whispercpp 1.3.4 → 1.3.5

data/ext/sources/examples/talk-llama/llama-context.cpp

@@ -1,5 +1,6 @@
 #include "llama-context.h"
 
+#include "llama-arch.h"
 #include "llama-impl.h"
 #include "llama-batch.h"
 #include "llama-io.h"
@@ -8,6 +9,7 @@
 #include "llama-model.h"
 
 #include <cinttypes>
+#include <cmath>
 #include <cstring>
 #include <limits>
 #include <stdexcept>
@@ -21,6 +23,8 @@ llama_context::llama_context(
               llama_context_params params) :
     model(model),
     balloc(std::make_unique<llama_batch_allocr>(model.hparams.n_pos_per_embd())) {
+    // TODO warning when creating llama_context with awkward ctx size that is not a power of 2,
+    //     may need to be backend-dependent
     LLAMA_LOG_INFO("%s: constructing llama_context\n", __func__);
 
     t_start_us = model.t_start_us;
@@ -56,6 +60,25 @@ llama_context::llama_context(
     cparams.cb_eval           = params.cb_eval;
     cparams.cb_eval_user_data = params.cb_eval_user_data;
 
+    // Initialize backend samplers here so they are part of the sampling graph
+    // before the reserve passes run later in this function. This avoids a later
+    // re-reserve when graph nodes change.
+    if (params.samplers != nullptr && params.n_samplers > 0) {
+        for (size_t i = 0; i < params.n_samplers; ++i) {
+            const auto & config = params.samplers[i];
+
+            if (llama_sampler_chain_get(config.sampler, -1) == nullptr) {
+                throw std::runtime_error("the backend samplers must be of type llama_sampler_chain");
+            }
+
+            if (set_sampler(config.seq_id, config.sampler)) {
+                const int n_samplers = llama_sampler_chain_n(config.sampler);
+
+                LLAMA_LOG_INFO("%s: setting backend sampler for seq_id %d (n = %d)\n", __func__, config.seq_id, n_samplers);
+            }
+        }
+    }
+
     auto rope_scaling_type = params.rope_scaling_type;
     if (rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED) {
         rope_scaling_type = hparams.rope_scaling_type_train;
@@ -69,6 +92,43 @@ llama_context::llama_context(
         cparams.yarn_ext_factor = rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_YARN ? 1.0f : 0.0f;
     }
 
+    if (cparams.yarn_ext_factor != 0) {
+        static auto get_mscale = [](float scale, float mscale) {
+            return scale <= 1.0f ? 1.0f : (0.1f * mscale * logf(scale) + 1.0f);
+        };
+
+        const float factor = 1.0f / cparams.rope_freq_scale;
+
+        // ref: https://github.com/huggingface/transformers/blob/6d00f6b0a5679c36510f203e4226e36f517c3032/src/transformers/modeling_rope_utils.py#L336-L348
+        if (hparams.rope_yarn_log_mul != 0.0f) {
+            // note: here we assume `mscale == 1.0f`
+            // TODO: start reading the actual value of mscale and handle the case where it is not 1.0f
+                  float mscale          = 1.0f;
+            const float mscale_all_dims = hparams.rope_yarn_log_mul;
+
+            // [TAG_DEEPSEEK2_YARN_LOG_MUL_FIX]
+            // special-case DEEPSEEK v2:
+            // https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite-Chat/blob/main/config.json#L42-L43
+            if (model.arch == LLM_ARCH_DEEPSEEK2 && mscale_all_dims != 1.0f) {
+                mscale = mscale_all_dims;
+            }
+
+            cparams.yarn_attn_factor = get_mscale(factor, mscale) / get_mscale(factor, mscale_all_dims);
+
+            LLAMA_LOG_WARN("%s: setting new yarn_attn_factor = %.4f (mscale == %.1f, mscale_all_dim = %.1f)\n",
+                    __func__, cparams.yarn_attn_factor, mscale, mscale_all_dims);
+        } else {
+            cparams.yarn_attn_factor = get_mscale(factor, 1.0f);
+        }
+
+        // when YARN is applied with yarn_ext_factor != 0.0f, we need to cancel this factor:
+        // https://github.com/ggml-org/llama.cpp/blob/a81a569577cc38b32558958b048228150be63eae/ggml/src/ggml-cpu/ops.cpp#L5541-L5544
+        //
+        // ref: https://github.com/ggml-org/llama.cpp/discussions/7416
+        //      https://github.com/ggml-org/llama.cpp/pull/17945
+        cparams.yarn_attn_factor *= 1.0f / (1.0f + 0.1f * logf(factor));
+    }
+
     cparams.yarn_attn_factor *= hparams.rope_attn_factor;
 
     if (cparams.pooling_type == LLAMA_POOLING_TYPE_UNSPECIFIED) {
@@ -90,14 +150,6 @@ llama_context::llama_context(
     // with causal attention, the batch size is limited by the context size
     cparams.n_batch = cparams.causal_attn ? std::min(cparams.n_ctx, params.n_batch) : params.n_batch;
 
-    // the batch has to be at least GGML_KQ_MASK_PAD because we will be padding the KQ_mask
-    // this is required by GPU kernels in order to avoid out-of-bounds accesses (e.g. ggml_flash_attn_ext)
-    // ref: https://github.com/ggerganov/llama.cpp/pull/5021
-    // TODO: this padding is not needed for the cache-less context so we should probably move it to llama_memory
-    if (cparams.n_batch < GGML_KQ_MASK_PAD) {
-        LLAMA_LOG_WARN("%s: n_batch is less than GGML_KQ_MASK_PAD - increasing to %d\n", __func__, GGML_KQ_MASK_PAD);
-        cparams.n_batch = GGML_KQ_MASK_PAD;
-    }
     cparams.n_ubatch = std::min(cparams.n_batch, params.n_ubatch == 0 ? params.n_batch : params.n_ubatch);
 
     cparams.op_offload = params.op_offload;
@@ -112,11 +164,28 @@ llama_context::llama_context(
         }
     }
 
-    const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max;
+    // ref: https://github.com/ggml-org/llama.cpp/pull/17046#discussion_r2503085732
+    cparams.n_ctx = GGML_PAD(cparams.n_ctx, 256);
+
+    if (cparams.kv_unified) {
+        cparams.n_ctx_seq = cparams.n_ctx;
+    } else {
+        cparams.n_ctx_seq = cparams.n_ctx / cparams.n_seq_max;
+        cparams.n_ctx_seq = GGML_PAD(cparams.n_ctx_seq, 256);
+
+        if (cparams.n_ctx_seq == 0) {
+            throw std::runtime_error("n_ctx_seq == 0");
+        }
+
+        if (cparams.n_ctx != cparams.n_ctx_seq * cparams.n_seq_max) {
+            cparams.n_ctx =  cparams.n_ctx_seq * cparams.n_seq_max;
+            LLAMA_LOG_WARN("%s: n_ctx is not divisible by n_seq_max - rounding down to %u\n", __func__, cparams.n_ctx);
+        }
+    }
 
