@novastera-oss/llamarn 0.2.1 → 0.2.2

package/cpp/llama.cpp/ggml/src/ggml-metal/ggml-metal.metal

@@ -856,6 +856,7 @@ kernel void kernel_tanh(
 constant float GELU_COEF_A     = 0.044715f;
 constant float GELU_QUICK_COEF = -1.702f;
 constant float SQRT_2_OVER_PI  = 0.79788456080286535587989211986876f;
+constant float SQRT_2_INV      = 0.70710678118654752440084436210484f;
 
 kernel void kernel_gelu(
     device const float * src0,
@@ -897,6 +898,42 @@ kernel void kernel_gelu_quick_4(
     dst[tpig] = x*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x)));
 }
 
+// based on Abramowitz and Stegun formula 7.1.26 or similar Hastings' approximation
+// ref: https://www.johndcook.com/blog/python_erf/
+constant float p_erf  = 0.3275911f;
+constant float a1_erf = 0.254829592f;
+constant float a2_erf = -0.284496736f;
+constant float a3_erf = 1.421413741f;
+constant float a4_erf = -1.453152027f;
+constant float a5_erf = 1.061405429f;
+
+template<typename T>
+T erf_approx(T x) {
+    T sign_x = sign(x);
+    x = fabs(x);
+    T t = 1.0f / (1.0f + p_erf * x);
+    T y = 1.0f - (((((a5_erf * t + a4_erf) * t) + a3_erf) * t + a2_erf) * t + a1_erf) * t * exp(-x * x);
+    return sign_x * y;
+}
+
+kernel void kernel_gelu_erf(
+    device const float * src0,
+    device       float * dst,
+    uint tpig[[thread_position_in_grid]]) {
+    device const float & x = src0[tpig];
+
+    dst[tpig] = 0.5f*x*(1.0f+erf_approx<float>(x*SQRT_2_INV));
+}
+
+kernel void kernel_gelu_erf_4(
+    device const float4 * src0,
+    device       float4 * dst,
+    uint tpig[[thread_position_in_grid]]) {
+    device const float4 & x = src0[tpig];
+
+    dst[tpig] = 0.5f*x*(1.0f+erf_approx<float4>(x*SQRT_2_INV));
+}
+
 kernel void kernel_silu(
         device const float * src0,
         device       float * dst,
@@ -2713,8 +2750,148 @@ kernel void kernel_rope_neox(
     }
 }
 
+template<typename T>
+kernel void kernel_rope_multi(
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
+
+    device const int32_t * pos = (device const int32_t *) src1;
+
+    const float inv_ndims = -1.f/args.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 < args.n_dims) {
+            const int ic = i0/2;
+
+            // mrope theta calculations
+            // note: the rest is the same as kernel_rope_neox
+            const int sect_dims = args.sect_0 + args.sect_1 + args.sect_2 + args.sect_3;
+            const int sec_w01   = args.sect_0 + args.sect_1;               // end of section 1
+            const int sec_w012  = args.sect_0 + args.sect_1 + args.sect_2; // end of section 2
+            const int sector    = ic % sect_dims;
+
+            float theta_base;
+            if (sector < args.sect_0) {
+                theta_base = (float) pos[i2];
+            } else if (sector < sec_w01) {
+                theta_base = (float) pos[i2 + args.ne02];
+            } else if (sector < sec_w012) {
+                theta_base = (float) pos[i2 + args.ne02 * 2];
+            } else {
+                theta_base = (float) pos[i2 + args.ne02 * 3];
+            }
+            // end of mrope
+
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
+
+            const float freq_factor = src2 != src0 ? ((device const float *) src2)[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + ic*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + ic*args.nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[args.n_dims/2];
+
+            dst_data[0]             = x0*cos_theta - x1*sin_theta;
+            dst_data[args.n_dims/2] = x0*sin_theta + x1*cos_theta;
+        } else {
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+template<typename T>
+kernel void kernel_rope_vision(
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
+
+    device const int32_t * pos = (device const int32_t *) src1;
+
+    const float inv_ndims = -1.f/args.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 < 2*args.n_dims) { // different from kernel_rope_multi
+            const int ic = i0/2;
+
+            // mrope theta calculations (only support 2 dimensions)
+            const int sect_dims = args.sect_0 + args.sect_1;
+            const int sector    = ic % sect_dims;
+
+            float p;
+            float theta_base;
+            if (sector < args.sect_1) {
+                p = (float) sector;
+                theta_base = (float) pos[i2];
+            } else {
+                p = (float) sector - args.sect_0;
+                theta_base = (float) pos[i2 + args.ne02];
+            }
+
+            const float theta = theta_base * pow(args.freq_base, 2.0f * inv_ndims * p);
+            // end of mrope
+
+            const float freq_factor = src2 != src0 ? ((device const float *) src2)[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + ic*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + ic*args.nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[args.n_dims]; // different from kernel_rope_multi
+
+            dst_data[0]           = x0*cos_theta - x1*sin_theta;
+            dst_data[args.n_dims] = x0*sin_theta + x1*cos_theta; // different from kernel_rope_multi
+        } else {
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
 typedef decltype(kernel_rope_norm<float>) kernel_rope_norm_t;
 typedef decltype(kernel_rope_neox<float>) kernel_rope_neox_t;
+typedef decltype(kernel_rope_multi<float>) kernel_rope_multi_t;
+typedef decltype(kernel_rope_vision<float>) kernel_rope_vision_t;
 
