llama-cpp-bin 9432.0.0__tar.gz → 9437.0.0__tar.gz

{llama_cpp_bin-9432.0.0 → llama_cpp_bin-9437.0.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: llama-cpp-bin
-Version: 9432.0.0
+Version: 9437.0.0
 Summary: llama.cpp server binary built from source
 Author-email: Vladyslav Tkachenko <vladlearns@gmail.com>
 License: AGPL-3.0-or-later

{llama_cpp_bin-9432.0.0 → llama_cpp_bin-9437.0.0}/llama.cpp/ggml/src/ggml-metal/ggml-metal-device.cpp

@@ -1732,6 +1732,8 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_rope(ggml_metal_
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_im2col(ggml_metal_library_t lib, const ggml_tensor * op) {
     assert(op->op == GGML_OP_IM2COL);
 
+    GGML_TENSOR_LOCALS(int64_t, ne0, op->src[0], ne);
+
     GGML_ASSERT(ggml_is_contiguous(op->src[1]));
     GGML_ASSERT(op->src[1]->type == GGML_TYPE_F32);
     GGML_ASSERT(op->type         == GGML_TYPE_F16 || op->type == GGML_TYPE_F32);
@@ -1739,7 +1741,11 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_im2col(ggml_meta
     char base[256];
     char name[256];
 
-    snprintf(base, 256, "kernel_im2col_%s", ggml_type_name(op->type));
+    if (ne00*ne01 <= 1024) {
+        snprintf(base, 256, "kernel_im2col_%s", ggml_type_name(op->type));
+    } else {
+        snprintf(base, 256, "kernel_im2col_ext_%s", ggml_type_name(op->type));
+    }
     snprintf(name, 256, "%s", base);
 
     ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);

{llama_cpp_bin-9432.0.0 → llama_cpp_bin-9437.0.0}/llama.cpp/ggml/src/ggml-metal/ggml-metal-ops.cpp

@@ -3635,16 +3635,26 @@ int ggml_metal_op_im2col(ggml_metal_op_t ctx, int idx) {
 
     auto pipeline = ggml_metal_library_get_pipeline_im2col(lib, op);
 
-    GGML_ASSERT(KH*KW <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
+    if (KH*KW <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
+        const uint64_t ntptg0 = std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)/(KH*KW), N);
 
-    const uint64_t ntptg0 = std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)/(KH*KW), N);
+        ggml_metal_encoder_set_pipeline(enc, pipeline);
+        ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), 1);
+        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         2);
 
-    ggml_metal_encoder_set_pipeline(enc, pipeline);
-    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), 1);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         2);
+        ggml_metal_encoder_dispatch_threadgroups(enc, IC, OH, OW, ntptg0, KH, KW);
+    } else {
+        const uint64_t n_threads = std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), N);
+        const int64_t  quotient  = N / n_threads + (N % n_threads > 0 ? 1 : 0);
+
+        ggml_metal_encoder_set_pipeline(enc, pipeline);
+        ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), 1);
+        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         2);
 
-    ggml_metal_encoder_dispatch_threadgroups(enc, IC, OH, OW, ntptg0, KH, KW);
+        ggml_metal_encoder_dispatch_threadgroups(enc, quotient * CHW, OH, OW, n_threads, 1, 1);
+    }
 
     return 1;
 }

{llama_cpp_bin-9432.0.0 → llama_cpp_bin-9437.0.0}/llama.cpp/ggml/src/ggml-metal/ggml-metal.metal

@@ -4696,59 +4696,59 @@ kernel void kernel_im2col(
 template [[host_name("kernel_im2col_f32")]] kernel im2col_t kernel_im2col<float>;
 template [[host_name("kernel_im2col_f16")]] kernel im2col_t kernel_im2col<half>;
 