     LLAMA_LOG_INFO("%s: n_seq_max     = %u\n",   __func__, cparams.n_seq_max);
     LLAMA_LOG_INFO("%s: n_ctx         = %u\n",   __func__, cparams.n_ctx);
-    LLAMA_LOG_INFO("%s: n_ctx_per_seq = %u\n",   __func__, n_ctx_per_seq);
+    LLAMA_LOG_INFO("%s: n_ctx_seq     = %u\n",   __func__, cparams.n_ctx_seq);
     LLAMA_LOG_INFO("%s: n_batch       = %u\n",   __func__, cparams.n_batch);
     LLAMA_LOG_INFO("%s: n_ubatch      = %u\n",   __func__, cparams.n_ubatch);
     LLAMA_LOG_INFO("%s: causal_attn   = %d\n",   __func__, cparams.causal_attn);
@@ -125,14 +194,14 @@ llama_context::llama_context(
     LLAMA_LOG_INFO("%s: freq_base     = %.1f\n", __func__, cparams.rope_freq_base);
     LLAMA_LOG_INFO("%s: freq_scale    = %g\n",   __func__, cparams.rope_freq_scale);
 
-    if (n_ctx_per_seq < hparams.n_ctx_train) {
-        LLAMA_LOG_WARN("%s: n_ctx_per_seq (%u) < n_ctx_train (%u) -- the full capacity of the model will not be utilized\n",
-                __func__, n_ctx_per_seq, hparams.n_ctx_train);
+    if (cparams.n_ctx_seq < hparams.n_ctx_train) {
+        LLAMA_LOG_WARN("%s: n_ctx_seq (%u) < n_ctx_train (%u) -- the full capacity of the model will not be utilized\n",
+                __func__, cparams.n_ctx_seq, hparams.n_ctx_train);
     }
 
-    if (n_ctx_per_seq > hparams.n_ctx_train) {
-        LLAMA_LOG_WARN("%s: n_ctx_per_seq (%u) > n_ctx_train (%u) -- possible training context overflow\n",
-                __func__, n_ctx_per_seq, hparams.n_ctx_train);
+    if (cparams.n_ctx_seq > hparams.n_ctx_train) {
+        LLAMA_LOG_WARN("%s: n_ctx_seq (%u) > n_ctx_train (%u) -- possible training context overflow\n",
+                __func__, cparams.n_ctx_seq, hparams.n_ctx_train);
     }
 
     if (!hparams.vocab_only) {
@@ -181,7 +250,10 @@ llama_context::llama_context(
         // graph outputs buffer
         {
             // resized during inference when a batch uses more outputs
-            if (output_reserve(params.n_seq_max) < params.n_seq_max) {
+            // Create a dummy batch for initialization.
+            llama_batch dummy_batch = {};
+            dummy_batch.n_tokens = 0;
+            if (output_reserve(params.n_seq_max, dummy_batch) < params.n_seq_max) {
                 throw std::runtime_error("failed to reserve initial output buffer");
             }
 
@@ -208,6 +280,7 @@ llama_context::llama_context(
 
         backend_buft.clear();
         backend_ptrs.clear();
+        backend_buf_exp_size.clear();
 
         for (auto & backend : backends) {
             auto * buft = ggml_backend_get_default_buffer_type(backend.get());
@@ -224,11 +297,15 @@ llama_context::llama_context(
 
             backend_buft.push_back(buft);
             backend_ptrs.push_back(backend.get());
+            backend_buf_exp_size.push_back(0);
         }
 
         LLAMA_LOG_DEBUG("%s: backend_ptrs.size() = %zu\n", __func__, backend_ptrs.size());
 
-        const size_t max_nodes = this->graph_max_nodes();
+        const uint32_t n_seqs = cparams.n_seq_max;
+        const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
+
+        const size_t max_nodes = this->graph_max_nodes(n_tokens);
 
         LLAMA_LOG_DEBUG("%s: max_nodes = %zu\n", __func__, max_nodes);
 
@@ -239,8 +316,8 @@ llama_context::llama_context(
         // enabling pipeline parallelism in the scheduler increases memory usage, so it is only done when necessary
         bool pipeline_parallel =
             model.n_devices() > 1 &&
-            model.params.n_gpu_layers > (int) model.hparams.n_layer &&
-            model.params.split_mode == LLAMA_SPLIT_MODE_LAYER &&
+            model.n_gpu_layers() > model.hparams.n_layer &&
+            model.split_mode() == LLAMA_SPLIT_MODE_LAYER &&
             cparams.offload_kqv &&
             !model.has_tensor_overrides();
 
@@ -268,9 +345,7 @@ llama_context::llama_context(
         if (pipeline_parallel) {
             LLAMA_LOG_INFO("%s: pipeline parallelism enabled (n_copies=%d)\n", __func__, ggml_backend_sched_get_n_copies(sched.get()));
         }
-    }
 
-    if (!hparams.vocab_only) {
         llama_memory_context_ptr mctx;
         if (memory) {
             LLAMA_LOG_DEBUG("%s: reserving full memory module\n", __func__);
@@ -282,9 +357,6 @@ llama_context::llama_context(
 
         cross.v_embd.clear();
 
-        const uint32_t n_seqs = cparams.kv_unified ? 1 : cparams.n_seq_max;
-        const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
-
         // avoid reserving graphs with zero outputs - assume one output per sequence
         n_outputs = n_seqs;
 
@@ -341,9 +413,17 @@ llama_context::llama_context(
 
         // reserve pp (prompt processing) graph first so that buffers are only allocated once
         {
-            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
+            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get(),
+                model.hparams.no_alloc, model.hparams.no_alloc ? backend_buf_exp_size.data() : nullptr);
             if (!gf) {
-                throw std::runtime_error("failed to allocate compute pp buffers");
+                if (pipeline_parallel) {
+                    LLAMA_LOG_WARN("%s: compute buffer allocation failed, retrying without pipeline parallelism\n", __func__);
+                    sched.reset(ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), max_nodes, false, cparams.op_offload));
+                    gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
+                }
+                if (!gf) {
+                    throw std::runtime_error("failed to allocate compute pp buffers");
+                }
             }
 
             n_splits_pp = ggml_backend_sched_get_n_splits(sched.get());
@@ -352,7 +432,7 @@ llama_context::llama_context(
 
         // reserve with tg (token generation) graph to get the number of splits and nodes
         {
-            auto * gf = graph_reserve(n_seqs, n_seqs, n_seqs, mctx.get());
+            auto * gf = graph_reserve(n_seqs, n_seqs, n_seqs, mctx.get(), model.hparams.no_alloc);
             if (!gf) {
                 throw std::runtime_error("failed to allocate compute tg buffers");
             }
@@ -367,7 +447,7 @@ llama_context::llama_context(
             //
             // auto * gf = graph_reserve(n_tokens, 1, n_tokens, mctx.get());
             //
-            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
+            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get(), model.hparams.no_alloc);
             if (!gf) {
                 throw std::runtime_error("failed to allocate compute pp buffers");
             }
@@ -376,11 +456,13 @@ llama_context::llama_context(
         for (size_t i = 0; i < backend_ptrs.size(); ++i) {
             ggml_backend_t             backend = backend_ptrs[i];
             ggml_backend_buffer_type_t buft    = backend_buft[i];
-            size_t size = ggml_backend_sched_get_buffer_size(sched.get(), backend);
-            if (size > 1) {
+            if (!model.hparams.no_alloc) {
+                backend_buf_exp_size[i] = ggml_backend_sched_get_buffer_size(sched.get(), backend);
+            }
+            if (backend_buf_exp_size[i] > 1) {
                 LLAMA_LOG_INFO("%s: %10s compute buffer size = %8.2f MiB\n", __func__,
                         ggml_backend_buft_name(buft),
-                        size / 1024.0 / 1024.0);
+                        backend_buf_exp_size[i] / 1024.0 / 1024.0);
             }
         }
 
