
-    int num_blocks_K = k / QK_K;
 
     global const ushort * weight_ptr = src0_q + gx_2;
     global const half   * d_ptr      = src0_d  + gx_2;
@@ -68,16 +69,16 @@ kernel void kernel_gemm_noshuffle_q4_k_f32(
         half4 d  = vload4(0, d_ptr  + sb_idx * m);
         half4 dm = vload4(0, dm_ptr + sb_idx * m);
 
-        global const uchar * sc0 = src0_s + (gx_2+0) * num_blocks_K * K_SCALE_SIZE + sb_idx * K_SCALE_SIZE;
-        global const uchar * sc1 = src0_s + (gx_2+1) * num_blocks_K * K_SCALE_SIZE + sb_idx * K_SCALE_SIZE;
-        global const uchar * sc2 = src0_s + (gx_2+2) * num_blocks_K * K_SCALE_SIZE + sb_idx * K_SCALE_SIZE;
-        global const uchar * sc3 = src0_s + (gx_2+3) * num_blocks_K * K_SCALE_SIZE + sb_idx * K_SCALE_SIZE;
+        global const uchar * sc0 = src0_s + sb_idx * K_SCALE_SIZE * m + (gx_2+0);
+        global const uchar * sc1 = sc0 + 1;
+        global const uchar * sc2 = sc0 + 2;
+        global const uchar * sc3 = sc0 + 3;
 
         uchar sv0, mn0, sv1, mn1, sv2, mn2, sv3, mn3;
-        get_scale_min_k4(sub_idx, sc0, &sv0, &mn0, mask_d6, mask_d4, mask_hi2);
-        get_scale_min_k4(sub_idx, sc1, &sv1, &mn1, mask_d6, mask_d4, mask_hi2);
-        get_scale_min_k4(sub_idx, sc2, &sv2, &mn2, mask_d6, mask_d4, mask_hi2);
-        get_scale_min_k4(sub_idx, sc3, &sv3, &mn3, mask_d6, mask_d4, mask_hi2);
+        get_scale_min_k4(sub_idx, sc0, m, &sv0, &mn0, mask_d6, mask_d4, mask_hi2);
+        get_scale_min_k4(sub_idx, sc1, m, &sv1, &mn1, mask_d6, mask_d4, mask_hi2);
+        get_scale_min_k4(sub_idx, sc2, m, &sv2, &mn2, mask_d6, mask_d4, mask_hi2);
+        get_scale_min_k4(sub_idx, sc3, m, &sv3, &mn3, mask_d6, mask_d4, mask_hi2);
 
         half4 scale = convert_half4(convert_float4(d)  * convert_float4((uchar4)(sv0, sv1, sv2, sv3)));
         half4 mval  = convert_half4(convert_float4(dm) * convert_float4((uchar4)(mn0, mn1, mn2, mn3)));