-// TODO: obsolete -- remove
-//typedef void (im2col_ext_t)(
-//        constant ggml_metal_kargs_im2col & args,
-//        device const float * x,
-//        device        char * dst,
-//        uint3 tgpig[[threadgroup_position_in_grid]],
-//        uint3  tgpg[[threadgroups_per_grid]],
-//        uint3 tpitg[[thread_position_in_threadgroup]],
-//        uint3   ntg[[threads_per_threadgroup]]);
-//
-//template <typename T>
-//kernel void kernel_im2col_ext(
-//        constant ggml_metal_kargs_im2col & args,
-//        device const float * x,
-//        device        char * dst,
-//        uint3 tgpig[[threadgroup_position_in_grid]],
-//        uint3  tgpg[[threadgroups_per_grid]],      // tgpg[0] = D x IC x KH x KW, CHW = IC x KH x KW
-//        uint3 tpitg[[thread_position_in_threadgroup]],
-//        uint3   ntg[[threads_per_threadgroup]]) {  // [M, 1, 1]
-//    const int64_t KHW = (int64_t)args.KHW;
-//
-//    const int64_t d   = tgpig[0] / args.CHW;
-//    const int64_t chw = tgpig[0] % args.CHW;
-//    const int64_t tgpig_0 = chw / KHW;  // 0 ~ (IC - 1)
-//    const int64_t HW = tgpig[0] % KHW;
-//
-//    const int64_t tpitg_0 = (d * ntg[0]) + tpitg[0];
-//    if (tpitg_0 >= args.N) {
-//        return;
-//    }
-//
-//    const int64_t tpitg_1 = HW / args.KW;
-//    const int64_t tpitg_2 = HW % args.KW;
-//
-//    const int64_t iiw = tgpig[2] * args.s0 + tpitg_2 * args.d0 - args.p0;
-//    const int64_t iih = tgpig[1] * args.s1 + tpitg_1 * args.d1 - args.p1;
-//
-//    const int64_t offset_dst =
-//        (tpitg_0 * tgpg[1] * tgpg[2] + tgpig[1] * tgpg[2] + tgpig[2]) * args.CHW +
-//        (tgpig_0 * KHW + tpitg_1 * args.KW + tpitg_2);
-//
-//    device T * pdst = (device T *) (dst);
-//
-//    if (iih < 0 || iih >= args.IH || iiw < 0 || iiw >= args.IW) {
-//        pdst[offset_dst] = 0.0f;
-//    } else {
-//        const int64_t offset_src = tpitg_0 * args.ofs0 + tgpig_0 * args.ofs1;
-//        pdst[offset_dst] = x[offset_src + iih * args.IW + iiw];
-//    }
-//}
-//
-//template [[host_name("kernel_im2col_ext_f32")]] kernel im2col_ext_t kernel_im2col_ext<float>;
-//template [[host_name("kernel_im2col_ext_f16")]] kernel im2col_ext_t kernel_im2col_ext<half>;
+// TODO: optimize
+typedef void (im2col_ext_t)(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]);
+
+template <typename T>
+kernel void kernel_im2col_ext(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],      // tgpg[0] = D x IC x KH x KW, CHW = IC x KH x KW
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {  // [M, 1, 1]
+    const int64_t KHW = (int64_t)args.KHW;
+
+    const int64_t d   = tgpig[0] / args.CHW;
+    const int64_t chw = tgpig[0] % args.CHW;
+    const int64_t tgpig_0 = chw / KHW;  // 0 ~ (IC - 1)
+    const int64_t HW = tgpig[0] % KHW;
+
+    const int64_t tpitg_0 = (d * ntg[0]) + tpitg[0];
+    if (tpitg_0 >= args.N) {
+        return;
+    }
+
+    const int64_t tpitg_1 = HW / args.KW;
+    const int64_t tpitg_2 = HW % args.KW;
+
+    const int64_t iiw = tgpig[2] * args.s0 + tpitg_2 * args.d0 - args.p0;
+    const int64_t iih = tgpig[1] * args.s1 + tpitg_1 * args.d1 - args.p1;
+
+    const int64_t offset_dst =
+        (tpitg_0 * tgpg[1] * tgpg[2] + tgpig[1] * tgpg[2] + tgpig[2]) * args.CHW +
+        (tgpig_0 * KHW + tpitg_1 * args.KW + tpitg_2);
+
+    device T * pdst = (device T *) (dst);
+
+    if (iih < 0 || iih >= args.IH || iiw < 0 || iiw >= args.IW) {
+        pdst[offset_dst] = 0.0f;
+    } else {
+        const int64_t offset_src = tpitg_0 * args.ofs0 + tgpig_0 * args.ofs1;
+        pdst[offset_dst] = x[offset_src + iih * args.IW + iiw];
+    }
+}
+
+template [[host_name("kernel_im2col_ext_f32")]] kernel im2col_ext_t kernel_im2col_ext<float>;
+template [[host_name("kernel_im2col_ext_f16")]] kernel im2col_ext_t kernel_im2col_ext<half>;
 
