feat: mpt wip

llmodel/mpt.cpp

@@ -20,17 +20,34 @@ static const size_t MB = 1024*1024;
 
 struct mpt_hparams {
     // FIXME: for mpt
-    int32_t n_vocab = 50400;
+    int32_t n_vocab = 50432;
     int32_t n_ctx   = 2048;
     int32_t n_embd  = 4096;
-    int32_t n_head  = 16;
+    int32_t n_head  = 32;
-    int32_t n_layer = 28;
+    int32_t n_layer = 32;
+    // this isn't used should we remove?
     int32_t n_rot   = 64;
     int32_t f16     = 1;
 };
 
 struct mpt_layer {
-    // FIXME
+    // normalization
+    struct ggml_tensor * ln_1_g;
+    struct ggml_tensor * ln_1_b;
+
+    // attention
+    struct ggml_tensor * c_attn_q_proj_w;
+    struct ggml_tensor * c_attn_k_proj_w;
+    struct ggml_tensor * c_attn_v_proj_w;
+
+    struct ggml_tensor * c_attn_proj_w;
+
+    // ff
+    struct ggml_tensor * c_mlp_fc_w;
+    struct ggml_tensor * c_mlp_fc_b;
+
+    struct ggml_tensor * c_mlp_proj_w;
+    struct ggml_tensor * c_mlp_proj_b;
 };
 
 struct mpt_buffer {
@@ -70,11 +87,20 @@ struct mpt_kv_cache {
 struct mpt_model {
     mpt_hparams hparams;
 
+    // normalization
+    struct ggml_tensor * ln_f_g;
+    struct ggml_tensor * ln_f_b;
+
+    struct ggml_tensor * wte; // position embedding
+    
+    // mpt does weight tying
+
+    std::vector<mpt_layer> layers;
+
     struct mpt_kv_cache kv_self;
     struct ggml_context * ctx;
     std::map<std::string, struct ggml_tensor *> tensors;
 
-    // FIXME
 
     mpt_buffer buf;
 
@@ -117,15 +143,320 @@ static bool kv_cache_init(
 }
 
 struct mpt_vocab {
-    // FIXME
+    using id    = int32_t;
+    using token = std::string;
+
+    std::map<token, id> token_to_id;
+    std::map<id, token> id_to_token;
 };
 
 // load the model's weights from a stream
 bool mpt_model_load(const std::string &fname, std::istream &fin, mpt_model & model, mpt_vocab & vocab) {
-    // FIXME
+    printf("%s: loading model from '%s' - please wait ...\n", __func__, fname.c_str());
+
+    // verify magic
+    // TODO: Do we really need this?
+    // {
+    //     uint32_t magic;
+    //     fin.read((char *) &magic, sizeof(magic));
+    //     if (magic != 0x67676d6c) {
+    //         fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
+    //         return false;
+    //     }
+    // }
+
+    // load hparams
+    {
+        auto & hparams = model.hparams;
+
+        fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        fin.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        fin.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        fin.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        fin.read((char *) &hparams.n_rot,   sizeof(hparams.n_rot));
+        fin.read((char *) &hparams.f16,     sizeof(hparams.f16));
+
+        printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
+        printf("%s: n_ctx   = %d\n", __func__, hparams.n_ctx);
+        printf("%s: n_embd  = %d\n", __func__, hparams.n_embd);
+        printf("%s: n_head  = %d\n", __func__, hparams.n_head);
+        printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
+        printf("%s: n_rot   = %d\n", __func__, hparams.n_rot);
+        printf("%s: f16     = %d\n", __func__, hparams.f16);
+    }
+    // load vocab
+    {
+        int32_t n_vocab = 0;
+        fin.read((char *) &n_vocab, sizeof(n_vocab));
+
+        if (n_vocab != model.hparams.n_vocab) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
+                    __func__, fname.c_str(), n_vocab, model.hparams.n_vocab);
             return false;
         }
 
