llama_cpp 0.16.0 → 0.16.2

data/vendor/tmp/llama.cpp/ggml-cuda/unary.cu

@@ -92,6 +92,15 @@ static __global__ void sqr_f32(const float * x, float * dst, const int k) {
     dst[i] = x[i] * x[i];
 }
 
+static __global__ void sqrt_f32(const float * x, float * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+    dst[i] = sqrtf(x[i]);
+}
+
 static void gelu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_GELU_BLOCK_SIZE - 1) / CUDA_GELU_BLOCK_SIZE;
     gelu_f32<<<num_blocks, CUDA_GELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
@@ -142,12 +151,19 @@ static void sqr_f32_cuda(const float * x, float * dst, const int k, cudaStream_t
     sqr_f32<<<num_blocks, CUDA_SQR_BLOCK_SIZE, 0, stream>>>(x, dst, k);
 }
 
+static void sqrt_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_SQRT_BLOCK_SIZE - 1) / CUDA_SQRT_BLOCK_SIZE;
+    sqrt_f32<<<num_blocks, CUDA_SQRT_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
 void ggml_cuda_op_gelu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const float * src0_d = (const float *)src0->data;
     float * dst_d = (float *)dst->data;
     cudaStream_t stream = ctx.stream();
 
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
@@ -160,6 +176,8 @@ void ggml_cuda_op_silu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     float * dst_d = (float *)dst->data;
     cudaStream_t stream = ctx.stream();
 
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
@@ -172,6 +190,8 @@ void ggml_cuda_op_gelu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
     float * dst_d = (float *)dst->data;
     cudaStream_t stream = ctx.stream();
 
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
@@ -184,6 +204,8 @@ void ggml_cuda_op_tanh(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     float * dst_d = (float *)dst->data;
     cudaStream_t stream = ctx.stream();
 
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
@@ -196,6 +218,8 @@ void ggml_cuda_op_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     float * dst_d = (float *)dst->data;
     cudaStream_t stream = ctx.stream();
 
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
@@ -208,6 +232,8 @@ void ggml_cuda_op_sigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     float * dst_d = (float *)dst->data;
     cudaStream_t stream = ctx.stream();
 
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
@@ -220,6 +246,8 @@ void ggml_cuda_op_hardsigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst
     float * dst_d = (float *)dst->data;
     cudaStream_t stream = ctx.stream();
 
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
@@ -232,6 +260,8 @@ void ggml_cuda_op_hardswish(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
     float * dst_d = (float *)dst->data;
     cudaStream_t stream = ctx.stream();
 
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
@@ -244,6 +274,8 @@ void ggml_cuda_op_leaky_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
     float * dst_d = (float *)dst->data;
     cudaStream_t stream = ctx.stream();
 
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
@@ -259,8 +291,24 @@ void ggml_cuda_op_sqr(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     float * dst_d = (float *)dst->data;
     cudaStream_t stream = ctx.stream();
 
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
     sqr_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
 }
+
+void ggml_cuda_op_sqrt(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    sqrt_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}

data/vendor/tmp/llama.cpp/ggml-cuda.cu

@@ -188,13 +188,15 @@ static ggml_cuda_device_info ggml_cuda_init() {
         info.default_tensor_split[id] = total_vram;
         total_vram += prop.totalGlobalMem;
 
+        info.devices[id].nsm   = prop.multiProcessorCount;
+        info.devices[id].smpb  = prop.sharedMemPerBlock;
 #if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+        info.devices[id].smpbo = prop.sharedMemPerBlock;
         info.devices[id].cc = 100*prop.major + 10*prop.minor + CC_OFFSET_AMD;
 #else
+        info.devices[id].smpbo = prop.sharedMemPerBlockOptin;
         info.devices[id].cc = 100*prop.major + 10*prop.minor;
 #endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-        info.devices[id].smpb = prop.sharedMemPerBlock;
-        info.devices[id].nsm  = prop.multiProcessorCount;
     }
 
     for (int id = 0; id < info.device_count; ++id) {
@@ -543,6 +545,10 @@ GGML_CALL static const char * ggml_backend_cuda_buffer_type_name(ggml_backend_bu
     return ctx->name.c_str();
 }
 
