cui-llama.rn 1.1.0 → 1.1.2

package/cpp/common.cpp

@@ -333,6 +333,10 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
 void gpt_params_parse_from_env(gpt_params & params) {
     // we only care about server-related params for now
     get_env("LLAMA_ARG_MODEL",            params.model);
+    get_env("LLAMA_ARG_MODEL_URL",        params.model_url);
+    get_env("LLAMA_ARG_MODEL_ALIAS",      params.model_alias);
+    get_env("LLAMA_ARG_HF_REPO",          params.hf_repo);
+    get_env("LLAMA_ARG_HF_FILE",          params.hf_file);
     get_env("LLAMA_ARG_THREADS",          params.n_threads);
     get_env("LLAMA_ARG_CTX_SIZE",         params.n_ctx);
     get_env("LLAMA_ARG_N_PARALLEL",       params.n_parallel);
@@ -347,6 +351,9 @@ void gpt_params_parse_from_env(gpt_params & params) {
     get_env("LLAMA_ARG_EMBEDDINGS",       params.embedding);
     get_env("LLAMA_ARG_FLASH_ATTN",       params.flash_attn);
     get_env("LLAMA_ARG_DEFRAG_THOLD",     params.defrag_thold);
+    get_env("LLAMA_ARG_CONT_BATCHING",    params.cont_batching);
+    get_env("LLAMA_ARG_HOST",             params.hostname);
+    get_env("LLAMA_ARG_PORT",             params.port);
 }
 
 bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
@@ -907,7 +914,7 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         }
         return true;
     }
-    if (arg == "-ngld" || arg == "--gpu-layers-draft" || arg == "--gpu-layers-draft") {
+    if (arg == "-ngld" || arg == "--gpu-layers-draft" || arg == "--n-gpu-layers-draft") {
         CHECK_ARG
         params.n_gpu_layers_draft = std::stoi(argv[i]);
         if (!llama_supports_gpu_offload()) {
@@ -1867,13 +1874,19 @@ std::string string_get_sortable_timestamp() {
 
 void string_replace_all(std::string & s, const std::string & search, const std::string & replace) {
     if (search.empty()) {
-        return; // Avoid infinite loop if 'search' is an empty string
+        return;
     }
+    std::string builder;
+    builder.reserve(s.length());
     size_t pos = 0;
-    while ((pos = s.find(search, pos)) != std::string::npos) {
-        s.replace(pos, search.length(), replace);
-        pos += replace.length();
-    }
+    size_t last_pos = 0;
+    while ((pos = s.find(search, last_pos)) != std::string::npos) {
+        builder.append(s, last_pos, pos - last_pos);
+        builder.append(replace);
+        last_pos = pos + search.length();
+    }
+    builder.append(s, last_pos, std::string::npos);
+    s = std::move(builder);
 }
 
 void string_process_escapes(std::string & input) {

package/cpp/ggml-aarch64.c

@@ -337,33 +337,18 @@ static size_t quantize_q4_0_nr_bl(const float * restrict src, void * restrict ds
 }
 
 size_t quantize_q4_0_4x4(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
-    if (!quant_weights) {
-        return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 4);
-    }
-    else {
-        assert(false);
-        return 0;
-    }
+    UNUSED(quant_weights);
+    return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 4);
 }
 
 size_t quantize_q4_0_4x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
-    if (!quant_weights) {
-        return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 8);
-    }
-    else {
-        assert(false);
-        return 0;
-    }
+    UNUSED(quant_weights);
+    return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 8);
 }
 
 size_t quantize_q4_0_8x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
-    if (!quant_weights) {
-        return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 8, 8);
-    }
-    else {
-        assert(false);
-        return 0;
-    }
+    UNUSED(quant_weights);
+    return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 8, 8);
 }
 