 template [[host_name("kernel_rope_norm_f32")]] kernel kernel_rope_norm_t kernel_rope_norm<float>;
 template [[host_name("kernel_rope_norm_f16")]] kernel kernel_rope_norm_t kernel_rope_norm<half>;
@@ -2722,6 +2899,12 @@ template [[host_name("kernel_rope_norm_f16")]] kernel kernel_rope_norm_t kernel_
 template [[host_name("kernel_rope_neox_f32")]] kernel kernel_rope_neox_t kernel_rope_neox<float>;
 template [[host_name("kernel_rope_neox_f16")]] kernel kernel_rope_neox_t kernel_rope_neox<half>;
 
+template [[host_name("kernel_rope_multi_f32")]] kernel kernel_rope_multi_t kernel_rope_multi<float>;
+template [[host_name("kernel_rope_multi_f16")]] kernel kernel_rope_multi_t kernel_rope_multi<half>;
+
+template [[host_name("kernel_rope_vision_f32")]] kernel kernel_rope_vision_t kernel_rope_vision<float>;
+template [[host_name("kernel_rope_vision_f16")]] kernel kernel_rope_vision_t kernel_rope_vision<half>;
+
 typedef void (im2col_t)(
         device const float * x,
         device        char * dst,
@@ -3109,7 +3292,7 @@ template<
     typename kd4x4_t, // key type in device memory
     short nl_k,
     void (*deq_k)(device const kd4x4_t *, short, thread k4x4_t &),
-    typename vd4x4_t, // key type in device memory
+    typename vd4x4_t, // value type in device memory
     short nl_v,
     void (*deq_v)(device const vd4x4_t *, short, thread v4x4_t &),
     short DK,        // K head size
@@ -3630,7 +3813,7 @@ template<
     typename kd4_t, // key type in device memory
     short nl_k,
     void (*deq_k_t4)(device const kd4_t *, short, thread k4_t &),
-    typename vd4_t, // key type in device memory
+    typename vd4_t, // value type in device memory
     short nl_v,
     void (*deq_v_t4)(device const vd4_t *, short, thread v4_t &),
     short DK,       // K head size
@@ -3741,6 +3924,11 @@ kernel void kernel_flash_attn_ext_vec(
                 sm[tiisg] = pm[ic + tiisg];
             }
 
+            // skip -INF blocks
+            if (simd_max(sm[tiisg]) == -INFINITY) {
+                continue;
+            }
+
             // Q*K^T
             {
                 // each simdgroup processes 1 query and NE (NW/NL) head elements
@@ -3973,6 +4161,16 @@ kernel void kernel_flash_attn_ext_vec(
 
 typedef decltype(kernel_flash_attn_ext_vec<FA_TYPES, half4, 1, dequantize_f16_t4, half4, 1, dequantize_f16_t4, 128, 128, 4>) flash_attn_ext_vec_t;
 
