llama_cpp 0.10.3 → 0.11.0

data/{ext/llama_cpp/src → vendor/tmp/llama.cpp}/ggml-backend.c

@@ -614,10 +614,14 @@ static void ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_c
 }
 
 static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
-    return true;
+    switch (op->op) {
+        case GGML_OP_MUL_MAT:
+            return op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == ggml_internal_get_type_traits(op->src[0]->type).vec_dot_type;
+        default:
+            return true;
+    }
 
     GGML_UNUSED(backend);
-    GGML_UNUSED(op);
 }
 
 static struct ggml_backend_i cpu_backend_i = {

data/{ext/llama_cpp/src → vendor/tmp/llama.cpp}/ggml-cuda.cu

@@ -119,7 +119,9 @@
 #define MIN_CC_DP4A   610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products
 #define CC_VOLTA      700
 #define CC_OFFSET_AMD 1000000
+#define CC_RDNA1      (CC_OFFSET_AMD + 1010)
 #define CC_RDNA2      (CC_OFFSET_AMD + 1030)
+#define CC_RDNA3      (CC_OFFSET_AMD + 1100)
 
 #define GGML_CUDA_MAX_NODES 8192
 
@@ -133,7 +135,6 @@
 
 // TODO: improve this to be correct for more hardware
 //       for example, currently fails for GeForce GTX 1660 which is TURING arch (> VOLTA) but does not have tensor cores
-//       probably other such cases, and not sure what happens on AMD hardware
 #if !defined(GGML_CUDA_FORCE_MMQ)
 #define CUDA_USE_TENSOR_CORES
 #endif
@@ -6662,7 +6663,7 @@ static void ggml_cuda_pool_free_leg(int device, void * ptr, size_t size) {
 // pool with virtual memory
 static CUdeviceptr g_cuda_pool_addr[GGML_CUDA_MAX_DEVICES] = {0};
 static size_t g_cuda_pool_used[GGML_CUDA_MAX_DEVICES] = {0};
-static const size_t CUDA_POOL_VMM_MAX_SIZE = 1ull << 36; // 64 GB
+static const size_t CUDA_POOL_VMM_MAX_SIZE = 1ull << 35; // 32 GB
 
 static void * ggml_cuda_pool_malloc_vmm(int device, size_t size, size_t * actual_size) {
     scoped_spin_lock lock(g_cuda_pool_lock);
@@ -7485,6 +7486,8 @@ static void ggml_cuda_op_dequantize_mul_mat_vec(
     const int64_t ne00 = src0->ne[0];
     const int64_t row_diff = row_high - row_low;
 
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+
     // on some GPUs it is faster to convert src1 to half and to use half precision intrinsics
 #ifdef GGML_CUDA_F16
     cuda_pool_alloc<half> src1_dfloat_a;
@@ -7577,6 +7580,7 @@ static void ggml_cuda_op_mul_mat_cublas(
     const int compute_capability = g_device_caps[id].cc;
 
     if (compute_capability >= CC_VOLTA && (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == src0->ne[1] && dst->op_params[0] == GGML_PREC_DEFAULT) {
+        //printf("this branch\n");
         // convert src0 and src1 to fp16, multiply as fp16, convert dst to fp32
         cuda_pool_alloc<half> src0_as_f16;
         if (src0->type != GGML_TYPE_F16) {
@@ -7614,9 +7618,9 @@ static void ggml_cuda_op_mul_mat_cublas(
 
         const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16);
         to_fp32_cuda(dst_f16.get(), dst_dd_i, row_diff*src1_ncols, stream);
-    }
-    else {
+    } else {
         cuda_pool_alloc<float> src0_ddq_as_f32;
+        cuda_pool_alloc<float> src1_ddq_as_f32;
 
