diff --git a/examples/talk-llama/llama-sampling.cpp b/examples/talk-llama/llama-sampling.cpp
index e255a8fc..25536eb6 100644
--- a/examples/talk-llama/llama-sampling.cpp
+++ b/examples/talk-llama/llama-sampling.cpp
@@ -63,6 +63,30 @@ static void llama_log_softmax(float * array, size_t size) {
 }
 */
 
+static void llama_sampler_temp_impl(llama_token_data_array * cur_p, float temp) {
+    if (temp <= 0.0f) {
+        // find the token with the highest logit and set the rest to -inf
+        size_t max_i = 0;
+        float  max_l = cur_p->data[0].logit;
+
+        for (size_t i = 1; i < cur_p->size; ++i) {
+            if (cur_p->data[i    ].logit > max_l) {
+                cur_p->data[max_i].logit = -INFINITY;
+                max_i = i;
+                max_l = cur_p->data[i].logit;
+            } else {
+                cur_p->data[i].logit = -INFINITY;
+            }
+        }
+
+        return;
+    }
+
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        cur_p->data[i].logit /= temp;
+    }
+}
+
 static void llama_sampler_softmax_impl(llama_token_data_array * cur_p) {
     GGML_ASSERT(cur_p->size > 0);
 
@@ -427,6 +451,9 @@ static const char * llama_sampler_dist_name(const struct llama_sampler * /*smpl*
 
 static void llama_sampler_dist_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
     auto * ctx = (llama_sampler_dist *) smpl->ctx;
+
+    llama_sampler_softmax_impl(cur_p);
+
     cur_p->selected = llama_sample_dist(cur_p, ctx->rng);
 }
 
@@ -912,9 +939,8 @@ static const char * llama_sampler_temp_name(const struct llama_sampler * /*smpl*
 
 static void llama_sampler_temp_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
     const auto * ctx = (llama_sampler_temp *) smpl->ctx;
-    for (size_t i = 0; i < cur_p->size; ++i) {
-        cur_p->data[i].logit /= ctx->temp;
-    }
+
+    llama_sampler_temp_impl(cur_p, ctx->temp);
 }
 
 static struct llama_sampler * llama_sampler_temp_clone(const struct llama_sampler * smpl) {
@@ -961,6 +987,7 @@ static void llama_sampler_temp_ext_apply(struct llama_sampler * smpl, llama_toke
     if (ctx->delta > 0) {
         const float min_temp = std::max(0.0f, ctx->temp - ctx->delta);
         const float max_temp = ctx->temp + ctx->delta;
+
         float exponent_val = ctx->exponent;
 
         // no need to do anything if there is only one (or zero) candidates
@@ -998,9 +1025,7 @@ static void llama_sampler_temp_ext_apply(struct llama_sampler * smpl, llama_toke
     #endif
 
         // Apply the dynamically calculated temperature scaling
-        for (size_t i = 0; i < cur_p->size; ++i) {
-            cur_p->data[i].logit /= dyn_temp;
-        }
+        llama_sampler_temp_impl(cur_p, dyn_temp);
 
         // Re-compute softmax probabilities after scaling logits with dynamic temperature
         const double max_l_double = cur_p->data[0].logit;
@@ -1024,9 +1049,7 @@ static void llama_sampler_temp_ext_apply(struct llama_sampler * smpl, llama_toke
         }
     #endif
     } else {
-        for (size_t i = 0; i < cur_p->size; ++i) {
-            cur_p->data[i].logit /= ctx->temp;
-        }
+        llama_sampler_temp_impl(cur_p, ctx->temp);
     }
 }
 
@@ -1059,6 +1082,101 @@ struct llama_sampler * llama_sampler_init_temp_ext(float temp, float delta, floa
     };
 }
 
+// xtc
+
+struct llama_sampler_xtc {
+    const float    probability;
+    const float    threshold;
+    const size_t   min_keep;
+
+    const uint32_t seed;
+    uint32_t       seed_cur;
+
+    std::mt19937   rng;
+};
+
+static const char * llama_sampler_xtc_name(const struct llama_sampler * /*smpl*/) {
+    return "xtc";
+}
+
+static void llama_sample_xtc_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    auto * ctx = (llama_sampler_xtc *) smpl->ctx;
+
+    if (ctx->probability <= 0.0f
+        || ctx->threshold > 0.5f
+        || cur_p->size < 2) {
+        return;
+    }
+
+    std::uniform_real_distribution<float> distribution(0.0f, 1.0f);
+    float chance = distribution(ctx->rng);
+    if (chance > ctx->probability) return;
+
+    // in case it's not sorted/recalculated yet
+    llama_sampler_softmax_impl(cur_p);
+
+    int pos_last = 0;
+
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        if (cur_p->data[i].p >= ctx->threshold) {
+            pos_last = i;
+        } else break;
+    }
+
+    if (cur_p->size - pos_last >= ctx->min_keep && pos_last > 0) {
+        cur_p->data += pos_last;
+        cur_p->size -= pos_last;
+    }
+}
+
+static struct llama_sampler * llama_sampler_xtc_clone(const struct llama_sampler * smpl) {
+    const auto * ctx = (const llama_sampler_xtc *) smpl->ctx;
+    auto * result = llama_sampler_init_xtc(ctx->probability, ctx->threshold, ctx->min_keep, ctx->seed);
+
+    // copy the state
+    {
+        auto * result_ctx = (llama_sampler_xtc *) result->ctx;
+
+        result_ctx->rng = ctx->rng;
+    }
+
+    return result;
+}
+
+static void llama_sampler_xtc_free(struct llama_sampler * smpl) {
+    delete (llama_sampler_xtc *) smpl->ctx;
+}
+
+static void llama_sampler_xtc_reset(struct llama_sampler * smpl) {
+    auto * ctx = (llama_sampler_xtc *) smpl->ctx;
+    ctx->seed_cur = get_rng_seed(ctx->seed);
+    ctx->rng.seed(ctx->seed_cur);
+}
+
+static struct llama_sampler_i llama_sampler_xtc_i = {
+    /* .name   = */ llama_sampler_xtc_name,
+    /* .accept = */ nullptr,
+    /* .apply  = */ llama_sample_xtc_apply,
+    /* .reset  = */ llama_sampler_xtc_reset,
+    /* .clone  = */ llama_sampler_xtc_clone,
+    /* .free   = */ llama_sampler_xtc_free,
+};
+
+struct llama_sampler * llama_sampler_init_xtc(float p, float t, size_t min_keep, uint32_t seed) {
+    auto seed_cur = get_rng_seed(seed);
+    return new llama_sampler {
+        /* .iface = */ &llama_sampler_xtc_i,
+        /* .ctx   = */ new llama_sampler_xtc {
+            /* .probability   = */ p,
+            /* .threshold     = */ t,
+            /* .min_keep      = */ min_keep,
+            /* .seed          = */ seed,
+            /* .seed_cur      = */ seed_cur,
+            /* .rng           = */ std::mt19937(seed_cur),
+        },
+    };
+}
+
 // mirostat
 
 struct llama_sampler_mirostat {
@@ -1565,6 +1683,397 @@ struct llama_sampler * llama_sampler_init_penalties(
     };
 }
 
+// DRY
+
+struct llama_sampler_dry {
+    int32_t total_context_size;
+
+    const float   dry_multiplier;
+    const float   dry_base;
+    const int32_t dry_allowed_length;
+    const int32_t dry_penalty_last_n;
+
+    std::unordered_multimap<llama_token, std::vector<llama_token>> dry_processed_breakers;
+    std::vector<int> dry_repeat_count;
+    std::unordered_map<llama_token, int> dry_max_token_repeat;
+    ring_buffer<llama_token> last_tokens;
+};
+
+// Ported from Koboldcpp, original PR: https://github.com/LostRuins/koboldcpp/pull/982 (Original author: pi6am)
+static void get_overlapping_token_sequences(const llama_vocab & vocab, const std::string& str, std::unordered_multimap<llama_token, std::vector<llama_token>>& token_sequences, int max_tail_len = -1) {
+    for (llama_token token_id = 0; token_id < (llama_token)vocab.n_vocab; token_id++) {
+        std::string word = llama_detokenize(vocab, {token_id}, true);
+        if (word.find(str) != std::string::npos) {
+            token_sequences.emplace(token_id, std::vector<llama_token>());
+        } else {
+            size_t word_len = word.size(), str_len = str.size();
+            size_t pos = -1;
+            while ((pos = word.find(str[0], pos + 1)) != std::string::npos) {
+                bool match = true;
+                size_t i;
+                for (i = 1; i < str_len && i + pos < word_len; ++i) {
+                    if (word[pos + i] != str[i]) {
+                        match = false;
+                        break;
+                    }
+                }
+                if (match) {
+                    std::vector<llama_token> tokenization = llama_tokenize_internal(vocab, str.substr(i), false, false);
+                    if (max_tail_len >= 0 && tokenization.size() > (size_t)max_tail_len) {
+                        tokenization.resize(max_tail_len);
+                    }
+
+                    // Ensure we don't already have a duplicate matching tokenization
+                    auto its = token_sequences.equal_range(token_id);
+                    bool found = false;
+                    for (auto it = its.first; it != its.second; ++it) {
+                        if (tokenization == it->second) {
+                            found = true;
+                            break;
+                        }
+                    }
+                    if (!found) {
+                        token_sequences.emplace(token_id, tokenization);
+                    }
+                }
+            }
+        }
+    }
+}
+
+static const char * llama_sampler_dry_name(const struct llama_sampler * /*smpl*/) {
+    return "dry";
+}
+
+static void llama_sampler_dry_accept(struct llama_sampler * smpl, llama_token token) {
+    auto * ctx = (llama_sampler_dry *) smpl->ctx;
+    if (ctx->dry_multiplier == 0.0f || ctx->dry_base < 1.0f || ctx->dry_penalty_last_n == 0) {
+        return;
+    }
+
+    ctx->last_tokens.push_back(token);
+}
+
+// Ported from Koboldcpp, original PR: https://github.com/LostRuins/koboldcpp/pull/982 (Original author: pi6am)
+static void llama_sampler_dry_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    auto * ctx = (llama_sampler_dry *) smpl->ctx;
+
+    if (ctx->dry_multiplier == 0.0f || ctx->dry_base < 1.0f || ctx->dry_penalty_last_n == 0) {
+        return;
+    }
+
+    int32_t effective_dry_penalty_last_n = (ctx->dry_penalty_last_n == -1) ? ctx->total_context_size : std::max(ctx->dry_penalty_last_n, 0);
+    int last_n_repeat = std::min(std::min((int)ctx->last_tokens.size(), effective_dry_penalty_last_n), ctx->total_context_size);
+
+    if (last_n_repeat <= ctx->dry_allowed_length) {
+        return;
+    }
+
+    ctx->dry_repeat_count.assign(last_n_repeat, 0);
+    ctx->dry_max_token_repeat.clear();
+
+    // Step 1: Look for restart sequences to limit the maximum repetition length.
+    // Work backwards through the context looking for any token that begins a restart sequence.
+    //
+    // The collection `restart_sequences` is a mapping from a "head" token to all "tail"
+    // sequences that together comprise a restart sequence. This allows us to quickly check
+    // whether each token is the head of a complete sequence. Most restart sequences are actually
+    // a single token, and for these the "tail" is an empty vector.
+    //
+    // If the token is a "head", test all restart sequences that begin with this token
+    // (there will often only be one sequence for each token, but if sequences like 'aaaq1' and
+    // 'aaa1' are used as restart strings, both could start with 'aaa' when tokenized). The
+    // longest matching sequence (if any) is used to limit the maximum repetition length.
+    //
+    // Note that in the case case of a short sequence contained in a longer one, this might fail to
+    // find the smallest value for `rep_limit`. For example, if 'amniotic' and 'ni' are both used as
+    // restart sequences, 'ni' will be found first, and since it's shorter it will fail to suppress
+    // 'otic'. This is a minor issue since fully contained restart sequences are likely to be rare.
+    //
+    // This is theoretically worst-case O(N^2) for arbitrary restart sequences, which is why we
+    // have already clamped the maximum tail sequence length when generating `restart_sequences`.
+    // With clamping, this scan is O(N) in the context length.
+
+    int rep_limit = last_n_repeat;
+    for (int i = 0; i < last_n_repeat; ++i) {
+        llama_token token = ctx->last_tokens.rat(i);
+        auto its = ctx->dry_processed_breakers.equal_range(token);
+        if (its.first == ctx->dry_processed_breakers.end()) {
+            continue;
+        }
+        int longest_match = -1;
+        for (auto it = its.first; it != its.second; ++it) {
+            // Note that (*it) does not contain the head character, so seq_len will be
+            // the restart sequence length minus 1.
+            // In the common case of a single-token restart sequence, (*it) will be empty
+            // and we will trivially match.
+            int seq_len = (int)it->second.size();
+            if (seq_len > longest_match && seq_len <= (int)i) {
+                bool match = true;
+                for (int offset = 0; offset < seq_len; ++offset) {
+                    // The -1 when indexing `last_tokens` is because we already matched the head.
+                    if (it->second[offset] != ctx->last_tokens.rat(i - offset - 1)) {
+                        match = false;
+                        break;
+                    }
+                }
+                if (match) {
+                    longest_match = seq_len;
+                }
+            }
+        }
+        if (longest_match >= 0) {
+            // We found a restart sequence starting `i` tokens from the end and continuing for
+            // `longest_match` tokens.
+            rep_limit = i - longest_match;
+            break;
+        }
+    }
+    if (rep_limit < ctx->dry_allowed_length) {
+        return;
+    }
+
+    // Step 2: Iterate in reverse over the last N tokens of the context, using the "Z-algorithm" (in
+    // the reverse direction) to efficiently compute the positions and lengths of suffixes appearing
+    // elsewhere in the context. We limit the suffix length to `rep_limit` to respect restart sequences.
+    //
+    // This algorithm is not currently documented on Wikipedia, but there is a clear description here:
+    // https://ivanyu.me/blog/2014/10/15/z-algorithm/
+    //
+    // The code below is adapted from the public domain implementation by the same author here:
+    // https://github.com/ivanyu/string-algorithms/blob/master/z_algorithm.py
+    //
+    // Example:
+    // Last N tokens: a b c c b c y a b c
+    // Repeat counts: 0 0 3 1 0 2 0 0 0 0
+    //                    ^
+    //   This `3` means that the last three tokens of the context (a b c) also appear here.
+    //
+    // This step is worst case O(N) since the Z-algorithm is linear, despite the appearance of nested
+    // for/while loops. This can be seen by observing that the `lt` and `rt` bounds are set after each
+    // repeated suffix is detected (i.e. after each while loop when n > 0). These bound variables
+    // ensure that the inner while loops only examine each token in the context once as the outer
+    // for loop iterates over the context.
+
+    {
+        const int last = last_n_repeat - 1;
+        int rt = 0, lt = 0;
+
+        for (int k = 1; k < last_n_repeat; ++k) {
+            if (k > rt) {
+                // If k is outside the current Z-box, do naive computation.
+                int n = 0;
+                while (n + k < last_n_repeat && ctx->last_tokens.rat(n) == ctx->last_tokens.rat(n+k)) {
+                    ++n;
+                }
+                ctx->dry_repeat_count[last - k] = std::min(n, rep_limit);
+                if (n > 0) {
+                    lt = k;
+                    rt = k+n-1;
+                }
+            } else {
+                // If k is inside the current Z-box, consider two cases.
+
+                int p = k - lt; // Pair index.
+                int right_part_len = rt - k + 1;
+
+                if (ctx->dry_repeat_count[last - p] < right_part_len) {
+                    int n = std::min(ctx->dry_repeat_count[last - p], rep_limit);
+                    ctx->dry_repeat_count[last - k] = n;
+                } else {
+                    int i = rt + 1;
+                    while (i < last_n_repeat && ctx->last_tokens.rat(i) == ctx->last_tokens.rat(i - k)) {
+                        i += 1;
+                    }
+
+                    int n = std::min(i - k, rep_limit);
+                    ctx->dry_repeat_count[last - k] = n;
+                    lt = k;
+                    rt = i - 1;
+                }
+            }
+        }
+    }
+
+    // Step 3: Iterate over dry_repeat_count and last_tokens, examining the maximum repeat length
+    // that would be generated by emitting each new token that would extend a sequence.
+    //
+    // Following the same example as above:
+    // Last N tokens: a b c c b c y a b c
+    // Repeat counts: 0 0 3 1 0 2 0 0 0 0
+    //
+    // For each non-zero, look ahead one token. This token, if emitted, would extend the repetition.
+    // c: 3 -> 4 (from `a b c` to `a b c c`)
+    // b: 1 -> 2 (from `c` to `c b`)
+    // y: 2 -> 3 (from `b c` to `b c y`)
+
+    for (int i = 0; i < last_n_repeat - 1; ++i) {
+        int repeat_len = ctx->dry_repeat_count[i];
+        if (repeat_len >= ctx->dry_allowed_length) {
+            // This token ends a repeat, so the next token would continue one.
+            // By convention, the value of `repeat_len` only includes the tokens currently
+            // in the context, not the new token that would be added.
+            llama_token token = ctx->last_tokens.rat(last_n_repeat - 2 - i);
+            // Track the maximum sequence ending in this token.
+            const auto& it = ctx->dry_max_token_repeat.find(token);
+            if (it == ctx->dry_max_token_repeat.end() || it->second < repeat_len) {
+                ctx->dry_max_token_repeat[token] = repeat_len;
+            }
+        }
+    }
+
+    // Step 4: Apply logit penalties based on the maximum repeat length for relevant tokens.
+
+    // Prevent floating point overflow in `pow(penalty_base, exponent)` by clamping to `max_exponent`.
+    // Compute it from `penalty_base` and the approximate log of `std::numeric_limits<float>::max()`
+    const float FLOAT_MAX_LOG = 88.7228391f;
+    int max_exponent = 0;
+    if (ctx->dry_base > 1.000001f) {
+        max_exponent = FLOAT_MAX_LOG / std::log(ctx->dry_base);
+    }
+
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        const auto& af_kvp = ctx->dry_max_token_repeat.find(cur_p->data[i].id);
+        if (af_kvp != ctx->dry_max_token_repeat.end()) {
+            // Check all sequence breakers starting with this token
+            auto range = ctx->dry_processed_breakers.equal_range(cur_p->data[i].id);
+            bool is_single_token_breaker = false;
+
+            for (auto it = range.first; it != range.second; ++it) {
+                if (it->second.empty()) {
+                    is_single_token_breaker = true;
+                    break;
+                }
+            }
+
+            // Apply penalty only if it's not a single-token sequence breaker
+            if (!is_single_token_breaker) {
+                int repeat_exp = af_kvp->second - ctx->dry_allowed_length;
+                if (max_exponent > 0 && repeat_exp > max_exponent) {
+                    repeat_exp = max_exponent;
+                }
+                float penalty = ctx->dry_multiplier * std::pow(ctx->dry_base, repeat_exp);
+                cur_p->data[i].logit -= penalty;
+            }
+        }
+    }
+
+    cur_p->sorted = false;
+}
+
+static void llama_sampler_dry_reset(struct llama_sampler * smpl) {
+    auto * ctx = (llama_sampler_dry *) smpl->ctx;
+    ctx->last_tokens.clear();
+    ctx->dry_repeat_count.clear();
+    ctx->dry_max_token_repeat.clear();
+}
+
+static struct llama_sampler * llama_sampler_dry_clone(const struct llama_sampler * smpl) {
+    const auto * ctx = (llama_sampler_dry *) smpl->ctx;
+
+    // nullptr is passed as vocab because it is only needed for raw sequence breaker processing, which we have already done and will be copying
+    auto * result = llama_sampler_init_dry(nullptr, ctx->dry_multiplier, ctx->dry_base, ctx->dry_allowed_length, ctx->dry_penalty_last_n, NULL, 0);
+    // Copy the state, including the processed breakers
+    {
+        auto * result_ctx = (llama_sampler_dry *) result->ctx;
+        result_ctx->dry_processed_breakers = ctx->dry_processed_breakers;
+        result_ctx->dry_repeat_count = ctx->dry_repeat_count;
+        result_ctx->dry_max_token_repeat = ctx->dry_max_token_repeat;
+        result_ctx->last_tokens = ctx->last_tokens;
+    }
+
+    return result;
+}
+
+static void llama_sampler_dry_free(struct llama_sampler * smpl) {
+    delete (llama_sampler_dry *) smpl->ctx;
+}
+
+static struct llama_sampler_i llama_sampler_dry_i = {
+    /* .name   = */ llama_sampler_dry_name,
+    /* .accept = */ llama_sampler_dry_accept,
+    /* .apply  = */ llama_sampler_dry_apply,
+    /* .reset  = */ llama_sampler_dry_reset,
+    /* .clone  = */ llama_sampler_dry_clone,
+    /* .free   = */ llama_sampler_dry_free,
+};
+
+struct llama_sampler * llama_sampler_init_dry_impl(const struct llama_vocab & vocab, int32_t context_size, float dry_multiplier, float dry_base, int32_t dry_allowed_length, int32_t dry_penalty_last_n, const char** seq_breakers, size_t num_breakers) {
+    int32_t effective_dry_penalty_last_n = (dry_penalty_last_n == -1) ? context_size : std::max(dry_penalty_last_n, 0);
+    std::unordered_multimap<llama_token, std::vector<llama_token>> processed_breakers;
+    const int MAX_CHAR_LEN = 40;
+    const int MAX_SEQ_LEN = 20;
+
+    const bool dry_enabled = (dry_multiplier != 0.0f && dry_base >= 1.0f && dry_penalty_last_n != 0);
+
+    if (dry_enabled && seq_breakers != nullptr && num_breakers > 0) {
+        // Process sequence breakers
+        for (size_t i = 0; i < num_breakers; ++i) {
+            if (seq_breakers[i] == nullptr || std::strlen(seq_breakers[i]) == 0) {
+                LLAMA_LOG_WARN("skipping null or empty DRY sequence breaker at index %zu\n", i);
+                continue;
+            }
+
+            std::string sequence_break(seq_breakers[i]);
+            if (sequence_break.empty()) {
+                LLAMA_LOG_WARN("skipping empty DRY sequence breaker\n");
+                continue;
+            }
+
+            if (sequence_break.size() > MAX_CHAR_LEN) {
+                LLAMA_LOG_WARN("truncating DRY sequence breaker to %d characters\n", MAX_CHAR_LEN);
+                sequence_break.resize(MAX_CHAR_LEN);
+            }
+
+            get_overlapping_token_sequences(vocab, sequence_break, processed_breakers, MAX_SEQ_LEN);
+        }
+    }
+
+    return new llama_sampler {
+        /* .iface = */ &llama_sampler_dry_i,
+        /* .ctx   = */ new llama_sampler_dry {
+            /* .total_context_size     = */ context_size,
+            /* .dry_multiplier         = */ dry_multiplier,
+            /* .dry_base               = */ dry_base,
+            /* .dry_allowed_length     = */ dry_allowed_length,
+            /* .dry_penalty_last_n     = */ dry_penalty_last_n,
+            /* .dry_processed_breakers = */ std::move(processed_breakers),
+            /* .dry_repeat_count       = */ dry_enabled ? std::vector<int>(effective_dry_penalty_last_n, 0) : std::vector<int>{},
+            /* .dry_max_token_repeat   = */ {},
+            /* .last_tokens            = */ dry_enabled ? ring_buffer<llama_token>(effective_dry_penalty_last_n) : ring_buffer<llama_token>(0),
+        },
+    };
+}
+
+// wrapper for test-sampling.cpp
+struct llama_sampler * llama_sampler_init_dry_testing(int32_t context_size, float dry_multiplier, float dry_base, int32_t dry_allowed_length, int32_t dry_penalty_last_n, const std::vector<std::vector<llama_token>>& seq_breakers) {
+    llama_vocab dummy_vocab;
+    auto * result = llama_sampler_init_dry_impl(dummy_vocab, context_size, dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n, NULL, 0);
+    auto * ctx = (llama_sampler_dry *) result->ctx;
+
+    // Process the token-based sequence breakers
+    ctx->dry_processed_breakers.clear();
+    if (seq_breakers.empty()) {
+        LLAMA_LOG_WARN("empty DRY sequence breakers list in llama_sampler_init_dry_testing\n");
+    } else {
+        for (const auto& breaker : seq_breakers) {
+            if (breaker.empty()) {
+                LLAMA_LOG_WARN("skipping DRY empty sequence breaker\n");
+                continue;
+            }
+            llama_token head_token = breaker[0];
+            std::vector<llama_token> tail_tokens(breaker.begin() + 1, breaker.end());
+            ctx->dry_processed_breakers.emplace(head_token, std::move(tail_tokens));
+        }
+
+        if (ctx->dry_processed_breakers.empty()) {
+            LLAMA_LOG_WARN("no valid DRY sequence breakers processed in llama_sampler_init_dry_testing\n");
+        }
+    }
+
+    return result;
+}
+
 // logit-bias
 
