whispercpp 1.3.3 → 1.3.4

data/ext/sources/examples/talk-llama/llama-arch.h

@@ -26,6 +26,7 @@ enum llm_arch {
     LLM_ARCH_NOMIC_BERT_MOE,
     LLM_ARCH_NEO_BERT,
     LLM_ARCH_JINA_BERT_V2,
+    LLM_ARCH_JINA_BERT_V3,
     LLM_ARCH_BLOOM,
     LLM_ARCH_STABLELM,
     LLM_ARCH_QWEN,
@@ -38,6 +39,7 @@ enum llm_arch {
     LLM_ARCH_PHI3,
     LLM_ARCH_PHIMOE,
     LLM_ARCH_PLAMO,
+    LLM_ARCH_PLAMO2,
     LLM_ARCH_CODESHELL,
     LLM_ARCH_ORION,
     LLM_ARCH_INTERNLM2,
@@ -47,8 +49,12 @@ enum llm_arch {
     LLM_ARCH_GEMMA2,
     LLM_ARCH_GEMMA3,
     LLM_ARCH_GEMMA3N,
+    LLM_ARCH_GEMMA_EMBEDDING,
     LLM_ARCH_STARCODER2,
     LLM_ARCH_MAMBA,
+    LLM_ARCH_MAMBA2,
+    LLM_ARCH_JAMBA,
+    LLM_ARCH_FALCON_H1,
     LLM_ARCH_XVERSE,
     LLM_ARCH_COMMAND_R,
     LLM_ARCH_COHERE2,
@@ -62,18 +68,22 @@ enum llm_arch {
     LLM_ARCH_DEEPSEEK2,
     LLM_ARCH_CHATGLM,
     LLM_ARCH_GLM4,
+    LLM_ARCH_GLM4_MOE,
     LLM_ARCH_BITNET,
     LLM_ARCH_T5,
     LLM_ARCH_T5ENCODER,
     LLM_ARCH_JAIS,
     LLM_ARCH_NEMOTRON,
+    LLM_ARCH_NEMOTRON_H,
     LLM_ARCH_EXAONE,
+    LLM_ARCH_EXAONE4,
     LLM_ARCH_RWKV6,
     LLM_ARCH_RWKV6QWEN2,
     LLM_ARCH_RWKV7,
     LLM_ARCH_ARWKV7,
     LLM_ARCH_GRANITE,
     LLM_ARCH_GRANITE_MOE,
+    LLM_ARCH_GRANITE_HYBRID,
     LLM_ARCH_CHAMELEON,
     LLM_ARCH_WAVTOKENIZER_DEC,
     LLM_ARCH_PLM,
@@ -81,6 +91,18 @@ enum llm_arch {
     LLM_ARCH_DOTS1,
     LLM_ARCH_ARCEE,
     LLM_ARCH_ERNIE4_5,
+    LLM_ARCH_ERNIE4_5_MOE,
+    LLM_ARCH_HUNYUAN_MOE,
+    LLM_ARCH_HUNYUAN_DENSE,
+    LLM_ARCH_SMOLLM3,
+    LLM_ARCH_OPENAI_MOE,
+    LLM_ARCH_LFM2,
+    LLM_ARCH_DREAM,
+    LLM_ARCH_SMALLTHINKER,
+    LLM_ARCH_LLADA,
+    LLM_ARCH_LLADA_MOE,
+    LLM_ARCH_SEED_OSS,
+    LLM_ARCH_GROVEMOE,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -108,6 +130,7 @@ enum llm_kv {
     LLM_KV_FEED_FORWARD_LENGTH,
     LLM_KV_EXPERT_FEED_FORWARD_LENGTH,
     LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH,
+    LLM_KV_EXPERT_CHUNK_FEED_FORWARD_LENGTH,
     LLM_KV_USE_PARALLEL_RESIDUAL,
     LLM_KV_TENSOR_DATA_LAYOUT,
     LLM_KV_EXPERT_COUNT,
@@ -116,11 +139,16 @@ enum llm_kv {
     LLM_KV_EXPERT_WEIGHTS_SCALE,
     LLM_KV_EXPERT_WEIGHTS_NORM,
     LLM_KV_EXPERT_GATING_FUNC,
+    LLM_KV_EXPERT_GROUP_SCALE,
+    LLM_KV_EXPERTS_PER_GROUP,
     LLM_KV_MOE_EVERY_N_LAYERS,
+    LLM_KV_NEXTN_PREDICT_LAYERS,
     LLM_KV_POOLING_TYPE,
     LLM_KV_LOGIT_SCALE,
     LLM_KV_DECODER_START_TOKEN_ID,
+    LLM_KV_DECODER_BLOCK_COUNT,
     LLM_KV_ATTN_LOGIT_SOFTCAPPING,
+    LLM_KV_ROUTER_LOGIT_SOFTCAPPING,
     LLM_KV_FINAL_LOGIT_SOFTCAPPING,
     LLM_KV_SWIN_NORM,
     LLM_KV_RESCALE_EVERY_N_LAYERS,
@@ -151,9 +179,10 @@ enum llm_kv {
     LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
     LLM_KV_ATTENTION_SLIDING_WINDOW,
     LLM_KV_ATTENTION_SCALE,
+    LLM_KV_ATTENTION_OUTPUT_SCALE,
+    LLM_KV_ATTENTION_TEMPERATURE_LENGTH,
     LLM_KV_ATTENTION_KEY_LENGTH_MLA,
     LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
-    LLM_KV_ATTENTION_LAYER_INDICES,
 