 template <typename TK>
 kernel void kernel_conv_2d(

{llama_cpp_bin-9432.0.0 → llama_cpp_bin-9437.0.0}/llama.cpp/ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -585,6 +585,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_convert_block_mxfp4_trans4_ns, kernel_restore_block_mxfp4_trans4_ns;
     cl_kernel kernel_convert_block_q8_0, kernel_restore_block_q8_0, kernel_restore_block_q8_0_trans;
     cl_kernel kernel_convert_block_q6_K_noshuffle, kernel_restore_block_q6_K_noshuffle;
+    cl_kernel kernel_convert_bf16_to_f16, kernel_convert_f16_to_bf16;
     cl_kernel kernel_mul_mat_q4_0_f32_8x_flat;
     cl_kernel kernel_convert_block_q4_0_noshuffle;
     cl_kernel kernel_restore_block_q4_0_noshuffle;
@@ -1175,6 +1176,8 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) {
         CL_CHECK((backend_ctx->kernel_restore_block_iq4_nl = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_iq4_nl", &err), err));
         CL_CHECK((backend_ctx->kernel_convert_block_iq4_nl_noshuffle = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_iq4_nl_noshuffle", &err), err));
         CL_CHECK((backend_ctx->kernel_restore_block_iq4_nl_noshuffle = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_iq4_nl_noshuffle", &err), err));
+        CL_CHECK((backend_ctx->kernel_convert_bf16_to_f16 = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_bf16_to_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_convert_f16_to_bf16 = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_f16_to_bf16", &err), err));
         GGML_LOG_CONT(".");
     }
 
@@ -5019,6 +5022,8 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
         case GGML_OP_MUL_MAT:
             if (op->src[0]->type == GGML_TYPE_F16) {
                 return true;
+            } else if (op->src[0]->type == GGML_TYPE_BF16) {
+                return true;
             } else if (op->src[0]->type == GGML_TYPE_F32) {
                 return op->src[1]->type == GGML_TYPE_F32;
             } else if (op->src[0]->type == GGML_TYPE_Q4_0  || op->src[0]->type == GGML_TYPE_Q4_1 ||
@@ -6828,6 +6833,40 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
     }
 #endif // GGML_OPENCL_SOA_Q
 
+    // convert bf16 to f16 and store as f16 in device buffer
+    if (tensor->type == GGML_TYPE_BF16) {
+        GGML_ASSERT(offset % sizeof(ggml_fp16_t) == 0 && size % sizeof(ggml_fp16_t) == 0
+            && "Offset and size must be multiples of 2 for bf16 tensors");
+
+        ggml_tensor_extra_cl * extra = (ggml_tensor_extra_cl *) tensor->extra;
+        GGML_ASSERT(extra);
+
+        cl_ulong n_elements = size / sizeof(ggml_fp16_t);
+        cl_ulong off_dst = (extra->offset + offset) / sizeof(ggml_fp16_t);
+
+        cl_int err;
+        cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
+            size, (void *) data, &err);
+        CL_CHECK(err);
+
+        cl_kernel kernel = backend_ctx->kernel_convert_bf16_to_f16;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &data_device));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_ulong), &off_dst));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &n_elements));
+
+        size_t global_work_size[] = { (size_t)CEIL_DIV(n_elements, 64)*64, 1, 1 };
+        size_t local_work_size[] = { 64, 1, 1 };
+
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clReleaseMemObject(data_device));
+        CL_CHECK(clReleaseEvent(evt));
+
+        return;
+    }
+
     ggml_tensor_extra_cl * extra = (ggml_tensor_extra_cl *) tensor->extra;
     GGML_ASSERT(extra);
 