@@ -396,9 +478,35 @@ llama_context::llama_context(
             LLAMA_LOG_INFO("%s: graph splits = %d (with bs=%d), %d (with bs=1)\n", __func__, n_splits_pp, n_tokens, n_splits_tg);
         }
     }
+
+    // Initialize the full vocabulary token ids for backend samplers.
+    {
+        const int n_vocab = model.vocab.n_tokens();
+
+        sampling.token_ids_full_vocab.resize(n_vocab);
+        for (int i = 0; i < n_vocab; ++i) {
+            sampling.token_ids_full_vocab[i] = i;
+        }
+    }
 }
 
 llama_context::~llama_context() {
+    if (!model.hparams.no_alloc) {
+        for (size_t i = 0; i < backend_ptrs.size(); ++i) {
+            ggml_backend_t             backend = backend_ptrs[i];
+            ggml_backend_buffer_type_t buft    = backend_buft[i];
+
+            const size_t size_exp = backend_buf_exp_size[i];
+            const size_t size_act = ggml_backend_sched_get_buffer_size(sched.get(), backend);
+            if (size_exp == size_act) {
+                LLAMA_LOG_DEBUG("%s: %10s compute buffer size is %8.4f MiB, matches expectation of %8.4f MiB\n",
+                    __func__, ggml_backend_buft_name(buft), size_act / (1024.0*1024.0), size_exp / (1024.0*1024.0));
+            } else {
+                LLAMA_LOG_WARN("%s: %10s compute buffer size of %8.4f MiB, does not match expectation of %8.4f MiB\n",
+                    __func__, ggml_backend_buft_name(buft), size_act / (1024.0*1024.0), size_exp / (1024.0*1024.0));
+            }
+        }
+    }
     ggml_opt_free(opt_ctx);
 }
 
@@ -448,8 +556,8 @@ uint32_t llama_context::n_ctx() const {
     return cparams.n_ctx;
 }
 
-uint32_t llama_context::n_ctx_per_seq() const {
-    return cparams.n_ctx / cparams.n_seq_max;
+uint32_t llama_context::n_ctx_seq() const {
+    return cparams.n_ctx_seq;
 }
 
 uint32_t llama_context::n_batch() const {
@@ -518,7 +626,7 @@ bool llama_context::memory_update(bool optimize) {
             throw std::runtime_error("failed to initialize memory context");
         }
 
-        const uint32_t n_seqs = cparams.kv_unified ? 1 : cparams.n_seq_max;
+        const uint32_t n_seqs = cparams.n_seq_max;
         const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
 
         auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
@@ -540,6 +648,35 @@ float * llama_context::get_logits() {
     return logits;
 }
 
+int64_t llama_context::output_resolve_row(int32_t i) const {
+    int64_t j = -1;
+
+    // support negative indices (last output row)
+    if (i < 0) {
+        j = n_outputs + i;
+        if (j < 0) {
+            throw std::runtime_error(format("negative index out of range [0, %d)", n_outputs));
+        }
+    } else if ((size_t) i >= output_ids.size()) {
+        throw std::runtime_error(format("out of range [0, %zu)", output_ids.size()));
+    } else {
+        // use output_ids to translate the batch token index into a row number
+        // that holds this token's data.
+        j = output_ids[i];
+    }
+
+    if (j < 0) {
+        // the batch token was not configured to output anything
+        throw std::runtime_error(format("batch.logits[%d] != true", i));
+    }
+
+    if (j >= n_outputs) {
+        throw std::runtime_error(format("corrupt output buffer (j=%" PRId64 ", n_outputs=%d)", j, n_outputs));
+    }
+
+    return j;
+}
+
 float * llama_context::get_logits_ith(int32_t i) {
     int64_t j = -1;
 
@@ -550,6 +687,7 @@ float * llama_context::get_logits_ith(int32_t i) {
             throw std::runtime_error("no logits");
         }
 
+        // TODO: use output_resolve_row()
         if (i < 0) {
             j = n_outputs + i;
             if (j < 0) {
@@ -586,6 +724,10 @@ float * llama_context::get_embeddings() {
     return embd;
 }
 
+llama_token * llama_context::get_sampled_tokens()  const{
+    return sampling.sampled;
+}
+
 float * llama_context::get_embeddings_ith(int32_t i) {
     int64_t j = -1;
 
@@ -596,6 +738,7 @@ float * llama_context::get_embeddings_ith(int32_t i) {
             throw std::runtime_error("no embeddings");
         }
 
+        // TODO: use output_resolve_row()
         if (i < 0) {
             j = n_outputs + i;
             if (j < 0) {
@@ -615,7 +758,8 @@ float * llama_context::get_embeddings_ith(int32_t i) {
             throw std::runtime_error(format("corrupt output buffer (j=%" PRId64 ", n_outputs=%d)", j, n_outputs));
         }
 
-        return embd + j*model.hparams.n_embd;
+        const uint32_t n_embd_out = model.hparams.get_n_embd_out();
+        return embd + j*n_embd_out;
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: invalid embeddings id %d, reason: %s\n", __func__, i, err.what());
 #ifndef NDEBUG
@@ -635,6 +779,136 @@ float * llama_context::get_embeddings_seq(llama_seq_id seq_id) {
     return it->second.data();
 }
 