+        std::string word;
+        for (int i = 0; i < n_vocab; i++) {
+            uint32_t len;
+            fin.read((char *) &len, sizeof(len));
+
+            word.resize(len);
+            fin.read((char *) word.data(), len);
+
+            vocab.token_to_id[word] = i;
+            vocab.id_to_token[i] = word;
+        }
+    }
+    // for the big tensors, we have the option to store the data in 16-bit floats or quantized
+    // in order to save memory and also to speed up the computation
+    ggml_type wtype = GGML_TYPE_COUNT;
+    switch (model.hparams.f16) {
+        case 0: wtype = GGML_TYPE_F32;  break;
+        case 1: wtype = GGML_TYPE_F16;  break;
+        case 2: wtype = GGML_TYPE_Q4_0; break;
+        case 3: wtype = GGML_TYPE_Q4_1; break;
+        default:
+                {
+                    fprintf(stderr, "%s: invalid model file '%s' (bad f16 value %d)\n",
+                            __func__, fname.c_str(), model.hparams.f16);
+                    return false;
+                }
+    }
+
+    const ggml_type wtype2 = GGML_TYPE_F32;
+
+    auto & ctx = model.ctx;
+
+    size_t ctx_size = 0;
+
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+        const int n_vocab = hparams.n_vocab;
+
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b
+
+        ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // wte
+
+        // with weight tying i don't think we need this
+        // ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype);         // lmh_g
+        // ctx_size +=        n_vocab*ggml_type_sizef(GGML_TYPE_F32); // lmh_b
+
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b
+
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_q_proj_w
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_k_proj_w
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_v_proj_w
+
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_proj_w
+
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_fc_w
+        ctx_size += n_layer*(       4*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_fc_b
+
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_proj_w
+        ctx_size += n_layer*(         n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b
+
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_k
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v
+
+        // TODO: what is this??
+        ctx_size += (5 + 10*n_layer)*256; // object overhead
+
+        printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
+    }
+
+    // create the ggml context
+    {
+        struct ggml_init_params params = {
+            .mem_size   = ctx_size,
+            .mem_buffer = NULL,
+        };
+
+        model.ctx = ggml_init(params);
+        if (!model.ctx) {
+            fprintf(stderr, "%s: ggml_init() failed\n", __func__);
+            return false;
+        }
+    }
+     // prepare memory for the weights
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+        const int n_vocab = hparams.n_vocab;
+
+        model.layers.resize(n_layer);
+
+        model.wte    = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+
+        model.ln_f_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+        model.ln_f_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+
+        // we don't need this because of weight tying
+        // model.lmh_g  = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+        // model.lmh_b  = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_vocab);
+
+        // map by name
+        model.tensors["transformer.wte.weight"] = model.wte;
+
+        model.tensors["transformer.ln_f.weight"] = model.ln_f_g;
+        model.tensors["transformer.ln_f.bias"]   = model.ln_f_b;
+
+        for (int i = 0; i < n_layer; ++i) {
+            auto & layer = model.layers[i];
+
+            layer.ln_1_g          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_1_b          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.c_attn_q_proj_w = ggml_new_tensor_2d(ctx, wtype,           n_embd,   n_embd);
+            layer.c_attn_k_proj_w = ggml_new_tensor_2d(ctx, wtype,           n_embd,   n_embd);
+            layer.c_attn_v_proj_w = ggml_new_tensor_2d(ctx, wtype,           n_embd,   n_embd);
+
+            layer.c_attn_proj_w   = ggml_new_tensor_2d(ctx, wtype,           n_embd,   n_embd);
+
+            layer.c_mlp_fc_w      = ggml_new_tensor_2d(ctx, wtype,           n_embd, 4*n_embd);
+            layer.c_mlp_fc_b      = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
+
+            layer.c_mlp_proj_w    = ggml_new_tensor_2d(ctx, wtype,         4*n_embd,   n_embd);
+            layer.c_mlp_proj_b    = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            // map by name
+            model.tensors["transformer.h." + std::to_string(i) + ".ln_1.weight"]          = layer.ln_1_g;
+            model.tensors["transformer.h." + std::to_string(i) + ".ln_1.bias"]            = layer.ln_1_b;
+