+template [[host_name("kernel_flash_attn_ext_vec_f16_h64")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4,             1, dequantize_f16_t4,  half4,       1, dequantize_f16_t4,  64, 64, 8>;
+#if defined(GGML_METAL_USE_BF16)
+template [[host_name("kernel_flash_attn_ext_vec_bf16_h64")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4,           1, dequantize_bf16_t4, bfloat4,     1, dequantize_bf16_t4, 64, 64, 8>;
+#endif
+template [[host_name("kernel_flash_attn_ext_vec_q4_0_h64")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_0,        8, dequantize_q4_0_t4, block_q4_0,  8, dequantize_q4_0_t4, 64, 64, 8>;
+template [[host_name("kernel_flash_attn_ext_vec_q4_1_h64")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_1,        8, dequantize_q4_1_t4, block_q4_1,  8, dequantize_q4_1_t4, 64, 64, 8>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_0_h64")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_0,        8, dequantize_q5_0_t4, block_q5_0,  8, dequantize_q5_0_t4, 64, 64, 8>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_1_h64")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_1,        8, dequantize_q5_1_t4, block_q5_1,  8, dequantize_q5_1_t4, 64, 64, 8>;
+template [[host_name("kernel_flash_attn_ext_vec_q8_0_h64")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q8_0,        8, dequantize_q8_0_t4, block_q8_0,  8, dequantize_q8_0_t4, 64, 64, 8>;
+
 template [[host_name("kernel_flash_attn_ext_vec_f16_h96")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4,             1, dequantize_f16_t4,  half4,       1, dequantize_f16_t4,  96, 96, 4>;
 #if defined(GGML_METAL_USE_BF16)
 template [[host_name("kernel_flash_attn_ext_vec_bf16_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4,           1, dequantize_bf16_t4, bfloat4,     1, dequantize_bf16_t4, 96, 96, 4>;

package/cpp/llama.cpp/ggml/src/ggml-musa/CMakeLists.txt

@@ -27,12 +27,15 @@ if (MUSAToolkit_FOUND)
 
     file(GLOB   GGML_HEADERS_MUSA "../ggml-cuda/*.cuh")
     list(APPEND GGML_HEADERS_MUSA "../../include/ggml-cuda.h")
+    list(APPEND GGML_HEADERS_MUSA "../ggml-musa/mudnn.cuh")
 
     file(GLOB   GGML_SOURCES_MUSA "../ggml-cuda/*.cu")
     file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-mma*.cu")
     list(APPEND GGML_SOURCES_MUSA ${SRCS})
     file(GLOB   SRCS "../ggml-cuda/template-instances/mmq*.cu")
     list(APPEND GGML_SOURCES_MUSA ${SRCS})
+    file(GLOB   SRCS "../ggml-musa/*.cu")
+    list(APPEND GGML_SOURCES_MUSA ${SRCS})
 
     if (GGML_CUDA_FA_ALL_QUANTS)
         file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*.cu")
@@ -62,7 +65,9 @@ if (MUSAToolkit_FOUND)
                             )
 
     # TODO: do not use CUDA definitions for MUSA
-    target_compile_definitions(ggml PUBLIC GGML_USE_CUDA)
+    if (NOT GGML_BACKEND_DL)
+        target_compile_definitions(ggml PUBLIC GGML_USE_CUDA)
+    endif()
 
     add_compile_definitions(GGML_USE_MUSA)
     add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
@@ -92,9 +97,10 @@ if (MUSAToolkit_FOUND)
     endif()
 
     if (GGML_STATIC)
+        # TODO: mudnn has not provided static libraries yet
         target_link_libraries(ggml-musa PRIVATE MUSA::musart_static MUSA::mublas_static)
     else()
-        target_link_libraries(ggml-musa PRIVATE MUSA::musart MUSA::mublas)
+        target_link_libraries(ggml-musa PRIVATE MUSA::musart MUSA::mublas mudnn)
     endif()
 
     if (GGML_CUDA_NO_VMM)