+llama_token llama_context::get_sampled_token_ith(int32_t idx) {
+    output_reorder();
+
+    if (sampling.sampled == nullptr) {
+        return LLAMA_TOKEN_NULL;
+    }
+
+    try {
+        const int64_t row = output_resolve_row(idx);
+        GGML_ASSERT(row < (int64_t) sampling.sampled_size);
+        return sampling.sampled[row];
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("%s: invalid backend sampled token id %d, reason: %s\n", __func__, idx, err.what());
+        return LLAMA_TOKEN_NULL;
+    }
+}
+
+float * llama_context::get_sampled_probs_ith(int32_t idx) {
+    output_reorder();
+
+    if (sampling.probs == nullptr) {
+        return nullptr;
+    }
+
+    try {
+        const int64_t row = output_resolve_row(idx);
+        if ((size_t) row >= sampling.probs_count.size() || sampling.probs_count[row] == 0) {
+            return nullptr;
+        }
+        return sampling.probs + row*model.vocab.n_tokens();
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("%s: invalid backend sampled probs id %d, reason: %s\n", __func__, idx, err.what());
+        return nullptr;
+    }
+}
+
+float * llama_context::get_sampled_logits_ith(int32_t idx) {
+    output_reorder();
+
+    if (sampling.logits == nullptr) {
+        return nullptr;
+    }
+
+    try {
+        const int64_t row = output_resolve_row(idx);
+        if ((size_t) row >= sampling.logits_count.size() || sampling.logits_count[row] == 0) {
+            return nullptr;
+        }
+        return sampling.logits + row*model.vocab.n_tokens();
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("%s: invalid backend sampled logits id %d, reason: %s\n", __func__, idx, err.what());
+        return nullptr;
+    }
+}
+
+const llama_token * llama_context::get_sampled_candidates_ith(int32_t idx) {
+    output_reorder();
+
+    try {
+        const int64_t row = output_resolve_row(idx);
+        if (sampling.candidates != nullptr &&
+            (size_t) row < sampling.candidates_count.size() &&
+            sampling.candidates_count[row] > 0) {
+            return sampling.candidates + row*model.vocab.n_tokens();
+        }
+    } catch (const std::exception & err) {
+        // fallback to full vocab list
+    }
+
+    return sampling.token_ids_full_vocab.data();
+}
+
+size_t llama_context::get_sampled_candidates_count(int32_t idx) {
+    output_reorder();
+
+    if (sampling.candidates == nullptr) {
+        return 0;
+    }
+
+    try {
+        const int64_t row = output_resolve_row(idx);
+        if ((size_t) row >= sampling.candidates_count.size()) {
+            return 0;
+        }
+        return sampling.candidates_count[row];
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("%s: invalid backend sampled candidates count id %d, reason: %s\n", __func__, idx, err.what());
+        return 0;
+    }
+}
+
+size_t llama_context::get_sampled_logits_count(int32_t idx) {
+    output_reorder();
+
+    if (sampling.logits == nullptr) {
+        return model.vocab.n_tokens();
+    }
+
+    try {
+        const int64_t row = output_resolve_row(idx);
+        if ((size_t) row >= sampling.logits_count.size()) {
+            return 0;
+        }
+        return sampling.logits_count[row];
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("%s: invalid backend sampled logits count id %d, reason: %s\n", __func__, idx, err.what());
+        return 0;
+    }
+}
+
+size_t llama_context::get_sampled_probs_count(int32_t idx) {
+    output_reorder();
+
+    if (sampling.probs == nullptr) {
+        return 0;
+    }
+
+    try {
+        const int64_t row = output_resolve_row(idx);
+        if ((size_t) row >= sampling.probs_count.size()) {
+            return 0;
+        }
+        return sampling.probs_count[row];
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("%s: invalid backend sampled probs count id %d, reason: %s\n", __func__, idx, err.what());
+        return 0;
+    }
+}
+
+
 void llama_context::attach_threadpool(
            ggml_threadpool_t threadpool,
            ggml_threadpool_t threadpool_batch) {
@@ -691,6 +965,42 @@ void llama_context::set_warmup(bool value) {
     cparams.warmup = value;
 }
 
+bool llama_context::set_sampler(llama_seq_id seq_id, llama_sampler * sampler) {
+    LLAMA_LOG_DEBUG("%s: seq_id = %d, sampler = %p\n", __func__, (int) seq_id, (void *) sampler);
+
+    const bool can_offload =
+        sampler &&
+        sampler->iface->backend_init &&
+        sampler->iface->backend_apply &&
+        llama_sampler_chain_n(sampler) > 0;
+
+    if (sampler && can_offload) {
+        ggml_backend_buffer_type_t buft = ggml_backend_dev_buffer_type(model.dev_output());
+        auto * host_buft = ggml_backend_dev_host_buffer_type(model.dev_output());
+        if (host_buft) {
+            buft = host_buft;
+        }
+
+        sampler->iface->backend_init(sampler, buft);
+
+        sampling.samplers[seq_id] = sampler;
+
+        return true;
+    }
+
+    if (sampler && !can_offload) {
+        LLAMA_LOG_WARN("%s: sampler '%s' for seq_id = %d, cannot be offloaded to the backend\n", __func__, llama_sampler_name(sampler), seq_id);
+
+        sampling.samplers.erase(seq_id);
+
+        return false;
+    }
+
+    sampling.samplers.erase(seq_id);
+
+    return true;
+}
+
 void llama_context::set_adapter_lora(
             llama_adapter_lora * adapter,
             float scale) {
@@ -803,7 +1113,7 @@ int llama_context::encode(const llama_batch & batch_inp) {
 
     const auto & hparams = model.hparams;
 
-    const int64_t n_embd  = hparams.n_embd;
+    const int64_t n_embd  = hparams.n_embd_inp();
     const int64_t n_vocab = model.vocab.n_tokens();
 
     // note: during encode, we always pass the full sequence starting from pos = 0
@@ -831,7 +1141,7 @@ int llama_context::encode(const llama_batch & batch_inp) {
     n_queued_tokens += n_tokens;
 
     // reserve output buffer
-    if (output_reserve(n_tokens) < n_tokens) {
+    if (output_reserve(n_tokens, batch_inp) < n_tokens) {
         LLAMA_LOG_ERROR("%s: could not reserve space for batch with %u outputs\n", __func__, n_tokens);
         return -2;
     };
@@ -885,9 +1195,10 @@ int llama_context::encode(const llama_batch & batch_inp) {
                 {
                     // extract token embeddings
                     GGML_ASSERT(embd != nullptr);
+                    const uint32_t n_embd_out = hparams.get_n_embd_out();
 
-                    GGML_ASSERT(n_tokens*n_embd <= (int64_t) embd_size);
-                    ggml_backend_tensor_get_async(backend_embd, t_embd, embd, 0, n_tokens*n_embd*sizeof(float));
+                    GGML_ASSERT(n_tokens*n_embd_out <= (int64_t) embd_size);
+                    ggml_backend_tensor_get_async(backend_embd, t_embd, embd, 0, n_tokens*n_embd_out*sizeof(float));
                 } break;
             case LLAMA_POOLING_TYPE_MEAN:
             case LLAMA_POOLING_TYPE_CLS:
@@ -955,6 +1266,112 @@ int llama_context::encode(const llama_batch & batch_inp) {
     return 0;
 }
 