+static bool ggml_backend_buft_is_cuda(ggml_backend_buffer_type_t buft) {
+    return buft->iface.get_name == ggml_backend_cuda_buffer_type_name;
+}
+
 GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
     ggml_backend_cuda_buffer_type_context * buft_ctx = (ggml_backend_cuda_buffer_type_context *)buft->context;
 
@@ -585,24 +591,12 @@ GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backen
     GGML_UNUSED(buft);
 }
 
-GGML_CALL static bool ggml_backend_cuda_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
-    if (!ggml_backend_is_cuda(backend)) {
-        return false;
-    }
-
-    ggml_backend_cuda_buffer_type_context * buft_ctx = (ggml_backend_cuda_buffer_type_context *)buft->context;
-    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
-
-    return buft_ctx->device == cuda_ctx->device;
-}
-
 static ggml_backend_buffer_type_i ggml_backend_cuda_buffer_type_interface = {
     /* .get_name         = */ ggml_backend_cuda_buffer_type_name,
     /* .alloc_buffer     = */ ggml_backend_cuda_buffer_type_alloc_buffer,
     /* .get_alignment    = */ ggml_backend_cuda_buffer_type_get_alignment,
     /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
     /* .get_alloc_size   = */ ggml_backend_cuda_buffer_type_get_alloc_size,
-    /* .supports_backend = */ ggml_backend_cuda_buffer_type_supports_backend,
     /* .is_host          = */ NULL,
 };
 
@@ -633,88 +627,22 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device) {
 
 // cuda split buffer
 
-static int64_t get_row_rounding(ggml_type type, const std::array<float, GGML_CUDA_MAX_DEVICES> & tensor_split) {
-    int64_t min_compute_capability = INT_MAX;
-    int64_t max_compute_capability = INT_MIN;
+static int64_t get_row_rounding(const std::array<float, GGML_CUDA_MAX_DEVICES> & tensor_split) {
+    int64_t row_rounding = 0;
     for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
-        if (tensor_split[id] < (id + 1 < ggml_backend_cuda_get_device_count() ? tensor_split[id + 1] : 1.0f)) {
-            if (min_compute_capability > ggml_cuda_info().devices[id].cc) {
-                min_compute_capability = ggml_cuda_info().devices[id].cc;
-            }
-            if (max_compute_capability < ggml_cuda_info().devices[id].cc) {
-                max_compute_capability = ggml_cuda_info().devices[id].cc;
-            }
+        if (tensor_split[id] >= (id + 1 < ggml_backend_cuda_get_device_count() ? tensor_split[id + 1] : 1.0f)) {
+            continue;
         }
-    }
 
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-    switch(type) {
-        case GGML_TYPE_Q4_0:
-        case GGML_TYPE_Q4_1:
-        case GGML_TYPE_Q5_0:
-        case GGML_TYPE_Q5_1:
-        case GGML_TYPE_Q8_0:
-            return max_compute_capability >= CC_RDNA2 ? 128 : 64;
-        case GGML_TYPE_F16:
-        case GGML_TYPE_F32:
-            return 1;
-        case GGML_TYPE_Q2_K:
-            return max_compute_capability >= CC_RDNA2 ? 128 : 32;
-        case GGML_TYPE_Q3_K:
-            return min_compute_capability < CC_RDNA2 ? 128 : 64;
-        case GGML_TYPE_Q4_K:
-        case GGML_TYPE_Q5_K:
-        case GGML_TYPE_Q6_K:
-        case GGML_TYPE_IQ2_XXS:
-        case GGML_TYPE_IQ2_XS:
-        case GGML_TYPE_IQ2_S:
-        case GGML_TYPE_IQ3_XXS:
-        case GGML_TYPE_IQ1_S:
-        case GGML_TYPE_IQ1_M:
-        case GGML_TYPE_IQ4_NL:
-        case GGML_TYPE_IQ4_XS:
-        case GGML_TYPE_IQ3_S:
-            return max_compute_capability >= CC_RDNA2 ? 128 : 64;
-        default:
-            GGML_ASSERT(false);
+        const int cc = ggml_cuda_info().devices[id].cc;
+        row_rounding = std::max(row_rounding, (int64_t)get_mmq_y_host(cc, get_mmq_x_max_host(cc)));
     }
-#else
-    switch(type) {
-        case GGML_TYPE_Q4_0:
-        case GGML_TYPE_Q4_1:
-            return max_compute_capability >= CC_VOLTA ? 128 : 64;
-        case GGML_TYPE_Q5_0:
-        case GGML_TYPE_Q5_1:
-        case GGML_TYPE_Q8_0:
-            return 64;
-        case GGML_TYPE_F16:
-        case GGML_TYPE_F32:
-            return 1;
-        case GGML_TYPE_Q2_K:
-        case GGML_TYPE_Q3_K:
-        case GGML_TYPE_Q4_K:
-        case GGML_TYPE_Q5_K:
-        case GGML_TYPE_IQ2_XXS:
-        case GGML_TYPE_IQ2_XS:
-        case GGML_TYPE_IQ2_S:
-        case GGML_TYPE_IQ3_XXS:
-        case GGML_TYPE_IQ1_S:
-        case GGML_TYPE_IQ1_M:
-        case GGML_TYPE_IQ4_NL:
-        case GGML_TYPE_IQ4_XS:
-        case GGML_TYPE_IQ3_S:
-            return max_compute_capability >= CC_VOLTA ? 128 : 64;
-        case GGML_TYPE_Q6_K:
-            return 64;
-        default:
-            GGML_ASSERT(false);
-    }
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+    return row_rounding;
 }
 