 void lm_ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {

package/cpp/ggml-metal.m

@@ -82,6 +82,8 @@ enum lm_ggml_metal_kernel_type {
     LM_GGML_METAL_KERNEL_TYPE_RMS_NORM,
     LM_GGML_METAL_KERNEL_TYPE_GROUP_NORM,
     LM_GGML_METAL_KERNEL_TYPE_NORM,
+    LM_GGML_METAL_KERNEL_TYPE_SSM_CONV_F32,
+    LM_GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32,
     LM_GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32,
     LM_GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16,
     LM_GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32,
@@ -542,6 +544,8 @@ static struct lm_ggml_backend_metal_context * lm_ggml_metal_init(int n_cb) {
         LM_GGML_METAL_ADD_KERNEL(LM_GGML_METAL_KERNEL_TYPE_RMS_NORM,                      rms_norm,                       ctx->support_simdgroup_reduction);
         LM_GGML_METAL_ADD_KERNEL(LM_GGML_METAL_KERNEL_TYPE_GROUP_NORM,                    group_norm,                     ctx->support_simdgroup_reduction);
         LM_GGML_METAL_ADD_KERNEL(LM_GGML_METAL_KERNEL_TYPE_NORM,                          norm,                           true);
+        LM_GGML_METAL_ADD_KERNEL(LM_GGML_METAL_KERNEL_TYPE_SSM_CONV_F32,                  ssm_conv_f32,                   true);
+        LM_GGML_METAL_ADD_KERNEL(LM_GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32,                  ssm_scan_f32,                   true);
         LM_GGML_METAL_ADD_KERNEL(LM_GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32,                mul_mv_f32_f32,                 ctx->support_simdgroup_reduction);
         LM_GGML_METAL_ADD_KERNEL(LM_GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16,                mul_mv_f16_f16,                 ctx->support_simdgroup_reduction);
         LM_GGML_METAL_ADD_KERNEL(LM_GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32,                mul_mv_f16_f32,                 ctx->support_simdgroup_reduction);
@@ -803,6 +807,9 @@ static bool lm_ggml_metal_supports_op(const struct lm_ggml_backend_metal_context
                 return false;
             }
             return ctx->support_simdgroup_mm; // TODO: over-restricted for vec-kernels
+        case LM_GGML_OP_SSM_CONV:
+        case LM_GGML_OP_SSM_SCAN:
+            return true;
         case LM_GGML_OP_MUL_MAT:
         case LM_GGML_OP_MUL_MAT_ID:
             return ctx->support_simdgroup_reduction &&
@@ -1538,6 +1545,121 @@ static enum lm_ggml_status lm_ggml_metal_graph_compute(
                             [encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                         }
                     } break;
+                case LM_GGML_OP_SSM_CONV:
+                    {
+                        LM_GGML_ASSERT(src0t == LM_GGML_TYPE_F32);
+                        LM_GGML_ASSERT(src1t == LM_GGML_TYPE_F32);
+
+                        LM_GGML_ASSERT(lm_ggml_is_contiguous(src0));
+                        LM_GGML_ASSERT(lm_ggml_is_contiguous(src1));
+
+                        id<MTLComputePipelineState> pipeline = ctx->kernels[LM_GGML_METAL_KERNEL_TYPE_SSM_CONV_F32].pipeline;
+
+                        [encoder setComputePipelineState:pipeline];
+                        [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];
+                        [encoder setBytes:&ne00    length:sizeof(ne00) atIndex:3];
+                        [encoder setBytes:&ne01    length:sizeof(ne01) atIndex:4];