+static std::map<llama_seq_id, uint32_t> build_seq_to_output_row(const llama_ubatch & ubatch, uint32_t row_offset) {
+    std::map<llama_seq_id, uint32_t> seq_to_row;
+    // how many output tokens we have seen so far for this ubatch.
+    uint32_t local = 0;
+    for (uint32_t i = 0; i < ubatch.n_tokens; ++i) {
+        // skip tokens that are not output.
+        if (!ubatch.output[i]) {
+            continue;
+        }
+
+        const llama_seq_id seq_id = ubatch.seq_id[i][0];
+        // row_offset is the number of output tokens before this ubatch.
+        seq_to_row[seq_id] = row_offset + local;
+        ++local;
+    }
+    return seq_to_row;
+}
+
+static void copy_tensor_async_ints(
+    const std::map<llama_seq_id, ggml_tensor*> & tensor_map,
+    llama_token * sampled,
+    size_t sampled_size,
+    const std::map<llama_seq_id, uint32_t> & seq_to_row,
+    ggml_backend_sched_t sched) {
+    if (sampled == nullptr) {
+        return;
+    }
+
+    for (const auto & [seq_id, tensor] : tensor_map) {
+        auto it = seq_to_row.find(seq_id);
+        if (it == seq_to_row.end()) {
+            continue;
+        }
+
+        const uint32_t row = it->second;
+        GGML_ASSERT(row < sampled_size);
+
+        GGML_ASSERT(ggml_is_contiguous(tensor) && "sampled tokens tensor must be contiguous for async copy");
+
+        ggml_backend_t backend = ggml_backend_sched_get_tensor_backend(sched, tensor);
+        ggml_backend_tensor_get_async(backend, tensor, sampled + row, 0, sizeof(sampled[row]));
+    }
+}
+
+static void copy_tensor_async_floats(
+    const std::map<llama_seq_id, ggml_tensor*> & tensor_map,
+    float * dst,
+    size_t stride,
+    std::vector<uint32_t> & counts,
+    const std::map<llama_seq_id, uint32_t> & seq_to_row,
+    ggml_backend_sched_t sched) {
+    if (dst == nullptr) {
+        return;
+    }
+
+    for (const auto & [seq_id, tensor] : tensor_map) {
+        auto it = seq_to_row.find(seq_id);
+        if (it == seq_to_row.end()) {
+            continue;
+        }
+
+        const uint32_t row = it->second;
+        GGML_ASSERT(row < counts.size());
+
+        GGML_ASSERT(ggml_is_contiguous(tensor) && "logits/probs tensor must be contiguous for async copy");
+
+        ggml_backend_t backend = ggml_backend_sched_get_tensor_backend(sched, tensor);
+        float * row_ptr = dst + (size_t) row * stride;
+        ggml_backend_tensor_get_async(backend, tensor, row_ptr, 0, ggml_nbytes(tensor));
+
+        // Update the actual number of logits/probabilities that were written for this row.
+        counts[row] = ggml_nelements(tensor);
+    }
+}
+
+static void copy_tensor_async_candidates(
+    const std::map<llama_seq_id, ggml_tensor*> & tensor_map,
+    llama_token * dst,
+    size_t stride,
+    std::vector<uint32_t> & counts,
+    const std::map<llama_seq_id, uint32_t> & seq_to_row,
+    ggml_backend_sched_t sched) {
+    if (dst == nullptr) {
+        return;
+    }
+
+    for (const auto & [seq_id, tensor] : tensor_map) {
+        auto it = seq_to_row.find(seq_id);
+        if (it == seq_to_row.end()) {
+            continue;
+        }
+
+        const uint32_t row = it->second;
+        GGML_ASSERT(row < counts.size());
+
+        GGML_ASSERT(ggml_is_contiguous(tensor) && "candidates tensor must be contiguous for async copy");
+
+        ggml_backend_t backend = ggml_backend_sched_get_tensor_backend(sched, tensor);
+        llama_token * row_ptr = dst + (size_t) row * stride;
+        ggml_backend_tensor_get_async(backend, tensor, row_ptr, 0, ggml_nbytes(tensor));
+
+        // Update the actual number of candidates that were written.
+        counts[row] = ggml_nelements(tensor);
+    }
+}
+
 int llama_context::decode(const llama_batch & batch_inp) {
     GGML_ASSERT((!batch_inp.token && batch_inp.embd) || (batch_inp.token && !batch_inp.embd)); // NOLINT
 
@@ -972,12 +1389,39 @@ int llama_context::decode(const llama_batch & batch_inp) {
     const auto & hparams = model.hparams;
 
     const int64_t n_vocab = vocab.n_tokens();
-    const int64_t n_embd  = hparams.n_embd;
+    const int64_t n_embd  = hparams.n_embd_inp();
 
     // when computing embeddings, all tokens are output
-    const bool output_all = cparams.embeddings;
+    const bool output_all   = cparams.embeddings;
+    const bool has_samplers = !sampling.samplers.empty();
+
+    const uint32_t n_seq_max = cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max;
+
+    // TODO: avoid this workaround in the future
+    if (has_samplers && batch_inp.logits) {
+        std::vector<int32_t> seq_output_count(n_seq_max, 0);
+
+        for (int32_t i = 0; i < batch_inp.n_tokens; ++i) {
+            if (batch_inp.logits[i] == 0) {
+                continue;
+            }
 
-    if (!balloc->init(batch_inp, vocab, memory.get(), n_embd, cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, output_all)) {
+            const int ns = batch_inp.n_seq_id ? batch_inp.n_seq_id[i] : 1;
+
+            for (int32_t s = 0; s < ns; ++s) {
+                const llama_seq_id seq_id = batch_inp.seq_id ? batch_inp.seq_id[i][s] : 0;
+
+                seq_output_count[seq_id]++;
+                if (seq_output_count[seq_id] > 1) {
+                    LLAMA_LOG_ERROR("%s: backend sampling requires at most one output token per sequence (seq_id %d had %d)\n",
+                            __func__, seq_id, seq_output_count[seq_id]);
+                    return -1;
+                }
+            }
+        }
+    }
+
+    if (!balloc->init(batch_inp, vocab, memory.get(), n_embd, n_seq_max, output_all)) {
         LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
         return -1;
     }
@@ -1058,7 +1502,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
     }
 
     // reserve output buffer
-    if (output_reserve(n_outputs_all) < n_outputs_all) {
+    if (output_reserve(n_outputs_all, balloc->get_batch()) < n_outputs_all) {
         LLAMA_LOG_ERROR("%s: could not reserve space for batch with %d outputs\n", __func__, n_outputs_all);
         return -2;
     };
@@ -1131,7 +1575,10 @@ int llama_context::decode(const llama_batch & batch_inp) {
         }
 
         // extract logits
-        if (t_logits && n_outputs > 0) {
+        // For multi-sequence batches that mix backend samplers and CPU sampler
+        // this is currently inefficient as we copy all logits even for the
+        // backend sampled tokens.
+        if (logits && t_logits && n_outputs > 0) {
             ggml_backend_t backend_res = ggml_backend_sched_get_tensor_backend(sched.get(), t_logits);
             GGML_ASSERT(backend_res != nullptr);
             GGML_ASSERT(logits != nullptr);
@@ -1146,7 +1593,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
         }
 
         // extract embeddings
-        if (t_embd && n_outputs > 0) {
+        if (embd && t_embd && n_outputs > 0) {
             ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(sched.get(), t_embd);
             GGML_ASSERT(backend_embd != nullptr);
 
@@ -1155,12 +1602,13 @@ int llama_context::decode(const llama_batch & batch_inp) {
                     {
                         // extract token embeddings
                         GGML_ASSERT(embd != nullptr);
-                        float * embd_out = embd + n_outputs_prev*n_embd;
+                        const uint32_t n_embd_out = hparams.get_n_embd_out();
+                        float * embd_out = embd + n_outputs_prev*n_embd_out;
 
                         if (n_outputs) {
                             GGML_ASSERT( n_outputs_prev + n_outputs <= n_outputs_all);
-                            GGML_ASSERT((n_outputs_prev + n_outputs)*n_embd <= (int64_t) embd_size);
-                            ggml_backend_tensor_get_async(backend_embd, t_embd, embd_out, 0, n_outputs*n_embd*sizeof(float));
+                            GGML_ASSERT((n_outputs_prev + n_outputs)*n_embd_out <= (int64_t) embd_size);
+                            ggml_backend_tensor_get_async(backend_embd, t_embd, embd_out, 0, n_outputs*n_embd_out*sizeof(float));
                         }
                     } break;
                 case LLAMA_POOLING_TYPE_MEAN:
@@ -1200,6 +1648,22 @@ int llama_context::decode(const llama_batch & batch_inp) {
             }
         }
 