         if (src0->type != GGML_TYPE_F32) {
             const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(src0->type);
@@ -7624,7 +7628,15 @@ static void ggml_cuda_op_mul_mat_cublas(
             src0_ddq_as_f32.alloc(row_diff*ne00);
             to_fp32_cuda(src0_dd_i, src0_ddq_as_f32.get(), row_diff*ne00, stream);
         }
+        if (src1->type != GGML_TYPE_F32) {
+            const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(src1->type);
+            GGML_ASSERT(to_fp32_cuda != nullptr);
+            src1_ddq_as_f32.alloc(src1_ncols*ne10);
+            to_fp32_cuda(src1_ddf_i, src1_ddq_as_f32.get(), src1_ncols*ne10, stream);
+        }
+
         const float * src0_ddf_i = src0->type == GGML_TYPE_F32 ? (const float *) src0_dd_i : src0_ddq_as_f32.get();
+        const float * src1_ddf1_i = src1->type == GGML_TYPE_F32 ? (const float *) src1_ddf_i : src1_ddq_as_f32.get();
 
         const float alpha = 1.0f;
         const float beta = 0.0f;
@@ -7633,9 +7645,9 @@ static void ggml_cuda_op_mul_mat_cublas(
         CUBLAS_CHECK(
             cublasSgemm(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N,
                     row_diff, src1_ncols, ne10,
-                    &alpha, src0_ddf_i, ne00,
-                            src1_ddf_i, ne10,
-                    &beta,  dst_dd_i,   ldc));
+                    &alpha, src0_ddf_i,  ne00,
+                            src1_ddf1_i, ne10,
+                    &beta,  dst_dd_i,    ldc));
     }
 
     (void) dst;
@@ -8035,6 +8047,7 @@ static void ggml_cuda_op_mul_mat(
 
     GGML_ASSERT(dst->backend != GGML_BACKEND_GPU_SPLIT);
     GGML_ASSERT(src1->backend != GGML_BACKEND_GPU_SPLIT);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32 || (src1->ne[2] == 1 && src1->ne[3] == 1));
 
     GGML_ASSERT(ne12 >= ne02 && ne12 % ne02 == 0);
 
@@ -8481,9 +8494,9 @@ static __global__ void k_compute_batched_ptrs(
     int64_t i03 = i13 / r3;
     int64_t i02 = i12 / r2;
 
-    ptrs_src[0*ne23 + i12 + i13*ne12] = (const char *) src0_as_f16 + i02*nb02   + i03*nb03;
-    ptrs_src[1*ne23 + i12 + i13*ne12] = (const char *) src1_as_f16 + i12*nb12/2 + i13*nb13/2;
-    ptrs_dst[0*ne23 + i12 + i13*ne12] = (      char *)         dst + i12*nbd2   + i13*nbd3;
+    ptrs_src[0*ne23 + i12 + i13*ne12] = (const char *) src0_as_f16 + i02*nb02 + i03*nb03;
+    ptrs_src[1*ne23 + i12 + i13*ne12] = (const char *) src1_as_f16 + i12*nb12 + i13*nb13;
+    ptrs_dst[0*ne23 + i12 + i13*ne12] = (      char *)         dst + i12*nbd2 + i13*nbd3;
 }
 
 static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -8492,28 +8505,10 @@ static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const
 
     GGML_ASSERT(src0->backend != GGML_BACKEND_GPU_SPLIT);
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-
-    const int64_t ne00 = src0->ne[0]; GGML_UNUSED(ne00);
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
-    const int64_t ne03 = src0->ne[3];
 
-    const int64_t nb01 = src0->nb[1];
-    const int64_t nb02 = src0->nb[2]; GGML_UNUSED(nb02);
-    const int64_t nb03 = src0->nb[3]; GGML_UNUSED(nb03);
-
-    const int64_t ne10 = src1->ne[0];
-    const int64_t ne11 = src1->ne[1];
-    const int64_t ne12 = src1->ne[2];
-    const int64_t ne13 = src1->ne[3];
-
-    const int64_t nb11 = src1->nb[1];
-    const int64_t nb12 = src1->nb[2]; GGML_UNUSED(nb12);
-    const int64_t nb13 = src1->nb[3]; GGML_UNUSED(nb13);
+    GGML_TENSOR_BINARY_OP_LOCALS
 