     LLM_KV_ROPE_DIMENSION_COUNT,
     LLM_KV_ROPE_DIMENSION_SECTIONS,
@@ -165,6 +194,10 @@ enum llm_kv {
     LLM_KV_ROPE_SCALING_ORIG_CTX_LEN,
     LLM_KV_ROPE_SCALING_FINETUNED,
     LLM_KV_ROPE_SCALING_YARN_LOG_MUL,
+    LLM_KV_ROPE_SCALING_YARN_EXT_FACTOR,
+    LLM_KV_ROPE_SCALING_YARN_ATTN_FACTOR,
+    LLM_KV_ROPE_SCALING_YARN_BETA_FAST,
+    LLM_KV_ROPE_SCALING_YARN_BETA_SLOW,
 
     LLM_KV_SPLIT_NO,
     LLM_KV_SPLIT_COUNT,
@@ -174,6 +207,7 @@ enum llm_kv {
     LLM_KV_SSM_CONV_KERNEL,
     LLM_KV_SSM_STATE_SIZE,
     LLM_KV_SSM_TIME_STEP_RANK,
+    LLM_KV_SSM_GROUP_COUNT,
     LLM_KV_SSM_DT_B_C_RMS,
 
     LLM_KV_WKV_HEAD_SIZE,
@@ -212,6 +246,9 @@ enum llm_kv {
 
     LLM_KV_ADAPTER_TYPE,
     LLM_KV_ADAPTER_LORA_ALPHA,
+    LLM_KV_ADAPTER_LORA_TASK_NAME,
+    LLM_KV_ADAPTER_LORA_PROMPT_PREFIX,
+    LLM_KV_ADAPTER_ALORA_INVOCATION_TOKENS,
 