+        // This flag indicates whether a backend sampler has actually sampled a specific
+        // token, or if it has produced probabilites. If true, we can skip the normal copying of logits and embeddings.
+        const bool has_sampled = !res->t_sampled.empty() || !res->t_sampled_probs.empty() || !res->t_sampled_logits.empty();
+
+        if (has_samplers && has_sampled) {
+            const auto seq_to_output_row = build_seq_to_output_row(ubatch, n_outputs_prev);
+            const auto stride = n_vocab;
+
+            // async copy the sampling data from the backend to the host
+            copy_tensor_async_ints(res->t_sampled, sampling.sampled, sampling.sampled_size, seq_to_output_row, sched.get());
+
+            copy_tensor_async_floats    (res->t_sampled_logits, sampling.logits,     stride, sampling.logits_count,     seq_to_output_row, sched.get());
+            copy_tensor_async_floats    (res->t_sampled_probs,  sampling.probs,      stride, sampling.probs_count,      seq_to_output_row, sched.get());
+            copy_tensor_async_candidates(res->t_candidates,     sampling.candidates, stride, sampling.candidates_count, seq_to_output_row, sched.get());
+        }
+
         n_outputs_prev += n_outputs;
     } while (mctx->next());
 
@@ -1224,7 +1688,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
 
         // make the outputs have the same order they had in the user-provided batch
         // note: this is mostly relevant for recurrent models atm
-        if (!sorted_output) {
+        if (!sorted_output && n_outputs > 1) {
             GGML_ASSERT((size_t) n_outputs == out_ids.size());
 
             // TODO: is there something more efficient which also minimizes swaps?
@@ -1263,15 +1727,15 @@ int llama_context::decode(const llama_batch & batch_inp) {
 // output
 //
 
-uint32_t llama_context::output_reserve(int32_t n_outputs) {
+uint32_t llama_context::output_reserve(int32_t n_outputs, const llama_batch & batch) {
     const auto & hparams = model.hparams;
     const auto & vocab   = model.vocab;
 
     const int64_t n_outputs_max = std::max<int64_t>(n_outputs, n_seq_max());
 
-    const auto n_batch = cparams.n_batch;
-    const auto n_vocab = vocab.n_tokens();
-    const auto n_embd  = hparams.n_embd;
+    const auto n_batch    = cparams.n_batch;
+    const auto n_vocab    = vocab.n_tokens();
+    const auto n_embd_out = hparams.get_n_embd_out();
 
     bool has_logits = true;
     bool has_embd   = cparams.embeddings;
@@ -1282,8 +1746,53 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
         has_embd   = true;
     }
 
-    logits_size = has_logits ? n_vocab*n_outputs_max : 0;
-    embd_size   = has_embd   ?  n_embd*n_outputs_max : 0;
+    // Check which sampling modes are needed for the current batch.
+    // TODO: avoid this branching by working with the worst-case
+    bool has_sampling = false;
+    bool cpu_logits   = false;
+
+    if (batch.logits) {
+        for (int32_t i = 0; i < batch.n_tokens; i++) {
+            if (!batch.logits[i]) {
+                continue;
+            }
+            for (int32_t j = 0; j < batch.n_seq_id[i]; j++) {
+                llama_seq_id seq_id = batch.seq_id[i][j];
+                if (sampling.samplers.find(seq_id) != sampling.samplers.end()) {
+                    has_sampling = true;
+                } else {
+                    cpu_logits = true;
+                }
+            }
+        }
+    } else {
+        // When batch.logits is nullptr (when loading state with a dummy batch),
+        // allocate CPU logits.
+        cpu_logits = true;
+    }
+
+    size_t backend_float_count = 0;
+    size_t backend_token_count = 0;
+
+    // Allocate CPU logits buffer only if needed by sequences in this batch
+    logits_size = (has_logits && cpu_logits) ? n_vocab*n_outputs_max : 0;
+    embd_size   = has_embd ? n_embd_out*n_outputs_max : 0;
+
+    // TODO: avoid this branching by working with the worst-case
+    if (!has_sampling) {
+        sampling.logits_size     = 0;
+        sampling.probs_size      = 0;
+        sampling.sampled_size    = 0;
+        sampling.candidates_size = 0;
+    } else {
+        sampling.logits_size     = n_vocab*n_outputs_max;
+        sampling.probs_size      = n_vocab*n_outputs_max;
+        sampling.sampled_size    =         n_outputs_max;
+        sampling.candidates_size = n_vocab*n_outputs_max;
+
+        backend_float_count = sampling.logits_size  + sampling.probs_size;
+        backend_token_count = sampling.sampled_size + sampling.candidates_size;
+    }
 
     if (output_ids.empty()) {
         // init, never resized afterwards
@@ -1291,7 +1800,9 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
     }
 
     const size_t prev_size = buf_output ? ggml_backend_buffer_get_size(buf_output.get()) : 0;
-    const size_t new_size  = (logits_size + embd_size) * sizeof(float);
+    const size_t new_size  =
+        (logits_size + embd_size + backend_float_count) * sizeof(float) +
+        (                          backend_token_count) * sizeof(llama_token);
 
     // alloc only when more than the current capacity is required
     // TODO: also consider shrinking the buffer
@@ -1299,8 +1810,11 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
         if (buf_output) {
 #ifndef NDEBUG
             // This doesn't happen often, but may be annoying in some cases (like the HellaSwag benchmark)
-            LLAMA_LOG_INFO("%s: reallocating output buffer from size %.02f MiB to %.02f MiB\n", __func__, prev_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
+            LLAMA_LOG_DEBUG("%s: reallocating output buffer from size %.02f MiB to %.02f MiB\n", __func__, prev_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
 #endif
+            synchronize();
+
+            // TODO: not needed?
             buf_output = nullptr;
             logits = nullptr;
             embd = nullptr;
@@ -1322,8 +1836,49 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
 
     float * output_base = (float *) ggml_backend_buffer_get_base(buf_output.get());
 
-    logits = has_logits ? output_base               : nullptr;
-    embd   = has_embd   ? output_base + logits_size : nullptr;
+    logits = nullptr;
+    embd   = nullptr;
+
+    size_t offset = 0;
+    uint8_t * base = (uint8_t *) output_base;
+
+    logits = (has_logits && cpu_logits) ? output_base : nullptr;
+    offset += logits_size * sizeof(float);
+
+    embd = has_embd ? (float *) (base + offset) : nullptr;
+    offset += embd_size * sizeof(float);
+
+    sampling.logits     = nullptr;
+    sampling.probs      = nullptr;
+    sampling.sampled    = nullptr;
+    sampling.candidates = nullptr;
+
+    if (has_sampling) {
+        sampling.logits = (float *) (base + offset);
+        offset += sampling.logits_size * sizeof(float);
+
+        sampling.probs = (float *) (base + offset);
+        offset += sampling.probs_size * sizeof(float);
+
+        sampling.sampled = (llama_token *) (base + offset);
+        offset += sampling.sampled_size * sizeof(llama_token);
+
+        sampling.candidates = (llama_token *) (base + offset);
+        offset += sampling.candidates_size * sizeof(llama_token);
+
+        // The count vectors keep track of the actual number of logits/probs/candidates
+        // copied from the backend for each output row.
+
+        sampling.logits_count.resize(n_outputs_max);
+        sampling.probs_count.resize(n_outputs_max);
+        sampling.candidates_count.resize(n_outputs_max);
+
+        std::fill(sampling.logits_count.begin(),     sampling.logits_count.end(),     0);
+        std::fill(sampling.probs_count.begin(),      sampling.probs_count.end(),      0);
+        std::fill(sampling.candidates_count.begin(), sampling.candidates_count.end(), 0);
+
+        std::fill_n(sampling.sampled, sampling.sampled_size, LLAMA_TOKEN_NULL);
+    }
 