 struct llama_sampler_logit_bias {
@@ -1644,6 +2153,229 @@ struct llama_sampler * llama_sampler_init_logit_bias(
     };
 }
 
+// infill
+
+//#define GGML_DEBUG_SAMPLER_INFILL
+
+struct llama_sampler_infill {
+    const struct llama_vocab * vocab;
+
+    std::vector<char> buf0;
+    std::vector<char> buf1;
+};
+
+static const char * llama_sampler_infill_name(const struct llama_sampler * /*smpl*/) {
+    return "infill";
+}
+
+static void llama_sampler_infill_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    auto * ctx = (llama_sampler_infill *) smpl->ctx;
+
+    llama_sampler_softmax_impl(cur_p);
+
+#if defined(GGML_DEBUG_SAMPLER_INFILL)
+#define LOG_DBG_CUR LLAMA_LOG_DEBUG
+#else
+#define LOG_DBG_CUR(...)
+#endif
+
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        LOG_DBG_CUR("%s: cur_p[%3zu] = { id: %6d, p: %.6f, logit: %6.3f }\n", __func__, i, cur_p->data[i].id, cur_p->data[i].p, cur_p->data[i].logit);
+    }
+
+    float p_txt_sum = 0.0f;
+    float p_eog_sum = 0.0f;
+
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        if (llama_token_is_eog_impl(*ctx->vocab, cur_p->data[i].id)) {
+            p_eog_sum += cur_p->data[i].p;
+        } else {
+            p_txt_sum += cur_p->data[i].p;
+        }
+    }
+
+    const float rat = p_eog_sum == 0.0 ? INFINITY : p_txt_sum / p_eog_sum; GGML_UNUSED(rat);
+
+    LOG_DBG_CUR("%s: p_txt_sum = %.2f, p_eog_sum = %.2f, rat = %.2f, n = %zu\n", __func__, p_txt_sum, p_eog_sum, rat, cur_p->size);
+
+    if (3*p_eog_sum*cur_p->size > p_txt_sum) {
+        LOG_DBG_CUR("%s: the ratio p_txt/p_eog = %.2f is too low -> sampling EOG\n", __func__, p_txt_sum/p_eog_sum);
+
+        // keep just the EOG tokens
+        const auto size_org = cur_p->size;
+
+        cur_p->size = 0;
+
+        float p_sum = 0.0f;
+
+        for (size_t i = 0; i < size_org; ++i) {
+            if (llama_token_is_eog_impl(*ctx->vocab, cur_p->data[i].id)) {
+                p_sum += cur_p->data[i].p;
+
+                cur_p->data[cur_p->size++] = cur_p->data[i];
+            }
+        }
+
+        // normalize probs
+        for (size_t i = 0; i < cur_p->size; ++i) {
+            cur_p->data[i].p /= p_sum;
+        }
+
+        return;
+    }
+
+    size_t n_combined = 0; GGML_UNUSED(n_combined);
+
+    // combine tokens with common prefix
+    for (size_t i0 = 0; i0 < cur_p->size; ++i0) {
+        for (size_t i1 = 0; i1 < cur_p->size; ++i1) {
+            if (cur_p->data[i0].logit == -INFINITY) {
+                break;
+            }
+
+            if (i0 == i1 || cur_p->data[i1].logit == -INFINITY) {
+                continue;
+            }
+
+            int len0 = llama_token_to_piece_impl(*ctx->vocab, cur_p->data[i0].id, ctx->buf0.data(), ctx->buf0.size(), 0, false);
+            if (len0 < 0) {
+                ctx->buf0.resize(len0);
+                len0 = llama_token_to_piece_impl(*ctx->vocab, cur_p->data[i0].id, ctx->buf0.data(), ctx->buf0.size(), 0, false);
+                assert(len0 > 0);
+            }
+
+            int len1 = llama_token_to_piece_impl(*ctx->vocab, cur_p->data[i1].id, ctx->buf1.data(), ctx->buf1.size(), 0, false);
+            if (len1 < 0) {
+                ctx->buf1.resize(len1);
+                len1 = llama_token_to_piece_impl(*ctx->vocab, cur_p->data[i1].id, ctx->buf1.data(), ctx->buf1.size(), 0, false);
+                assert(len1 > 0);
+            }
+
+            // token i0 is a prefix of token i1
+            if (len0 > 0 && len0 <= len1 && memcmp(ctx->buf0.data(), ctx->buf1.data(), len0) == 0) {
+                int dst = i0;
+                int src = i1;
+
+                // merge into the token with higher probability
+                if (cur_p->data[i1].p > cur_p->data[i0].p) {
+                    std::swap(dst, src);
+                }
+
+                cur_p->data[dst].p += cur_p->data[src].p;
+                cur_p->data[src].logit = -INFINITY;
+                cur_p->data[src].p     = 0.0f;
+
+                n_combined++;
+            }
+        }
+    }
+
+    size_t n_non_eog = 0;
+
+    size_t size_org = cur_p->size;
+
+    float p_sum = 0.0f;
+    float thold = 0.2f;
+
+    cur_p->size = 0;
+
+    LOG_DBG_CUR("%s: n_combined = %zu, applying thold = %.3f\n", __func__, n_combined, thold);
+
+    for (size_t i = 0; i < size_org; ++i) {
+        const bool is_eog = llama_token_is_eog_impl(*ctx->vocab, cur_p->data[i].id);
+
+        if (cur_p->data[i].p < thold && !is_eog) {
+            continue;
+        }
+
+        if (!is_eog) {
+            ++n_non_eog;
+        }
+
+        p_sum += cur_p->data[i].p;
+
+        // keep this token
+        cur_p->data[cur_p->size++] = cur_p->data[i];
+    }
+
+    LOG_DBG_CUR("%s: n_non_eog = %zu\n", __func__, n_non_eog);
+
+    // if no non-EOG tokens are left -> reduce cur_p to single EOT token
+    if (n_non_eog == 0) {
+        cur_p->size = 1;
+        cur_p->data[0].id = llama_token_eot_impl(*ctx->vocab);
+        cur_p->data[0].logit = 1.0f;
+
+        return;
+    }
+
+    // normalize probs
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        cur_p->data[i].p /= p_sum;
+
+        LOG_DBG_CUR("%s: cur_p[%3zu] = { id: %6d, p: %.6f, logit: %6.3f }\n", __func__, i, cur_p->data[i].id, cur_p->data[i].p, cur_p->data[i].logit);
+    }
+
+    size_org = cur_p->size;
+    p_sum = 0.0f;
+    thold = 1.0/(n_non_eog + 1);
+
+    cur_p->size = 0;
+
+    LOG_DBG_CUR("%s: applying thold = %.3f\n", __func__, thold);
+
+    for (size_t i = 0; i < size_org; ++i) {
+        const bool is_eog = llama_token_is_eog_impl(*ctx->vocab, cur_p->data[i].id);
+
+        if (cur_p->data[i].p < thold && !is_eog) {
+            continue;
+        }
+
+        p_sum += cur_p->data[i].p;
+
+        cur_p->data[cur_p->size++] = cur_p->data[i];
+    }
+
+    // normalize probs
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        cur_p->data[i].p /= p_sum;
+
+        LOG_DBG_CUR("%s: cur_p[%3zu] = { id: %6d, p: %.6f, logit: %6.3f }\n", __func__, i, cur_p->data[i].id, cur_p->data[i].p, cur_p->data[i].logit);
+    }
+
+#undef LOG_DBG_CUR
+}
+
+static struct llama_sampler * llama_sampler_infill_clone(const struct llama_sampler * smpl) {
+    const auto * ctx = (const llama_sampler_infill *) smpl->ctx;
+    return llama_sampler_init_infill_impl(*ctx->vocab);
+}
+
+static void llama_sampler_infill_free(struct llama_sampler * smpl) {
+    delete (llama_sampler_infill *) smpl->ctx;
+}
+
+static struct llama_sampler_i llama_sampler_infill_i = {
+    /* .name   = */ llama_sampler_infill_name,
+    /* .accept = */ nullptr,
+    /* .apply  = */ llama_sampler_infill_apply,
+    /* .reset  = */ nullptr,
+    /* .clone  = */ llama_sampler_infill_clone,
+    /* .free   = */ llama_sampler_infill_free,
+};
+
+struct llama_sampler * llama_sampler_init_infill_impl(
+        const struct llama_vocab & vocab) {
+    return new llama_sampler {
+        /* .iface = */ &llama_sampler_infill_i,
+        /* .ctx   = */ new llama_sampler_infill {
+            /* .vocab = */ &vocab,
+            /* .buf0 = */ std::vector<char>(512),
+            /* .buf1 = */ std::vector<char>(512),
+        },
+    };
+}
+
 // utils
 
 uint32_t llama_sampler_get_seed(const struct llama_sampler * smpl) {
diff --git a/examples/talk-llama/llama-sampling.h b/examples/talk-llama/llama-sampling.h
index d90b1471..919f6fdf 100644
--- a/examples/talk-llama/llama-sampling.h
+++ b/examples/talk-llama/llama-sampling.h
@@ -4,8 +4,6 @@
 
 #include "llama-grammar.h"
 
-#include <unordered_map>
-
 struct llama_vocab;
 struct llama_grammar;
 
@@ -27,3 +25,24 @@ struct llama_sampler * llama_sampler_init_grammar_impl(
         const struct llama_vocab & vocab,
                       const char * grammar_str,
                       const char * grammar_root);
+
+struct llama_sampler * llama_sampler_init_infill_impl(
+        const struct llama_vocab & vocab);
+
+struct llama_sampler * llama_sampler_init_dry_impl(
+        const struct llama_vocab &  vocab,
+                         int32_t    context_size,
+                           float    dry_multiplier,
+                           float    dry_base,
+                         int32_t    dry_allowed_length,
+                         int32_t    dry_penalty_last_n,
+                      const char ** seq_breakers,
+                          size_t    num_breakers);
+
+struct llama_sampler * llama_sampler_init_dry_testing(
+                         int32_t   context_size,
+                           float   dry_multiplier,
+                           float   dry_base,
+                         int32_t   dry_allowed_length,
+                         int32_t   dry_penalty_last_n,
+  const std::vector<std::vector<llama_token>>& seq_breakers);
diff --git a/examples/talk-llama/llama-vocab.cpp b/examples/talk-llama/llama-vocab.cpp
index d2f34ddd..d1dc9627 100644
--- a/examples/talk-llama/llama-vocab.cpp
+++ b/examples/talk-llama/llama-vocab.cpp
@@ -221,7 +221,7 @@ struct llm_tokenizer_spm_session {
         }
 
         // seed the work queue with all possible 2-character tokens.
-        for (size_t i = 1; i < symbols.size(); ++i) {
+        for (int i = 1; i < (int) symbols.size(); ++i) {
             try_add_bigram(i - 1, i);
         }
 
@@ -563,7 +563,7 @@ struct llm_tokenizer_bpe_session {
                 index++;
                 symbols.emplace_back(sym);
             }
-            for (size_t i = 1; i < symbols.size(); ++i) {
+            for (int i = 1; i < (int) symbols.size(); ++i) {
                 add_new_bigram(i - 1, i);
             }
 
@@ -1663,6 +1663,14 @@ llama_token llama_token_eos_impl(const struct llama_vocab & vocab) {
     return vocab.special_eos_id;
 }
 
+llama_token llama_token_eot_impl(const struct llama_vocab & vocab) {
+    return vocab.special_eot_id;
+}
+
+llama_token llama_token_eom_impl(const struct llama_vocab & vocab) {
+    return vocab.special_eom_id;
+}
+
 llama_token llama_token_cls_impl(const struct llama_vocab & vocab) {
     return vocab.special_cls_id;
 }
@@ -1688,23 +1696,39 @@ bool llama_add_eos_token_impl(const struct llama_vocab & vocab) {
 }
 
 llama_token llama_token_prefix_impl(const struct llama_vocab & vocab) {
-    return vocab.special_prefix_id;
+    return vocab.special_fim_pre_id;
 }
 
 llama_token llama_token_middle_impl(const struct llama_vocab & vocab) {
-    return vocab.special_middle_id;
+    return vocab.special_fim_mid_id;
 }
 
 llama_token llama_token_suffix_impl(const struct llama_vocab & vocab) {
-    return vocab.special_suffix_id;
+    return vocab.special_fim_suf_id;
 }
 
-llama_token llama_token_eot_impl(const struct llama_vocab & vocab) {
-    return vocab.special_eot_id;
+llama_token llama_token_fim_pre_impl(const struct llama_vocab & vocab) {
+    return vocab.special_fim_pre_id;
 }
 
-llama_token llama_token_eom_impl(const struct llama_vocab & vocab) {
-    return vocab.special_eom_id;
+llama_token llama_token_fim_suf_impl(const struct llama_vocab & vocab) {
+    return vocab.special_fim_suf_id;
+}
+
+llama_token llama_token_fim_mid_impl(const struct llama_vocab & vocab) {
+    return vocab.special_fim_mid_id;
+}
+
+llama_token llama_token_fim_pad_impl(const struct llama_vocab & vocab) {
+    return vocab.special_fim_pad_id;
+}
+
+llama_token llama_token_fim_rep_impl(const struct llama_vocab & vocab) {
+    return vocab.special_fim_rep_id;
+}
+
+llama_token llama_token_fim_sep_impl(const struct llama_vocab & vocab) {
+    return vocab.special_fim_sep_id;
 }
 
 int32_t llama_tokenize_impl(
@@ -1942,3 +1966,19 @@ int32_t llama_detokenize_impl(
 
     return total <= text_len_max ? total : -total;
 }
+
+std::string llama_detokenize(const struct llama_vocab & vocab, const std::vector<llama_token> & tokens, bool special) {
+    std::string text;
+    text.resize(std::max(text.capacity(), tokens.size()));
+    int32_t n_chars = llama_detokenize_impl(vocab, tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
+    if (n_chars < 0) {
+        text.resize(-n_chars);
+        n_chars = llama_detokenize_impl(vocab, tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
+        GGML_ASSERT(n_chars <= (int32_t)text.size());  // whitespace trimming is performed after per-token detokenization
+    }
+
+    text.resize(n_chars);
+
+    // NOTE: the original tokenizer decodes bytes after collecting the pieces.
+    return text;
+}
diff --git a/examples/talk-llama/llama-vocab.h b/examples/talk-llama/llama-vocab.h
index 069bdc42..4bb16d2e 100644
--- a/examples/talk-llama/llama-vocab.h
+++ b/examples/talk-llama/llama-vocab.h
@@ -37,20 +37,26 @@ struct llama_vocab {
     std::map<std::pair<std::string, std::string>, int> bpe_ranks;
 
     // default LLaMA special tokens
+    // TODO: should we set all of these to LLAMA_TOKEN_NULL?
     id special_bos_id  = 1;
     id special_eos_id  = 2;
+    id special_eot_id  = LLAMA_TOKEN_NULL;
+    id special_eom_id  = LLAMA_TOKEN_NULL;
     id special_unk_id  = 0;
-    id special_sep_id  = -1;
-    id special_pad_id  = -1;
-    id special_cls_id  = -1;
-    id special_mask_id = -1;
+    id special_sep_id  = LLAMA_TOKEN_NULL;
+    id special_pad_id  = LLAMA_TOKEN_NULL;
+    id special_cls_id  = LLAMA_TOKEN_NULL;
+    id special_mask_id = LLAMA_TOKEN_NULL;
 
-    id linefeed_id       = 13;
-    id special_prefix_id = -1;
-    id special_suffix_id = -1;
-    id special_middle_id = -1;
-    id special_eot_id    = -1; // TODO: move above after "eos_id", and here add "file separator" token
-    id special_eom_id    = -1;
+    id linefeed_id = 13;
+
+    // fim tokens
+    id special_fim_pre_id = LLAMA_TOKEN_NULL;
+    id special_fim_suf_id = LLAMA_TOKEN_NULL;
+    id special_fim_mid_id = LLAMA_TOKEN_NULL;
+    id special_fim_pad_id = LLAMA_TOKEN_NULL;
+    id special_fim_rep_id = LLAMA_TOKEN_NULL; // repo
+    id special_fim_sep_id = LLAMA_TOKEN_NULL; // file separator
 