+                        [encoder setBytes:&ne02    length:sizeof(ne02) atIndex:5];
+                        [encoder setBytes:&nb00    length:sizeof(nb00) atIndex:6];
+                        [encoder setBytes:&nb01    length:sizeof(nb01) atIndex:7];
+                        [encoder setBytes:&nb02    length:sizeof(nb02) atIndex:8];
+                        [encoder setBytes:&ne10    length:sizeof(ne10) atIndex:9];
+                        [encoder setBytes:&ne11    length:sizeof(ne11) atIndex:10];
+                        [encoder setBytes:&nb10    length:sizeof(nb10) atIndex:11];
+                        [encoder setBytes:&nb11    length:sizeof(nb11) atIndex:12];
+                        [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:13];
+                        [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:14];
+                        [encoder setBytes:&ne2     length:sizeof(ne2)  atIndex:15];
+                        [encoder setBytes:&nb0     length:sizeof(nb0)  atIndex:16];
+                        [encoder setBytes:&nb1     length:sizeof(nb1)  atIndex:17];
+                        [encoder setBytes:&nb2     length:sizeof(nb2)  atIndex:18];
+
+                        [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne1, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                    } break;
+                case LM_GGML_OP_SSM_SCAN:
+                    {
+                        struct lm_ggml_tensor * src3 = gf->nodes[i]->src[3];
+                        struct lm_ggml_tensor * src4 = gf->nodes[i]->src[4];
+                        struct lm_ggml_tensor * src5 = gf->nodes[i]->src[5];
+
+                        LM_GGML_ASSERT(src3);
+                        LM_GGML_ASSERT(src4);
+                        LM_GGML_ASSERT(src5);
+
+                        size_t offs_src3 = 0;
+                        size_t offs_src4 = 0;
+                        size_t offs_src5 = 0;
+
+                        id<MTLBuffer> id_src3 = src3 ? lm_ggml_metal_get_buffer(src3, &offs_src3) : nil;
+                        id<MTLBuffer> id_src4 = src4 ? lm_ggml_metal_get_buffer(src4, &offs_src4) : nil;
+                        id<MTLBuffer> id_src5 = src5 ? lm_ggml_metal_get_buffer(src5, &offs_src5) : nil;
+
+                        const int64_t  ne30 = src3->ne[0]; LM_GGML_UNUSED(ne30);
+                        const int64_t  ne31 = src3->ne[1]; LM_GGML_UNUSED(ne31);
+
+                        const uint64_t nb30 = src3->nb[0];
+                        const uint64_t nb31 = src3->nb[1];
+
+                        const int64_t  ne40 = src4->ne[0]; LM_GGML_UNUSED(ne40);
+                        const int64_t  ne41 = src4->ne[1]; LM_GGML_UNUSED(ne41);
+                        const int64_t  ne42 = src4->ne[2]; LM_GGML_UNUSED(ne42);
+
+                        const uint64_t nb40 = src4->nb[0];
+                        const uint64_t nb41 = src4->nb[1];
+                        const uint64_t nb42 = src4->nb[2];
+
+                        const int64_t  ne50 = src5->ne[0]; LM_GGML_UNUSED(ne50);
+                        const int64_t  ne51 = src5->ne[1]; LM_GGML_UNUSED(ne51);
+                        const int64_t  ne52 = src5->ne[2]; LM_GGML_UNUSED(ne52);
+
+                        const uint64_t nb50 = src5->nb[0];
+                        const uint64_t nb51 = src5->nb[1];