-    const int64_t ne1 = ggml_nelements(src1);
-    const int64_t ne  = ggml_nelements(dst);
+    const int64_t ne_dst = ggml_nelements(dst);
 
     ggml_cuda_set_device(g_main_device);
     cudaStream_t main_stream = g_cudaStreams[g_main_device][0];
@@ -8522,7 +8517,7 @@ static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const
 
     ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
     void * src0_ddq = src0_extra->data_device[g_main_device];
-    half * src0_as_f16 = (half *) src0_ddq;
+    half * src0_f16 = (half *) src0_ddq;
 
     ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
     float * src1_ddf = (float *) src1_extra->data_device[g_main_device];
@@ -8531,11 +8526,15 @@ static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const
     float * dst_ddf = (float *) dst_extra->data_device[g_main_device];
 
     // convert src1 to fp16
-    const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
-    GGML_ASSERT(to_fp16_cuda != nullptr);
-
-    cuda_pool_alloc<half> src1_as_f16(ne1);
-    to_fp16_cuda(src1_ddf, src1_as_f16.get(), ne1, main_stream);
+    cuda_pool_alloc<half> src1_f16_alloc;
+    if (src1->type != GGML_TYPE_F16) {
+        const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
+        const int64_t ne_src1 = ggml_nelements(src1);
+        src1_f16_alloc.alloc(ne_src1);
+        GGML_ASSERT(to_fp16_cuda != nullptr);
+        to_fp16_cuda(src1_ddf, src1_f16_alloc.get(), ne_src1, main_stream);
+    }
+    half * src1_f16 = src1->type == GGML_TYPE_F16 ? (half *) src1_ddf : src1_f16_alloc.get();
 
     cuda_pool_alloc<half> dst_f16;
     char * dst_t;
@@ -8557,7 +8556,7 @@ static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const
     const void * beta  = &beta_f16;
 
     if (dst->op_params[0] == GGML_PREC_DEFAULT) {
-        dst_t = (char *) dst_f16.alloc(ne);
+        dst_t = (char *) dst_f16.alloc(ne_dst);
 
         nbd2 /= sizeof(float) / sizeof(half);
         nbd3 /= sizeof(float) / sizeof(half);
@@ -8604,9 +8603,9 @@ static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const
         CUBLAS_CHECK(
         cublasGemmStridedBatchedEx(g_cublas_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N,
                 ne01, ne11, ne10,
-                alpha, (const char *) src0_as_f16,       CUDA_R_16F,   nb01/sizeof(half),  src0->nb[2]/sizeof(half),  // strideA
-                       (const char *) src1_as_f16.get(), CUDA_R_16F,   nb11/sizeof(float), src1->nb[2]/sizeof(float), // strideB
-                beta,  (      char *)       dst_t,       cu_data_type, ne01,                dst->nb[2]/sizeof(float), // strideC
+                alpha, (const char *) src0_f16, CUDA_R_16F,   nb01/nb00, nb02/nb00,  // strideA
+                       (const char *) src1_f16, CUDA_R_16F,   nb11/nb10, nb12/nb10,  // strideB
+                beta,  (      char *)    dst_t, cu_data_type, ne01,       nb2/nb0,   // strideC
                 ne12*ne13,
                 cu_compute_type,
                 CUBLAS_GEMM_DEFAULT_TENSOR_OP));
@@ -8619,12 +8618,13 @@ static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const
 
         dim3 block_dims(ne13, ne12);
         k_compute_batched_ptrs<<<1, block_dims, 0, main_stream>>>(
-                src0_as_f16, src1_as_f16.get(), dst_t,
+                src0_f16, src1_f16, dst_t,
                 ptrs_src.get(), ptrs_dst.get(),
                 ne12, ne13,
                 ne23,
                 nb02, nb03,
-                nb12, nb13,
+                src1->type == GGML_TYPE_F16 ? nb12 : nb12/2,
+                src1->type == GGML_TYPE_F16 ? nb13 : nb13/2,
                 nbd2, nbd3,
                 r2, r3);
         CUDA_CHECK(cudaGetLastError());
@@ -8632,8 +8632,8 @@ static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const
         CUBLAS_CHECK(
         cublasGemmBatchedEx(g_cublas_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N,
                 ne01, ne11, ne10,
-                alpha, (const void **) (ptrs_src.get() + 0*ne23), CUDA_R_16F,   nb01/sizeof(half),
-                       (const void **) (ptrs_src.get() + 1*ne23), CUDA_R_16F,   nb11/sizeof(float),
+                alpha, (const void **) (ptrs_src.get() + 0*ne23), CUDA_R_16F,   nb01/nb00,
+                       (const void **) (ptrs_src.get() + 1*ne23), CUDA_R_16F,   nb11/nb10,
                 beta,  (      void **) (ptrs_dst.get() + 0*ne23), cu_data_type, ne01,
                 ne23,
                 cu_compute_type,
@@ -8643,7 +8643,7 @@ static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const
 