     // set of all tokens that cause "end of generation"
     std::set<id> special_eog_ids;
@@ -104,19 +110,26 @@ bool llama_token_is_control_impl(const struct llama_vocab & vocab, llama_token t
 
 llama_token llama_token_bos_impl(const struct llama_vocab & vocab);
 llama_token llama_token_eos_impl(const struct llama_vocab & vocab);
+llama_token llama_token_eot_impl(const struct llama_vocab & vocab);
+llama_token llama_token_eom_impl(const struct llama_vocab & vocab);
 llama_token llama_token_cls_impl(const struct llama_vocab & vocab);
 llama_token llama_token_sep_impl(const struct llama_vocab & vocab);
 llama_token llama_token_nl_impl (const struct llama_vocab & vocab);
 llama_token llama_token_pad_impl(const struct llama_vocab & vocab);
 
-bool llama_add_bos_token_impl(const struct llama_vocab & vocab);
-bool llama_add_eos_token_impl(const struct llama_vocab & vocab);
-
 llama_token llama_token_prefix_impl(const struct llama_vocab & vocab);
 llama_token llama_token_middle_impl(const struct llama_vocab & vocab);
 llama_token llama_token_suffix_impl(const struct llama_vocab & vocab);
-llama_token llama_token_eot_impl   (const struct llama_vocab & vocab);
-llama_token llama_token_eom_impl   (const struct llama_vocab & vocab);
+
+llama_token llama_token_fim_pre_impl(const struct llama_vocab & vocab);
+llama_token llama_token_fim_suf_impl(const struct llama_vocab & vocab);
+llama_token llama_token_fim_mid_impl(const struct llama_vocab & vocab);
+llama_token llama_token_fim_pad_impl(const struct llama_vocab & vocab);
+llama_token llama_token_fim_rep_impl(const struct llama_vocab & vocab);
+llama_token llama_token_fim_sep_impl(const struct llama_vocab & vocab);
+
+bool llama_add_bos_token_impl(const struct llama_vocab & vocab);
+bool llama_add_eos_token_impl(const struct llama_vocab & vocab);
 
 int32_t llama_tokenize_impl(
         const struct llama_vocab & vocab,
@@ -136,6 +149,12 @@ int32_t llama_token_to_piece_impl(
                          int32_t   lstrip,
                             bool   special);
 
+// check if token0 is contained as a prefix in token1
+bool llama_token_is_prefix_impl(
+        const struct llama_vocab & vocab,
+                     llama_token   token0,
+                     llama_token   token1);
+
 int32_t llama_detokenize_impl(
         const struct llama_vocab & vocab,
                const llama_token * tokens,
@@ -144,3 +163,8 @@ int32_t llama_detokenize_impl(
                          int32_t   text_len_max,
                             bool   remove_special,
                             bool   unparse_special);
+
+std::string llama_detokenize(
+        const struct llama_vocab & vocab,
+  const std::vector<llama_token> & tokens,
+                            bool   special);
diff --git a/examples/talk-llama/llama.cpp b/examples/talk-llama/llama.cpp
index 3443b068..53979e83 100644
--- a/examples/talk-llama/llama.cpp
+++ b/examples/talk-llama/llama.cpp
@@ -8,26 +8,16 @@
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
 
-#ifdef GGML_USE_RPC
-#  include "ggml-rpc.h"
-#endif
-
-#if defined(GGML_USE_VULKAN)
-#  include "ggml-vulkan.h"
-#elif defined(GGML_USE_SYCL)
-#  include "ggml-sycl.h"
-#elif defined(GGML_USE_KOMPUTE)
+#if defined(GGML_USE_KOMPUTE)
 #   include "ggml-kompute.h"
-#elif defined(GGML_USE_CANN)
-#   include "ggml-cann.h"
 #endif
 
-#ifdef GGML_USE_BLAS
-#  include "ggml-blas.h"
+#ifndef __AMX_INT8__
+#undef GGML_USE_AMX
 #endif
 
-#ifdef GGML_USE_METAL
-#  include "ggml-metal.h"
+#ifdef GGML_USE_AMX
+#  include "ggml-amx.h"
 #endif
 
 // TODO: replace with ggml API call
@@ -357,6 +347,8 @@ enum llm_kv {
     LLM_KV_TOKENIZER_MERGES,
     LLM_KV_TOKENIZER_BOS_ID,
     LLM_KV_TOKENIZER_EOS_ID,
+    LLM_KV_TOKENIZER_EOT_ID,
+    LLM_KV_TOKENIZER_EOM_ID,
     LLM_KV_TOKENIZER_UNK_ID,
     LLM_KV_TOKENIZER_SEP_ID,
     LLM_KV_TOKENIZER_PAD_ID,
@@ -369,14 +361,20 @@ enum llm_kv {
     LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP,
     LLM_KV_TOKENIZER_HF_JSON,
     LLM_KV_TOKENIZER_RWKV,
-    LLM_KV_TOKENIZER_PREFIX_ID,
-    LLM_KV_TOKENIZER_SUFFIX_ID,
-    LLM_KV_TOKENIZER_MIDDLE_ID,
-    LLM_KV_TOKENIZER_EOT_ID,
-    LLM_KV_TOKENIZER_EOM_ID,
+    LLM_KV_TOKENIZER_FIM_PRE_ID,
+    LLM_KV_TOKENIZER_FIM_SUF_ID,
+    LLM_KV_TOKENIZER_FIM_MID_ID,
+    LLM_KV_TOKENIZER_FIM_PAD_ID,
+    LLM_KV_TOKENIZER_FIM_REP_ID,
+    LLM_KV_TOKENIZER_FIM_SEP_ID,
 
     LLM_KV_ADAPTER_TYPE,
     LLM_KV_ADAPTER_LORA_ALPHA,
+
+    // deprecated:
+    LLM_KV_TOKENIZER_PREFIX_ID,
+    LLM_KV_TOKENIZER_SUFFIX_ID,
+    LLM_KV_TOKENIZER_MIDDLE_ID,
 };
 
 static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
@@ -434,57 +432,65 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ATTENTION_SLIDING_WINDOW,         "%s.attention.sliding_window"         },
     { LLM_KV_ATTENTION_SCALE,                  "%s.attention.scale"                  },
 
-    { LLM_KV_ROPE_DIMENSION_COUNT,          "%s.rope.dimension_count"                 },
-    { LLM_KV_ROPE_FREQ_BASE,                "%s.rope.freq_base"                       },
-    { LLM_KV_ROPE_SCALE_LINEAR,             "%s.rope.scale_linear"                    },
-    { LLM_KV_ROPE_SCALING_TYPE,             "%s.rope.scaling.type"                    },
-    { LLM_KV_ROPE_SCALING_FACTOR,           "%s.rope.scaling.factor"                  },
-    { LLM_KV_ROPE_SCALING_ATTN_FACTOR,      "%s.rope.scaling.attn_factor"             },
-    { LLM_KV_ROPE_SCALING_ORIG_CTX_LEN,     "%s.rope.scaling.original_context_length" },
-    { LLM_KV_ROPE_SCALING_FINETUNED,        "%s.rope.scaling.finetuned"               },
-    { LLM_KV_ROPE_SCALING_YARN_LOG_MUL,     "%s.rope.scaling.yarn_log_multiplier"     },
+    { LLM_KV_ROPE_DIMENSION_COUNT,             "%s.rope.dimension_count"                 },
+    { LLM_KV_ROPE_FREQ_BASE,                   "%s.rope.freq_base"                       },
+    { LLM_KV_ROPE_SCALE_LINEAR,                "%s.rope.scale_linear"                    },
+    { LLM_KV_ROPE_SCALING_TYPE,                "%s.rope.scaling.type"                    },
+    { LLM_KV_ROPE_SCALING_FACTOR,              "%s.rope.scaling.factor"                  },
+    { LLM_KV_ROPE_SCALING_ATTN_FACTOR,         "%s.rope.scaling.attn_factor"             },
+    { LLM_KV_ROPE_SCALING_ORIG_CTX_LEN,        "%s.rope.scaling.original_context_length" },
+    { LLM_KV_ROPE_SCALING_FINETUNED,           "%s.rope.scaling.finetuned"               },
+    { LLM_KV_ROPE_SCALING_YARN_LOG_MUL,        "%s.rope.scaling.yarn_log_multiplier"     },
 
-    { LLM_KV_SPLIT_NO,                      "split.no"            },
-    { LLM_KV_SPLIT_COUNT,                   "split.count"         },
-    { LLM_KV_SPLIT_TENSORS_COUNT,           "split.tensors.count" },
+    { LLM_KV_SPLIT_NO,                         "split.no"            },
+    { LLM_KV_SPLIT_COUNT,                      "split.count"         },
+    { LLM_KV_SPLIT_TENSORS_COUNT,              "split.tensors.count" },
 
-    { LLM_KV_SSM_CONV_KERNEL,               "%s.ssm.conv_kernel"    },
-    { LLM_KV_SSM_INNER_SIZE,                "%s.ssm.inner_size"     },
-    { LLM_KV_SSM_STATE_SIZE,                "%s.ssm.state_size"     },
-    { LLM_KV_SSM_TIME_STEP_RANK,            "%s.ssm.time_step_rank" },
-    { LLM_KV_SSM_DT_B_C_RMS,                "%s.ssm.dt_b_c_rms" },
+    { LLM_KV_SSM_CONV_KERNEL,                  "%s.ssm.conv_kernel"    },
+    { LLM_KV_SSM_INNER_SIZE,                   "%s.ssm.inner_size"     },
+    { LLM_KV_SSM_STATE_SIZE,                   "%s.ssm.state_size"     },
+    { LLM_KV_SSM_TIME_STEP_RANK,               "%s.ssm.time_step_rank" },
+    { LLM_KV_SSM_DT_B_C_RMS,                   "%s.ssm.dt_b_c_rms"     },
 
-    { LLM_KV_WKV_HEAD_SIZE,                 "%s.wkv.head_size" },
+    { LLM_KV_WKV_HEAD_SIZE,                    "%s.wkv.head_size" },
 
-    { LLM_KV_TOKENIZER_MODEL,                "tokenizer.ggml.model"                    },
-    { LLM_KV_TOKENIZER_PRE,                  "tokenizer.ggml.pre"                      },
-    { LLM_KV_TOKENIZER_LIST,                 "tokenizer.ggml.tokens"                   },
-    { LLM_KV_TOKENIZER_TOKEN_TYPE,           "tokenizer.ggml.token_type"               },
-    { LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT,     "tokenizer.ggml.token_type_count"         },
-    { LLM_KV_TOKENIZER_SCORES,               "tokenizer.ggml.scores"                   },
-    { LLM_KV_TOKENIZER_MERGES,               "tokenizer.ggml.merges"                   },
-    { LLM_KV_TOKENIZER_BOS_ID,               "tokenizer.ggml.bos_token_id"             },
-    { LLM_KV_TOKENIZER_EOS_ID,               "tokenizer.ggml.eos_token_id"             },
-    { LLM_KV_TOKENIZER_UNK_ID,               "tokenizer.ggml.unknown_token_id"         },
-    { LLM_KV_TOKENIZER_SEP_ID,               "tokenizer.ggml.seperator_token_id"       },
-    { LLM_KV_TOKENIZER_PAD_ID,               "tokenizer.ggml.padding_token_id"         },
-    { LLM_KV_TOKENIZER_CLS_ID,               "tokenizer.ggml.cls_token_id"             },
-    { LLM_KV_TOKENIZER_MASK_ID,              "tokenizer.ggml.mask_token_id"            },
-    { LLM_KV_TOKENIZER_ADD_BOS,              "tokenizer.ggml.add_bos_token"            },
-    { LLM_KV_TOKENIZER_ADD_EOS,              "tokenizer.ggml.add_eos_token"            },
-    { LLM_KV_TOKENIZER_ADD_PREFIX,           "tokenizer.ggml.add_space_prefix"         },
-    { LLM_KV_TOKENIZER_REMOVE_EXTRA_WS,      "tokenizer.ggml.remove_extra_whitespaces" },
-    { LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP, "tokenizer.ggml.precompiled_charsmap"     },
-    { LLM_KV_TOKENIZER_HF_JSON,              "tokenizer.huggingface.json"              },
-    { LLM_KV_TOKENIZER_RWKV,                 "tokenizer.rwkv.world"                    },
-    { LLM_KV_TOKENIZER_PREFIX_ID,            "tokenizer.ggml.prefix_token_id"          },
-    { LLM_KV_TOKENIZER_SUFFIX_ID,            "tokenizer.ggml.suffix_token_id"          },
-    { LLM_KV_TOKENIZER_MIDDLE_ID,            "tokenizer.ggml.middle_token_id"          },
-    { LLM_KV_TOKENIZER_EOT_ID,               "tokenizer.ggml.eot_token_id"             },
-    { LLM_KV_TOKENIZER_EOM_ID,               "tokenizer.ggml.eom_token_id"             },
+    { LLM_KV_TOKENIZER_MODEL,                  "tokenizer.ggml.model"                    },
+    { LLM_KV_TOKENIZER_PRE,                    "tokenizer.ggml.pre"                      },
+    { LLM_KV_TOKENIZER_LIST,                   "tokenizer.ggml.tokens"                   },
+    { LLM_KV_TOKENIZER_TOKEN_TYPE,             "tokenizer.ggml.token_type"               },
+    { LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT,       "tokenizer.ggml.token_type_count"         },
+    { LLM_KV_TOKENIZER_SCORES,                 "tokenizer.ggml.scores"                   },
+    { LLM_KV_TOKENIZER_MERGES,                 "tokenizer.ggml.merges"                   },
+    { LLM_KV_TOKENIZER_BOS_ID,                 "tokenizer.ggml.bos_token_id"             },
+    { LLM_KV_TOKENIZER_EOS_ID,                 "tokenizer.ggml.eos_token_id"             },
+    { LLM_KV_TOKENIZER_EOT_ID,                 "tokenizer.ggml.eot_token_id"             },
+    { LLM_KV_TOKENIZER_EOM_ID,                 "tokenizer.ggml.eom_token_id"             },
+    { LLM_KV_TOKENIZER_UNK_ID,                 "tokenizer.ggml.unknown_token_id"         },
+    { LLM_KV_TOKENIZER_SEP_ID,                 "tokenizer.ggml.seperator_token_id"       },
+    { LLM_KV_TOKENIZER_PAD_ID,                 "tokenizer.ggml.padding_token_id"         },
+    { LLM_KV_TOKENIZER_CLS_ID,                 "tokenizer.ggml.cls_token_id"             },
+    { LLM_KV_TOKENIZER_MASK_ID,                "tokenizer.ggml.mask_token_id"            },
+    { LLM_KV_TOKENIZER_ADD_BOS,                "tokenizer.ggml.add_bos_token"            },
+    { LLM_KV_TOKENIZER_ADD_EOS,                "tokenizer.ggml.add_eos_token"            },
+    { LLM_KV_TOKENIZER_ADD_PREFIX,             "tokenizer.ggml.add_space_prefix"         },
+    { LLM_KV_TOKENIZER_REMOVE_EXTRA_WS,        "tokenizer.ggml.remove_extra_whitespaces" },
+    { LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP,   "tokenizer.ggml.precompiled_charsmap"     },
+    { LLM_KV_TOKENIZER_HF_JSON,                "tokenizer.huggingface.json"              },
+    { LLM_KV_TOKENIZER_RWKV,                   "tokenizer.rwkv.world"                    },
+    { LLM_KV_TOKENIZER_FIM_PRE_ID,             "tokenizer.ggml.fim_pre_token_id"         },
+    { LLM_KV_TOKENIZER_FIM_SUF_ID,             "tokenizer.ggml.fim_suf_token_id"         },
+    { LLM_KV_TOKENIZER_FIM_MID_ID,             "tokenizer.ggml.fim_mid_token_id"         },
+    { LLM_KV_TOKENIZER_FIM_PAD_ID,             "tokenizer.ggml.fim_pad_token_id"         },
+    { LLM_KV_TOKENIZER_FIM_REP_ID,             "tokenizer.ggml.fim_rep_token_id"         },
+    { LLM_KV_TOKENIZER_FIM_SEP_ID,             "tokenizer.ggml.fim_sep_token_id"         },
 
-    { LLM_KV_ADAPTER_TYPE,                  "adapter.type"       },
-    { LLM_KV_ADAPTER_LORA_ALPHA,            "adapter.lora.alpha" },
+    { LLM_KV_ADAPTER_TYPE,                     "adapter.type"       },
+    { LLM_KV_ADAPTER_LORA_ALPHA,               "adapter.lora.alpha" },
+
+    // deprecated
+    { LLM_KV_TOKENIZER_PREFIX_ID,              "tokenizer.ggml.prefix_token_id" },
+    { LLM_KV_TOKENIZER_SUFFIX_ID,              "tokenizer.ggml.suffix_token_id" },
+    { LLM_KV_TOKENIZER_MIDDLE_ID,              "tokenizer.ggml.middle_token_id" },
 };
 
 struct LLM_KV {
@@ -2412,7 +2418,7 @@ struct llama_hparams {
 
     // needed by encoder-decoder models (e.g. T5, FLAN-T5)
     // ref: https://github.com/ggerganov/llama.cpp/pull/8141
-    llama_token dec_start_token_id = -1;
+    llama_token dec_start_token_id = LLAMA_TOKEN_NULL;
 
     enum llama_pooling_type      pooling_type            = LLAMA_POOLING_TYPE_NONE;
     enum llama_rope_type         rope_type               = LLAMA_ROPE_TYPE_NONE;
@@ -2941,9 +2947,6 @@ struct llama_sbatch_seq {
     llama_seq_id * seq_id;
     size_t offset;
     size_t length;
-
-    // helper for smoother batch API transition -- can be deprecated in the future
-    llama_seq_id all_seq_id; // used if seq_id == NULL
 };
 
 // sequence-length-aware batch splitting
@@ -3038,30 +3041,18 @@ struct llama_sbatch {
         } else {
             ubatch.embd = nullptr;
         }
-        // from here on, the else branches are deprecated;
-        // they are helpers for smoother batch API transition
-        if (batch->pos) {
-            if (ubatch.equal_seqs) {
-                for (size_t i = 0; i < length; ++i) {
-                    ubatch.pos[ubatch.n_tokens + i] = batch->pos[ids[seq.offset + i]];
-                }
-            } else {
-                // simple split
-                ubatch.pos = batch->pos + seq.offset;
+        if (ubatch.equal_seqs) {
+            for (size_t i = 0; i < length; ++i) {
+                ubatch.pos[ubatch.n_tokens + i] = batch->pos[ids[seq.offset + i]];
             }
         } else {
-            for (size_t i = 0; i < length; ++i) {
-                llama_pos bi = ids[seq.offset + i];
-                ubatch.pos[ubatch.n_tokens + i] = batch->all_pos_0 + (bi * batch->all_pos_1);
-            }
+            // simple split
+            ubatch.pos = batch->pos + seq.offset;
         }
         if (ubatch.equal_seqs) {
             ubatch.n_seq_id[ubatch.n_seqs] = seq.n_seq_id;
             if (seq.seq_id) {
                 ubatch.seq_id[ubatch.n_seqs] = seq.seq_id;
-            } else {
-                GGML_ASSERT(seq.n_seq_id == 1);
-                ubatch.seq_id[ubatch.n_seqs] = &seq.all_seq_id;
             }
         } else {
             // simple split
@@ -3074,10 +3065,6 @@ struct llama_sbatch {
             }
             if (batch->seq_id) {
                 ubatch.seq_id = batch->seq_id + seq.offset;
-            } else {
-                for (size_t i = 0; i < length; ++i) {
-                    ubatch.seq_id[ubatch.n_seqs + i] = &seq.all_seq_id;
-                }
             }
         }
         if (logits_all) {
@@ -3196,7 +3183,6 @@ struct llama_sbatch {
             s.seq_id = nullptr;
             s.offset = 0;
             s.length = n_tokens;
-            s.all_seq_id = batch.all_seq_id;
             return;
         }
         std::sort(ids.begin(), ids.end(),
@@ -3219,7 +3205,7 @@ struct llama_sbatch {
                     if (batch.pos) {
                         return batch.pos[a] < batch.pos[b];
                     }
-                    // no pos, sort by id (assuming batch.all_pos_1 is positive)
+                    // no pos, sort by id
                     return a < b;
                 }
                 // shared prompts go first
@@ -3229,30 +3215,25 @@ struct llama_sbatch {
         // init seq
         llama_sbatch_seq * last_seq = nullptr;
 