 static void get_row_split(int64_t * row_low, int64_t * row_high, const ggml_tensor * tensor, const std::array<float, GGML_CUDA_MAX_DEVICES> & tensor_split, int id) {
     const int64_t nrows = ggml_nrows(tensor);
-    const int64_t rounding = get_row_rounding(tensor->type, tensor_split);
+    const int64_t rounding = get_row_rounding(tensor_split);
 
     *row_low = id == 0 ? 0 : nrows*tensor_split[id];
     *row_low -= *row_low % rounding;
@@ -929,6 +857,10 @@ GGML_CALL static const char * ggml_backend_cuda_split_buffer_type_name(ggml_back
     GGML_UNUSED(buft);
 }
 
+static bool ggml_backend_buft_is_cuda_split(ggml_backend_buffer_type_t buft) {
+    return buft->iface.get_name == ggml_backend_cuda_split_buffer_type_name;
+}
+
 GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_split_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
     // since we don't know the exact split after rounding, we cannot allocate the device buffers at this point
     // instead, we allocate them for each tensor separately in init_tensor
@@ -972,12 +904,6 @@ GGML_CALL static size_t ggml_backend_cuda_split_buffer_type_get_alloc_size(ggml_
     return total_size;
 }
 
-GGML_CALL static bool ggml_backend_cuda_split_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
-    return ggml_backend_is_cuda(backend);
-
-    GGML_UNUSED(buft);
-}
-
 GGML_CALL static bool ggml_backend_cuda_split_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
     return false;
 
@@ -990,7 +916,6 @@ static ggml_backend_buffer_type_i ggml_backend_cuda_split_buffer_type_interface
     /* .get_alignment    = */ ggml_backend_cuda_split_buffer_type_get_alignment,
     /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
     /* .get_alloc_size   = */ ggml_backend_cuda_split_buffer_type_get_alloc_size,
-    /* .supports_backend = */ ggml_backend_cuda_split_buffer_type_supports_backend,
     /* .is_host          = */ ggml_backend_cuda_split_buffer_type_is_host,
 };
 
@@ -1090,7 +1015,6 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type() {
             /* .get_alignment    = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
             /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
             /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
-            /* .supports_backend = */ ggml_backend_cpu_buffer_type()->iface.supports_backend,
             /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
         },
         /* .context  = */ nullptr,
@@ -1413,10 +1337,30 @@ static void ggml_cuda_set_peer_access(const int n_tokens, int main_device) {
     GGML_UNUSED(main_device);
 }
 