+            model.tensors["transformer.h." + std::to_string(i) + ".attn.q_proj.weight"]   = layer.c_attn_q_proj_w;
+            model.tensors["transformer.h." + std::to_string(i) + ".attn.k_proj.weight"]   = layer.c_attn_k_proj_w;
+            model.tensors["transformer.h." + std::to_string(i) + ".attn.v_proj.weight"]   = layer.c_attn_v_proj_w;
+
+            model.tensors["transformer.h." + std::to_string(i) + ".attn.out_proj.weight"] = layer.c_attn_proj_w;
+
+            model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_in.weight"]     = layer.c_mlp_fc_w;
+            model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_in.bias"]       = layer.c_mlp_fc_b;
+
+            model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_out.weight"]    = layer.c_mlp_proj_w;
+            model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_out.bias"]      = layer.c_mlp_proj_b;
+        }
+
+        // key + value memory
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+
+        const int n_mem      = n_layer*n_ctx;
+        const int n_elements = n_embd*n_mem;
+
+        if (!kv_cache_init(hparams, model.kv_self, GGML_TYPE_F32, model.hparams.n_ctx)) {
+            fprintf(stderr, "%s: kv_cache_init() failed for self-attention cache\n", __func__);
+            ggml_free(ctx);
+            return false;
+        }
+
+        const size_t memory_size = ggml_nbytes(model.kv_self.k) + ggml_nbytes(model.kv_self.v);
+        printf("%s: kv self size  = %7.2f MB\n", __func__, memory_size / 1024.0 / 1024.0);
+    }
+// load weights
+    {
+        int n_tensors = 0;
+        size_t total_size = 0;
+
+        printf("%s: ", __func__);
+
+        while (true) {
+            int32_t n_dims;
+            int32_t length;
+            int32_t ftype;
+
+            fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+            fin.read(reinterpret_cast<char *>(&length), sizeof(length));
+            fin.read(reinterpret_cast<char *>(&ftype),  sizeof(ftype));
+
+            if (fin.eof()) {
+                break;
+            }
+
+            int32_t nelements = 1;
+            int32_t ne[2] = { 1, 1 };
+            for (int i = 0; i < n_dims; ++i) {
+                fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+                nelements *= ne[i];
+            }
+
+            std::string name(length, 0);
+            fin.read(&name[0], length);
+
+            if (model.tensors.find(name.data()) == model.tensors.end()) {
+                fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data());
+                return false;
+            }
+
+            auto tensor = model.tensors[name.data()];
+            if (ggml_nelements(tensor) != nelements) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
+                return false;
+            }
+
+            if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
+                fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%lu, %lu], expected [%d, %d]\n",
+                        __func__, name.data(), tensor->ne[0], tensor->ne[1], ne[0], ne[1]);
+                return false;
+            }
+
+            if (0) {
+                static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", };
+                printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.data(), ne[0], ne[1], ftype_str[ftype], ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor));
+            }
+
+            size_t bpe = 0;
+
+            switch (ftype) {
+                case 0: bpe = ggml_type_size(GGML_TYPE_F32);  break;
+                case 1: bpe = ggml_type_size(GGML_TYPE_F16);  break;
+                case 2: bpe = ggml_type_size(GGML_TYPE_Q4_0); assert(ne[0] % 64 == 0); break;
+                case 3: bpe = ggml_type_size(GGML_TYPE_Q4_1); assert(ne[0] % 64 == 0); break;
+                default:
+                        {
+                            fprintf(stderr, "%s: unknown ftype %d in model file\n", __func__, ftype);
+                            return false;
+                        }
+            };
+
+            if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
+                        __func__, name.data(), ggml_nbytes(tensor), nelements*bpe);
+                return false;
+            }
+
+            fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
+
+            //printf("%42s - [%5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ftype == 0 ? "float" : "f16", ggml_nbytes(tensor)/1024.0/1024.0);
+            total_size += ggml_nbytes(tensor);
+            if (++n_tensors % 8 == 0) {
+                printf(".");
+                fflush(stdout);
+            }
+        }
+
+        printf(" done\n");
+
+        printf("%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size/1024.0/1024.0, n_tensors);
+    }
+
+    return true;
+    }
+}
+
+
 // load the model's weights from a file path
 bool gptj_model_load(const std::string & fname, mpt_model & model, mpt_vocab & vocab) {
 