     // set all ids as invalid (negative)
     std::fill(output_ids.begin(), output_ids.end(), -1);
@@ -1352,6 +1907,40 @@ void llama_context::output_reorder() {
                 std::swap(embd[i0*n_embd + k], embd[i1*n_embd + k]);
             }
         }
+
+        if (sampling.logits && sampling.logits_size > 0) {
+            for (uint64_t k = 0; k < n_vocab; ++k) {
+                std::swap(sampling.logits[i0*n_vocab + k], sampling.logits[i1*n_vocab + k]);
+            }
+        }
+
+        if (sampling.probs && sampling.probs_size > 0) {
+            for (uint64_t k = 0; k < n_vocab; ++k) {
+                std::swap(sampling.probs[i0*n_vocab + k], sampling.probs[i1*n_vocab + k]);
+            }
+        }
+
+        if (sampling.candidates && sampling.candidates_size > 0) {
+            for (uint64_t k = 0; k < n_vocab; ++k) {
+                std::swap(sampling.candidates[i0*n_vocab + k], sampling.candidates[i1*n_vocab + k]);
+            }
+        }
+
+        if (sampling.sampled && sampling.sampled_size > 0) {
+            std::swap(sampling.sampled[i0], sampling.sampled[i1]);
+        }
+
+        if (!sampling.logits_count.empty()) {
+            std::swap(sampling.logits_count[i0], sampling.logits_count[i1]);
+        }
+
+        if (!sampling.probs_count.empty()) {
+            std::swap(sampling.probs_count[i0], sampling.probs_count[i1]);
+        }
+
+        if (!sampling.candidates_count.empty()) {
+            std::swap(sampling.candidates_count[i0], sampling.candidates_count[i1]);
+        }
     }
 
     output_swaps.clear();
@@ -1361,21 +1950,27 @@ void llama_context::output_reorder() {
 // graph
 //
 
-uint32_t llama_context::graph_max_nodes() const {
-    return std::max<uint32_t>(1024u, 8u*model.n_tensors());
+uint32_t llama_context::graph_max_nodes(uint32_t n_tokens) const {
+    if (model.arch == LLM_ARCH_QWEN3NEXT) {
+        return std::max<uint32_t>(n_tokens * 40, 32u * model.n_tensors());
+    }
+    uint32_t res = std::max<uint32_t>(1024u, 8u*model.n_tensors());
+    res += model.n_lora_nodes;
+    return res;
 }
 
 llm_graph_result * llama_context::get_gf_res_reserve() const {
     return static_cast<llm_graph_result *>(gf_res_reserve.get());
 }
 
-ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx, bool split_only) {
+ggml_cgraph * llama_context::graph_reserve(
+        uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx, bool split_only, size_t * sizes) {
     LLAMA_LOG_DEBUG("%s: reserving a graph for ubatch with n_tokens = %4u, n_seqs = %2u, n_outputs = %4u\n", __func__, n_tokens, n_seqs, n_outputs);
     GGML_ASSERT(n_outputs >= 1);
 
     if (n_tokens % n_seqs != 0) {
         n_tokens = ((n_tokens + (n_seqs - 1)) / n_seqs) * n_seqs; // round to next multiple of n_seqs
-        n_outputs = std::min(n_outputs, n_tokens);
+        n_outputs = std::max(n_outputs, n_tokens);
 
         LLAMA_LOG_DEBUG("%s: making n_tokens a multiple of n_seqs - n_tokens = %u, n_seqs = %u, n_outputs = %u\n", __func__, n_tokens, n_seqs, n_outputs);
     }
@@ -1394,6 +1989,15 @@ ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, u
     llama_batch_allocr balloc(model.hparams.n_pos_per_embd());
     llama_ubatch ubatch = balloc.ubatch_reserve(n_tokens/n_seqs, n_seqs);
 
+    // set one output token per sequence in order to activate all backend samplers
+    std::vector<llama_seq_id> seq_ids(n_seqs);
+    for (uint32_t i = 0; i < n_seqs; ++i) {
+        seq_ids[i] = i;
+        ubatch.n_seq_id[i] = 1;
+        ubatch.seq_id[i] = &seq_ids[i];
+        ubatch.output[i] = true;
+    }
+
     auto * res = gf_res_reserve.get();
 
     const auto gparams = graph_params(res, ubatch, mctx, LLM_GRAPH_TYPE_DEFAULT);
@@ -1406,8 +2010,13 @@ ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, u
 
     // initialize scheduler with the specified graph
     if (split_only) {
-        ggml_backend_sched_split_graph(sched.get(), gf);
+        if (sizes) {
+            ggml_backend_sched_reserve_size(sched.get(), gf, sizes);
+        } else {
+            ggml_backend_sched_split_graph(sched.get(), gf);
+        }
     } else if (!ggml_backend_sched_reserve(sched.get(), gf)) {
+        GGML_ASSERT(!sizes);
         LLAMA_LOG_ERROR("%s: failed to allocate compute buffers\n", __func__);
         return nullptr;
     }
@@ -1419,7 +2028,7 @@ llm_graph_params llama_context::graph_params(
                         llm_graph_result * res,
                       const llama_ubatch & ubatch,
             const llama_memory_context_i * mctx,
-            llm_graph_type   gtype) const {
+                          llm_graph_type   gtype) const {
     return {
         /*.arch        =*/ model.arch,
         /*.hparams     =*/ model.hparams,
@@ -1432,6 +2041,7 @@ llm_graph_params llama_context::graph_params(
         /*.loras       =*/ &loras,
         /*.mctx        =*/ mctx,
         /*.cross       =*/ &cross,
+        /*.samplers    =*/ sampling.samplers,
         /*.n_outputs   =*/ n_outputs,
         /*.cb          =*/ graph_get_cb(),
         /*.res         =*/ res,
@@ -1484,7 +2094,7 @@ llm_graph_cb llama_context::graph_get_cb() const {
 
         // norm may be automatically assigned to the backend of the previous layer, increasing data transfer between backends
         // FIXME: fix in ggml_backend_sched
-        const bool full_offload = model.params.n_gpu_layers > (int) model.hparams.n_layer;
+        const bool full_offload = model.n_gpu_layers() > model.hparams.n_layer;
         if (ubatch.n_tokens < 32 || full_offload) {
             if (il != -1 && strcmp(name, "norm") == 0) {
                 const auto & dev_layer = model.dev_layer(il);
@@ -1887,6 +2497,9 @@ size_t llama_context::state_write_data(llama_io_write_i & io) {
         }
     }
 
+    // TODO: handle sampling buffers and samplers state ?
+    //       https://github.com/ggml-org/llama.cpp/pull/17004
+
     if (memory != nullptr) {
         LLAMA_LOG_DEBUG("%s: - writing memory module\n", __func__);
         memory->state_write(io);
@@ -1919,7 +2532,10 @@ size_t llama_context::state_read_data(llama_io_read_i & io) {
         auto n_outputs = this->n_outputs;
         io.read_to(&n_outputs, sizeof(n_outputs));
 