-        if (batch.n_seq_id != nullptr && batch.seq_id != nullptr) {
-            for (size_t i = 0; i < n_tokens; ++i) {
-                const size_t bi = ids[i];
-                const int32_t n_seqs = batch.n_seq_id[bi];
-                llama_seq_id * seq_ids = batch.seq_id[bi];
-                if (last_seq != nullptr) {
-                    bool same = n_seqs == last_seq->n_seq_id;
-                    for (int32_t j = 0; same && j < n_seqs; ++j) {
-                        if (seq_ids[j] != last_seq->seq_id[j]) {
-                            same = false;
-                        }
-                    }
-                    if (same) {
-                        last_seq->length += 1;
-                        continue;
+        for (size_t i = 0; i < n_tokens; ++i) {
+            const size_t bi = ids[i];
+            const int32_t n_seqs = batch.n_seq_id[bi];
+            llama_seq_id * seq_ids = batch.seq_id[bi];
+            if (last_seq != nullptr) {
+                bool same = n_seqs == last_seq->n_seq_id;
+                for (int32_t j = 0; same && j < n_seqs; ++j) {
+                    if (seq_ids[j] != last_seq->seq_id[j]) {
+                        same = false;
                     }
                 }
-                llama_sbatch_seq new_seq = {n_seqs, seq_ids, i, 1, batch.all_seq_id};
-                seq.push_back(new_seq);
-                last_seq = &seq.back();
+                if (same) {
+                    last_seq->length += 1;
+                    continue;
+                }
             }
-        } else {
-            llama_sbatch_seq new_seq = {1, nullptr, 0, n_tokens, batch.all_seq_id};
+            llama_sbatch_seq new_seq = {n_seqs, seq_ids, i, 1};
             seq.push_back(new_seq);
+            last_seq = &seq.back();
         }
         // keep shared prompts first at the end, then sort by length descending.
         std::sort(seq.begin(), seq.end(),
@@ -3292,12 +3273,8 @@ struct llama_context {
     std::unordered_map<struct llama_lora_adapter *, float> lora_adapters;
 
     std::vector<ggml_backend_t> backends;
-#ifdef GGML_USE_METAL
-    ggml_backend_t backend_metal = nullptr;
-#endif
-#ifdef GGML_USE_BLAS
-    ggml_backend_t backend_blas = nullptr;
-#endif
+    std::vector<std::pair<ggml_backend_t, ggml_backend_set_n_threads_t>> set_n_threads_fns;
+
     ggml_backend_t backend_cpu = nullptr;
 
     ggml_threadpool_t threadpool       = nullptr;
@@ -3420,16 +3397,6 @@ static int llama_get_device_count(const llama_model & model) {
     count += (int) model.rpc_servers.size();
 #endif
 
-#if defined(GGML_USE_METAL)
-    count += 1;
-#elif defined(GGML_USE_SYCL)
-    count += ggml_backend_sycl_get_device_count();
-#elif defined(GGML_USE_VULKAN)
-    count += ggml_backend_vk_get_device_count();
-#elif defined(GGML_USE_CANN)
-    count += ggml_backend_cann_get_device_count();
-#endif
-
     return count;
 
     GGML_UNUSED(model);
@@ -3447,20 +3414,8 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_cpu(const llama_mode
         }
     }
 
-#if defined(GGML_USE_SYCL)
-    if (host_buffer) {
-        buft = ggml_backend_sycl_host_buffer_type();
-    }
-#elif defined(GGML_USE_CANN)
-    if (host_buffer) {
-        buft = ggml_backend_cann_host_buffer_type();
-    }
-#elif defined(GGML_USE_CPU_HBM)
+#if defined(GGML_USE_CPU_HBM)
     buft = ggml_backend_cpu_hbm_buffer_type();
-#elif defined(GGML_USE_VULKAN)
-    if (host_buffer) {
-        buft = ggml_backend_vk_host_buffer_type();
-    }
 #endif
 
     if (buft == nullptr) {
@@ -3474,30 +3429,13 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_cpu(const llama_mode
 static ggml_backend_buffer_type_t llama_default_buffer_type_offload(const llama_model & model, int device) {
     ggml_backend_buffer_type_t buft = nullptr;
 
-#if defined(GGML_USE_RPC)
-    int rpc_count = (int)model.rpc_servers.size();
-    if (device < rpc_count) {
-        const char * endpoint = model.rpc_servers[device].c_str();
-        return ggml_backend_rpc_buffer_type(endpoint);
-    }
-    device -= rpc_count;
-#endif
-
     if (device < (int)model.devices.size()) {
         return ggml_backend_dev_buffer_type(model.devices[device]);
     }
     device -= (int)model.devices.size();
 
-#if defined(GGML_USE_METAL)
-    buft = ggml_backend_metal_buffer_type();
-#elif defined(GGML_USE_VULKAN)
-    buft = ggml_backend_vk_buffer_type(device);
-#elif defined(GGML_USE_SYCL)
-    buft = ggml_backend_sycl_buffer_type(device);
-#elif defined(GGML_USE_KOMPUTE)
+#if defined(GGML_USE_KOMPUTE)
     buft = ggml_backend_kompute_buffer_type(device);
-#elif defined(GGML_USE_CANN)
-    buft = ggml_backend_cann_buffer_type(device);
 #endif
 
     if (buft == nullptr) {
@@ -3524,12 +3462,6 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_split(const llama_mo
         }
     }
 
-#ifdef GGML_USE_SYCL
-    if (ggml_backend_sycl_get_device_count() > 1) {
-        buft = ggml_backend_sycl_split_buffer_type(tensor_split);
-    }
-#endif
-
     if (buft == nullptr) {
         buft = llama_default_buffer_type_offload(model, fallback_gpu);
     }
@@ -3539,18 +3471,6 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_split(const llama_mo
 }
 
 static size_t llama_get_device_memory(const llama_model & model, int device) {
-#if defined(GGML_USE_RPC)
-    int rpc_count = (int)model.rpc_servers.size();
-    if (device < rpc_count) {
-        size_t total;
-        size_t free;
-        const char * endpoint = model.rpc_servers[device].c_str();
-        ggml_backend_rpc_get_device_memory(endpoint, &free, &total);
-        return free;
-    }
-    device = device - rpc_count;
-#endif
-
     if (device < (int)model.devices.size()) {
         ggml_backend_dev_t dev = model.devices[device];
         size_t total;
@@ -3559,24 +3479,14 @@ static size_t llama_get_device_memory(const llama_model & model, int device) {
         return free;
     }
 
-#if defined(GGML_USE_SYCL)
-    size_t total;
-    size_t free;
-    ggml_backend_sycl_get_device_memory(device, &free, &total);
-    return free;
-#elif defined(GGML_USE_VULKAN)
-    size_t total;
-    size_t free;
-    ggml_backend_vk_get_device_memory(device, &free, &total);
-    return free;
-#elif defined(GGML_USE_CANN)
-    size_t total;
-    size_t free;
-    ggml_backend_cann_get_device_memory(device, &free, &total);
-    return free;
-#else
+    if (model.devices.size() > 0) {
+        ggml_backend_reg_t reg = ggml_backend_dev_backend_reg(model.devices[0]);
+        LLAMA_LOG_WARN("%s: failed to get free memmory of device:%d of backend:%s, for device id is out of range.\n", __func__, device, ggml_backend_reg_name(reg));
+    } else {
+        LLAMA_LOG_WARN("%s: failed to get free memmory of device, no devices in inputted model.\n", __func__);
+    }
     return 1;
-#endif
+
     GGML_UNUSED(model);
     GGML_UNUSED(device);
 }
@@ -5267,6 +5177,57 @@ struct llama_model_loader {
     }
 };
 
+// temporary allocate memory for the input batch if needed
+static const llama_seq_id batch_default_seq_id = 0;
+struct llama_batch_allocr {
+    std::array<llama_seq_id, 1> seq_id_0 = {batch_default_seq_id};
+    std::vector<llama_pos>      pos;
+    std::vector<int32_t>        n_seq_id;
+    std::vector<llama_seq_id *> seq_id;
+    std::vector<int8_t>         logits;
+    struct llama_batch          batch;
+    // optionally fulfill the batch returned by llama_batch_get_one
+    llama_batch_allocr(llama_context & ctx, struct llama_batch in_batch) {
+        batch = in_batch;
+        GGML_ASSERT(batch.n_tokens > 0);
+        if (!batch.pos) {
+            // determine the last position in KV cache
+            llama_pos last_pos = -1;
+            for (const auto & cell : ctx.kv_self.cells) {
+                if (cell.has_seq_id(batch_default_seq_id)) {
+                    last_pos = std::max(last_pos, cell.pos);
+                }
+            }
+            last_pos++; // next position
+            pos.resize(batch.n_tokens);
+            for (int32_t i = 0; i < batch.n_tokens; i++) {
+                pos[i] = i+last_pos;
+            }
+            batch.pos = pos.data();
+        }
+        if (!batch.n_seq_id) {
+            n_seq_id.resize(batch.n_tokens);
+            for (int32_t i = 0; i < batch.n_tokens; i++) {
+                n_seq_id[i] = seq_id_0.size();
+            }
+            batch.n_seq_id = n_seq_id.data();
+        }
+        if (!batch.seq_id) {
+            seq_id.resize(batch.n_tokens + 1);
+            seq_id[batch.n_tokens] = NULL;
+            for (int32_t i = 0; i < batch.n_tokens; i++) {
+                seq_id[i] = seq_id_0.data();
+            }
+            batch.seq_id = seq_id.data();
+        }
+        if (!batch.logits) {
+            logits.resize(batch.n_tokens);
+            logits[logits.size() - 1] = true;
+            batch.logits = logits.data();
+        }
+    }
+};
+
 template<>
 bool llama_model_loader::get_key(const enum llm_kv kid, enum llama_pooling_type & result, const bool required) {
     uint32_t tmp;
@@ -6209,14 +6170,14 @@ static void llm_load_vocab(
             vocab.type = LLAMA_VOCAB_TYPE_NONE;
 
             // default special tokens
-            vocab.special_bos_id  = -1;
-            vocab.special_eos_id  = -1;
-            vocab.special_unk_id  = -1;
-            vocab.special_sep_id  = -1;
-            vocab.special_pad_id  = -1;
-            vocab.special_cls_id  = -1;
-            vocab.special_mask_id = -1;
-            vocab.linefeed_id     = -1;
+            vocab.special_bos_id  = LLAMA_TOKEN_NULL;
+            vocab.special_eos_id  = LLAMA_TOKEN_NULL;
+            vocab.special_unk_id  = LLAMA_TOKEN_NULL;
+            vocab.special_sep_id  = LLAMA_TOKEN_NULL;
+            vocab.special_pad_id  = LLAMA_TOKEN_NULL;
+            vocab.special_cls_id  = LLAMA_TOKEN_NULL;
+            vocab.special_mask_id = LLAMA_TOKEN_NULL;
+            vocab.linefeed_id     = LLAMA_TOKEN_NULL;
 
             // read vocab size from metadata
             if (!ml.get_key(LLM_KV_VOCAB_SIZE, vocab.n_vocab, false)) {
@@ -6233,16 +6194,16 @@ static void llm_load_vocab(
             vocab.special_bos_id  = 1;
             vocab.special_eos_id  = 2;
             vocab.special_unk_id  = 0;
-            vocab.special_sep_id  = -1;
-            vocab.special_pad_id  = -1;
-            vocab.special_cls_id  = -1;
-            vocab.special_mask_id = -1;
+            vocab.special_sep_id  = LLAMA_TOKEN_NULL;
+            vocab.special_pad_id  = LLAMA_TOKEN_NULL;
+            vocab.special_cls_id  = LLAMA_TOKEN_NULL;
+            vocab.special_mask_id = LLAMA_TOKEN_NULL;
         } else if (tokenizer_model == "bert") {
             vocab.type = LLAMA_VOCAB_TYPE_WPM;
 
             // default special tokens
-            vocab.special_bos_id  = -1;
-            vocab.special_eos_id  = -1;
+            vocab.special_bos_id  = LLAMA_TOKEN_NULL;
+            vocab.special_eos_id  = LLAMA_TOKEN_NULL;
             vocab.special_unk_id  = 100;
             vocab.special_sep_id  = 102;
             vocab.special_pad_id  = 0;
@@ -6278,22 +6239,22 @@ static void llm_load_vocab(
             // default special tokens
             vocab.special_bos_id  = 11;
             vocab.special_eos_id  = 11;
-            vocab.special_unk_id  = -1;
-            vocab.special_sep_id  = -1;
-            vocab.special_pad_id  = -1;
-            vocab.special_cls_id  = -1;
-            vocab.special_mask_id = -1;
+            vocab.special_unk_id  = LLAMA_TOKEN_NULL;
+            vocab.special_sep_id  = LLAMA_TOKEN_NULL;
+            vocab.special_pad_id  = LLAMA_TOKEN_NULL;
+            vocab.special_cls_id  = LLAMA_TOKEN_NULL;
+            vocab.special_mask_id = LLAMA_TOKEN_NULL;
         } else if (tokenizer_model == "t5") {
             vocab.type = LLAMA_VOCAB_TYPE_UGM;
 
             // default special tokens
-            vocab.special_bos_id  = -1;
+            vocab.special_bos_id  = LLAMA_TOKEN_NULL;
             vocab.special_eos_id  = 1;
             vocab.special_unk_id  = 2;
-            vocab.special_sep_id  = -1;
+            vocab.special_sep_id  = LLAMA_TOKEN_NULL;
             vocab.special_pad_id  = 0;
-            vocab.special_cls_id  = -1;
-            vocab.special_mask_id = -1;
+            vocab.special_cls_id  = LLAMA_TOKEN_NULL;
+            vocab.special_mask_id = LLAMA_TOKEN_NULL;
 
             const int precompiled_charsmap_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP).c_str());
             if (precompiled_charsmap_keyidx != -1) {
@@ -6316,11 +6277,11 @@ static void llm_load_vocab(
             vocab.type = LLAMA_VOCAB_TYPE_RWKV;
 
             // default special tokens
-            vocab.special_bos_id = -1;
-            vocab.special_eos_id = -1;
-            vocab.special_unk_id = -1;
-            vocab.special_sep_id = -1;
-            vocab.special_pad_id = -1;
+            vocab.special_bos_id = LLAMA_TOKEN_NULL;
+            vocab.special_eos_id = LLAMA_TOKEN_NULL;
+            vocab.special_unk_id = LLAMA_TOKEN_NULL;
+            vocab.special_sep_id = LLAMA_TOKEN_NULL;
+            vocab.special_pad_id = LLAMA_TOKEN_NULL;
         } else {
             throw std::runtime_error(format("unknown tokenizer: '%s'", tokenizer_model.c_str()));
         }
@@ -6404,7 +6365,7 @@ static void llm_load_vocab(
             } else if (
                 tokenizer_pre == "chatglm-bpe") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_CHATGLM4;
-                vocab.special_bos_id  = -1;
+                vocab.special_bos_id = LLAMA_TOKEN_NULL;
             } else if (
                 tokenizer_pre == "viking") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_VIKING;
@@ -6530,44 +6491,6 @@ static void llm_load_vocab(
 
     // determine the newline token: LLaMA "<0x0A>" == 10 == '\n', Falcon 193 == '\n'
     if (vocab.type == LLAMA_VOCAB_TYPE_SPM) {
-        // For Fill-In-the-Middle (FIM)/infill models which where converted
-        // prior to support of FIM special tokens in GGUF, the following
-        // will allow those models to continue to work. The general names
-        // of the known models are currently CodeLlama (LLM_ARCH_LLAMA) and
-        // CodeGemma (LLM_ARCH_GEMMA). This can potentially be removed once
-        // new versions of these models have been published.
-        std::string gen_name;
-        ml.get_key(LLM_KV_GENERAL_NAME, gen_name, false);
-
-        std::transform(gen_name.begin(), gen_name.end(), gen_name.begin(),
-            [](unsigned char c){ return std::tolower(c); });
-
-        if (gen_name.find("code") != std::string::npos) {
-            if (model.arch == LLM_ARCH_LLAMA
-              && 32010 < vocab.id_to_token.size()
-              && vocab.id_to_token[32007].text.find("<PRE>") != std::string::npos
-              && vocab.id_to_token[32008].text.find("<SUF>") != std::string::npos
-              && vocab.id_to_token[32009].text.find("<MID>") != std::string::npos
-              && vocab.id_to_token[32010].text.find("<EOT>") != std::string::npos) {
-                vocab.special_prefix_id = 32007;
-                vocab.special_suffix_id = 32008;
-                vocab.special_middle_id = 32009;
-                vocab.special_eot_id    = 32010;
-            } else if (model.arch == LLM_ARCH_GEMMA
-              && 107 < vocab.id_to_token.size()
-              && vocab.id_to_token[67].text == "<|fim_prefix|>"
-              && vocab.id_to_token[69].text == "<|fim_suffix|>"
-              && vocab.id_to_token[68].text == "<|fim_middle|>"
-              && vocab.id_to_token[107].text == "<end_of_turn>") {
-                vocab.special_prefix_id = 67;
-                vocab.special_suffix_id = 69;
-                vocab.special_middle_id = 68;
-                // TODO: this is not EOT, it is "file separator" token, needs fix
-                //       https://huggingface.co/google/codegemma-7b-it/blob/9b1d9231388358c04d90bd003458f5070d97db44/tokenizer_config.json#L565-L572
-                //vocab.special_eot_id    = 70;
-                vocab.special_eot_id    = 107;
-            }
-        }
         try {
             vocab.linefeed_id = llama_byte_to_token_impl(vocab, '\n');
         } catch (const std::exception & e) {
@@ -6595,18 +6518,26 @@ static void llm_load_vocab(
     // special tokens
     {
         const std::vector<std::pair<enum llm_kv, int32_t &>> special_token_types = {
-            { LLM_KV_TOKENIZER_BOS_ID,    vocab.special_bos_id    },
-            { LLM_KV_TOKENIZER_EOS_ID,    vocab.special_eos_id    },
-            { LLM_KV_TOKENIZER_UNK_ID,    vocab.special_unk_id    },
-            { LLM_KV_TOKENIZER_SEP_ID,    vocab.special_sep_id    },
-            { LLM_KV_TOKENIZER_PAD_ID,    vocab.special_pad_id    },
-            { LLM_KV_TOKENIZER_CLS_ID,    vocab.special_cls_id    },
-            { LLM_KV_TOKENIZER_MASK_ID,   vocab.special_mask_id   },
-            { LLM_KV_TOKENIZER_PREFIX_ID, vocab.special_prefix_id },
-            { LLM_KV_TOKENIZER_SUFFIX_ID, vocab.special_suffix_id },
-            { LLM_KV_TOKENIZER_MIDDLE_ID, vocab.special_middle_id },
-            { LLM_KV_TOKENIZER_EOT_ID,    vocab.special_eot_id    },
-            { LLM_KV_TOKENIZER_EOM_ID,    vocab.special_eom_id    },
+            { LLM_KV_TOKENIZER_BOS_ID,     vocab.special_bos_id     },
+            { LLM_KV_TOKENIZER_EOS_ID,     vocab.special_eos_id     },
+            { LLM_KV_TOKENIZER_EOT_ID,     vocab.special_eot_id     },
+            { LLM_KV_TOKENIZER_EOM_ID,     vocab.special_eom_id     },
+            { LLM_KV_TOKENIZER_UNK_ID,     vocab.special_unk_id     },
+            { LLM_KV_TOKENIZER_SEP_ID,     vocab.special_sep_id     },
+            { LLM_KV_TOKENIZER_PAD_ID,     vocab.special_pad_id     },
+            { LLM_KV_TOKENIZER_CLS_ID,     vocab.special_cls_id     },
+            { LLM_KV_TOKENIZER_MASK_ID,    vocab.special_mask_id    },
+            { LLM_KV_TOKENIZER_FIM_PRE_ID, vocab.special_fim_pre_id },
+            { LLM_KV_TOKENIZER_FIM_SUF_ID, vocab.special_fim_suf_id },
+            { LLM_KV_TOKENIZER_FIM_MID_ID, vocab.special_fim_mid_id },
+            { LLM_KV_TOKENIZER_FIM_PAD_ID, vocab.special_fim_pad_id },
+            { LLM_KV_TOKENIZER_FIM_REP_ID, vocab.special_fim_rep_id },
+            { LLM_KV_TOKENIZER_FIM_SEP_ID, vocab.special_fim_sep_id },
+
+            // deprecated
+            { LLM_KV_TOKENIZER_PREFIX_ID, vocab.special_fim_pre_id },
+            { LLM_KV_TOKENIZER_SUFFIX_ID, vocab.special_fim_suf_id },
+            { LLM_KV_TOKENIZER_MIDDLE_ID, vocab.special_fim_mid_id },
         };
 
         for (const auto & it : special_token_types) {
@@ -6637,46 +6568,140 @@ static void llm_load_vocab(
             }
         }
 