package/cpp/llama.cpp/ggml/src/ggml-musa/mudnn.cu

@@ -0,0 +1,112 @@
+#include <mutex>
+#include <mudnn.h>
+
+#include "mudnn.cuh"
+
+namespace mudnn = musa::dnn;
+
+// Returns a human-readable error string for mudnn::Status
+const char* mudnnGetErrorString(mudnn::Status err) {
+    switch (err) {
+        case mudnn::Status::SUCCESS:
+            return "Success";
+        case mudnn::Status::INVALID_PARAMETER:
+            return "Invalid parameter";
+        case mudnn::Status::NOT_INITIALIZED:
+            return "Not initialized";
+        case mudnn::Status::ALLOC_FAILED:
+            return "Allocation failed";
+        case mudnn::Status::NOT_SUPPORTED:
+            return "Not supported";
+        case mudnn::Status::INTERNAL_ERROR:
+            return "Internal error";
+        case mudnn::Status::ARCH_MISMATCH:
+            return "Architecture mismatch";
+        case mudnn::Status::EXECUTION_FAILED:
+            return "Execution failed";
+        default:
+            return "Unknown mudnn status";
+    }
+}
+
+// Error checking macro for MUDNN calls
+#define MUDNN_CHECK(err) CUDA_CHECK_GEN(err, mudnn::Status::SUCCESS, mudnnGetErrorString)
+
+namespace {
+    // Thread-safe cache for mudnn::Handle objects per device
+    std::unordered_map<int, std::unique_ptr<mudnn::Handle>> handle_cache;
+    std::mutex handle_cache_mutex;
+
+    mudnn::Handle* get_cached_handle(int device_id) {
+        std::lock_guard<std::mutex> lock(handle_cache_mutex);
+        auto it = handle_cache.find(device_id);
+        if (it != handle_cache.end()) {
+            return it->second.get();
+        }
+        auto handle = std::make_unique<mudnn::Handle>(device_id);
+        mudnn::Handle* handle_ptr = handle.get();
+        handle_cache[device_id] = std::move(handle);
+        return handle_ptr;
+    }
+}
+
+// Extracts dimensions and strides from a ggml_tensor
+int get_ggml_dims_and_strides(const ggml_tensor* tensor,
+                              std::vector<int64_t>& dims,
+                              std::vector<int64_t>& strides) {
+    const int ndims = ggml_n_dims(tensor);
+    const size_t element_size = ggml_element_size(tensor);
+
+    dims.resize(ndims);
+    strides.resize(ndims);
+
+    for (int i = 0; i < ndims; ++i) {
+        dims[i] = tensor->ne[i];
+        strides[i] = tensor->nb[i] / static_cast<int64_t>(element_size);
+    }
+    return ndims;
+}
+
+// Converts ggml_type to mudnn::Tensor::Type
+mudnn::Tensor::Type ggml_type_to_mudnn_type(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_F32:
+            return mudnn::Tensor::Type::FLOAT;
+        case GGML_TYPE_F16:
+            return mudnn::Tensor::Type::HALF;
+
+        // TODO: Add support for other types
+
+        default:
+            MUDNN_CHECK(mudnn::Status::NOT_SUPPORTED);
+    }
+
+    return mudnn::Tensor::Type::FLOAT; // Default fallback
+}
+
+// Asynchronous memory copy using mudnn::Unary::IDENTITY
+musaError_t mudnnMemcpyAsync(ggml_backend_cuda_context& ctx, const ggml_tensor* dst, const ggml_tensor* src) {
+    mudnn::Tensor tensor_dst, tensor_src;
+
+    MUDNN_CHECK(tensor_dst.SetType(ggml_type_to_mudnn_type(dst->type)));
+    MUDNN_CHECK(tensor_src.SetType(ggml_type_to_mudnn_type(src->type)));
+
+    std::vector<int64_t> dims, strides;
+    const int ndims = get_ggml_dims_and_strides(src, dims, strides);
+
+    MUDNN_CHECK(tensor_dst.SetNdInfo(ndims, dims.data(), strides.data()));
+    MUDNN_CHECK(tensor_src.SetNdInfo(ndims, dims.data(), strides.data()));
+    MUDNN_CHECK(tensor_dst.SetAddr(dst->data));
+    MUDNN_CHECK(tensor_src.SetAddr(src->data));
+
+    mudnn::Unary op;
+    MUDNN_CHECK(op.SetMode(mudnn::Unary::Mode::IDENTITY));
+    MUDNN_CHECK(op.SetAlpha(0.0f));
+    MUDNN_CHECK(op.SetBeta(0.0f));
+
+    mudnn::Handle* handle = get_cached_handle(ctx.device);
+    MUDNN_CHECK(handle->SetStream(ctx.stream()));
+    MUDNN_CHECK(op.Run(*handle, tensor_dst, tensor_src));
+
+    return musaSuccess;
+}

package/cpp/llama.cpp/ggml/src/ggml-musa/mudnn.cuh

@@ -0,0 +1,12 @@
+#pragma once
+
+#include "../include/ggml.h"
+#include "../ggml-cuda/common.cuh"
+
+// Asynchronously copies data from src tensor to dst tensor using the provided context.
+// Returns a musaError_t indicating success or failure.
+musaError_t mudnnMemcpyAsync(
+    ggml_backend_cuda_context &ctx,
+    const ggml_tensor *dst,
+    const ggml_tensor *src
+);

package/cpp/llama.cpp/ggml/src/ggml-opencl/CMakeLists.txt

@@ -55,14 +55,17 @@ endfunction()
 
 set(GGML_OPENCL_KERNELS
     add
+    argsort
     clamp
     cpy
     cvt
     diag_mask_inf
+    div
     gelu
     gemv_noshuffle_general
     gemv_noshuffle
     get_rows
+    group_norm
     im2col_f32
     im2col_f16
     mul_mat_Ab_Bi_8x4
@@ -83,11 +86,14 @@ set(GGML_OPENCL_KERNELS
     rms_norm
     rope
     scale
+    sigmoid
     silu
     softmax_4_f32
     softmax_4_f16
     softmax_f32
     softmax_f16
+    sub
+    sum_rows
     transpose
 )