+                        const uint64_t nb52 = src5->nb[2];
+
+                        const int64_t d_state      = ne00;
+                        const int64_t d_inner      = ne01;
+                        const int64_t n_seq_tokens = ne11;
+                        const int64_t n_seqs       = ne02;
+
+                        id<MTLComputePipelineState> pipeline = ctx->kernels[LM_GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32].pipeline;
+
+                        [encoder setComputePipelineState:pipeline];
+                        [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                        [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
+                        [encoder setBuffer:id_src2 offset:offs_src2 atIndex:2];
+                        [encoder setBuffer:id_src3 offset:offs_src3 atIndex:3];
+                        [encoder setBuffer:id_src4 offset:offs_src4 atIndex:4];
+                        [encoder setBuffer:id_src5 offset:offs_src5 atIndex:5];
+                        [encoder setBuffer:id_dst  offset:offs_dst  atIndex:6];
+
+                        [encoder setBytes:&d_state      length:sizeof(d_state)      atIndex:7];
+                        [encoder setBytes:&d_inner      length:sizeof(d_inner)      atIndex:8];
+                        [encoder setBytes:&n_seq_tokens length:sizeof(n_seq_tokens) atIndex:9];
+                        [encoder setBytes:&n_seqs       length:sizeof(n_seqs)       atIndex:10];
+
+                        [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:11];
+                        [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:12];
+                        [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:13];
+                        [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:14];
+                        [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:15];
+                        [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:16];
+                        [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:17];
+                        [encoder setBytes:&nb20 length:sizeof(nb20) atIndex:18];
+                        [encoder setBytes:&nb21 length:sizeof(nb21) atIndex:19];
+                        [encoder setBytes:&nb22 length:sizeof(nb22) atIndex:20];
+                        [encoder setBytes:&nb30 length:sizeof(nb30) atIndex:21];
+                        [encoder setBytes:&nb31 length:sizeof(nb31) atIndex:22];
+                        [encoder setBytes:&nb40 length:sizeof(nb40) atIndex:23];
+                        [encoder setBytes:&nb41 length:sizeof(nb41) atIndex:24];
+                        [encoder setBytes:&nb42 length:sizeof(nb42) atIndex:25];
+                        [encoder setBytes:&nb50 length:sizeof(nb50) atIndex:26];
+                        [encoder setBytes:&nb51 length:sizeof(nb51) atIndex:27];
+                        [encoder setBytes:&nb52 length:sizeof(nb52) atIndex:28];
+
+                        [encoder dispatchThreadgroups:MTLSizeMake(d_inner, n_seqs, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                    } break;
                 case LM_GGML_OP_MUL_MAT:
                     {
                         LM_GGML_ASSERT(ne00 == ne10);
@@ -2624,9 +2746,14 @@ static enum lm_ggml_status lm_ggml_metal_graph_compute(
 