-        // find EOT token: "<|eot_id|>", "<|im_end|>", "<end_of_turn>", etc.
-        //
-        // TODO: convert scripts should provide this token through the KV metadata LLAMA_KV_TOKENIZER_EOT_ID
-        //       for now, we apply this workaround to find the EOT token based on its text
-        if (vocab.special_eot_id == -1) {
-            for (const auto & t : vocab.token_to_id) {
+        // auto-detect special tokens by text
+        // TODO: convert scripts should provide these tokens through the KV metadata LLM_KV_TOKENIZER_...
+        //       for now, we apply this workaround to find the tokens based on their text
+
+        for (const auto & t : vocab.token_to_id) {
+            // find EOT token: "<|eot_id|>", "<|im_end|>", "<end_of_turn>", etc.
+            if (vocab.special_eot_id == LLAMA_TOKEN_NULL) {
                 if (false
-                        // TODO: gemma "<end_of_turn>" is exported as a normal token, so the following check does not work
-                        //       need to fix convert script
-                        //vocab.id_to_token[t.second].type == LLAMA_TOKEN_TYPE_CONTROL &&
                         || t.first == "<|eot_id|>"
                         || t.first == "<|im_end|>"
                         || t.first == "<|end|>"
                         || t.first == "<end_of_turn>"
                         || t.first == "<|endoftext|>"
                         || t.first == "<EOT>"
+                        || t.first == "<｜end▁of▁sentence｜>" // DeepSeek
                    ) {
                     vocab.special_eot_id = t.second;
                     if ((vocab.id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
-                        LLAMA_LOG_WARN("%s: control-looking token: '%s' was not control-type; this is probably a bug in the model. its type will be overridden\n",
-                                __func__, t.first.c_str());
+                        LLAMA_LOG_WARN("%s: control-looking token: %6d '%s' was not control-type; this is probably a bug in the model. its type will be overridden\n",
+                                __func__, t.second, t.first.c_str());
                         vocab.id_to_token[t.second].attr = LLAMA_TOKEN_ATTR_CONTROL;
                     }
-                    break;
                 }
             }
-        }
 
-        // find EOM token: "<|eom_id|>"
-        //
-        // TODO: convert scripts should provide this token through the KV metadata LLAMA_KV_TOKENIZER_EOM_ID
-        //       for now, we apply this workaround to find the EOM token based on its text
-        if (vocab.special_eom_id == -1) {
-            const auto & t = vocab.token_to_id.find("<|eom_id|>");
-            if (t != vocab.token_to_id.end()) {
-                vocab.special_eom_id = t->second;
-                if ((vocab.id_to_token[t->second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
-                    LLAMA_LOG_WARN("%s: control-looking token: '%s' was not control-type; this is probably a bug in the model. its type will be overridden\n",
-                        __func__, t->first.c_str());
-                    vocab.id_to_token[t->second].attr = LLAMA_TOKEN_ATTR_CONTROL;
+            // find EOM token: "<|eom_id|>"
+            if (vocab.special_eom_id == LLAMA_TOKEN_NULL) {
+                if (false
+                        || t.first == "<|eom_id|>"
+                        ) {
+                    vocab.special_eom_id = t.second;
+                    if ((vocab.id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
+                        LLAMA_LOG_WARN("%s: control-looking token: %6d '%s' was not control-type; this is probably a bug in the model. its type will be overridden\n",
+                                __func__, t.second, t.first.c_str());
+                        vocab.id_to_token[t.second].attr = LLAMA_TOKEN_ATTR_CONTROL;
+                    }
+                }
+            }
+
+            // find FIM_PRE token: "<|fim_prefix|>", "<fim-prefix>", "<PRE>", etc.
+            if (vocab.special_fim_pre_id == LLAMA_TOKEN_NULL) {
+                if (false
+                        || t.first == "<|fim_prefix|>"  // Qwen
+                        || t.first == "<fim-prefix>"
+                        || t.first == "<｜fim▁begin｜>" // DeepSeek
+                        || t.first == "<PRE>"
+                        ) {
+                    vocab.special_fim_pre_id = t.second;
+                    if ((vocab.id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
+                        LLAMA_LOG_WARN("%s: control-looking token: %6d '%s' was not control-type; this is probably a bug in the model. its type will be overridden\n",
+                                __func__, t.second, t.first.c_str());
+                        vocab.id_to_token[t.second].attr = LLAMA_TOKEN_ATTR_CONTROL;
+                    }
+                }
+            }
+
+            // find FIM_SUF token: "<|fim_suffix|>", "<fim-suffix>", "<SUF>", etc.
+            if (vocab.special_fim_suf_id == LLAMA_TOKEN_NULL) {
+                if (false
+                        || t.first == "<|fim_suffix|>" // Qwen
+                        || t.first == "<fim-suffix>"
+                        || t.first == "<｜fim▁hole｜>" // DeepSeek
+                        || t.first == "<SUF>"
+                        ) {
+                    vocab.special_fim_suf_id = t.second;
+                    if ((vocab.id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
+                        LLAMA_LOG_WARN("%s: control-looking token: %6d '%s' was not control-type; this is probably a bug in the model. its type will be overridden\n",
+                                __func__, t.second, t.first.c_str());
+                        vocab.id_to_token[t.second].attr = LLAMA_TOKEN_ATTR_CONTROL;
+                    }
+                }
+            }
+
+            // find FIM_MID token: "<|fim_middle|>", "<fim-middle>", "<MID>", etc.
+            if (vocab.special_fim_mid_id == LLAMA_TOKEN_NULL) {
+                if (false
+                        || t.first == "<|fim_middle|>" // Qwen
+                        || t.first == "<fim-middle>"
+                        || t.first == "<｜fim▁end｜>"  // DeepSeek
+                        || t.first == "<MID>"
+                        ) {
+                    vocab.special_fim_mid_id = t.second;
+                    if ((vocab.id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
+                        LLAMA_LOG_WARN("%s: control-looking token: %6d '%s' was not control-type; this is probably a bug in the model. its type will be overridden\n",
+                                __func__, t.second, t.first.c_str());
+                        vocab.id_to_token[t.second].attr = LLAMA_TOKEN_ATTR_CONTROL;
+                    }
+                }
+            }
+
+            // find FIM_PAD token: "<|fim_pad|>", "<fim-pad>", "<PAD>", etc.
+            if (vocab.special_fim_pad_id == LLAMA_TOKEN_NULL) {
+                if (false
+                        || t.first == "<|fim_pad|>" // Qwen
+                        || t.first == "<fim-pad>"
+                        || t.first == "<PAD>"
+                        ) {
+                    vocab.special_fim_pad_id = t.second;
+                    if ((vocab.id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
+                        LLAMA_LOG_WARN("%s: control-looking token: %6d '%s' was not control-type; this is probably a bug in the model. its type will be overridden\n",
+                                __func__, t.second, t.first.c_str());
+                        vocab.id_to_token[t.second].attr = LLAMA_TOKEN_ATTR_CONTROL;
+                    }
+                }
+            }
+
+            // find FIM_REP token: "<|fim_repo|>", "<fim-repo>", "<REP>", etc.
+            if (vocab.special_fim_rep_id == LLAMA_TOKEN_NULL) {
+                if (false
+                        || t.first == "<|fim_repo|>"  // Qwen
+                        || t.first == "<|repo_name|>"
+                        || t.first == "<fim-repo>"
+                        || t.first == "<REPO>"
+                        ) {
+                    vocab.special_fim_rep_id = t.second;
+                    if ((vocab.id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
+                        LLAMA_LOG_WARN("%s: control-looking token: %6d '%s' was not control-type; this is probably a bug in the model. its type will be overridden\n",
+                                __func__, t.second, t.first.c_str());
+                        vocab.id_to_token[t.second].attr = LLAMA_TOKEN_ATTR_CONTROL;
+                    }
+                }
+            }
+
+            // find FIM_SEP token: "<|file_sep|>"
+            if (vocab.special_fim_sep_id == LLAMA_TOKEN_NULL) {
+                if (false
+                        || t.first == "<|file_sep|>" // Qwen
+                        ) {
+                    vocab.special_fim_sep_id = t.second;
+                    if ((vocab.id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
+                        LLAMA_LOG_WARN("%s: control-looking token: %6d '%s' was not control-type; this is probably a bug in the model. its type will be overridden\n",
+                                __func__, t.second, t.first.c_str());
+                        vocab.id_to_token[t.second].attr = LLAMA_TOKEN_ATTR_CONTROL;
+                    }
                 }
             }
         }
@@ -6685,6 +6710,19 @@ static void llm_load_vocab(
         // this is currently determined based on the token text, which is obviously not ideal
         // ref: https://github.com/ggerganov/llama.cpp/issues/9606
         vocab.special_eog_ids.clear();
+
+        if (vocab.special_fim_pad_id != LLAMA_TOKEN_NULL && vocab.special_eog_ids.count(vocab.special_fim_pad_id) == 0) {
+            vocab.special_eog_ids.insert(vocab.special_fim_pad_id);
+        }
+
+        if (vocab.special_fim_rep_id != LLAMA_TOKEN_NULL && vocab.special_eog_ids.count(vocab.special_fim_rep_id) == 0) {
+            vocab.special_eog_ids.insert(vocab.special_fim_rep_id);
+        }
+
+        if (vocab.special_fim_sep_id != LLAMA_TOKEN_NULL && vocab.special_eog_ids.count(vocab.special_fim_sep_id) == 0) {
+            vocab.special_eog_ids.insert(vocab.special_fim_sep_id);
+        }
+
         for (const auto & t : vocab.token_to_id) {
             if (false
                     || t.first == "<|eot_id|>"
@@ -6697,24 +6735,31 @@ static void llm_load_vocab(
                ) {
                 vocab.special_eog_ids.insert(t.second);
                 if ((vocab.id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
-                    LLAMA_LOG_WARN("%s: control-looking token: '%s' was not control-type; this is probably a bug in the model. its type will be overridden\n",
-                            __func__, t.first.c_str());
+                    LLAMA_LOG_WARN("%s: control-looking token: %6d '%s' was not control-type; this is probably a bug in the model. its type will be overridden\n",
+                            __func__, t.second, t.first.c_str());
                     vocab.id_to_token[t.second].attr = LLAMA_TOKEN_ATTR_CONTROL;
                 }
+            } else {
+                // token is control, but not marked as EOG -> print a debug log
+                if (vocab.id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL && vocab.special_eog_ids.count(t.second) == 0) {
+                    LLAMA_LOG_DEBUG("%s: control token: %6d '%s' is not marked as EOG\n",
+                            __func__, t.second, t.first.c_str());
+                }
             }
         }
 
-        if (vocab.special_eos_id != -1 && vocab.special_eog_ids.count(vocab.special_eos_id) == 0) {
+        // sanity checks
+        if (vocab.special_eos_id != LLAMA_TOKEN_NULL && vocab.special_eog_ids.count(vocab.special_eos_id) == 0) {
             vocab.special_eog_ids.insert(vocab.special_eos_id);
             LLAMA_LOG_WARN("%s: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect\n", __func__);
         }
 
-        if (vocab.special_eot_id != -1 && vocab.special_eog_ids.count(vocab.special_eot_id) == 0) {
+        if (vocab.special_eot_id != LLAMA_TOKEN_NULL && vocab.special_eog_ids.count(vocab.special_eot_id) == 0) {
             vocab.special_eog_ids.insert(vocab.special_eot_id);
             LLAMA_LOG_WARN("%s: special_eot_id is not in special_eog_ids - the tokenizer config may be incorrect\n", __func__);
         }
 
-        if (vocab.special_eom_id != -1 && vocab.special_eog_ids.count(vocab.special_eom_id) == 0) {
+        if (vocab.special_eom_id != LLAMA_TOKEN_NULL && vocab.special_eog_ids.count(vocab.special_eom_id) == 0) {
             vocab.special_eog_ids.insert(vocab.special_eom_id);
             LLAMA_LOG_WARN("%s: special_eom_id is not in special_eog_ids - the tokenizer config may be incorrect\n", __func__);
         }
@@ -6908,20 +6953,24 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
     LLAMA_LOG_INFO("%s: general.name     = %s\n",    __func__, model.name.c_str());
 
     // special tokens
-    if (vocab.special_bos_id    != -1) { LLAMA_LOG_INFO( "%s: BOS token        = %d '%s'\n", __func__, vocab.special_bos_id,  vocab.id_to_token[vocab.special_bos_id].text.c_str() );  }
-    if (vocab.special_eos_id    != -1) { LLAMA_LOG_INFO( "%s: EOS token        = %d '%s'\n", __func__, vocab.special_eos_id,  vocab.id_to_token[vocab.special_eos_id].text.c_str() );  }
-    if (vocab.special_unk_id    != -1) { LLAMA_LOG_INFO( "%s: UNK token        = %d '%s'\n", __func__, vocab.special_unk_id,  vocab.id_to_token[vocab.special_unk_id].text.c_str() );  }
-    if (vocab.special_sep_id    != -1) { LLAMA_LOG_INFO( "%s: SEP token        = %d '%s'\n", __func__, vocab.special_sep_id,  vocab.id_to_token[vocab.special_sep_id].text.c_str() );  }
-    if (vocab.special_pad_id    != -1) { LLAMA_LOG_INFO( "%s: PAD token        = %d '%s'\n", __func__, vocab.special_pad_id,  vocab.id_to_token[vocab.special_pad_id].text.c_str() );  }
-    if (vocab.special_cls_id    != -1) { LLAMA_LOG_INFO( "%s: CLS token        = %d '%s'\n", __func__, vocab.special_cls_id,  vocab.id_to_token[vocab.special_cls_id].text.c_str() );  }
-    if (vocab.special_mask_id   != -1) { LLAMA_LOG_INFO( "%s: MASK token       = %d '%s'\n", __func__, vocab.special_mask_id, vocab.id_to_token[vocab.special_mask_id].text.c_str() ); }
+    if (vocab.special_bos_id  != -1)    { LLAMA_LOG_INFO( "%s: BOS token        = %d '%s'\n", __func__, vocab.special_bos_id,     vocab.id_to_token[vocab.special_bos_id].text.c_str() );  }
+    if (vocab.special_eos_id  != -1)    { LLAMA_LOG_INFO( "%s: EOS token        = %d '%s'\n", __func__, vocab.special_eos_id,     vocab.id_to_token[vocab.special_eos_id].text.c_str() );  }
+    if (vocab.special_eot_id  != -1)    { LLAMA_LOG_INFO( "%s: EOT token        = %d '%s'\n", __func__, vocab.special_eot_id,     vocab.id_to_token[vocab.special_eot_id].text.c_str() );  }
+    if (vocab.special_eom_id  != -1)    { LLAMA_LOG_INFO( "%s: EOM token        = %d '%s'\n", __func__, vocab.special_eom_id,     vocab.id_to_token[vocab.special_eom_id].text.c_str() );  }
+    if (vocab.special_unk_id  != -1)    { LLAMA_LOG_INFO( "%s: UNK token        = %d '%s'\n", __func__, vocab.special_unk_id,     vocab.id_to_token[vocab.special_unk_id].text.c_str() );  }
+    if (vocab.special_sep_id  != -1)    { LLAMA_LOG_INFO( "%s: SEP token        = %d '%s'\n", __func__, vocab.special_sep_id,     vocab.id_to_token[vocab.special_sep_id].text.c_str() );  }
+    if (vocab.special_pad_id  != -1)    { LLAMA_LOG_INFO( "%s: PAD token        = %d '%s'\n", __func__, vocab.special_pad_id,     vocab.id_to_token[vocab.special_pad_id].text.c_str() );  }
+    if (vocab.special_cls_id  != -1)    { LLAMA_LOG_INFO( "%s: CLS token        = %d '%s'\n", __func__, vocab.special_cls_id,     vocab.id_to_token[vocab.special_cls_id].text.c_str() );  }
+    if (vocab.special_mask_id != -1)    { LLAMA_LOG_INFO( "%s: MASK token       = %d '%s'\n", __func__, vocab.special_mask_id,    vocab.id_to_token[vocab.special_mask_id].text.c_str() ); }
 
-    if (vocab.linefeed_id       != -1) { LLAMA_LOG_INFO( "%s: LF token         = %d '%s'\n", __func__, vocab.linefeed_id,       vocab.id_to_token[vocab.linefeed_id].text.c_str() );       }
-    if (vocab.special_prefix_id != -1) { LLAMA_LOG_INFO( "%s: PRE token        = %d '%s'\n", __func__, vocab.special_prefix_id, vocab.id_to_token[vocab.special_prefix_id].text.c_str() ); }
-    if (vocab.special_suffix_id != -1) { LLAMA_LOG_INFO( "%s: SUF token        = %d '%s'\n", __func__, vocab.special_suffix_id, vocab.id_to_token[vocab.special_suffix_id].text.c_str() ); }
-    if (vocab.special_middle_id != -1) { LLAMA_LOG_INFO( "%s: MID token        = %d '%s'\n", __func__, vocab.special_middle_id, vocab.id_to_token[vocab.special_middle_id].text.c_str() ); }
-    if (vocab.special_eot_id    != -1) { LLAMA_LOG_INFO( "%s: EOT token        = %d '%s'\n", __func__, vocab.special_eot_id,    vocab.id_to_token[vocab.special_eot_id].text.c_str() );    }
-    if (vocab.special_eom_id    != -1) { LLAMA_LOG_INFO( "%s: EOM token        = %d '%s'\n", __func__, vocab.special_eom_id,    vocab.id_to_token[vocab.special_eom_id].text.c_str() );    }
+    if (vocab.linefeed_id != -1)        { LLAMA_LOG_INFO( "%s: LF token         = %d '%s'\n", __func__, vocab.linefeed_id,        vocab.id_to_token[vocab.linefeed_id].text.c_str() ); }
+
+    if (vocab.special_fim_pre_id != -1) { LLAMA_LOG_INFO( "%s: FIM PRE token    = %d '%s'\n", __func__, vocab.special_fim_pre_id, vocab.id_to_token[vocab.special_fim_pre_id].text.c_str() ); }
+    if (vocab.special_fim_suf_id != -1) { LLAMA_LOG_INFO( "%s: FIM SUF token    = %d '%s'\n", __func__, vocab.special_fim_suf_id, vocab.id_to_token[vocab.special_fim_suf_id].text.c_str() ); }
+    if (vocab.special_fim_mid_id != -1) { LLAMA_LOG_INFO( "%s: FIM MID token    = %d '%s'\n", __func__, vocab.special_fim_mid_id, vocab.id_to_token[vocab.special_fim_mid_id].text.c_str() ); }
+    if (vocab.special_fim_pad_id != -1) { LLAMA_LOG_INFO( "%s: FIM PAD token    = %d '%s'\n", __func__, vocab.special_fim_pad_id, vocab.id_to_token[vocab.special_fim_pad_id].text.c_str() ); }
+    if (vocab.special_fim_rep_id != -1) { LLAMA_LOG_INFO( "%s: FIM REP token    = %d '%s'\n", __func__, vocab.special_fim_rep_id, vocab.id_to_token[vocab.special_fim_rep_id].text.c_str() ); }
+    if (vocab.special_fim_sep_id != -1) { LLAMA_LOG_INFO( "%s: FIM SEP token    = %d '%s'\n", __func__, vocab.special_fim_sep_id, vocab.id_to_token[vocab.special_fim_sep_id].text.c_str() ); }
 
     for (const auto & id : vocab.special_eog_ids) {
         LLAMA_LOG_INFO( "%s: EOG token        = %d '%s'\n", __func__, id, vocab.id_to_token[id].text.c_str() );
@@ -6987,7 +7036,14 @@ static bool llm_load_tensors(
 