+static cudaError_t ggml_cuda_Memcpy2DPeerAsync(
+    void * dst, int dstDevice, size_t dpitch, void * src, int srcDevice, size_t spitch, size_t width, size_t height, cudaStream_t stream) {
+
+#if !defined(GGML_USE_HIPBLAS)
+    // cudaMemcpy2DAsync may fail with copies between vmm pools of different devices
+    cudaMemcpy3DPeerParms p = {};
+    p.dstDevice = dstDevice;
+    p.dstPtr = make_cudaPitchedPtr(dst, dpitch, dpitch, height);
+    p.srcDevice = srcDevice;
+    p.srcPtr = make_cudaPitchedPtr(src, spitch, spitch, height);
+    p.extent = make_cudaExtent(width, height, 1);
+    return cudaMemcpy3DPeerAsync(&p, stream);
+#else
+    // HIP does not support cudaMemcpy3DPeerAsync or vmm pools
+    GGML_UNUSED(dstDevice);
+    GGML_UNUSED(srcDevice);
+    return cudaMemcpy2DAsync(dst, dpitch, src, spitch, width, height, cudaMemcpyDeviceToDevice, stream);
+#endif // !defined(GGML_USE_HIPBLAS)
+}
+
 static void ggml_cuda_op_mul_mat(
     ggml_backend_cuda_context & ctx,
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, ggml_cuda_op_mul_mat_t op,
-    const bool convert_src1_to_q8_1) {
+    quantize_cuda_t quantize_src1) {
 
     const int64_t ne00 = src0->ne[0];
     const int64_t ne01 = src0->ne[1];
@@ -1473,7 +1417,9 @@ static void ggml_cuda_op_mul_mat(
     }
 
     struct dev_data {
-        ggml_cuda_pool_alloc<char>  src0_dd_alloc;
+        int cc;
+
+        ggml_cuda_pool_alloc<char>   src0_dd_alloc;
         ggml_cuda_pool_alloc<float> src1_ddf_alloc;
         ggml_cuda_pool_alloc<char>  src1_ddq_alloc;
         ggml_cuda_pool_alloc<float>   dst_dd_alloc;
@@ -1492,6 +1438,8 @@ static void ggml_cuda_op_mul_mat(
     int used_devices = 0;
 
     for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
+        dev[id].cc = ggml_cuda_info().devices[id].cc;
+
         // by default, use all rows
         dev[id].row_low  = 0;
         dev[id].row_high = ne01;
@@ -1499,7 +1447,7 @@ static void ggml_cuda_op_mul_mat(
         // for multi GPU, get the row boundaries from tensor split
         // and round to mul_mat_q tile sizes
         if (split) {
-            const int64_t rounding = get_row_rounding(src0->type, tensor_split);
+            const int64_t rounding = get_row_rounding(tensor_split);
 
             if (id != 0) {
                 dev[id].row_low  = ne01*tensor_split[id];
@@ -1542,11 +1490,15 @@ static void ggml_cuda_op_mul_mat(
             dev[id].src1_ddf = dev[id].src1_ddf_alloc.alloc(ctx.pool(id), ggml_nelements(src1));
         }
 
-        if (convert_src1_to_q8_1) {
-            dev[id].src1_ddq = dev[id].src1_ddq_alloc.alloc(ctx.pool(id), nrows1*src1_padded_col_size*q8_1_ts/q8_1_bs);
+        if (quantize_src1) {
+            size_t src_1_ddq_size = nrows1*src1_padded_col_size*q8_1_ts/q8_1_bs;
+            if (quantize_src1 == quantize_mmq_q8_1_cuda) {
+                src_1_ddq_size += get_mmq_x_max_host(dev[id].cc)*sizeof(block_q8_1_mmq);
+            }
+            dev[id].src1_ddq = dev[id].src1_ddq_alloc.alloc(ctx.pool(id), src_1_ddq_size);
 
             if (src1_on_device && src1_is_contiguous) {
-                quantize_row_q8_1_cuda(dev[id].src1_ddf, dev[id].src1_ddq, ne10, nrows1, src1_padded_col_size, stream);
+                quantize_src1(dev[id].src1_ddf, dev[id].src1_ddq, ne10, ne11, ne12*ne13, src1_padded_col_size, src0->type, stream);
                 CUDA_CHECK(cudaGetLastError());
             }
         }
@@ -1592,7 +1544,12 @@ static void ggml_cuda_op_mul_mat(
                 const int64_t i03 = i0 / ne12;
                 const int64_t i02 = i0 % ne12;
 
-                const size_t src1_ddq_i_offset = (i0*ne11 + src1_col_0) * src1_padded_col_size*q8_1_ts/q8_1_bs;
+                size_t src1_ddq_i_offset = i0*ne11 * src1_padded_col_size*q8_1_ts/q8_1_bs;
+                if (quantize_src1 == quantize_mmq_q8_1_cuda) {
+                    src1_ddq_i_offset += src1_col_0 * sizeof(block_q8_1_mmq);
+                } else {
+                    src1_ddq_i_offset += src1_col_0 * src1_padded_col_size*q8_1_ts/q8_1_bs;
+                }
 