@@ -147,9 +478,224 @@ bool mpt_eval(
         const std::vector<int>           & embd_inp,
               std::vector<float>         & embd_w,
               size_t                     & mem_per_token) {
-    // FIXME
+    const int N = embd_inp.size();
+
+    const auto & hparams = model.hparams;
+
+    const int n_embd  = hparams.n_embd;
+    const int n_layer = hparams.n_layer;
+    const int n_ctx   = hparams.n_ctx;
+    const int n_head  = hparams.n_head;
+    const int n_vocab = hparams.n_vocab;
+    const int n_rot   = hparams.n_rot;
+
+    const int d_key = n_embd/n_head;
+
+    static size_t buf_size = 1024u*MB;
+    if (!model.buf.addr || model.buf.size < buf_size)
+        model.buf.resize(buf_size);
+
+    if (mem_per_token > 0 && mem_per_token*N > model.buf.size) {
+        const size_t buf_size_new = 1.1*(mem_per_token*N); // add 10% to account for ggml object overhead
+        printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, model.buf.size, buf_size_new);
+
+        // reallocate
+        model.buf.resize(buf_size_new);
+        if (model.buf.addr == nullptr) {
+            fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, model.buf.size);
             return false;
         }
+    }
+
+    struct ggml_init_params params = {
+        .mem_size   = model.buf.size,
+        .mem_buffer = model.buf.addr,
+    };
+
+    struct ggml_context * ctx0 = ggml_init(params);
+    struct ggml_cgraph gf = { .n_threads = n_threads };
+
+    struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
+
+    // wte
+    struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.wte, embd);
+
+    for (int il = 0; il < n_layer; ++il) {
+        struct ggml_tensor * cur;
+
+        // norm
+        {
+            cur = ggml_norm(ctx0, inpL);
+
+            // cur = ln_1_g*cur + ln_1_b
+            cur = ggml_add(ctx0,
+                    ggml_mul(ctx0,
+                        ggml_repeat(ctx0, model.layers[il].ln_1_g, cur),
+                        cur),
+                    ggml_repeat(ctx0, model.layers[il].ln_1_b, cur));
+        }
+
+        struct ggml_tensor * inpSA = cur;
+
+        // self-attention
+        {
+            struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].c_attn_q_proj_w, cur);
+            struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].c_attn_k_proj_w, cur);
+            struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].c_attn_v_proj_w, cur);
+
+            // store key and value to memory
+            if (N >= 1) {
+                struct ggml_tensor * k = ggml_view_1d(ctx0, model.kv_self.k, N*n_embd, (ggml_element_size(model.kv_self.k)*n_embd)*(il*n_ctx + n_past));
+                struct ggml_tensor * v = ggml_view_1d(ctx0, model.kv_self.v, N*n_embd, (ggml_element_size(model.kv_self.v)*n_embd)*(il*n_ctx + n_past));
+
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
+            }
+            // we need to replace rope with alibi 
+
+            // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
+            struct ggml_tensor * Q =
+                ggml_permute(ctx0,
+                            ggml_cpy(ctx0,
+                                Qcur,
+                                ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd/n_head, n_head, N)),
+                        0, 2, 1, 3);
+
+            // K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
+            struct ggml_tensor * K =
+                ggml_permute(ctx0,
+                            ggml_reshape_3d(ctx0,
+                                ggml_view_1d(ctx0, model.kv_self.k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.kv_self.k)*n_embd),
+                                n_embd/n_head, n_head, n_past + N),
+                        0, 2, 1, 3);
+
+            // K * Q
+            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+
+            // KQ_scaled = KQ / sqrt(n_embd/n_head)
+            struct ggml_tensor * KQ_scaled =