     if (dst->op_params[0] == GGML_PREC_DEFAULT) {
         const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16);
-        to_fp32_cuda(dst_f16.get(), dst_ddf, ne, main_stream);
+        to_fp32_cuda(dst_f16.get(), dst_ddf, ne_dst, main_stream);
     }
 }
 
@@ -8662,11 +8662,25 @@ static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1
         }
     }
 
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+
+    const bool fp16_performance_good = min_compute_capability >= CC_RDNA1;
+    bool               use_mul_mat_q = ggml_is_quantized(src0->type);
 #ifdef CUDA_USE_TENSOR_CORES
-    const bool use_tensor_cores = true;
+    use_mul_mat_q = use_mul_mat_q && min_compute_capability < CC_RDNA3;
+#endif // CUDA_USE_TENSOR_CORES
+
 #else
-    const bool use_tensor_cores = false;
-#endif
+
+    const bool fp16_performance_good = min_compute_capability >= CC_VOLTA;
+    bool               use_mul_mat_q = min_compute_capability >= MIN_CC_DP4A && ggml_is_quantized(src0->type);
+#ifdef CUDA_USE_TENSOR_CORES
+    // when tensor cores are available, use them for large batch size
+    // ref: https://github.com/ggerganov/llama.cpp/pull/3776
+    use_mul_mat_q = use_mul_mat_q && !(fp16_performance_good && src1->ne[1] > MMQ_MAX_BATCH_SIZE);
+#endif // CUDA_USE_TENSOR_CORES
+
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
 
     // debug helpers
     //printf("src0: %8d %8d %8d %8d\n", src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3]);
@@ -8676,19 +8690,19 @@ static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1
     //printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
     //printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
 
-    if (!split && all_on_device && !use_tensor_cores && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
+    if (!split && all_on_device && !fp16_performance_good && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
         // KQ single-batch
         ggml_cuda_mul_mat_vec_p021(src0, src1, dst);
-    } else if (!split && all_on_device && !use_tensor_cores && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
+    } else if (!split && all_on_device && !fp16_performance_good && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
         // KQV single-batch
         ggml_cuda_mul_mat_vec_nc(src0, src1, dst);
-    } else if (!split && all_on_device && use_tensor_cores && src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1)) {
+    } else if (!split && all_on_device && fp16_performance_good && src0->type == GGML_TYPE_F16 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1)) {
         // KQ + KQV multi-batch
         ggml_cuda_mul_mat_mat_batched_cublas(src0, src1, dst);
     } else if (src0->type == GGML_TYPE_F32) {
         ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, false);
     } else if (ggml_is_quantized(src0->type) || src0->type == GGML_TYPE_F16) {
-        if (src1->ne[1] == 1 && src0->ne[0] % GGML_CUDA_DMMV_X == 0) {
+        if (src1->ne[1] == 1 && src0->ne[0] % GGML_CUDA_DMMV_X == 0 && src1->type == GGML_TYPE_F32) {
 #ifdef GGML_CUDA_FORCE_DMMV
             const bool use_mul_mat_vec_q = false;
 #else
@@ -8702,14 +8716,6 @@ static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1
                 ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, false);
             }
         } else {
-            bool use_mul_mat_q = min_compute_capability >= MIN_CC_DP4A && ggml_is_quantized(src0->type);
-
-            // when tensor cores are available, use them for large batch size
-            // ref: https://github.com/ggerganov/llama.cpp/pull/3776
-            if (use_tensor_cores && min_compute_capability >= CC_VOLTA && src1->ne[1] > MMQ_MAX_BATCH_SIZE) {
-                use_mul_mat_q = false;
-            }
-
             if (use_mul_mat_q) {
                 ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_q, true);
             } else {
@@ -10033,14 +10039,19 @@ static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, const ggml_ten
                 }
                 return false;
             } break;
+        case GGML_OP_DUP:
+        case GGML_OP_REPEAT:
+        case GGML_OP_CONCAT:
+            {
+                ggml_type src0_type = op->src[0]->type;
+                return src0_type != GGML_TYPE_I32 && src0_type != GGML_TYPE_I16;
+            } break;
         case GGML_OP_NONE:
         case GGML_OP_RESHAPE:
         case GGML_OP_VIEW:
         case GGML_OP_PERMUTE:
         case GGML_OP_TRANSPOSE:
         case GGML_OP_NORM:
-        case GGML_OP_REPEAT:
-        case GGML_OP_DUP:
         case GGML_OP_ADD:
         case GGML_OP_MUL:
         case GGML_OP_DIV:
@@ -10057,7 +10068,6 @@ static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, const ggml_ten
         case GGML_OP_SUM_ROWS:
         case GGML_OP_ARGSORT:
         case GGML_OP_ACC:
-        case GGML_OP_CONCAT:
         case GGML_OP_GROUP_NORM:
         case GGML_OP_UPSCALE:
         case GGML_OP_PAD:

data/{ext/llama_cpp/src → vendor/tmp/llama.cpp}/ggml-impl.h

@@ -5,6 +5,7 @@
 // GGML internal header
 
 #include <assert.h>
+#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
 #include <stddef.h>
 #include <stdbool.h>
 #include <string.h> // memcpy

data/{ext/llama_cpp/src → vendor/tmp/llama.cpp}/ggml-metal.m

@@ -87,6 +87,7 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(get_rows_q4_K);
     GGML_METAL_DECL_KERNEL(get_rows_q5_K);
     GGML_METAL_DECL_KERNEL(get_rows_q6_K);
+    GGML_METAL_DECL_KERNEL(get_rows_i32);
     GGML_METAL_DECL_KERNEL(rms_norm);
     GGML_METAL_DECL_KERNEL(group_norm);
     GGML_METAL_DECL_KERNEL(norm);
@@ -257,13 +258,14 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
 #endif
         NSError * error = nil;
-        NSString * libPath = [bundle pathForResource:@"default" ofType:@"metallib"];
+        NSString * libPath = [bundle pathForResource:@"ggml" ofType:@"metallib"];
         if (libPath != nil) {
+            // pre-compiled library found
             NSURL * libURL = [NSURL fileURLWithPath:libPath];
             GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [libPath UTF8String]);
             ctx->library = [ctx->device newLibraryWithURL:libURL error:&error];
         } else {
-            GGML_METAL_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__);
+            GGML_METAL_LOG_INFO("%s: ggml.metallib not found, loading from source\n", __func__);
 
             NSString * sourcePath;
             NSString * ggmlMetalPathResources = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
@@ -291,6 +293,13 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
             options = [MTLCompileOptions new];
             options.preprocessorMacros = @{ @"QK_K" : @(64) };
 #endif
+            // try to disable fast-math
+            // NOTE: this seems to have no effect whatsoever
+            //       instead, in order to disable fast-math, we have to build ggml.metallib from the command line
+            //       using xcrun -sdk macosx metal -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
+            //       and go through the "pre-compiled library found" path above
+            //[options setFastMathEnabled:false];
+
             ctx->library = [ctx->device newLibraryWithSource:src options:options error:&error];
         }
 