     LLM_KV_POSNET_EMBEDDING_LENGTH,
     LLM_KV_POSNET_BLOCK_COUNT,
@@ -221,6 +258,8 @@ enum llm_kv {
 
     LLM_KV_CLASSIFIER_OUTPUT_LABELS,
 
+    LLM_KV_SHORTCONV_L_CACHE,
+
     // deprecated:
     LLM_KV_TOKENIZER_PREFIX_ID,
     LLM_KV_TOKENIZER_SUFFIX_ID,
@@ -247,6 +286,7 @@ enum llm_tensor {
     LLM_TENSOR_ATTN_OUT_NORM,
     LLM_TENSOR_ATTN_POST_NORM,
     LLM_TENSOR_ATTN_ROT_EMBD,
+    LLM_TENSOR_ATTN_SINKS,
     LLM_TENSOR_FFN_GATE_INP,
     LLM_TENSOR_FFN_GATE_INP_SHEXP,
     LLM_TENSOR_FFN_NORM,
@@ -265,6 +305,9 @@ enum llm_tensor {
     LLM_TENSOR_FFN_DOWN_SHEXP,
     LLM_TENSOR_FFN_GATE_SHEXP,
     LLM_TENSOR_FFN_UP_SHEXP,
+    LLM_TENSOR_FFN_DOWN_CHEXPS,
+    LLM_TENSOR_FFN_GATE_CHEXPS,
+    LLM_TENSOR_FFN_UP_CHEXPS,
     LLM_TENSOR_FFN_EXP_PROBS_B,
     LLM_TENSOR_ATTN_Q_NORM,
     LLM_TENSOR_ATTN_K_NORM,
@@ -291,8 +334,12 @@ enum llm_tensor {
     LLM_TENSOR_SSM_CONV1D,
     LLM_TENSOR_SSM_X,
     LLM_TENSOR_SSM_DT,
+    LLM_TENSOR_SSM_DT_NORM,
     LLM_TENSOR_SSM_A,
+    LLM_TENSOR_SSM_B_NORM,
+    LLM_TENSOR_SSM_C_NORM,
     LLM_TENSOR_SSM_D,
+    LLM_TENSOR_SSM_NORM,
     LLM_TENSOR_SSM_OUT,
     LLM_TENSOR_TIME_MIX_W0,
     LLM_TENSOR_TIME_MIX_W1,
@@ -386,6 +433,15 @@ enum llm_tensor {
     LLM_TENSOR_POS_NET_ATTN_K,
     LLM_TENSOR_POS_NET_ATTN_V,
     LLM_TENSOR_POS_NET_ATTN_OUT,
+    LLM_TENSOR_SHORTCONV_CONV,
+    LLM_TENSOR_SHORTCONV_INPROJ,
+    LLM_TENSOR_SHORTCONV_OUTPROJ,
+    LLM_TENSOR_NEXTN_EH_PROJ,
+    LLM_TENSOR_NEXTN_EMBED_TOKENS,
+    LLM_TENSOR_NEXTN_ENORM,
+    LLM_TENSOR_NEXTN_HNORM,
+    LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD,
+    LLM_TENSOR_NEXTN_SHARED_HEAD_NORM,
 };
 
 enum llm_tensor_layer {
@@ -462,3 +518,4 @@ const llm_tensor_info & llm_tensor_info_for(llm_tensor tensor);
 
 bool llm_arch_is_recurrent(const llm_arch & arch);
 bool llm_arch_is_hybrid   (const llm_arch & arch);
+bool llm_arch_is_diffusion(const llm_arch & arch);

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

@@ -27,6 +27,7 @@ bool llama_batch_allocr::init(
         const llama_vocab & vocab,
         const llama_memory_i * memory,
         uint32_t n_embd,
+        uint32_t n_seq_max,
         bool output_all) {
     clear();
 
@@ -40,6 +41,11 @@ bool llama_batch_allocr::init(
     // validate input batch
     //
 
+    if (n_seq_max > LLAMA_MAX_SEQ) {
+        LLAMA_LOG_ERROR("%s: n_seq_max = %d > %d\n", __func__, n_seq_max, LLAMA_MAX_SEQ);
+        return false;
+    }
+
     if (batch.token) {
         for (int32_t i = 0; i < batch.n_tokens; ++i) {
             if (batch.token[i] < 0 || (uint32_t) batch.token[i] >= vocab.n_tokens()) {
@@ -52,8 +58,8 @@ bool llama_batch_allocr::init(
     if (batch.seq_id) {
         for (int32_t i = 0; i < batch.n_tokens; ++i) {
             for (int32_t s = 0; s < batch.n_seq_id[i]; ++s) {
-                if (batch.seq_id && (batch.seq_id[i][s] < 0 || batch.seq_id[i][s] >= LLAMA_MAX_SEQ)) {
-                    LLAMA_LOG_ERROR("%s: invalid seq_id[%d][%d] = %d > %d\n", __func__, i, s, batch.seq_id[i][s], LLAMA_MAX_SEQ);
+                if (batch.seq_id && (batch.seq_id[i][s] < 0 || batch.seq_id[i][s] >= (llama_seq_id) n_seq_max)) {
+                    LLAMA_LOG_ERROR("%s: invalid seq_id[%d][%d] = %d >= %d\n", __func__, i, s, batch.seq_id[i][s], (llama_seq_id) n_seq_max);
                     return false;
                 }
             }
@@ -86,7 +92,7 @@ bool llama_batch_allocr::init(
 