+                ggml_scale(ctx0,
+                        KQ,
+                        ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
+                        );
+
+            struct ggml_tensor * KQ_scaled_alibi = ggml_alibi(ctx0, KQ_scaled, n_past, n_head);
+
+            // KQ_masked = mask_past(KQ_scaled)
+            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled_alibi, n_past);
+
+            // KQ = soft_max(KQ_masked)
+            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
+
+            // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
+            struct ggml_tensor * V_trans =
+                ggml_cpy(ctx0,
+                        ggml_permute(ctx0,
+                            ggml_reshape_3d(ctx0,
+                                ggml_view_1d(ctx0, model.kv_self.v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.kv_self.v)*n_embd),
+                                n_embd/n_head, n_head, n_past + N),
+                            1, 2, 0, 3),
+                        ggml_new_tensor_3d(ctx0, model.kv_self.v->type, n_past + N, n_embd/n_head, n_head));
+
+            // KQV = transpose(V) * KQ_soft_max
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
+
+            // KQV_merged = KQV.permute(0, 2, 1, 3)
+            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+
+            // cur = KQV_merged.contiguous().view(n_embd, N)
+            cur = ggml_cpy(ctx0,
+                    KQV_merged,
+                    ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
+
+            // projection (no bias)
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].c_attn_proj_w,
+                    cur);
+        }
+
+        struct ggml_tensor * inpFF = cur;
+
+        // feed-forward network
+        // this is independent of the self-attention result, so it could be done in parallel to the self-attention
+        {
+            // note here we pass inpSA instead of cur
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].c_mlp_fc_w,
+                    inpSA);
+
+            cur = ggml_add(ctx0,
+                    ggml_repeat(ctx0, model.layers[il].c_mlp_fc_b, cur),
+                    cur);
+
+            // RELU activation
+            cur = ggml_relu(ctx0, cur);
+
+            // projection
+            // cur = proj_w*cur + proj_b
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].c_mlp_proj_w,
+                    cur);
+
+            cur = ggml_add(ctx0,
+                    ggml_repeat(ctx0, model.layers[il].c_mlp_proj_b, cur),
+                    cur);
+        }
+
+        // self-attention + FF
+        cur  = ggml_add(ctx0, cur, inpFF);
+
+        // input for next layer
+        inpL = ggml_add(ctx0, cur, inpL);
+    }
+
+    // norm
+    {
+        inpL = ggml_norm(ctx0, inpL);
+
+        // inpL = ln_f_g*inpL + ln_f_b
+        inpL = ggml_add(ctx0,
+                ggml_mul(ctx0,
+                    ggml_repeat(ctx0, model.ln_f_g, inpL),
+                    inpL),
+                ggml_repeat(ctx0, model.ln_f_b, inpL));
+    }
+
+    // lm_head with weight tying
+    {
+        inpL = ggml_mul_mat(ctx0, model.wte, inpL);
+
+    }
+
+    // logits -> probs
+    //inpL = ggml_soft_max(ctx0, inpL);
+
+    // run the computation
+    ggml_build_forward_expand(&gf, inpL);
+    ggml_graph_compute       (ctx0, &gf);
+
+    //if (n_past%100 == 0) {
+    //    ggml_graph_print   (&gf);
+    //    ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
+    //}
+
+    //embd_w.resize(n_vocab*N);
+    //memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
+
+    // return result for just the last token
+    embd_w.resize(n_vocab);
+    memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
+
+    if (mem_per_token == 0) {
+        mem_per_token = ggml_used_mem(ctx0)/N;
+    }
+    //printf("used_mem = %zu\n", ggml_used_mem(ctx0));
+
+    ggml_free(ctx0);
+
+    return true;
+}
 