@@ -7676,6 +7715,41 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,
     }
 #endif // GGML_OPENCL_SOA_Q
 
+    if (tensor->type == GGML_TYPE_BF16) {
+        GGML_ASSERT(offset % sizeof(ggml_fp16_t) == 0 && size % sizeof(ggml_fp16_t) == 0
+            && "Offset and size must be multiples of 2 for bf16 tensors");
+
+        ggml_tensor_extra_cl * extra = (ggml_tensor_extra_cl *) tensor->extra;
+        GGML_ASSERT(extra);
+
+        cl_ulong n_elements = size / sizeof(ggml_fp16_t);
+        cl_ulong off_src = (extra->offset + tensor->view_offs + offset) / sizeof(ggml_fp16_t);
+
+        cl_int err;
+        cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE, size, NULL, &err);
+        CL_CHECK(err);
+
+        cl_kernel kernel = backend_ctx->kernel_convert_f16_to_bf16;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &off_src));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &data_device));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &n_elements));
+
+        size_t global_work_size[] = { (size_t)CEIL_DIV(n_elements, 64)*64, 1, 1 };
+        size_t local_work_size[] = { 64, 1, 1 };
+
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clReleaseEvent(evt));
+
+        CL_CHECK(clEnqueueReadBuffer(
+            queue, data_device, CL_TRUE, 0, size, data, 0, NULL, NULL));
+        CL_CHECK(clReleaseMemObject(data_device));
+
+        return;
+    }
+
     ggml_tensor_extra_cl * extra = (ggml_tensor_extra_cl *) tensor->extra;
 
     CL_CHECK(clEnqueueReadBuffer(
@@ -8165,6 +8239,7 @@ static void ggml_cl_copy_to_contiguous(ggml_backend_t backend, const ggml_tensor
             kernel = backend_ctx->kernel_cpy_f32_f32;
             break;
         case GGML_TYPE_F16:
+        case GGML_TYPE_BF16: // stored as f16 on device
             kernel = backend_ctx->kernel_cpy_f16_f16;
             break;
         default:
@@ -11125,7 +11200,8 @@ static bool ggml_cl_can_use_adreno_xmem_gemm_f16_f32(
     if (backend_ctx->gpu_family != GPU_FAMILY::ADRENO) {
         return false;
     }
-    if (src0->type != GGML_TYPE_F16 || src1->type != GGML_TYPE_F32 || dst->type != GGML_TYPE_F32) {
+    if ((src0->type != GGML_TYPE_F16 && src0->type != GGML_TYPE_BF16) ||
+        src1->type != GGML_TYPE_F32 || dst->type != GGML_TYPE_F32) {
         return false;
     }
     if (!ggml_is_contiguous(src0) || !ggml_is_contiguous(src1) || !ggml_is_contiguous(dst)) {
@@ -12843,7 +12919,8 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
     GGML_ASSERT(dst);
     GGML_ASSERT(dst->extra);
 
-    const enum ggml_type src0t = src0->type;
+    // bf16 is stored as f16 on device
+    const enum ggml_type src0t = (src0->type == GGML_TYPE_BF16) ? GGML_TYPE_F16 : src0->type;
     const enum ggml_type src1t = src1->type;
 
     ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;