@@ -369,6 +378,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(get_rows_q4_K);
         GGML_METAL_ADD_KERNEL(get_rows_q5_K);
         GGML_METAL_ADD_KERNEL(get_rows_q6_K);
+        GGML_METAL_ADD_KERNEL(get_rows_i32);
         GGML_METAL_ADD_KERNEL(rms_norm);
         GGML_METAL_ADD_KERNEL(group_norm);
         GGML_METAL_ADD_KERNEL(norm);
@@ -491,6 +501,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_DEL_KERNEL(get_rows_q4_K);
     GGML_METAL_DEL_KERNEL(get_rows_q5_K);
     GGML_METAL_DEL_KERNEL(get_rows_q6_K);
+    GGML_METAL_DEL_KERNEL(get_rows_i32);
     GGML_METAL_DEL_KERNEL(rms_norm);
     GGML_METAL_DEL_KERNEL(group_norm);
     GGML_METAL_DEL_KERNEL(norm);
@@ -1230,7 +1241,7 @@ void ggml_metal_graph_compute(
                                 // not sure how to avoid this
                                 // TODO: make a simpler cpy_bytes kernel
 
-                                const int nth = MIN(1024, ne00);
+                                const int nth = MIN((int) ctx->pipeline_cpy_f32_f32.maxTotalThreadsPerThreadgroup, ne00);
 
                                 [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f32];
                                 [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
@@ -1285,7 +1296,7 @@ void ggml_metal_graph_compute(
                             [encoder setBytes:&pnb3 length:sizeof(pnb3) atIndex:26];
                             [encoder setBytes:&offs length:sizeof(offs) atIndex:27];
 
-                            const int nth = MIN(1024, ne0);
+                            const int nth = MIN((int) ctx->pipeline_add.maxTotalThreadsPerThreadgroup, ne00);
 
                             [encoder dispatchThreadgroups:MTLSizeMake(ne11, ne12, ne13) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
@@ -1649,6 +1660,10 @@ void ggml_metal_graph_compute(
                                         }
                                 };
 
+                                if (ggml_is_quantized(src0t)) {
+                                    GGML_ASSERT(ne00 >= nth0*nth1);
+                                }
+
                                 [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                                 [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
                                 [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
@@ -1707,6 +1722,9 @@ void ggml_metal_graph_compute(
                             // TODO: make this more general
                             GGML_ASSERT(n_as <= 8);
 
+                            // max size of the src1ids array in the kernel stack
+                            GGML_ASSERT(ne11 <= 512);
+
                             struct ggml_tensor * src2 = gf->nodes[i]->src[2];
 
                             const int64_t  ne20 = src2 ? src2->ne[0] : 0;
@@ -1724,9 +1742,6 @@ void ggml_metal_graph_compute(
                             GGML_ASSERT(!ggml_is_transposed(src2));
                             GGML_ASSERT(!ggml_is_transposed(src1));
 
-                            GGML_ASSERT(ne20 % 32 == 0);
-                            // !!!!!!!!! TODO: this assert is probably required but not sure!
-                            //GGML_ASSERT(ne20 >= 64);
                             GGML_ASSERT(src1t == GGML_TYPE_F32);
 
                             const uint r2 = ne12/ne22;
@@ -1734,22 +1749,22 @@ void ggml_metal_graph_compute(
 
                             // find the break-even point where the matrix-matrix kernel becomes more efficient compared
                             // to the matrix-vector kernel
-                            int ne11_mm_min = 1;
+                            int ne11_mm_min = n_as;
 
                             const int idx = ((int32_t *) dst->op_params)[0];
 
                             // batch size
                             GGML_ASSERT(ne01 == ne11);
 