     // assign cpu layers
     for (int i = 0; i < i_gpu_start; ++i) {
+#ifdef GGML_USE_AMX
+        model.buft_layer[i] = {
+            ggml_backend_amx_buffer_type(),
+            llama_default_buffer_type_cpu(model, true)
+        };
+#else
         model.buft_layer[i] = llama_default_buffer_type_cpu(model, true);
+#endif
     }
 
     if (split_mode == LLAMA_SPLIT_MODE_LAYER) {
@@ -8918,48 +8974,40 @@ static bool llm_load_tensors(
         llama_buf_map bufs;
         bufs.reserve(n_max_backend_buffer);
 
-        // only the mmap region containing the tensors in the model is mapped to the backend buffer
-        // this is important for metal with apple silicon: if the entire model could be mapped to a metal buffer, then we could just use metal for all layers
-        // this allows using partial offloading when the model size exceeds the metal buffer size, but not the RAM size
-        if (ml.use_mmap && use_mmap_buffer && buft == llama_default_buffer_type_cpu(model, true)) {
+        // check if this backend device supports buffer_from_host_ptr
+        // when using a host buffer as the CPU bakcend buffer, use the CPU device to prioritize using buffer_from_host_ptr over the host buffer
+        ggml_backend_dev_t dev = ggml_backend_buft_get_device(buft == llama_default_buffer_type_cpu(model, true) ? ggml_backend_cpu_buffer_type() : buft);
+        bool buffer_from_host_ptr_supported = false;
+        if (dev) {
+            ggml_backend_dev_props props;
+            ggml_backend_dev_get_props(dev, &props);
+            buffer_from_host_ptr_supported = props.caps.buffer_from_host_ptr;
+        }
+
+        if (ml.use_mmap && use_mmap_buffer && buffer_from_host_ptr_supported) {
             for (uint32_t idx = 0; idx < ml.files.size(); idx++) {
+                // only the mmap region containing the tensors in the model is mapped to the backend buffer
+                // this is important for metal with apple silicon: if the entire model could be mapped to a metal buffer, then we could just use metal for all layers
+                // this allows using partial offloading when the model size exceeds the metal buffer size, but not the RAM size
                 void * addr = nullptr;
-                size_t first, last;
+                size_t first, last; // NOLINT
                 ml.get_mapping_range(&first, &last, &addr, idx, ctx);
                 if (first >= last) {
                     continue;
                 }
-                ggml_backend_buffer_t buf = ggml_backend_cpu_buffer_from_ptr((char *) addr + first, last - first);
-                if (buf == nullptr) {
-                    throw std::runtime_error("unable to allocate backend CPU buffer");
-                }
-                model.bufs.push_back(buf);
-                bufs.emplace(idx, buf);
-            }
-        }
-#ifdef GGML_USE_METAL
-        else if (ml.use_mmap && use_mmap_buffer && buft == ggml_backend_metal_buffer_type()) {
-            for (uint32_t idx = 0; idx < ml.files.size(); idx++) {
                 const size_t max_size = ggml_get_max_tensor_size(ctx);
-                void * addr = nullptr;
-                size_t first, last;
-                ml.get_mapping_range(&first, &last, &addr, idx, ctx);
-                if (first >= last) {
-                    continue;
-                }
-                ggml_backend_buffer_t buf = ggml_backend_metal_buffer_from_ptr((char *) addr + first, last - first, max_size);
+                ggml_backend_buffer_t buf = ggml_backend_dev_buffer_from_host_ptr(dev, (char *) addr + first, last - first, max_size);
                 if (buf == nullptr) {
-                    throw std::runtime_error("unable to allocate backend metal buffer");
+                    throw std::runtime_error(format("unable to allocate %s buffer", ggml_backend_buft_name(buft)));
                 }
                 model.bufs.push_back(buf);
                 bufs.emplace(idx, buf);
             }
         }
-#endif
         else {
             ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
             if (buf == nullptr) {
-                throw std::runtime_error("unable to allocate backend buffer");
+                throw std::runtime_error(format("unable to allocate %s buffer", ggml_backend_buft_name(buft)));
             }
             model.bufs.push_back(buf);
             if (use_mlock && ggml_backend_buffer_is_host(buf)) {
@@ -9570,20 +9618,16 @@ static struct ggml_tensor * llm_build_kqv(
         cur = ggml_flash_attn_ext(ctx, q, k, v, kq_mask, kq_scale, hparams.f_max_alibi_bias,
                                   hparams.attn_soft_cap ? hparams.f_attn_logit_softcapping : 0.0f);
 
-        if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX || model.arch == LLM_ARCH_GEMMA2) {
-            ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
-        }
+        ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
 
         cur = ggml_reshape_2d(ctx, cur, n_embd_head_v*n_head, n_tokens);
     } else {
         struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q);
         cb(kq, "kq", il);
 
-        if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX || model.arch == LLM_ARCH_QWEN2 || model.arch == LLM_ARCH_NEMOTRON || model.arch == LLM_ARCH_CHATGLM) {
-            // for this arch, we need to perform the KQ multiplication with F32 precision, otherwise we get NaNs
-            // ref: https://github.com/ggerganov/llama.cpp/pull/4490#issuecomment-1859055847
-            ggml_mul_mat_set_prec(kq, GGML_PREC_F32);
-        }
+        // note: this op tends to require high floating point range
+        //       while for some models F16 is enough, for others it is not, so we default to F32 here
+        ggml_mul_mat_set_prec(kq, GGML_PREC_F32);
 
         if (model.arch == LLM_ARCH_GROK) {
             // need to do the following:
@@ -9592,9 +9636,6 @@ static struct ggml_tensor * llm_build_kqv(
             // kq = 30 * tanh(kq / 30)
             // before the softmax below
 
-            //try from phi2
-            //ggml_mul_mat_set_prec(kq, GGML_PREC_F32);
-
             kq = ggml_tanh(ctx, ggml_scale(ctx, kq, 0.08838834764831845f/30.0f));
             kq = ggml_scale(ctx, kq, 30);
         }
@@ -10020,7 +10061,7 @@ struct llm_build_context {
           llama_context  & lctx;
     const llama_hparams  & hparams;
     const llama_cparams  & cparams;
-    const llama_ubatch   & batch;
+    const llama_ubatch   & ubatch;
     const llama_kv_cache & kv_self;
 
     const int64_t n_embd;
@@ -10066,14 +10107,14 @@ struct llm_build_context {
     // TODO: consider making the entire interface noexcept
     llm_build_context(
         llama_context  & lctx,
-    const llama_ubatch & batch,
+    const llama_ubatch & ubatch,
     const llm_build_cb & cb,
                   bool   worst_case) :
         model            (lctx.model),
         lctx             (lctx),
         hparams          (model.hparams),
         cparams          (lctx.cparams),
-        batch            (batch),
+        ubatch           (ubatch),
         kv_self          (lctx.kv_self),
         n_embd           (hparams.n_embd),
         n_layer          (hparams.n_layer),
@@ -10095,7 +10136,7 @@ struct llm_build_context {
         beta_slow        (cparams.yarn_beta_slow),
         norm_eps         (hparams.f_norm_eps),
         norm_rms_eps     (hparams.f_norm_rms_eps),
-        n_tokens         (batch.n_tokens),
+        n_tokens         (ubatch.n_tokens),
         n_kv             (worst_case ? kv_self.size : kv_self.n),
         n_outputs        (worst_case ? n_tokens : lctx.n_outputs),
         n_outputs_enc    (worst_case ? n_tokens : lctx.embd_enc.size() / hparams.n_embd),
@@ -10464,7 +10505,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -10624,7 +10665,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = model.type == MODEL_7B ? build_inp_pos() : nullptr;
@@ -10739,7 +10780,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -10843,7 +10884,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -10965,7 +11006,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // multiply by embedding_multiplier_scale of 78.38367176906169
         inpL = ggml_scale(ctx0, inpL, 78.38367176906169f);
@@ -11123,7 +11164,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -11245,7 +11286,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -11348,7 +11389,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
         struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
@@ -11450,7 +11491,7 @@ struct llm_build_context {
         }
 
         // construct input embeddings (token, type, position)
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // token types are hardcoded to zero ("Sentence A")
         struct ggml_tensor * type_row0 = ggml_view_1d(ctx0, model.type_embd, n_embd, 0);
@@ -11637,7 +11678,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
         struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
@@ -11739,7 +11780,7 @@ struct llm_build_context {
         struct ggml_tensor * pos;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
         struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
@@ -11877,7 +11918,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -12027,7 +12068,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -12140,7 +12181,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -12255,7 +12296,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -12400,7 +12441,7 @@ struct llm_build_context {
         struct ggml_tensor * ffn_output;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -12519,7 +12560,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -12647,7 +12688,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -12752,7 +12793,7 @@ struct llm_build_context {
         struct ggml_tensor * pos;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -12857,7 +12898,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -12967,7 +13008,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -13085,7 +13126,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -13212,7 +13253,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // scale the input embeddings
         inpL = ggml_scale(ctx0, inpL, scale_embd);
@@ -13356,7 +13397,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // scale the input embeddings
         inpL = ggml_scale(ctx0, inpL, scale_embd);
@@ -13557,7 +13598,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
         cb(inpL, "inp_scaled", -1);
@@ -13665,7 +13706,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
         cb(inpL, "inp_scaled", -1);
@@ -13803,7 +13844,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -13919,7 +13960,7 @@ struct llm_build_context {
         struct ggml_tensor * inpL;
 
         // {n_embd, n_tokens}
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         struct ggml_tensor * state_copy = build_inp_s_copy();
         struct ggml_tensor * state_mask = build_inp_s_mask();
@@ -13931,7 +13972,7 @@ struct llm_build_context {
                     LLM_NORM_RMS, cb, il);
             cb(cur, "attn_norm", il);
 
-            cur = llm_build_mamba(ctx0, lctx, batch, gf, cur,
+            cur = llm_build_mamba(ctx0, lctx, ubatch, gf, cur,
                     state_copy, state_mask,
                     kv_head, n_kv, cb, il);
 
@@ -13977,7 +14018,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -14134,7 +14175,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -14262,7 +14303,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -14381,7 +14422,7 @@ struct llm_build_context {
 
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -14508,7 +14549,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -14653,7 +14694,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -14794,7 +14835,7 @@ struct llm_build_context {
         struct ggml_tensor * inpL;
 
         // {n_embd, n_tokens}
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -15009,7 +15050,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -15163,7 +15204,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         GGML_ASSERT(lctx.is_encoding);
         struct ggml_tensor * pos_bucket_enc = llm_build_pos_bucket(false);
@@ -15295,7 +15336,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         GGML_ASSERT(!lctx.is_encoding);
         GGML_ASSERT(n_outputs_enc > 0 && "call llama_encode() first");
@@ -15497,7 +15538,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
         struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
@@ -15589,7 +15630,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -15703,7 +15744,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -15827,7 +15868,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -15947,11 +15988,11 @@ struct llm_build_context {
         // Token shift state dimensions should be 2 * n_emb
         GGML_ASSERT(n_embd == hparams.n_embd_k_s() / 2);
 
-        const int64_t n_seqs = batch.n_seqs;
-        const int64_t n_seq_tokens = batch.n_seq_tokens;
-        const int64_t n_tokens = batch.n_tokens;
+        const int64_t n_seqs = ubatch.n_seqs;
+        const int64_t n_seq_tokens = ubatch.n_seq_tokens;
+        const int64_t n_tokens = ubatch.n_tokens;
         GGML_ASSERT(n_seqs != 0);
-        GGML_ASSERT(batch.equal_seqs);
+        GGML_ASSERT(ubatch.equal_seqs);
         GGML_ASSERT(n_tokens == n_seq_tokens * n_seqs);
 
         struct ggml_tensor * cur;
@@ -15959,7 +16000,7 @@ struct llm_build_context {
         struct ggml_tensor * state_copy = build_inp_s_copy();
         struct ggml_tensor * state_mask = build_inp_s_mask();
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
         inpL = llm_build_norm(ctx0, inpL, hparams, model.tok_norm, model.tok_norm_b, LLM_NORM, cb, -1);
 
         for (int il = 0; il < n_layer; ++il) {
@@ -16044,9 +16085,11 @@ struct llm_build_context {
         cur = ggml_get_rows(ctx0, cur, inp_out_ids);
 
         cur = llm_build_norm(ctx0, cur, hparams, model.output_norm, model.output_norm_b, LLM_NORM, cb, -1);
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_norm", -1);
 
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
         cb(cur, "result_output", -1);
+
         ggml_build_forward_expand(gf, cur);
 
         return gf;
@@ -16071,7 +16114,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -16267,7 +16310,7 @@ static struct ggml_cgraph * llama_build_graph_k_shift(llama_context & lctx) {
 
 static struct ggml_cgraph * llama_build_graph(
          llama_context & lctx,
-    const llama_ubatch & batch,
+    const llama_ubatch & ubatch,
                   bool   worst_case) {
     const auto & model = lctx.model;
 
@@ -16289,7 +16332,7 @@ static struct ggml_cgraph * llama_build_graph(
         // 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 = lctx.model.n_gpu_layers > (int)lctx.model.hparams.n_layer;
-        if (batch.n_tokens < 32 || full_offload) {
+        if (ubatch.n_tokens < 32 || full_offload) {
             if (il != -1 && strcmp(name, "norm") == 0) {
                 for (auto * backend : lctx.backends) {
                     if (ggml_backend_supports_buft(backend, lctx.model.buft_layer[il].buft) &&
@@ -16304,7 +16347,7 @@ static struct ggml_cgraph * llama_build_graph(
 
     struct ggml_cgraph * result = NULL;
 
-    struct llm_build_context llm(lctx, batch, cb, worst_case);
+    struct llm_build_context llm(lctx, ubatch, cb, worst_case);
 
     llm.init();
 
@@ -16555,7 +16598,7 @@ static int32_t llama_relative_position_bucket(llama_pos x, llama_pos y, uint64_t
     return relative_bucket;
 }
 
-static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
+static void llama_set_inputs(llama_context & lctx, const llama_ubatch & ubatch) {
     //
     // set input data
     //
@@ -16564,28 +16607,28 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
     const auto & cparams = lctx.cparams;
     const auto & kv_self = lctx.kv_self;
 
-    if (batch.token) {
-        const int64_t n_tokens = batch.n_tokens;
+    if (ubatch.token) {
+        const int64_t n_tokens = ubatch.n_tokens;
 
-        ggml_backend_tensor_set(lctx.inp_tokens, batch.token, 0, n_tokens*ggml_element_size(lctx.inp_tokens));
+        ggml_backend_tensor_set(lctx.inp_tokens, ubatch.token, 0, n_tokens*ggml_element_size(lctx.inp_tokens));
     }
 
-    if (batch.embd) {
+    if (ubatch.embd) {
         const int64_t n_embd   = hparams.n_embd;
-        const int64_t n_tokens = batch.n_tokens;
+        const int64_t n_tokens = ubatch.n_tokens;
 
-        ggml_backend_tensor_set(lctx.inp_embd, batch.embd, 0, n_tokens*n_embd*ggml_element_size(lctx.inp_embd));
+        ggml_backend_tensor_set(lctx.inp_embd, ubatch.embd, 0, n_tokens*n_embd*ggml_element_size(lctx.inp_embd));
     }
 
-    if (batch.pos && lctx.inp_pos) {
-        const int64_t n_tokens = batch.n_tokens;
+    if (ubatch.pos && lctx.inp_pos) {
+        const int64_t n_tokens = ubatch.n_tokens;
 
-        ggml_backend_tensor_set(lctx.inp_pos, batch.pos, 0, n_tokens*ggml_element_size(lctx.inp_pos));
+        ggml_backend_tensor_set(lctx.inp_pos, ubatch.pos, 0, n_tokens*ggml_element_size(lctx.inp_pos));
     }
 
     if (hparams.causal_attn || cparams.pooling_type == LLAMA_POOLING_TYPE_NONE) {
         GGML_ASSERT(lctx.inp_out_ids && "every model that can must skip unused outputs");
-        const int64_t n_tokens = batch.n_tokens;
+        const int64_t n_tokens = ubatch.n_tokens;
 
         GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_out_ids->buffer));
         int32_t * data = (int32_t *) lctx.inp_out_ids->data;
@@ -16594,10 +16637,10 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
             for (int i = 0; i < n_tokens; ++i) {
                 data[i] = i;
             }
-        } else if (batch.output) {
+        } else if (ubatch.output) {
             int32_t n_outputs = 0;
             for (int i = 0; i < n_tokens; ++i) {
-                if (batch.output[i]) {
+                if (ubatch.output[i]) {
                     data[n_outputs++] = i;
                 }
             }
@@ -16622,9 +16665,9 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
         // NOTE: hparams.causal_attn indicates the model is capable of generation and uses the kv cache.
         if (cparams.causal_attn && !lctx.is_encoding) {
             const int64_t n_kv         = kv_self.n;
-            const int64_t n_tokens     = batch.n_tokens;
-            const int64_t n_seq_tokens = batch.n_seq_tokens;
-            const int64_t n_seqs       = batch.n_seqs;
+            const int64_t n_tokens     = ubatch.n_tokens;
+            const int64_t n_seq_tokens = ubatch.n_seq_tokens;
+            const int64_t n_seqs       = ubatch.n_seqs;
 
 
             float * data     = nullptr;
@@ -16641,14 +16684,14 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
             }
 
             // For causal attention, use only the previous KV cells
-            // of the correct sequence for each token of the batch.
+            // of the correct sequence for each token of the ubatch.
             // It's assumed that if a token in the batch has multiple sequences, they are equivalent.
             for (int h = 0; h < 1; ++h) {
                 for (int s = 0; s < n_seqs; ++s) {
-                    const llama_seq_id seq_id = batch.seq_id[s][0];
+                    const llama_seq_id seq_id = ubatch.seq_id[s][0];
 
                     for (int j = 0; j < n_seq_tokens; ++j) {
-                        const llama_pos pos = batch.pos[s*n_seq_tokens + j];
+                        const llama_pos pos = ubatch.pos[s*n_seq_tokens + j];
 
                         for (int i = 0; i < n_kv; ++i) {
                             float f;
@@ -16694,9 +16737,9 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
                 }
             }
         } else {
-            const int64_t n_tokens     = batch.n_tokens;
-            const int64_t n_seq_tokens = batch.n_seq_tokens;
-            const int64_t n_seqs       = batch.n_seqs;
+            const int64_t n_tokens     = ubatch.n_tokens;
+            const int64_t n_seq_tokens = ubatch.n_seq_tokens;
+            const int64_t n_seqs       = ubatch.n_seqs;
             // when using kv cache, the mask needs to match the kv cache size
             const int64_t n_stride = hparams.causal_attn && !lctx.is_encoding ? kv_self.n : n_tokens;
 
@@ -16706,7 +16749,7 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
 
             for (int h = 0; h < 1; ++h) {
                 for (int s1 = 0; s1 < n_seqs; ++s1) {
-                    const llama_seq_id seq_id = batch.seq_id[s1][0];
+                    const llama_seq_id seq_id = ubatch.seq_id[s1][0];
 
                     for (int j = 0; j < n_seq_tokens; ++j) {
                         const int32_t tj = s1*n_seq_tokens + j;
@@ -16716,10 +16759,10 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
                                 const int32_t ti = s0*n_seq_tokens + i;
                                 float f = -INFINITY;
 
-                                for (int s = 0; s < batch.n_seq_id[s0]; ++s) {
-                                    if (batch.seq_id[s0][s] == seq_id) {
+                                for (int s = 0; s < ubatch.n_seq_id[s0]; ++s) {
+                                    if (ubatch.seq_id[s0][s] == seq_id) {
                                         if (hparams.use_alibi) {
-                                            f = -std::abs(batch.pos[ti] - batch.pos[tj]);
+                                            f = -std::abs(ubatch.pos[ti] - ubatch.pos[tj]);
                                         } else {
                                             f = 0.0f;
                                         }
@@ -16741,9 +16784,9 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
     }
 
     if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_MEAN) {
-        const int64_t n_tokens     = batch.n_tokens;
-        const int64_t n_seq_tokens = batch.n_seq_tokens;
-        const int64_t n_seqs       = batch.n_seqs;
+        const int64_t n_tokens     = ubatch.n_tokens;
+        const int64_t n_seq_tokens = ubatch.n_seq_tokens;
+        const int64_t n_seqs       = ubatch.n_seqs;
 
         GGML_ASSERT(lctx.inp_mean);
         GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_mean->buffer));
@@ -16754,12 +16797,12 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
         std::vector<uint64_t> sum(n_tokens, 0);
 
         for (int s = 0; s < n_seqs; ++s) {
-            const llama_seq_id seq_id = batch.seq_id[s][0];
+            const llama_seq_id seq_id = ubatch.seq_id[s][0];
 