                 // for split tensors the data begins at i0 == i0_offset_low
                 char  *  src0_dd_i =  dev[id].src0_dd + (i0/i02_divisor) * (ne01*ne00*src0_ts)/src0_bs;
@@ -1609,10 +1566,17 @@ static void ggml_cuda_op_mul_mat(
                 // copy src0, src1 to device if necessary
                 if (src1_is_contiguous) {
                     if (id != ctx.device) {
-                        if (convert_src1_to_q8_1) {
+                        if (quantize_src1) {
                             char * src1_ddq_i_source = dev[ctx.device].src1_ddq + src1_ddq_i_offset;
-                            CUDA_CHECK(cudaMemcpyPeerAsync(src1_ddq_i, id, src1_ddq_i_source, ctx.device,
-                                                            src1_ncols*src1_padded_col_size*q8_1_ts/q8_1_bs, stream));
+                            if (quantize_src1 == quantize_mmq_q8_1_cuda) {
+                                const size_t pitch = ne11*sizeof(block_q8_1_mmq);
+                                const size_t width = src1_ncols*sizeof(block_q8_1_mmq);
+                                const size_t height = src1_padded_col_size/(4*QK8_1);
+                                CUDA_CHECK(ggml_cuda_Memcpy2DPeerAsync(src1_ddq_i, id, pitch, src1_ddq_i_source, ctx.device, pitch, width, height, stream));
+                            } else {
+                                CUDA_CHECK(cudaMemcpyPeerAsync(
+                                    src1_ddq_i, id, src1_ddq_i_source, ctx.device, src1_ncols*src1_padded_col_size*q8_1_ts/q8_1_bs, stream));
+                            }
                         } else {
                             float * src1_ddf_i_source = (float *) src1->data;
                             src1_ddf_i_source += (i0*ne11 + src1_col_0) * ne10;
@@ -1627,8 +1591,8 @@ static void ggml_cuda_op_mul_mat(
                     GGML_ASSERT(false);
                 }
 
-                if (convert_src1_to_q8_1 && !src1_is_contiguous) {
-                    quantize_row_q8_1_cuda(src1_ddf_i, src1_ddq_i, ne10, src1_ncols, src1_padded_col_size, stream);
+                if (quantize_src1 && !src1_is_contiguous) {
+                    quantize_src1(src1_ddf_i, src1_ddq_i, ne10, src1_ncols, 1, src1_padded_col_size, src0->type, stream);
                     CUDA_CHECK(cudaGetLastError());
                 }
 
@@ -1653,22 +1617,8 @@ static void ggml_cuda_op_mul_mat(
                         float * dhf_dst_i = (float *) ((char *) dst_off_device + i02*nb2 + i03*nb3);
                         GGML_ASSERT(dst->nb[1] == ne0*sizeof(float));
                         dhf_dst_i += src1_col_0*ne0 + dev[id].row_low;
-#if !defined(GGML_USE_HIPBLAS)
-                        // cudaMemcpy2DAsync may fail with copies between vmm pools of different devices
-                        cudaMemcpy3DPeerParms p = {};
-                        p.dstDevice = ctx.device;
-                        p.dstPtr = make_cudaPitchedPtr(dhf_dst_i, ne0*sizeof(float), row_diff, src1_ncols);
-                        p.srcDevice = id;
-                        p.srcPtr = make_cudaPitchedPtr(dst_dd_i, row_diff*sizeof(float), row_diff, src1_ncols);
-                        p.extent = make_cudaExtent(row_diff*sizeof(float), src1_ncols, 1);
-                        CUDA_CHECK(cudaMemcpy3DPeerAsync(&p, stream));
-#else
-                        // HIP does not support cudaMemcpy3DPeerAsync or vmm pools
-                        CUDA_CHECK(cudaMemcpy2DAsync(dhf_dst_i, ne0*sizeof(float),
-                                                        dst_dd_i, row_diff*sizeof(float),
-                                                        row_diff*sizeof(float), src1_ncols,
-                                                        cudaMemcpyDeviceToDevice, stream));
-#endif
+                        CUDA_CHECK(ggml_cuda_Memcpy2DPeerAsync(
+                            dhf_dst_i, ctx.device, ne0*sizeof(float), dst_dd_i, id, row_diff*sizeof(float), row_diff*sizeof(float), src1_ncols, stream));
                     } else {
                         float * dhf_dst_i = (float *) ((char *) dst_off_device + i02*nb2 + i03*nb3);
                         GGML_ASSERT(dst->nb[1] == ne0*sizeof(float));
@@ -2007,13 +1957,13 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
         // KQ + KQV multi-batch
         ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
     } else if (use_dequantize_mul_mat_vec) {
-        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, false);
+        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, nullptr);
     } else if (use_mul_mat_vec_q) {
-        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_vec_q, true);
+        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_vec_q, quantize_row_q8_1_cuda);
     } else if (use_mul_mat_q) {
-        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_q, true);
+        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_q, quantize_mmq_q8_1_cuda);
     } else {
-        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_cublas, false);
+        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_cublas, nullptr);
     }
 }
 