{llama_cpp_bin-9432.0.0 → llama_cpp_bin-9437.0.0}/llama.cpp/src/llama-model.cpp

@@ -410,16 +410,16 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
     auto get_tensor_config = [&]() -> tensor_config {
         // standard attention
         if (std::regex_match(tensor_name, pattern_q_weight) || std::regex_match(tensor_name, pattern_kv_weight)) {
-            return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_1, "attn_output.weight");
+            return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_1, "attn_output.weight", "ssm_out.weight");
         }
         if (std::regex_match(tensor_name, pattern_q_bias) || std::regex_match(tensor_name, pattern_kv_bias)) {
-            return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_0, "attn_output.weight");
+            return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_0, "attn_output.weight", "ssm_out.weight");
         }
         if (std::regex_match(tensor_name, pattern_qkv_weight)) {
-            return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_1);
+            return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_1, "attn_output.weight", "ssm_out.weight");
         }
         if ( std::regex_match(tensor_name, pattern_qkv_bias)) {
-            return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_0);
+            return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_0, "attn_output.weight", "ssm_out.weight");
         }
         if (std::regex_match(tensor_name, pattern_qk_norm)) {
             return get_tensor_config_impl(tensor->ne[1] == 1 ? GGML_BACKEND_SPLIT_AXIS_MIRRORED : GGML_BACKEND_SPLIT_AXIS_1, "attn_output.weight");
@@ -435,7 +435,7 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
         }
 
         if (std::regex_match(tensor_name, pattern_attn_gate_weight)) {
-            return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_1);
+            return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_1, "attn_output.weight", "ssm_out.weight");
         }
         if (std::regex_match(tensor_name, pattern_ssm_dt) || std::regex_match(tensor_name, pattern_ssm_a)) {
             return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_0, "ssm_out.weight");

{llama_cpp_bin-9432.0.0 → llama_cpp_bin-9437.0.0}/llama.cpp/tests/test-backend-ops.cpp

@@ -7812,6 +7812,8 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 2, 2560}, {3, 3, 2, 2560}, 1, 1, 1, 1, 1, 1, true));
     test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {5, 5, 1, 32}, {3, 4, 1, 32}, 1, 1, 0, 0, 1, 1, true));
     test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32, {2, 2, 1536, 729}, {2, 2, 1536, 4096}, 1, 1, 0, 0, 1, 1, true));
+    test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {128, 128, 1, 2}, {32, 33, 1, 2}, 1, 1, 1, 1, 1, 1, true));
+    test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {128, 128, 2, 1}, {33, 34, 2, 1}, 1, 1, 1, 1, 1, 1, true));
 
     // im2col 3D
     test_cases.emplace_back(new test_im2col_3d(GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32));

{llama_cpp_bin-9432.0.0 → llama_cpp_bin-9437.0.0}/llama.cpp/tools/llama-bench/llama-bench.cpp

@@ -19,6 +19,7 @@
 #include <vector>
 #include <unordered_set>
 
+#include "arg.h"
 #include "build-info.h"
 #include "common.h"
 #include "download.h"
@@ -275,9 +276,11 @@ static std::string pair_str(const std::pair<int, int> & p) {
     return buf;
 }
 
-static std::vector<int> parse_int_range(const std::string & s) {
+static std::vector<int> parse_int_range(const std::string & s, bool allow_negative = false) {
     // first[-last[(+|*)step]]
-    std::regex range_regex(R"(^(\d+)(?:-(\d+)(?:([\+|\*])(\d+))?)?(?:,|$))");
+    std::regex range_regex(allow_negative
+        ? R"(^(-?\d+)(?:-(\d+)(?:([\+|\*])(\d+))?)?(?:,|$))"
+        : R"(^(\d+)(?:-(\d+)(?:([\+|\*])(\d+))?)?(?:,|$))");
 