-            // TODO: adapt limits to n_seqs when batch.equal_seqs is true
+            // TODO: adapt limits to n_seqs when ubatch.equal_seqs is true
             GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == MEAN");
 
-            sum[seq_id] += batch.n_seq_tokens;
+            sum[seq_id] += ubatch.n_seq_tokens;
         }
 
         std::vector<float> div(n_tokens, 0.0f);
@@ -16771,7 +16814,7 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
         }
 
         for (int s = 0; s < n_seqs; ++s) {
-            const llama_seq_id seq_id = batch.seq_id[s][0];
+            const llama_seq_id seq_id = ubatch.seq_id[s][0];
 
             for (int i = 0; i < n_seq_tokens; ++i) {
                 data[seq_id*n_tokens + s*n_seq_tokens + i] = div[seq_id];
@@ -16782,9 +16825,9 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
     if (cparams.embeddings && (
                 cparams.pooling_type == LLAMA_POOLING_TYPE_CLS ||
                 cparams.pooling_type == LLAMA_POOLING_TYPE_RANK)) {
-        const int64_t n_tokens     = batch.n_tokens;
-        const int64_t n_seq_tokens = batch.n_seq_tokens;
-        const int64_t n_seqs       = batch.n_seqs;
+        const int64_t n_tokens     = ubatch.n_tokens;
+        const int64_t n_seq_tokens = ubatch.n_seq_tokens;
+        const int64_t n_seqs       = ubatch.n_seqs;
 
         GGML_ASSERT(lctx.inp_cls);
         GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_cls->buffer));
@@ -16793,13 +16836,13 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
         memset(lctx.inp_cls->data, 0, n_tokens * ggml_element_size(lctx.inp_cls));
 
         for (int s = 0; s < n_seqs; ++s) {
-            const llama_seq_id seq_id = batch.seq_id[s][0];
+            const llama_seq_id seq_id = ubatch.seq_id[s][0];
 
-            // TODO: adapt limits to n_seqs when batch.equal_seqs is true
+            // TODO: adapt limits to n_seqs when ubatch.equal_seqs is true
             GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == CLS or RANK");
 
             for (int i = 0; i < n_seq_tokens; ++i) {
-                const llama_pos pos = batch.pos[s*n_seq_tokens + i];
+                const llama_pos pos = ubatch.pos[s*n_seq_tokens + i];
 
                 if (pos == 0) {
                     data[seq_id] = s*n_seq_tokens + i;
@@ -16809,9 +16852,9 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
     }
 
     if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_LAST) {
-        const int64_t n_tokens     = batch.n_tokens;
-        const int64_t n_seq_tokens = batch.n_seq_tokens;
-        const int64_t n_seqs       = batch.n_seqs;
+        const int64_t n_tokens     = ubatch.n_tokens;
+        const int64_t n_seq_tokens = ubatch.n_seq_tokens;
+        const int64_t n_seqs       = ubatch.n_seqs;
 
         GGML_ASSERT(lctx.inp_cls);
         GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_cls->buffer));
@@ -16823,13 +16866,13 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
         std::vector<int> last_row(n_tokens, -1);
 
         for (int s = 0; s < n_seqs; ++s) {
-            const llama_seq_id seq_id = batch.seq_id[s][0];
+            const llama_seq_id seq_id = ubatch.seq_id[s][0];
 
-            // TODO: adapt limits to n_seqs when batch.equal_seqs is true
+            // TODO: adapt limits to n_seqs when ubatch.equal_seqs is true
             GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == LAST");
 
             for (int i = 0; i < n_seq_tokens; ++i) {
-                const llama_pos pos = batch.pos[s*n_seq_tokens + i];
+                const llama_pos pos = ubatch.pos[s*n_seq_tokens + i];
 
                 if (pos >= last_pos[seq_id]) {
                     last_pos[seq_id] = pos;
@@ -16891,10 +16934,10 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
     }
 
     if (lctx.inp_pos_bucket) {
-        const int64_t n_tokens = batch.n_tokens;
+        const int64_t n_tokens = ubatch.n_tokens;
 
         GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_pos_bucket->buffer));
-        GGML_ASSERT(!batch.equal_seqs); // TODO: use batch.n_seqs instead of failing
+        GGML_ASSERT(!ubatch.equal_seqs); // TODO: use ubatch.n_seqs instead of failing
 
         int32_t * data = (int32_t *) lctx.inp_pos_bucket->data;
 
@@ -16903,7 +16946,7 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
             for (int h = 0; h < 1; ++h) {
                 for (int j = 0; j < n_tokens; ++j) {
                     for (int i = 0; i < n_kv; ++i) {
-                        data[h*(n_kv*n_tokens) + j*n_kv + i] = llama_relative_position_bucket(lctx.kv_self.cells[i].pos, batch.pos[j], hparams.n_rel_attn_bkts, lctx.is_encoding);
+                        data[h*(n_kv*n_tokens) + j*n_kv + i] = llama_relative_position_bucket(lctx.kv_self.cells[i].pos, ubatch.pos[j], hparams.n_rel_attn_bkts, lctx.is_encoding);
                     }
                 }
             }
@@ -16911,7 +16954,7 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
             for (int h = 0; h < 1; ++h) {
                 for (int j = 0; j < n_tokens; ++j) {
                     for (int i = 0; i < n_tokens; ++i) {
-                        data[h*(n_tokens*n_tokens) + j*n_tokens + i] = llama_relative_position_bucket(batch.pos[i], batch.pos[j], hparams.n_rel_attn_bkts, lctx.is_encoding);
+                        data[h*(n_tokens*n_tokens) + j*n_tokens + i] = llama_relative_position_bucket(ubatch.pos[i], ubatch.pos[j], hparams.n_rel_attn_bkts, lctx.is_encoding);
                     }
                 }
             }
@@ -16927,10 +16970,10 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
 
     if (!lctx.is_encoding && lctx.inp_KQ_mask_cross) {
         const int64_t n_output_enc = lctx.embd_enc.size() / hparams.n_embd;
-        const int64_t n_tokens = batch.n_tokens;
+        const int64_t n_tokens = ubatch.n_tokens;
 
         GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_KQ_mask_cross->buffer));
-        GGML_ASSERT(!batch.equal_seqs); // TODO: use batch.n_seqs instead of failing
+        GGML_ASSERT(!ubatch.equal_seqs); // TODO: use ubatch.n_seqs instead of failing
 
         float * data = (float *) lctx.inp_KQ_mask_cross->data;
 
@@ -16938,8 +16981,8 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
             for (int j = 0; j < n_tokens; ++j) {
                 for (int i = 0; i < n_output_enc; ++i) {
                     float f = -INFINITY;
-                    for (int s = 0; s < batch.n_seq_id[j]; ++s) {
-                        const llama_seq_id seq_id = batch.seq_id[j][s];
+                    for (int s = 0; s < ubatch.n_seq_id[j]; ++s) {
+                        const llama_seq_id seq_id = ubatch.seq_id[j][s];
                         if (lctx.seq_ids_enc[i].find(seq_id) != lctx.seq_ids_enc[i].end()) {
                             f = 0.0f;
                         }
@@ -17068,17 +17111,19 @@ static void llama_graph_compute(
                     int   n_threads,
         ggml_threadpool * threadpool) {
     if (lctx.backend_cpu != nullptr) {
-        ggml_backend_cpu_set_n_threads(lctx.backend_cpu, n_threads);
         ggml_backend_cpu_set_threadpool(lctx.backend_cpu, threadpool);
         ggml_backend_cpu_set_abort_callback(lctx.backend_cpu, lctx.abort_callback, lctx.abort_callback_data);
     }
-#ifdef GGML_USE_BLAS
-    if (lctx.backend_blas != nullptr) {
-        ggml_backend_blas_set_n_threads(lctx.backend_blas, n_threads);
-    }
-#endif
 
-    ggml_backend_sched_graph_compute_async(lctx.sched, gf);
+    // set the number of threads for all the backends
+    for (const auto & set_n_threads_fn : lctx.set_n_threads_fns) {
+        set_n_threads_fn.second(set_n_threads_fn.first, n_threads);
+    }
+
+    auto err = ggml_backend_sched_graph_compute_async(lctx.sched, gf);
+    if (err != GGML_STATUS_SUCCESS) {
+        LLAMA_LOG_ERROR("%s: ggml_backend_sched_graph_compute_async failed with error %d\n", __func__, err);
+    }
 
     // fprintf(stderr, "splits: %d\n", ggml_backend_sched_get_n_splits(lctx.sched));
 }
@@ -17094,26 +17139,30 @@ static void llama_graph_compute(
 //
 static int llama_decode_internal(
          llama_context & lctx,
-           llama_batch   batch_all) { // TODO: rename back to batch
+           llama_batch   inp_batch) {
 
     lctx.is_encoding = false;
-    const uint32_t n_tokens_all = batch_all.n_tokens;
 
-    if (n_tokens_all == 0) {
+    if (inp_batch.n_tokens == 0) {
         LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__);
         return -1;
     }
 
+    // temporary allocate memory for the input batch if needed
+    llama_batch_allocr batch_allocr(lctx, inp_batch);
+    const llama_batch & batch = batch_allocr.batch;
+    const uint32_t n_tokens_all = batch.n_tokens;
+
     const auto & model   = lctx.model;
     const auto & hparams = model.hparams;
     const auto & cparams = lctx.cparams;
 
-    GGML_ASSERT((!batch_all.token && batch_all.embd) || (batch_all.token && !batch_all.embd)); // NOLINT
+    GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT
 
-    if (batch_all.token) {
+    if (batch.token) {
         for (uint32_t i = 0; i < n_tokens_all; ++i) {
-            if (batch_all.token[i] < 0 || (uint32_t)batch_all.token[i] >= model.vocab.n_vocab) {
-                LLAMA_LOG_ERROR("%s: invalid token[%d] = %d\n", __func__, i, batch_all.token[i]);
+            if (batch.token[i] < 0 || (uint32_t)batch.token[i] >= model.vocab.n_vocab) {
+                LLAMA_LOG_ERROR("%s: invalid token[%d] = %d\n", __func__, i, batch.token[i]);
                 return -1;
             }
         }
@@ -17144,9 +17193,9 @@ static int llama_decode_internal(
     lctx.embd_seq.clear();
 
     // count outputs
-    if (batch_all.logits && !embd_pooled) {
+    if (batch.logits && !embd_pooled) {
         for (uint32_t i = 0; i < n_tokens_all; ++i) {
-            n_outputs += batch_all.logits[i] != 0;
+            n_outputs += batch.logits[i] != 0;
         }
     } else if (lctx.logits_all || embd_pooled) {
         n_outputs = n_tokens_all;
@@ -17155,7 +17204,7 @@ static int llama_decode_internal(
         n_outputs = 1;
     }
 
-    lctx.sbatch.from_batch(batch_all, n_embd,
+    lctx.sbatch.from_batch(batch, n_embd,
         /* simple_split */ !kv_self.recurrent,
         /* logits_all   */ n_outputs == n_tokens_all);
 
@@ -17408,17 +17457,20 @@ static int llama_decode_internal(
 //
 static int llama_encode_internal(
          llama_context & lctx,
-           llama_batch   batch) {
+           llama_batch   inp_batch) {
 
     lctx.is_encoding = true;
 
-    const uint32_t n_tokens = batch.n_tokens;
-
-    if (n_tokens == 0) {
+    if (inp_batch.n_tokens == 0) {
         LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__);
         return -1;
     }
 
+    // temporary allocate memory for the input batch if needed
+    llama_batch_allocr batch_allocr(lctx, inp_batch);
+    const llama_batch & batch = batch_allocr.batch;
+    const uint32_t n_tokens = batch.n_tokens;
+
     const auto & model   = lctx.model;
     const auto & hparams = model.hparams;
     const auto & cparams = lctx.cparams;
@@ -17894,10 +17946,9 @@ static void llama_tensor_dequantize_internal(
     }
     float * f32_output = (float *) output.data();
 
-    ggml_type_traits_t qtype;
+    const ggml_type_traits * qtype = ggml_get_type_traits(tensor->type);
     if (ggml_is_quantized(tensor->type)) {
-        qtype = ggml_internal_get_type_traits(tensor->type);
-        if (qtype.to_float == NULL) {
+        if (qtype->to_float == NULL) {
             throw std::runtime_error(format("type %s unsupported for integer quantization: no dequantization available", ggml_type_name(tensor->type)));
         }
     } else if (tensor->type != GGML_TYPE_F16 &&
@@ -17911,7 +17962,7 @@ static void llama_tensor_dequantize_internal(
         } else if (tensor->type == GGML_TYPE_BF16) {
             ggml_bf16_to_fp32_row((ggml_bf16_t *)tensor->data, f32_output, nelements);
         } else if (ggml_is_quantized(tensor->type)) {
-            qtype.to_float(tensor->data, f32_output, nelements);
+            qtype->to_float(tensor->data, f32_output, nelements);
         } else {
             GGML_ABORT("fatal error"); // unreachable
         }
@@ -17947,7 +17998,7 @@ static void llama_tensor_dequantize_internal(
             } else if (typ == GGML_TYPE_BF16) {
                 ggml_bf16_to_fp32_row((ggml_bf16_t *)inbuf, outbuf, nels);
             } else {
-                qtype.to_float(inbuf, outbuf, nels);
+                qtype->to_float(inbuf, outbuf, nels);
             }
         };
         workers.emplace_back(compute, tensor->type, (uint8_t *) tensor->data + in_buff_offs, f32_output + out_buff_offs, thr_elems);
@@ -19041,16 +19092,20 @@ bool llama_supports_mlock(void) {
 }
 
 bool llama_supports_gpu_offload(void) {
-#if defined(GGML_USE_METAL)   || defined(GGML_USE_VULKAN) || \
-    defined(GGML_USE_SYCL) || defined(GGML_USE_KOMPUTE) || defined(GGML_USE_RPC)
+#if defined(GGML_USE_KOMPUTE)
     // Defined when llama.cpp is compiled with support for offloading model layers to GPU.
     return true;
 #else
     return ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU) != nullptr ||
-        ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU_FULL) != nullptr;
+           ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU_FULL) != nullptr ||
+           llama_supports_rpc();
 #endif
 }
 
+bool llama_supports_rpc(void) {
+    return ggml_backend_reg_by_name("RPC") != nullptr;
+}
+
 void llama_backend_init(void) {
     ggml_time_init();
 
@@ -19125,14 +19180,56 @@ struct llama_model * llama_load_model_from_file(
         model->rpc_servers.push_back(servers);
     }
 
+    // add RPC devices
+    if (!model->rpc_servers.empty()) {
+        ggml_backend_reg_t rpc_reg = ggml_backend_reg_by_name("RPC");
+        if (!rpc_reg) {
+            LLAMA_LOG_ERROR("%s: failed to find RPC backend\n", __func__);
+            llama_free_model(model);
+            return nullptr;
+        }
+
+        // ggml_backend_dev_t ggml_backend_rpc_add_device(const char * endpoint);
+        using ggml_backend_rpc_add_device_t = ggml_backend_dev_t (*)(const char *);
+        ggml_backend_rpc_add_device_t ggml_backend_rpc_add_device_fn = (ggml_backend_rpc_add_device_t) ggml_backend_reg_get_proc_address(rpc_reg, "ggml_backend_rpc_add_device");
+        if (!ggml_backend_rpc_add_device_fn) {
+            LLAMA_LOG_ERROR("%s: failed to find RPC device add function\n", __func__);
+            llama_free_model(model);
+            return nullptr;
+        }
+
+        for (const std::string & server : model->rpc_servers) {
+            ggml_backend_dev_t dev = ggml_backend_rpc_add_device_fn(server.c_str());
+            if (dev) {
+                model->devices.push_back(dev);
+            } else {
+                LLAMA_LOG_ERROR("%s: failed to add RPC device for server '%s'\n", __func__, server.c_str());
+                llama_free_model(model);
+                return nullptr;
+            }
+        }
+    }
+
     // create list of devices to use with this model
     // currently, we use all available devices
     // TODO: rework API to give user more control over device selection
     for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
         ggml_backend_dev_t dev = ggml_backend_dev_get(i);
-        // skip the CPU backend since it is handled separately
-        if (ggml_backend_dev_type(dev) != GGML_BACKEND_DEVICE_TYPE_CPU_FULL) {
-            model->devices.push_back(dev);
+        switch (ggml_backend_dev_type(dev)) {
+            case GGML_BACKEND_DEVICE_TYPE_CPU:
+            case GGML_BACKEND_DEVICE_TYPE_CPU_FULL:
+                // skip CPU backends since they are `handled separately
+                break;
+
+            case GGML_BACKEND_DEVICE_TYPE_GPU:
+            case GGML_BACKEND_DEVICE_TYPE_GPU_FULL:
+            {
+                size_t free, total; // NOLINT
+                ggml_backend_dev_memory(dev, &free, &total);
+                LLAMA_LOG_INFO("%s: using device %s (%s) - %zu MiB free\n", __func__, ggml_backend_dev_name(dev), ggml_backend_dev_description(dev), free/1024/1024);
+                model->devices.push_back(dev);
+                break;
+            }
         }
     }
 
@@ -19144,7 +19241,7 @@ struct llama_model * llama_load_model_from_file(
         } else if (status == -2) {
             LLAMA_LOG_INFO("%s: cancelled model load\n", __func__);
         }
-        delete model;
+        llama_free_model(model);
         return nullptr;
     }
 
@@ -19184,7 +19281,7 @@ struct llama_context * llama_new_context_with_model(
         params.flash_attn = false;
     }
 
-    if (params.type_v != GGML_TYPE_F16 && !params.flash_attn) {
+    if (ggml_is_quantized(params.type_v) && !params.flash_attn) {
         LLAMA_LOG_ERROR("%s: V cache quantization requires flash_attn\n", __func__);
         return nullptr;
     }
@@ -19327,81 +19424,7 @@ struct llama_context * llama_new_context_with_model(
             main_gpu -= (int)model->devices.size();
         }
 