         // initialize the starting position for each sequence based on the positions in the memory
         llama_pos p0[LLAMA_MAX_SEQ];
-        for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+        for (uint32_t s = 0; s < n_seq_max; ++s) {
             if (!memory) {
                 // if no memory -> start from 0
                 p0[s] = 0;
@@ -143,13 +149,16 @@ bool llama_batch_allocr::init(
     // compute stats
     //
 
-    this->n_embd = n_embd;
+    this->n_embd    = n_embd;
+    this->n_seq_max = n_seq_max;
 
     // count the outputs in this batch
     for (int32_t i = 0; i < batch.n_tokens; ++i) {
         n_outputs += batch.logits[i] != 0;
     }
 
+    has_cpl = false;
+
     // determine coupled sequences
     // these are pairs of sequences that have at least one token in the input batch that is assigned to both of them
     for (int32_t i = 0; i < batch.n_tokens; ++i) {
@@ -166,6 +175,8 @@ bool llama_batch_allocr::init(
 
                 // note: tracking the other way around is not necessary for now
                 //seq_cpl[s0][s1] = true;
+
+                has_cpl = true;
             }
         }
     }
@@ -187,7 +198,7 @@ bool llama_batch_allocr::init(
             seq_set_map[cur].push_back(i);
         }
 
-        for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+        for (uint32_t s = 0; s < n_seq_max; ++s) {
             if (seq_set_unq.test(s)) {
                 seq_idx[s] = seq_id_unq.size();
                 seq_id_unq.push_back(s);
@@ -199,7 +210,7 @@ bool llama_batch_allocr::init(
         LLAMA_LOG_DEBUG("%s: input batch info:\n", __func__);
 
         llama_ubatch ubatch {
-            /*.equal_seqs   =*/ false,
+            /*.b_equal_seqs =*/ false,
             /*.n_tokens     =*/ (uint32_t) batch.n_tokens,
             /*.n_seq_tokens =*/ (uint32_t) 1,
             /*.n_seqs       =*/ (uint32_t) batch.n_tokens,
@@ -212,6 +223,7 @@ bool llama_batch_allocr::init(
             /*.seq_id_unq   =*/ this->seq_id_unq.data(),
             /*.seq_idx      =*/ this->seq_idx.data(),
             /*.output       =*/ batch.logits,
+            /*.data         =*/ {},
         };
 
         ubatch_print(ubatch, debug);
@@ -239,7 +251,7 @@ bool llama_batch_allocr::init(
     // consistency checks
     //
 
-    for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+    for (uint32_t s = 0; s < n_seq_max; ++s) {
         if (seq_pos[s].empty()) {
             continue;
         }
@@ -282,8 +294,8 @@ bool llama_batch_allocr::init(
     }
 
     if (memory) {
-        for (int32_t s0 = 0; s0 < LLAMA_MAX_SEQ; ++s0) {
-            for (int32_t s1 = 0; s1 < LLAMA_MAX_SEQ; ++s1) {
+        for (uint32_t s0 = 0; s0 < n_seq_max; ++s0) {
+            for (uint32_t s1 = 0; s1 < n_seq_max; ++s1) {
                 if (seq_cpl[s0][s1]) {
                     if (memory->seq_pos_min(s0) != memory->seq_pos_min(s1) ||
                         memory->seq_pos_max(s0) != memory->seq_pos_max(s1)) {
@@ -314,12 +326,12 @@ bool llama_batch_allocr::init(
     //
     {
         seq_set_t cur_seq_set[LLAMA_MAX_SEQ];
-        for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+        for (uint32_t s = 0; s < n_seq_max; ++s) {
             cur_seq_set[s].set();
         }
 
         llama_pos cur_seq_pos[LLAMA_MAX_SEQ];
-        for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+        for (uint32_t s = 0; s < n_seq_max; ++s) {
             cur_seq_pos[s] = -1;
         }
 
@@ -355,39 +367,38 @@ llama_ubatch llama_batch_allocr::ubatch_reserve(uint32_t n_seq_tokens, uint32_t
     clear();
     split_reset();
 
-    ubatches.emplace_back();
-
-    auto & ubatch = ubatches.back();
+    auto udata = std::make_shared<llama_ubatch::data_t>();
 