-                            const int64_t _ne1 = 1; // kernel_mul_mm_impl needs a reference in constant memory
-
                             // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
                             // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
                             // !!!
                             // TODO: for now, always use mat-vec kernels until we figure out how to improve the
                             //       indirect matrix multiplication
                             // !!!
-                            if ([ctx->device supportsFamily:MTLGPUFamilyApple7] && _ne1 > ne11_mm_min) {
+                            if ([ctx->device supportsFamily:MTLGPUFamilyApple7] &&
+                                ne20 % 32 == 0 && ne20 >= 64 &&
+                                ne11 > ne11_mm_min) {
                                 switch (src2->type) {
                                     case GGML_TYPE_F32:  [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_f32_f32];  break;
                                     case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_f16_f32];  break;
@@ -1779,14 +1794,15 @@ void ggml_metal_graph_compute(
                                 [encoder setBytes:&nb11    length:sizeof(nb11) atIndex:11];
                                 [encoder setBytes:&nb12    length:sizeof(nb12) atIndex:12];
                                 [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:13];
-                                [encoder setBytes:&_ne1    length:sizeof(_ne1) atIndex:14];
+                                [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:14];
                                 [encoder setBytes:&nb1     length:sizeof(nb1)  atIndex:15];
                                 [encoder setBytes:&r2      length:sizeof(r2)   atIndex:16];
                                 [encoder setBytes:&r3      length:sizeof(r3)   atIndex:17];
                                 [encoder setBytes:&idx     length:sizeof(idx)  atIndex:18];
                                 // TODO: how to make this an array? read Metal docs
-                                for (int j = 0; j < n_as; ++j) {
-                                    struct ggml_tensor * src_cur = dst->src[2 + j];
+                                for (int j = 0; j < 8; ++j) {
+                                    // NOTE: this is done like this to avoid uninitialized kernel arguments when n_as < 8
+                                    struct ggml_tensor * src_cur = dst->src[2 + (j % n_as)];
 
                                     size_t offs_src_cur = 0;
                                     id<MTLBuffer> id_src_cur = ggml_metal_get_buffer(ctx, src_cur, &offs_src_cur);
@@ -1796,8 +1812,7 @@ void ggml_metal_graph_compute(
 
                                 [encoder setThreadgroupMemoryLength:8192 atIndex:0];
 
-                                // TODO: processing one row at a time (ne11 -> 1) is not efficient
-                                [encoder dispatchThreadgroups:MTLSizeMake( (_ne1 + 31)/32, (ne21 + 63)/64, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne11 + 31)/32, (ne21 + 63)/64, n_as*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
                             } else {
                                 int nth0 = 32;
                                 int nth1 = 1;
@@ -1880,11 +1895,17 @@ void ggml_metal_graph_compute(
                                         } break;
                                     default:
                                         {
-                                            GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src0t);
+                                            GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src2t);
                                             GGML_ASSERT(false && "not implemented");
                                         }
                                 };
 
+                                if (ggml_is_quantized(src2t)) {
+                                    GGML_ASSERT(ne20 >= nth0*nth1);
+                                }
+
+                                const int64_t _ne1 = 1; // kernels needs a reference in constant memory
+
                                 [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                                 [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
                                 [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
@@ -1909,8 +1930,9 @@ void ggml_metal_graph_compute(
                                 [encoder setBytes:&r3   length:sizeof(r3)   atIndex:21];
                                 [encoder setBytes:&idx  length:sizeof(idx)  atIndex:22];
                                 // TODO: how to make this an array? read Metal docs
-                                for (int j = 0; j < n_as; ++j) {
-                                    struct ggml_tensor * src_cur = dst->src[2 + j];
+                                for (int j = 0; j < 8; ++j) {
+                                    // NOTE: this is done like this to avoid uninitialized kernel arguments when n_as < 8
+                                    struct ggml_tensor * src_cur = dst->src[2 + (j % n_as)];
 
                                     size_t offs_src_cur = 0;
                                     id<MTLBuffer> id_src_cur = ggml_metal_get_buffer(ctx, src_cur, &offs_src_cur);
@@ -1959,6 +1981,7 @@ void ggml_metal_graph_compute(
                                 case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_K]; break;
                                 case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q5_K]; break;
                                 case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q6_K]; break;
+                                case GGML_TYPE_I32:  [encoder setComputePipelineState:ctx->pipeline_get_rows_i32];  break;
                                 default: GGML_ASSERT(false && "not implemented");
                             }
 
@@ -2229,7 +2252,7 @@ void ggml_metal_graph_compute(
                             [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:17];
                             [encoder setBytes:&sf   length:sizeof(sf)   atIndex:18];
 
-                            const int nth = MIN(1024, ne0);
+                            const int nth = MIN((int) ctx->pipeline_upscale_f32.maxTotalThreadsPerThreadgroup, ne0);
 
                             [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;