     std::smatch match;
     std::string::const_iterator search_start(s.cbegin());
@@ -337,7 +340,7 @@ struct cmd_params {
     std::vector<llama_split_mode>    split_mode;
     std::vector<int>                 main_gpu;
     std::vector<bool>                no_kv_offload;
-    std::vector<bool>                flash_attn;
+    std::vector<llama_flash_attn_type> flash_attn;
     std::vector<std::vector<ggml_backend_dev_t>> devices;
     std::vector<std::vector<float>>  tensor_split;
     std::vector<std::vector<llama_model_tensor_buft_override>> tensor_buft_overrides;
@@ -376,12 +379,12 @@ static const cmd_params cmd_params_defaults = {
     /* cpu_mask             */ { "0x0" },
     /* cpu_strict           */ { false },
     /* poll                 */ { 50 },
-    /* n_gpu_layers         */ { 99 },
+    /* n_gpu_layers         */ { -1 },
     /* n_cpu_moe            */ { 0 },
     /* split_mode           */ { LLAMA_SPLIT_MODE_LAYER },
     /* main_gpu             */ { 0 },
     /* no_kv_offload        */ { false },
-    /* flash_attn           */ { false },
+    /* flash_attn           */ { LLAMA_FLASH_ATTN_TYPE_AUTO },
     /* devices              */ { {} },
     /* tensor_split         */ { std::vector<float>(llama_max_devices(), 0.0f) },
     /* tensor_buft_overrides*/ { std::vector<llama_model_tensor_buft_override>{ { nullptr, nullptr } } },
@@ -451,7 +454,7 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -sm, --split-mode <none|layer|row|tensor>   (default: %s)\n", join(transform_to_str(cmd_params_defaults.split_mode, split_mode_str), ",").c_str());
     printf("  -mg, --main-gpu <i>                         (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
     printf("  -nkvo, --no-kv-offload <0|1>                (default: %s)\n", join(cmd_params_defaults.no_kv_offload, ",").c_str());
-    printf("  -fa, --flash-attn <0|1>                     (default: %s)\n", join(cmd_params_defaults.flash_attn, ",").c_str());
+    printf("  -fa, --flash-attn <on|off|auto>             (default: %s)\n", join(transform_to_str(cmd_params_defaults.flash_attn, llama_flash_attn_type_name), ",").c_str());
     printf("  -dev, --device <dev0/dev1/...>              (default: auto)\n");
     printf("  -mmp, --mmap <0|1>                          (default: %s)\n", join(cmd_params_defaults.use_mmap, ",").c_str());
     printf("  -dio, --direct-io <0|1>                     (default: %s)\n", join(cmd_params_defaults.use_direct_io, ",").c_str());
@@ -710,7 +713,7 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                     invalid_param = true;
                     break;
                 }
-                auto p = parse_int_range(argv[i]);
+                auto p = parse_int_range(argv[i], /*allow_negative=*/true);
                 params.n_gpu_layers.insert(params.n_gpu_layers.end(), p.begin(), p.end());
             } else if (arg == "-ncmoe" || arg == "--n-cpu-moe") {
                 if (++i >= argc) {
@@ -793,8 +796,27 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                     invalid_param = true;
                     break;
                 }
-                auto p = string_split<bool>(argv[i], split_delim);
-                params.flash_attn.insert(params.flash_attn.end(), p.begin(), p.end());
+                auto p = string_split<std::string>(argv[i], split_delim);
+
+                std::vector<llama_flash_attn_type> types;
+                for (const auto & v : p) {
+                    llama_flash_attn_type type;
+                    if (common_arg_utils::is_truthy(v)) {
+                        type = LLAMA_FLASH_ATTN_TYPE_ENABLED;
+                    } else if (common_arg_utils::is_falsey(v)) {
+                        type = LLAMA_FLASH_ATTN_TYPE_DISABLED;
+                    } else if (common_arg_utils::is_autoy(v)) {
+                        type = LLAMA_FLASH_ATTN_TYPE_AUTO;
+                    } else {
+                        invalid_param = true;
+                        break;
+                    }
+                    types.push_back(type);
+                }