-        if (n_outputs > output_reserve(n_outputs)) {
+        // Create a dummy batch for state loading.
+        llama_batch dummy_batch = {};
+        dummy_batch.n_tokens = 0;
+        if (n_outputs > output_reserve(n_outputs, dummy_batch)) {
             throw std::runtime_error("could not reserve outputs");
         }
 
@@ -1973,6 +2589,9 @@ size_t llama_context::state_read_data(llama_io_read_i & io) {
         }
     }
 
+    // TODO: handle sampling buffers and samplers state ?
+    //       https://github.com/ggml-org/llama.cpp/pull/17004
+
     if (memory) {
         LLAMA_LOG_DEBUG("%s: - reading memory module\n", __func__);
 
@@ -2029,15 +2648,26 @@ void llama_context::perf_reset() {
 
 std::map<ggml_backend_buffer_type_t, llama_memory_breakdown_data> llama_context::memory_breakdown() const {
     std::map<ggml_backend_buffer_type_t, llama_memory_breakdown_data> ret;
-    for (const auto & buft_size : model.memory_breakdown()) {
-        ret[buft_size.first].model += buft_size.second;
+    for (const auto & [buft, size] : model.memory_breakdown()) {
+        ret[buft].model += size;
     }
-    for (const auto & buft_size : memory->memory_breakdown()) {
-        ret[buft_size.first].context += buft_size.second;
+    if (memory) {
+        for (const auto & [buft, size] : memory->memory_breakdown()) {
+            ret[buft].context += size;
+        }
     }
-    for (const auto & backend_ptr : backends) {
-        ggml_backend_t backend = backend_ptr.get();
-        ret[ggml_backend_sched_get_buffer_type(sched.get(), backend)].compute += ggml_backend_sched_get_buffer_size(sched.get(), backend);
+    if (model.hparams.no_alloc) {
+        for (size_t i = 0; i < backends.size(); ++i) {
+            ggml_backend_t             backend = backends[i].get();
+            ggml_backend_buffer_type_t buft    = ggml_backend_sched_get_buffer_type(sched.get(), backend);
+            ret[buft].compute += backend_buf_exp_size[i];
+        }
+    } else {
+        for (const auto & backend_ptr : backends) {
+            ggml_backend_t             backend = backend_ptr.get();
+            ggml_backend_buffer_type_t buft    = ggml_backend_sched_get_buffer_type(sched.get(), backend);
+            ret[buft].compute += ggml_backend_sched_get_buffer_size(sched.get(), backend);
+        }
     }
     return ret;
 }
@@ -2130,7 +2760,7 @@ void llama_context::opt_epoch_iter(
             batch.logits  [pos_batch]    = true;
         }
 
-        if (!balloc->init(batch, model.vocab, nullptr, model.hparams.n_embd, cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, true)) {
+        if (!balloc->init(batch, model.vocab, nullptr, model.hparams.n_embd_inp(), cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, true)) {
             LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
             return;
         }
@@ -2150,7 +2780,7 @@ void llama_context::opt_epoch_iter(
         }
 
         // reserve output buffer
-        if (output_reserve(n_outputs_all) < n_outputs_all) {
+        if (output_reserve(n_outputs_all, balloc->get_batch()) < n_outputs_all) {
             LLAMA_LOG_ERROR("%s: could not reserve space for batch with %d outputs\n", __func__, n_outputs_all);
             GGML_ABORT("TODO: handle this error");
         };
@@ -2295,6 +2925,8 @@ llama_context_params llama_context_default_params() {
         /*.op_offload                  =*/ true,
         /*.swa_full                    =*/ true,
         /*.kv_unified                  =*/ false,
+        /*.sampler                     =*/ nullptr,
+        /*.n_sampler                   =*/ 0,
     };
 
     return result;
@@ -2346,6 +2978,13 @@ llama_context * llama_init_from_model(
         return nullptr;
     }
 
+    if (params.pooling_type != LLAMA_POOLING_TYPE_UNSPECIFIED &&
+        params.pooling_type != model->hparams.pooling_type) {
+        //user-specified pooling-type is different from the model default
+        LLAMA_LOG_WARN("%s: model default pooling_type is [%d], but [%d] was specified\n", __func__,
+                       model->hparams.pooling_type, params.pooling_type);
+    }
+
     try {
         auto * ctx = new llama_context(*model, params);
         return ctx;
@@ -2371,6 +3010,10 @@ uint32_t llama_n_ctx(const llama_context * ctx) {
     return ctx->n_ctx();
 }
 
+uint32_t llama_n_ctx_seq(const llama_context * ctx) {
+    return ctx->n_ctx_seq();
+}
+
 uint32_t llama_n_batch(const llama_context * ctx) {
     return ctx->n_batch();
 }
@@ -2443,7 +3086,15 @@ float * llama_get_logits(llama_context * ctx) {
 float * llama_get_logits_ith(llama_context * ctx, int32_t i) {
     ctx->synchronize();
 
-    return ctx->get_logits_ith(i);
+    float * res = nullptr;
+
+    res = ctx->get_sampled_logits_ith(i);
+
+    if (!res) {
+        res = ctx->get_logits_ith(i);
+    }
+
+    return res;
 }
 
 float * llama_get_embeddings(llama_context * ctx) {
@@ -2464,6 +3115,52 @@ float * llama_get_embeddings_seq(llama_context * ctx, llama_seq_id seq_id) {
     return ctx->get_embeddings_seq(seq_id);
 }
 
+bool llama_set_sampler(llama_context * ctx, llama_seq_id seq_id, llama_sampler * smpl) {
+    return ctx->set_sampler(seq_id, smpl);
+}
+
+llama_token llama_get_sampled_token_ith(llama_context * ctx, int32_t i) {
+    ctx->synchronize();
+
+    return ctx->get_sampled_token_ith(i);
+}
+
+float * llama_get_sampled_probs_ith(llama_context * ctx, int32_t i) {
+    ctx->synchronize();
+
+    return ctx->get_sampled_probs_ith(i);
+}
+
+float * llama_get_sampled_logits_ith(llama_context * ctx, int32_t i) {
+    ctx->synchronize();
+
+    return ctx->get_sampled_logits_ith(i);
+}
+
+llama_token * llama_get_sampled_candidates_ith(llama_context * ctx, int32_t i) {
+    ctx->synchronize();
+
+    return const_cast<llama_token *>(ctx->get_sampled_candidates_ith(i));
+}
+
+uint32_t llama_get_sampled_candidates_count_ith(llama_context * ctx, int32_t i) {
+    ctx->synchronize();
+
+    return static_cast<uint32_t>(ctx->get_sampled_candidates_count(i));
+}
+
+uint32_t llama_get_sampled_logits_count_ith(llama_context * ctx, int32_t i) {
+    ctx->synchronize();
+
+    return static_cast<uint32_t>(ctx->get_sampled_logits_count(i));
+}
+
+uint32_t llama_get_sampled_probs_count_ith(llama_context * ctx, int32_t i) {
+    ctx->synchronize();
+
+    return static_cast<uint32_t>(ctx->get_sampled_probs_count(i));
+}
+
 // llama adapter API
 
 int32_t llama_set_adapter_lora(