-#if defined(GGML_USE_RPC)
-        if (model->n_gpu_layers > 0) {
-            for (const auto & endpoint : model->rpc_servers) {
-                ggml_backend_t backend = ggml_backend_rpc_init(endpoint.c_str());
-                if (backend == nullptr) {
-                    LLAMA_LOG_ERROR("%s: failed to initialize RPC to '%s'\n", __func__, endpoint.c_str());
-                    llama_free(ctx);
-                    return nullptr;
-                }
-                ctx->backends.push_back(backend);
-            }
-        }
-        if (main_gpu >= (int)model->rpc_servers.size()) {
-            main_gpu -= (int)model->rpc_servers.size();
-        }
-#endif
-
-#if defined(GGML_USE_METAL)
-        if (model->n_gpu_layers > 0) {
-            ctx->backend_metal = ggml_backend_metal_init();
-            if (ctx->backend_metal == nullptr) {
-                LLAMA_LOG_ERROR("%s: failed to initialize Metal backend\n", __func__);
-                llama_free(ctx);
-                return nullptr;
-            }
-            ctx->backends.push_back(ctx->backend_metal);
-        }
-#elif defined(GGML_USE_VULKAN)
-        if (model->split_mode == LLAMA_SPLIT_MODE_ROW) {
-            LLAMA_LOG_ERROR("%s: Row split not supported. Failed to initialize Vulkan backend\n", __func__);
-            llama_free(ctx);
-            return nullptr;
-        }
-        if (model->split_mode == LLAMA_SPLIT_MODE_NONE) {
-            ggml_backend_t backend = ggml_backend_vk_init(main_gpu);
-            if (backend == nullptr) {
-                LLAMA_LOG_ERROR("%s: failed to initialize Vulkan backend\n", __func__);
-                llama_free(ctx);
-                return nullptr;
-            }
-            ctx->backends.push_back(backend);
-        } else {
-            for (int device = 0; device < ggml_backend_vk_get_device_count(); ++device) {
-                ggml_backend_t backend = ggml_backend_vk_init(device);
-                if (backend == nullptr) {
-                    LLAMA_LOG_ERROR("%s: failed to initialize Vulkan%d backend\n", __func__, device);
-                    llama_free(ctx);
-                    return nullptr;
-                }
-                ctx->backends.push_back(backend);
-            }
-        }
-#elif defined(GGML_USE_SYCL)
-        // with split_mode LLAMA_SPLIT_MODE_NONE or LLAMA_SPLIT_MODE_ROW, only the main GPU backend is used
-        if (model->split_mode == LLAMA_SPLIT_MODE_NONE || model->split_mode == LLAMA_SPLIT_MODE_ROW) {
-            ggml_backend_t backend = ggml_backend_sycl_init(main_gpu);
-            if (backend == nullptr) {
-                LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d backend\n", __func__, main_gpu);
-                llama_free(ctx);
-                return nullptr;
-            }
-            ctx->backends.push_back(backend);
-        } else {
-            // LLAMA_SPLIT_LAYER requires a backend for each GPU
-            for (int i = 0; i < ggml_backend_sycl_get_device_count(); ++i) {
-                ggml_backend_t backend = ggml_backend_sycl_init(i);
-                if (backend == nullptr) {
-                    LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d for No.%d backend\n", __func__, i, i);
-                    llama_free(ctx);
-                    return nullptr;
-                }
-                ctx->backends.push_back(backend);
-            }
-        }
-#elif defined(GGML_USE_KOMPUTE)
+#if defined(GGML_USE_KOMPUTE)
         if (model->n_gpu_layers > 0) {
             auto * backend = ggml_backend_kompute_init(main_gpu);
             if (backend == nullptr) {
@@ -19411,40 +19434,21 @@ struct llama_context * llama_new_context_with_model(
             }
             ctx->backends.push_back(backend);
         }
-#elif defined(GGML_USE_CANN)
-        // with split_mode LLAMA_SPLIT_MODE_NONE or LLAMA_SPLIT_MODE_ROW, only the main GPU backend is used
-        // TODO: ggml_backend_cann is not support split tensor now, just leave code here.
-        if (model->split_mode == LLAMA_SPLIT_MODE_NONE || model->split_mode == LLAMA_SPLIT_MODE_ROW) {
-            ggml_backend_t backend = ggml_backend_cann_init(main_gpu);
-            if (backend == nullptr) {
-                LLAMA_LOG_ERROR("%s: failed to initialize CANN%d backend\n", __func__, main_gpu);
-                llama_free(ctx);
-                return nullptr;
-            }
-            ctx->backends.push_back(backend);
-        } else {
-            // LLAMA_SPLIT_MODE_LAYER requires a backend for each GPU
-            // TODO: currently, CANN can't use multi-gpus, just leave code here for further cann version.
-            for (int32_t device = 0; device < ggml_backend_cann_get_device_count(); ++device) {
-                ggml_backend_t backend = ggml_backend_cann_init(device);
+#endif
+
+        // add other backends (such as BLAS)
+        for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
+            ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+            if (ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_CPU) {
+                ggml_backend_t backend = ggml_backend_dev_init(dev, nullptr);
                 if (backend == nullptr) {
-                    LLAMA_LOG_ERROR("%s: failed to initialize CANN%d backend\n", __func__, device);
+                    LLAMA_LOG_ERROR("%s: failed to initialize %s backend\n", __func__, ggml_backend_dev_name(dev));
                     llama_free(ctx);
                     return nullptr;
                 }
                 ctx->backends.push_back(backend);
             }
         }
-#endif
-
-#ifdef GGML_USE_BLAS
-        ctx->backend_blas = ggml_backend_blas_init();
-        if (ctx->backend_blas == nullptr) {
-            LLAMA_LOG_WARN("%s: failed to initialize BLAS backend\n", __func__);
-        } else {
-            ctx->backends.push_back(ctx->backend_blas);
-        }
-#endif
 
         ctx->backend_cpu = ggml_backend_cpu_init();
         if (ctx->backend_cpu == nullptr) {
@@ -19454,6 +19458,18 @@ struct llama_context * llama_new_context_with_model(
         }
         ctx->backends.push_back(ctx->backend_cpu);
 
+        // create a list of the set_n_threads functions in the backends
+        for (auto * backend : ctx->backends) {
+            ggml_backend_dev_t dev = ggml_backend_get_device(backend);
+            ggml_backend_reg_t reg = dev ? ggml_backend_dev_backend_reg(dev) : nullptr;
+            if (reg) {
+                auto ggml_backend_set_n_threads_fn = (ggml_backend_set_n_threads_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_n_threads");
+                if (ggml_backend_set_n_threads_fn) {
+                    ctx->set_n_threads_fns.emplace_back(backend, ggml_backend_set_n_threads_fn);
+                }
+            }
+        }
+
         if (!llama_kv_cache_init(ctx->kv_self, ctx, type_k, type_v, kv_size, cparams.offload_kqv)) {
             LLAMA_LOG_ERROR("%s: llama_kv_cache_init() failed for self-attention cache\n", __func__);
             llama_free(ctx);
@@ -19473,7 +19489,7 @@ struct llama_context * llama_new_context_with_model(
             }
 
             LLAMA_LOG_INFO("%s: KV self size  = %7.2f MiB, K (%s): %7.2f MiB, V (%s): %7.2f MiB\n", __func__,
-                (float)(memory_size_k + memory_size_v) / (1024.0f * 1024.0f),
+                      (float)(memory_size_k + memory_size_v) / (1024.0f * 1024.0f),
                 ggml_type_name(type_k), (float)memory_size_k / (1024.0f * 1024.0f),
                 ggml_type_name(type_v), (float)memory_size_v / (1024.0f * 1024.0f));
         }
@@ -21069,9 +21085,7 @@ void llama_set_causal_attn(struct llama_context * ctx, bool causal_attn) {
 
 struct llama_batch llama_batch_get_one(
              llama_token * tokens,
-                 int32_t   n_tokens,
-               llama_pos   pos_0,
-            llama_seq_id   seq_id) {
+                 int32_t   n_tokens) {
     return {
         /*n_tokens       =*/ n_tokens,
         /*tokens         =*/ tokens,
@@ -21080,9 +21094,6 @@ struct llama_batch llama_batch_get_one(
         /*n_seq_id       =*/ nullptr,
         /*seq_id         =*/ nullptr,
         /*logits         =*/ nullptr,
-        /*all_pos_0      =*/ pos_0,
-        /*all_pos_1      =*/ 1,
-        /*all_seq_id     =*/ seq_id,
     };
 }
 
@@ -21095,9 +21106,6 @@ struct llama_batch llama_batch_init(int32_t n_tokens_alloc, int32_t embd, int32_
         /*n_seq_id       =*/ nullptr,
         /*seq_id         =*/ nullptr,
         /*logits         =*/ nullptr,
-        /*all_pos_0      =*/ 0,
-        /*all_pos_1      =*/ 0,
-        /*all_seq_id     =*/ 0,
     };
 
     if (embd) {
@@ -21137,7 +21145,7 @@ int32_t llama_encode(
         struct llama_context * ctx,
           struct llama_batch   batch) {
     const int ret = llama_encode_internal(*ctx, batch);
-    if (ret < 0) {
+    if (ret != 0) {
         LLAMA_LOG_ERROR("%s: failed to encode, ret = %d\n", __func__, ret);
     }
 
@@ -21148,7 +21156,7 @@ int32_t llama_decode(
         struct llama_context * ctx,
           struct llama_batch   batch) {
     const int ret = llama_decode_internal(*ctx, batch);
-    if (ret < 0) {
+    if (ret != 0) {
         LLAMA_LOG_ERROR("%s: failed to decode, ret = %d\n", __func__, ret);
     }
 
@@ -21327,6 +21335,10 @@ llama_token llama_token_eos(const struct llama_model * model) {
     return llama_token_eos_impl(model->vocab);
 }
 
+llama_token llama_token_eot(const struct llama_model * model) {
+    return llama_token_eot_impl(model->vocab);
+}
+
 llama_token llama_token_cls(const struct llama_model * model) {
     return llama_token_cls_impl(model->vocab);
 }
@@ -21363,8 +21375,28 @@ llama_token llama_token_suffix(const struct llama_model * model) {
     return llama_token_suffix_impl(model->vocab);
 }
 
-llama_token llama_token_eot(const struct llama_model * model) {
-    return llama_token_eot_impl(model->vocab);
+llama_token llama_token_fim_pre(const struct llama_model * model) {
+    return llama_token_fim_pre_impl(model->vocab);
+}
+
+llama_token llama_token_fim_suf(const struct llama_model * model) {
+    return llama_token_fim_suf_impl(model->vocab);
+}
+
+llama_token llama_token_fim_mid(const struct llama_model * model) {
+    return llama_token_fim_mid_impl(model->vocab);
+}
+
+llama_token llama_token_fim_pad(const struct llama_model * model) {
+    return llama_token_fim_pad_impl(model->vocab);
+}
+
+llama_token llama_token_fim_rep(const struct llama_model * model) {
+    return llama_token_fim_rep_impl(model->vocab);
+}
+
+llama_token llama_token_fim_sep(const struct llama_model * model) {
+    return llama_token_fim_sep_impl(model->vocab);
 }
 
 //
@@ -21664,6 +21696,16 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "[|assistant|]";
         }
+    } else if (tmpl == "rwkv-world" || tmpl_contains("rwkv-world")) {
+        // this template requires the model to have "\n\n" as EOT token
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "user") {
+                ss << "User: " << message->content << "\n\nAssistant:";
+            } else {
+                ss << message->content << "\n\n";
+            }
+        }
     } else {
         // template not supported
         return -1;
@@ -21722,6 +21764,14 @@ struct llama_sampler * llama_sampler_init_grammar(const struct llama_model * mod
     return llama_sampler_init_grammar_impl(model->vocab, grammar_str, grammar_root);
 }
 
+struct llama_sampler * llama_sampler_init_infill(const struct llama_model * model) {
+    return llama_sampler_init_infill_impl(model->vocab);
+}
+
+struct llama_sampler * llama_sampler_init_dry(const struct llama_model * model, float dry_multiplier, float dry_base, int32_t dry_allowed_length, int32_t dry_penalty_last_n, const char** seq_breakers, size_t num_breakers) {
+    return llama_sampler_init_dry_impl(model->vocab, llama_n_ctx_train(model), dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n, seq_breakers, num_breakers);
+}
+
 //
 // model split
 //
@@ -21761,6 +21811,7 @@ const char * llama_print_system_info(void) {
     s += "AVX512_VBMI = " + std::to_string(ggml_cpu_has_avx512_vbmi()) + " | ";
     s += "AVX512_VNNI = " + std::to_string(ggml_cpu_has_avx512_vnni()) + " | ";
     s += "AVX512_BF16 = " + std::to_string(ggml_cpu_has_avx512_bf16()) + " | ";
+    s += "AMX_INT8 = "    + std::to_string(ggml_cpu_has_amx_int8())    + " | ";
     s += "FMA = "         + std::to_string(ggml_cpu_has_fma())         + " | ";
     s += "NEON = "        + std::to_string(ggml_cpu_has_neon())        + " | ";
     s += "SVE = "         + std::to_string(ggml_cpu_has_sve())         + " | ";
diff --git a/examples/talk-llama/llama.h b/examples/talk-llama/llama.h
index 7cae1bbe..b2d1e7d5 100644
--- a/examples/talk-llama/llama.h
+++ b/examples/talk-llama/llama.h
@@ -217,6 +217,7 @@ extern "C" {
 
     typedef struct llama_token_data_array {
         // TODO: consider SoA
+        // NOTE: this pointer can be modified by the samplers
         llama_token_data * data;
         size_t size;
         int64_t selected; // this is the index in the data array (i.e. not the token id)
@@ -232,8 +233,11 @@ extern "C" {
     // - token  : the token ids of the input (used when embd is NULL)
     // - embd   : token embeddings (i.e. float vector of size n_embd) (used when token is NULL)
     // - pos    : the positions of the respective token in the sequence
+    //            (if set to NULL, the token position will be tracked automatically by llama_decode)
     // - seq_id : the sequence to which the respective token belongs
+    //            (if set to NULL, the sequence ID will be assumed to be 0)
     // - logits : if zero, the logits (and/or the embeddings) for the respective token will not be output
+    //            (if set to NULL, only the logits for last token will be returned)
     //
     typedef struct llama_batch {
         int32_t n_tokens;
@@ -244,15 +248,6 @@ extern "C" {
         int32_t      *  n_seq_id;
         llama_seq_id ** seq_id;
         int8_t       *  logits; // TODO: rename this to "output"
-
-        // NOTE: helpers for smooth API transition - can be deprecated in the future
-        //       for future-proof code, use the above fields instead and ignore everything below
-        //
-        // pos[i] = all_pos_0 + i*all_pos_1
-        //
-        llama_pos    all_pos_0;  // used if pos == NULL
-        llama_pos    all_pos_1;  // used if pos == NULL
-        llama_seq_id all_seq_id; // used if seq_id == NULL
     } llama_batch;
 
     enum llama_model_kv_override_type {
@@ -433,6 +428,7 @@ extern "C" {
     LLAMA_API bool llama_supports_mmap       (void);
     LLAMA_API bool llama_supports_mlock      (void);
     LLAMA_API bool llama_supports_gpu_offload(void);
+    LLAMA_API bool llama_supports_rpc        (void);
 
     LLAMA_API uint32_t llama_n_ctx      (const struct llama_context * ctx);
     LLAMA_API uint32_t llama_n_batch    (const struct llama_context * ctx);
@@ -775,15 +771,15 @@ extern "C" {
     // Decoding
     //
 
-    // Return batch for single sequence of tokens starting at pos_0
+    // Return batch for single sequence of tokens
+    // The sequence ID will be fixed to 0
+    // The position of the tokens will be tracked automatically by llama_decode
     //
     // NOTE: this is a helper function to facilitate transition to the new batch API - avoid using it
     //
     LLAMA_API struct llama_batch llama_batch_get_one(
                   llama_token * tokens,
-                      int32_t   n_tokens,
-                    llama_pos   pos_0,
-                 llama_seq_id   seq_id);
+                      int32_t   n_tokens);
 
     // Allocates a batch of tokens on the heap that can hold a maximum of n_tokens
     // Each token can be assigned up to n_seq_max sequence ids
@@ -896,6 +892,7 @@ extern "C" {
     // Special tokens
     LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence
     LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence
+    LLAMA_API llama_token llama_token_eot(const struct llama_model * model); // end-of-turn
     LLAMA_API llama_token llama_token_cls(const struct llama_model * model); // classification
     LLAMA_API llama_token llama_token_sep(const struct llama_model * model); // sentence separator
     LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line
@@ -904,11 +901,17 @@ extern "C" {
     LLAMA_API bool llama_add_bos_token(const struct llama_model * model);
     LLAMA_API bool llama_add_eos_token(const struct llama_model * model);
 
-    // Codellama infill tokens
-    LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix
-    LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle
-    LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix
-    LLAMA_API llama_token llama_token_eot   (const struct llama_model * model); // End of infill middle
+    // infill tokens
+    DEPRECATED(LLAMA_API llama_token llama_token_prefix(const struct llama_model * model), "use llama_token_fim_pre instead");
+    DEPRECATED(LLAMA_API llama_token llama_token_middle(const struct llama_model * model), "use llama_token_fim_mid instead");
+    DEPRECATED(LLAMA_API llama_token llama_token_suffix(const struct llama_model * model), "use llama_token_fim_suf instead");
+
+    LLAMA_API llama_token llama_token_fim_pre(const struct llama_model * model);
+    LLAMA_API llama_token llama_token_fim_suf(const struct llama_model * model);
+    LLAMA_API llama_token llama_token_fim_mid(const struct llama_model * model);
+    LLAMA_API llama_token llama_token_fim_pad(const struct llama_model * model);
+    LLAMA_API llama_token llama_token_fim_rep(const struct llama_model * model);
+    LLAMA_API llama_token llama_token_fim_sep(const struct llama_model * model);
 
     //
     // Tokenization
@@ -1067,12 +1070,13 @@ extern "C" {
 
     // available samplers:
 
-    LLAMA_API struct llama_sampler * llama_sampler_init_greedy     (void);
-    LLAMA_API struct llama_sampler * llama_sampler_init_dist       (uint32_t seed);
+    LLAMA_API struct llama_sampler * llama_sampler_init_greedy(void);
+    LLAMA_API struct llama_sampler * llama_sampler_init_dist  (uint32_t seed);
 
     /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
     /// NOTE: Avoid using on the full vocabulary as the sorting can become slow. For example, apply top-k or top-p sampling first.
-    LLAMA_API struct llama_sampler * llama_sampler_init_softmax    (void);
+    DEPRECATED(LLAMA_API struct llama_sampler * llama_sampler_init_softmax    (void),
+        "will be removed in the future (see https://github.com/ggerganov/llama.cpp/pull/9896#discussion_r1800920915)");
 
     /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
     LLAMA_API struct llama_sampler * llama_sampler_init_top_k      (int32_t k);
@@ -1088,11 +1092,16 @@ extern "C" {
 
     /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
     LLAMA_API struct llama_sampler * llama_sampler_init_typical    (float   p, size_t min_keep);
+
+    /// #details Updates the logits l_i` = l_i/t. When t <= 0.0f, the maximum logit is kept at it's original value, the rest are set to -inf
     LLAMA_API struct llama_sampler * llama_sampler_init_temp       (float   t);
 
     /// @details Dynamic temperature implementation (a.k.a. entropy) described in the paper https://arxiv.org/abs/2309.02772.
     LLAMA_API struct llama_sampler * llama_sampler_init_temp_ext   (float   t, float   delta, float exponent);
 
+    /// @details XTC sampler as described in https://github.com/oobabooga/text-generation-webui/pull/6335
+    LLAMA_API struct llama_sampler * llama_sampler_init_xtc        (float   p, float   t,     size_t min_keep, uint32_t seed);
+
     /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
     /// @param tau  The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
@@ -1132,11 +1141,43 @@ extern "C" {
                                 bool   penalize_nl,     // consider newlines as a repeatable token
                                 bool   ignore_eos);     // ignore the end-of-sequence token
 
+    ///  @details DRY sampler, designed by p-e-w, as described in: https://github.com/oobabooga/text-generation-webui/pull/5677, porting Koboldcpp implementation authored by pi6am: https://github.com/LostRuins/koboldcpp/pull/982
+    LLAMA_API struct llama_sampler *    llama_sampler_init_dry(
+            const struct llama_model *  model,
+                               float    dry_multiplier,
+                               float    dry_base,
+                             int32_t    dry_allowed_length,
+                             int32_t    dry_penalty_last_n,
+                          const char ** seq_breakers,
+                              size_t    num_breakers);
+
     LLAMA_API struct llama_sampler * llama_sampler_init_logit_bias(
                              int32_t   n_vocab,
                              int32_t   n_logit_bias,
               const llama_logit_bias * logit_bias);
 
+    // this sampler is meant to be used for fill-in-the-middle infilling
+    // it's supposed to be used after top_k + top_p sampling
+    //
+    // 1. if the sum of the EOG probs times the number of candidates is higher than the sum of the other probs -> pick EOG
+    // 2. combine probs of tokens that have the same prefix
+    //
+    // example:
+    //
+    // - before:
+    //   "hel":   0.5
+    //   "hell":  0.2
+    //   "hello": 0.1
+    //   "dummy": 0.1
+    //
+    // - after:
+    //   "hel":   0.8
+    //   "dummy": 0.1
+    //
+    // 3. discard non-EOG tokens with low prob
+    // 4. if no tokens are left -> pick EOT
+    //
+    LLAMA_API struct llama_sampler * llama_sampler_init_infill(const struct llama_model * model);
 
     // Returns the seed used by the sampler if applicable, LLAMA_DEFAULT_SEED otherwise
     LLAMA_API uint32_t llama_sampler_get_seed(const struct llama_sampler * smpl);
diff --git a/examples/talk-llama/unicode-data.cpp b/examples/talk-llama/unicode-data.cpp
index 07424bba..04dcd7fc 100644
--- a/examples/talk-llama/unicode-data.cpp
+++ b/examples/talk-llama/unicode-data.cpp
@@ -2311,7 +2311,7 @@ const std::unordered_set<uint32_t> unicode_set_whitespace = {
 0x003000,
 };
 
-// list is always in ascending order, to enable binary searh
+// list is always in ascending order, to enable binary search
 const std::initializer_list<std::pair<uint32_t, uint32_t>> unicode_map_lowercase = {
 {0x000041, 0x000061},
 {0x000042, 0x000062},
@@ -3748,7 +3748,7 @@ const std::initializer_list<std::pair<uint32_t, uint32_t>> unicode_map_lowercase
 {0x01E921, 0x01E943},
 };
 
-// list is always in ascending order, to enable binary searh
+// list is always in ascending order, to enable binary search
 const std::initializer_list<std::pair<uint32_t, uint32_t>> unicode_map_uppercase = {
 {0x000061, 0x000041},
 {0x000062, 0x000042},