@@ -2780,7 +2730,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
                 case GGML_UNARY_OP_HARDSWISH:
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_TANH:
-                    return true;
+                    return ggml_is_contiguous(op->src[0]);
                 default:
                     return false;
             }
@@ -2919,6 +2869,20 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
     GGML_UNUSED(backend);
 }
 
+GGML_CALL static bool ggml_backend_cuda_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+    if (ggml_backend_buft_is_cuda_split(buft)) {
+        return true;
+    }
+
+    if (ggml_backend_buft_is_cuda(buft)) {
+        ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
+        ggml_backend_cuda_buffer_type_context * buft_ctx = (ggml_backend_cuda_buffer_type_context *)buft->context;
+        return buft_ctx->device == cuda_ctx->device;
+    }
+
+    return false;
+}
+
 GGML_CALL static bool ggml_backend_cuda_offload_op(ggml_backend_t backend, const ggml_tensor * op) {
     const int min_batch_size = 32;
 
@@ -2991,9 +2955,11 @@ static ggml_backend_i ggml_backend_cuda_interface = {
     /* .synchronize             = */ ggml_backend_cuda_synchronize,
     /* .graph_plan_create       = */ NULL,
     /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_update       = */ NULL,
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_cuda_graph_compute,
     /* .supports_op             = */ ggml_backend_cuda_supports_op,
+    /* .supports_buft           = */ ggml_backend_cuda_supports_buft,
     /* .offload_op              = */ ggml_backend_cuda_offload_op,
     /* .event_new               = */ ggml_backend_cuda_event_new,
     /* .event_free              = */ ggml_backend_cuda_event_free,

data/vendor/tmp/llama.cpp/ggml-impl.h

@@ -17,7 +17,7 @@
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 
-#if defined(_WIN32)
+#if defined(_MSC_VER)
 
 #define m512bh(p) p
 #define m512i(p) p

data/vendor/tmp/llama.cpp/ggml-kompute.cpp

@@ -1340,7 +1340,7 @@ static bool ggml_vk_supports_op(const struct ggml_tensor * op) {
                 case GGML_UNARY_OP_RELU:
                 case GGML_UNARY_OP_GELU:
                 case GGML_UNARY_OP_SILU:
-                    return true;
+                    return ggml_is_contiguous(op->src[0]);
                 default:
                     ;
             }
@@ -1902,18 +1902,12 @@ static size_t ggml_backend_vk_buffer_type_get_max_size(ggml_backend_buffer_type_
     return ctx->max_alloc;
 }
 
-static bool ggml_backend_kompute_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
-    GGML_UNUSED(buft);
-    return ggml_backend_is_kompute(backend);
-}
-
 static ggml_backend_buffer_type_i ggml_backend_kompute_buffer_type_interface = {
     /* .get_name         = */ ggml_backend_kompute_buffer_type_get_name,
     /* .alloc_buffer     = */ ggml_backend_kompute_buffer_type_alloc_buffer,
     /* .get_alignment    = */ ggml_backend_kompute_buffer_type_get_alignment,
     /* .get_max_size     = */ ggml_backend_vk_buffer_type_get_max_size,
     /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
-    /* .supports_backend = */ ggml_backend_kompute_buffer_type_supports_backend,
     /* .is_host          = */ NULL,
 };
 