 std::vector<int> mpt_tokenize(const mpt_vocab & vocab, const std::string & text) {
     // FIXME
@@ -161,20 +707,6 @@ const std::string mpt_token_to_str(const mpt_vocab & vocab, int token) {
     return std::string();
 }
 
-int mpt_sample_top_k_top_p(
-        const mpt_vocab & vocab,
-        const int32_t * last_n_tokens_data,
-        int   last_n_tokens_size,
-        const std::vector<float> logits,
-        int    top_k,
-        double top_p,
-        double temp,
-        float repeat_penalty,
-        std::mt19937 & rng)
-{
-    // FIXME
-    return 0;
-}
 
 #define MPT_MAX_RNG_STATE 64*1024
 
@@ -237,6 +769,103 @@ size_t mpt_copy_state_data(const mpt_model &model, const std::mt19937 &rng, uint
     return written;
 }
 
+mpt_vocab::id mpt_sample_top_k_top_p(
+        const mpt_vocab & vocab,
+        const int32_t * last_n_tokens_data,
+        int   last_n_tokens_size,
+        const std::vector<float> logits,
+        int    top_k,
+        double top_p,
+        double temp,
+        float repeat_penalty,
+        std::mt19937 & rng) {
+    int n_logits = vocab.id_to_token.size();
+
+    const auto last_n_tokens = std::vector<int32_t>(last_n_tokens_data, last_n_tokens_data + last_n_tokens_size);
+    const auto * plogits = logits.data() + logits.size() - n_logits;
+
+    std::vector<std::pair<double, mpt_vocab::id>> logits_id;
+    logits_id.reserve(n_logits);
+
+    {
+        const float scale = 1.0f/temp;
+        for (int i = 0; i < n_logits; ++i) {
+            // repetition penalty from ctrl paper (https://arxiv.org/abs/1909.05858)
+            // credit https://github.com/facebookresearch/llama/compare/main...shawwn:llama:main
+            if (std::find(last_n_tokens.begin(), last_n_tokens.end(), i) != last_n_tokens.end()) {
+                // if score < 0 then repetition penalty has to multiplied to reduce the previous token probability
+                if (plogits[i] < 0.0f) {
+                    logits_id.push_back(std::make_pair(plogits[i]*scale*repeat_penalty, i));
+                } else {
+                    logits_id.push_back(std::make_pair(plogits[i]*scale/repeat_penalty, i));
+                }
+            } else {
+                logits_id.push_back(std::make_pair(plogits[i]*scale, i));
+            }
+        }
+    }
+
+    // find the top K tokens
+    std::partial_sort(
+            logits_id.begin(),
+            logits_id.begin() + top_k, logits_id.end(),
+            [](const std::pair<double, mpt_vocab::id> & a, const std::pair<double, mpt_vocab::id> & b) {
+        return a.first > b.first;
+    });
+
+    logits_id.resize(top_k);
+
+    double maxl = -INFINITY;
+    for (const auto & kv : logits_id) {
+        maxl = std::max(maxl, kv.first);
+    }
+
+    // compute probs for the top K tokens
+    std::vector<double> probs;
+    probs.reserve(logits_id.size());
+
+    double sum = 0.0;
+    for (const auto & kv : logits_id) {
+        double p = exp(kv.first - maxl);
+        probs.push_back(p);
+        sum += p;
+    }
+
+    // normalize the probs
+    for (auto & p : probs) {
+        p /= sum;
+    }
+
+    if (top_p < 1.0f) {
+        double cumsum = 0.0f;
+        for (int i = 0; i < top_k; i++) {
+            cumsum += probs[i];
+            if (cumsum >= top_p) {
+                top_k = i + 1;
+                probs.resize(top_k);
+                logits_id.resize(top_k);
+                break;
+            }
+        }
+
+        cumsum = 1.0/cumsum;
+        for (int i = 0; i < (int) probs.size(); i++) {
+            probs[i] *= cumsum;
+        }
+    }
+
+    //printf("\n");
+    //for (int i = 0; i < (int) probs.size(); i++) {
+    //    printf("%d: '%s' %f\n", i, vocab.id_to_token.at(logits_id[i].second).c_str(), probs[i]);
+    //}
+    //exit(0);
+
+    std::discrete_distribution<> dist(probs.begin(), probs.end());
+    int idx = dist(rng);
+
+    return logits_id[idx].second;
+}
+
 size_t mpt_set_state_data(mpt_model *model, std::mt19937 *rng, const uint8_t *src)
 {
     const uint8_t * in = src;