+                if (invalid_param) {
+                    break;
+                }
+                params.flash_attn.insert(params.flash_attn.end(), types.begin(), types.end());
             } else if (arg == "-mmp" || arg == "--mmap") {
                 if (++i >= argc) {
                     invalid_param = true;
@@ -1138,7 +1160,7 @@ struct cmd_params_instance {
     llama_split_mode   split_mode;
     int                main_gpu;
     bool               no_kv_offload;
-    bool               flash_attn;
+    llama_flash_attn_type flash_attn;
     std::vector<ggml_backend_dev_t> devices;
     std::vector<float> tensor_split;
     std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
@@ -1222,7 +1244,7 @@ struct cmd_params_instance {
         cparams.type_k          = type_k;
         cparams.type_v          = type_v;
         cparams.offload_kqv     = !no_kv_offload;
-        cparams.flash_attn_type = flash_attn ? LLAMA_FLASH_ATTN_TYPE_ENABLED : LLAMA_FLASH_ATTN_TYPE_DISABLED;
+        cparams.flash_attn_type = flash_attn;
         cparams.embeddings      = embeddings;
         cparams.op_offload      = !no_op_offload;
         cparams.swa_full        = false;
@@ -1400,7 +1422,7 @@ struct test {
     llama_split_mode         split_mode;
     int                      main_gpu;
     bool                     no_kv_offload;
-    bool                     flash_attn;
+    llama_flash_attn_type    flash_attn;
     std::vector<ggml_backend_dev_t> devices;
     std::vector<float>       tensor_split;
     std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
@@ -1522,10 +1544,10 @@ struct test {
             field == "poll" || field == "model_size" || field == "model_n_params" || field == "n_gpu_layers" ||
             field == "main_gpu" || field == "n_prompt" || field == "n_gen" || field == "n_depth" || field == "avg_ns" ||
             field == "stddev_ns" || field == "no_op_offload" || field == "n_cpu_moe" ||
-            field == "fit_target" || field == "fit_min_ctx") {
+            field == "fit_target" || field == "fit_min_ctx" || field == "flash_attn") {
             return INT;
         }
-        if (field == "f16_kv" || field == "no_kv_offload" || field == "cpu_strict" || field == "flash_attn" ||
+        if (field == "f16_kv" || field == "no_kv_offload" || field == "cpu_strict" ||
             field == "use_mmap" || field == "use_direct_io" || field == "embeddings" || field == "no_host") {
             return BOOL;
         }
@@ -1594,7 +1616,7 @@ struct test {
                                             split_mode_str(split_mode),
                                             std::to_string(main_gpu),
                                             std::to_string(no_kv_offload),
-                                            std::to_string(flash_attn),
+                                            std::to_string((int) flash_attn),
                                             devices_to_string(devices),
                                             tensor_split_str,
                                             tensor_buft_overrides_str,
@@ -1779,7 +1801,7 @@ struct markdown_printer : public printer {
             return 6;
         }
         if (field == "flash_attn") {
-            return 2;
+            return 3;
         }
         if (field == "devices") {
             return -12;

{llama_cpp_bin-9432.0.0 → llama_cpp_bin-9437.0.0}/src/llama_cpp_bin/_version.py

@@ -18,7 +18,7 @@ version_tuple: tuple[int | str, ...]
 commit_id: str | None
 __commit_id__: str | None
 
-__version__ = version = '9432.0.0'
-__version_tuple__ = version_tuple = (9432, 0, 0)
+__version__ = version = '9437.0.0'
+__version_tuple__ = version_tuple = (9437, 0, 0)
 
-__commit_id__ = commit_id = 'g3a09ef5f3'
+__commit_id__ = commit_id = 'ga6ed30703'

{llama_cpp_bin-9432.0.0 → llama_cpp_bin-9437.0.0}/src/llama_cpp_bin.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: llama-cpp-bin
-Version: 9432.0.0
+Version: 9437.0.0
 Summary: llama.cpp server binary built from source
 Author-email: Vladyslav Tkachenko <vladlearns@gmail.com>
 License: AGPL-3.0-or-later