@@ -1973,6 +1967,11 @@ static bool ggml_backend_kompute_supports_op(ggml_backend_t backend, const struc
     return ggml_vk_supports_op(op);
 }
 
+static bool ggml_backend_kompute_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+    GGML_UNUSED(backend);
+    return buft->iface.get_name == ggml_backend_kompute_buffer_type_get_name;
+}
+
 static struct ggml_backend_i kompute_backend_i = {
     /* .get_name                = */ ggml_backend_kompute_name,
     /* .free                    = */ ggml_backend_kompute_free,
@@ -1983,9 +1982,11 @@ static struct ggml_backend_i kompute_backend_i = {
     /* .synchronize             = */ NULL,
     /* .graph_plan_create       = */ NULL,
     /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_update       = */ NULL,
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_kompute_graph_compute,
     /* .supports_op             = */ ggml_backend_kompute_supports_op,
+    /* .supports_buft           = */ ggml_backend_kompute_supports_buft,
     /* .offload_op              = */ NULL,
     /* .event_new               = */ NULL,
     /* .event_free              = */ NULL,

data/vendor/tmp/llama.cpp/ggml-metal.m

@@ -735,6 +735,12 @@ static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_tensor * t, size_t * offs
 }
 
 static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const struct ggml_tensor * op) {
+    for (size_t i = 0, n = 3; i < n; ++i) {
+        if (op->src[i] != NULL && op->src[i]->type == GGML_TYPE_BF16) {
+            return false;
+        }
+    }
+
     switch (op->op) {
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(op)) {
@@ -744,7 +750,7 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
                 case GGML_UNARY_OP_GELU:
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_SILU:
-                    return true;
+                    return ggml_is_contiguous(op->src[0]);
                 default:
                     return false;
             }
@@ -1862,9 +1868,10 @@ static enum ggml_status ggml_metal_graph_compute(
                         // ne21 = n_rows
                         const int dst_rows = ne20*ne21;
                         const int dst_rows_min = n_as;
+                        const int dst_rows_max = (ctx->device.maxThreadgroupMemoryLength - 32 - 8192)/4;
 
                         // max size of the rowids array in the kernel shared buffer
-                        GGML_ASSERT(dst_rows <= 2048);
+                        GGML_ASSERT(dst_rows <= dst_rows_max);
 
                         // 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
@@ -3044,12 +3051,6 @@ GGML_CALL static size_t ggml_backend_metal_buffer_type_get_max_size(ggml_backend
     UNUSED(buft);
 }
 
-GGML_CALL static bool ggml_backend_metal_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
-    return ggml_backend_is_metal(backend) || ggml_backend_is_cpu(backend);
-
-    UNUSED(buft);
-}
-
 GGML_CALL static bool ggml_backend_metal_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
     return true;
 
@@ -3064,7 +3065,6 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) {
             /* .get_alignment    = */ ggml_backend_metal_buffer_type_get_alignment,
             /* .get_max_size     = */ ggml_backend_metal_buffer_type_get_max_size,
             /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
-            /* .supports_backend = */ ggml_backend_metal_buffer_type_supports_backend,
             /* .is_host          = */ ggml_backend_metal_buffer_type_is_host,
         },
         /* .context = */ NULL,
@@ -3179,6 +3179,12 @@ GGML_CALL static bool ggml_backend_metal_supports_op(ggml_backend_t backend, con
     return ggml_metal_supports_op(metal_ctx, op);
 }
 
+GGML_CALL static bool ggml_backend_metal_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+    return buft->iface.get_name == ggml_backend_metal_buffer_type_get_name;
+
+    UNUSED(backend);
+}
+
 static struct ggml_backend_i ggml_backend_metal_i = {
     /* .get_name                = */ ggml_backend_metal_name,
     /* .free                    = */ ggml_backend_metal_free,
@@ -3189,9 +3195,11 @@ static struct ggml_backend_i ggml_backend_metal_i = {
     /* .synchronize             = */ NULL,
     /* .graph_plan_create       = */ NULL,
     /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_update       = */ NULL,
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_metal_graph_compute,
     /* .supports_op             = */ ggml_backend_metal_supports_op,
+    /* .supports_buft           = */ ggml_backend_metal_supports_buft,
     /* .offload_op              = */ NULL,
     /* .event_new               = */ NULL,
     /* .event_free              = */ NULL,