                         float scale;
                         float max_bias;
+                        float logit_softcap;
+                        memcpy(&scale,         ((int32_t *) dst->op_params) + 0, sizeof(scale));
+                        memcpy(&max_bias,      ((int32_t *) dst->op_params) + 1, sizeof(max_bias));
+                        memcpy(&logit_softcap, ((int32_t *) dst->op_params) + 2, sizeof(logit_softcap));
 
-                        memcpy(&scale,    ((int32_t *) dst->op_params) + 0, sizeof(scale));
-                        memcpy(&max_bias, ((int32_t *) dst->op_params) + 1, sizeof(max_bias));
+                        if (logit_softcap != 0.0f) {
+                            scale /= logit_softcap;
+                        }
 
                         const uint32_t n_head      = src0->ne[2];
                         const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
@@ -2677,30 +2804,31 @@ static enum lm_ggml_status lm_ggml_metal_graph_compute(
                         } else {
                             [encoder setBuffer:id_src0     offset:offs_src0           atIndex:3];
                         }
-                        [encoder setBuffer:id_dst      offset:offs_dst            atIndex:4];
-                        [encoder setBytes:&ne01        length:sizeof( int64_t)    atIndex:5];
-                        [encoder setBytes:&ne02        length:sizeof( int64_t)    atIndex:6];
-                        [encoder setBytes:&ne03        length:sizeof( int64_t)    atIndex:7];
-                        [encoder setBytes:&nb01        length:sizeof(uint64_t)    atIndex:8];
-                        [encoder setBytes:&nb02        length:sizeof(uint64_t)    atIndex:9];
-                        [encoder setBytes:&nb03        length:sizeof(uint64_t)    atIndex:10];
-                        [encoder setBytes:&ne11        length:sizeof( int64_t)    atIndex:11];
-                        [encoder setBytes:&ne12        length:sizeof( int64_t)    atIndex:12];
-                        [encoder setBytes:&ne13        length:sizeof( int64_t)    atIndex:13];
-                        [encoder setBytes:&nb11        length:sizeof(uint64_t)    atIndex:14];
-                        [encoder setBytes:&nb12        length:sizeof(uint64_t)    atIndex:15];
-                        [encoder setBytes:&nb13        length:sizeof(uint64_t)    atIndex:16];
-                        [encoder setBytes:&nb21        length:sizeof(uint64_t)    atIndex:17];
-                        [encoder setBytes:&nb22        length:sizeof(uint64_t)    atIndex:18];
-                        [encoder setBytes:&nb23        length:sizeof(uint64_t)    atIndex:19];
-                        [encoder setBytes:&nb31        length:sizeof(uint64_t)    atIndex:20];
-                        [encoder setBytes:&ne1         length:sizeof( int64_t)    atIndex:21];
-                        [encoder setBytes:&ne2         length:sizeof( int64_t)    atIndex:22];
-                        [encoder setBytes:&scale       length:sizeof(   float)    atIndex:23];
-                        [encoder setBytes:&max_bias    length:sizeof(   float)    atIndex:24];
-                        [encoder setBytes:&m0          length:sizeof(m0)          atIndex:25];
-                        [encoder setBytes:&m1          length:sizeof(m1)          atIndex:26];
-                        [encoder setBytes:&n_head_log2 length:sizeof(n_head_log2) atIndex:27];
+                        [encoder setBuffer:id_dst        offset:offs_dst              atIndex:4];
+                        [encoder setBytes:&ne01          length:sizeof( int64_t)      atIndex:5];
+                        [encoder setBytes:&ne02          length:sizeof( int64_t)      atIndex:6];
+                        [encoder setBytes:&ne03          length:sizeof( int64_t)      atIndex:7];
+                        [encoder setBytes:&nb01          length:sizeof(uint64_t)      atIndex:8];
+                        [encoder setBytes:&nb02          length:sizeof(uint64_t)      atIndex:9];
+                        [encoder setBytes:&nb03          length:sizeof(uint64_t)      atIndex:10];
+                        [encoder setBytes:&ne11          length:sizeof( int64_t)      atIndex:11];
+                        [encoder setBytes:&ne12          length:sizeof( int64_t)      atIndex:12];
+                        [encoder setBytes:&ne13          length:sizeof( int64_t)      atIndex:13];
+                        [encoder setBytes:&nb11          length:sizeof(uint64_t)      atIndex:14];
+                        [encoder setBytes:&nb12          length:sizeof(uint64_t)      atIndex:15];
+                        [encoder setBytes:&nb13          length:sizeof(uint64_t)      atIndex:16];
+                        [encoder setBytes:&nb21          length:sizeof(uint64_t)      atIndex:17];
+                        [encoder setBytes:&nb22          length:sizeof(uint64_t)      atIndex:18];
+                        [encoder setBytes:&nb23          length:sizeof(uint64_t)      atIndex:19];
+                        [encoder setBytes:&nb31          length:sizeof(uint64_t)      atIndex:20];
+                        [encoder setBytes:&ne1           length:sizeof( int64_t)      atIndex:21];
+                        [encoder setBytes:&ne2           length:sizeof( int64_t)      atIndex:22];
+                        [encoder setBytes:&scale         length:sizeof(   float)      atIndex:23];
+                        [encoder setBytes:&max_bias      length:sizeof(   float)      atIndex:24];
+                        [encoder setBytes:&m0            length:sizeof(m0)            atIndex:25];
+                        [encoder setBytes:&m1            length:sizeof(m1)            atIndex:26];
+                        [encoder setBytes:&n_head_log2   length:sizeof(n_head_log2)   atIndex:27];
+                        [encoder setBytes:&logit_softcap length:sizeof(logit_softcap) atIndex:28];
 
                         if (!use_vec_kernel) {
                             // half8x8 kernel