-    ubatch.token     .resize(n_tokens);
-    ubatch.embd      .clear();
-    ubatch.pos       .resize(n_tokens);
-    ubatch.n_seq_id  .resize(n_tokens);
-    ubatch.seq_id    .resize(n_tokens);
-    ubatch.seq_id_unq.resize(0);
-    ubatch.seq_idx   .resize(LLAMA_MAX_SEQ, -1);
-    ubatch.output    .resize(n_tokens);
+    udata->token     .resize(n_tokens);
+    udata->embd      .clear();
+    udata->pos       .resize(n_tokens);
+    udata->n_seq_id  .resize(n_tokens);
+    udata->seq_id    .resize(n_tokens);
+    udata->seq_id_unq.resize(0);
+    udata->seq_idx   .resize(LLAMA_MAX_SEQ, -1);
+    udata->output    .resize(n_tokens);
 
     for (uint32_t s = 0; s < n_seqs; ++s) {
-        ubatch.seq_idx[s] = s;
-        ubatch.seq_id_unq.push_back(s);
+        udata->seq_idx[s] = s;
+        udata->seq_id_unq.push_back(s);
     }
 
     llama_ubatch res {
-        /*.equal_seqs   =*/ true,
+        /*.b_equal_seqs =*/ true,
         /*.n_tokens     =*/ n_tokens,
         /*.n_seq_tokens =*/ n_seq_tokens,
         /*.n_seqs       =*/ n_seqs,
         /*.n_seqs_unq   =*/ n_seqs,
 
-        /*.token        =*/ ubatch.token.data(),
+        /*.token        =*/ udata->token.data(),
         /*.embd         =*/ nullptr,
-        /*.pos          =*/ ubatch.pos.data(),
-        /*.n_seq_id     =*/ ubatch.n_seq_id.data(),
-        /*.seq_id       =*/ ubatch.seq_id.data(),
-        /*.seq_id_unq   =*/ ubatch.seq_id_unq.data(),
-        /*.seq_idx      =*/ ubatch.seq_idx.data(),
-        /*.output       =*/ ubatch.output.data(),
+        /*.pos          =*/ udata->pos.data(),
+        /*.n_seq_id     =*/ udata->n_seq_id.data(),
+        /*.seq_id       =*/ udata->seq_id.data(),
+        /*.seq_id_unq   =*/ udata->seq_id_unq.data(),
+        /*.seq_idx      =*/ udata->seq_idx.data(),
+        /*.output       =*/ udata->output.data(),
+        /*.data         =*/ std::move(udata),
     };
 
     return res;
@@ -405,6 +416,10 @@ uint32_t llama_batch_allocr::get_n_outputs() const {
     return n_outputs;
 }
 
+uint32_t llama_batch_allocr::get_n_used() const {
+    return n_used;
+}
+
 std::vector<int32_t> & llama_batch_allocr::get_out_ids() {
     return out_ids;
 }
@@ -420,10 +435,10 @@ llama_pos llama_batch_allocr::seq_pos_max(llama_seq_id seq_id) const {
 void llama_batch_allocr::split_reset() {
     out_ids.clear();
 
+    n_used = 0;
+
     used.clear();
     used.resize(get_n_tokens(), false);
-
-    ubatches.clear();
 }
 
 llama_ubatch llama_batch_allocr::split_simple(uint32_t n_ubatch) {
@@ -444,6 +459,7 @@ llama_ubatch llama_batch_allocr::split_simple(uint32_t n_ubatch) {
         idxs.push_back(cur_idx);
 
         used[cur_idx] = true;
+        ++n_used;
 
         ++cur_idx;
 
@@ -459,9 +475,17 @@ llama_ubatch llama_batch_allocr::split_simple(uint32_t n_ubatch) {
     return ubatch_add(idxs, idxs.size(), false);
 }
 
-llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch) {
+llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch, bool sequential) {
+    if (sequential && has_cpl) {
+        LLAMA_LOG_ERROR("%s: sequential split is not supported when there are coupled sequences in the input batch (you may need to use the -kvu flag)\n", __func__);
+
+        return {};
+    }
+
     std::vector<seq_set_t> cur_seq_set;
 
+    llama_seq_id last_seq_id = -1;
+
     // determine the non-overlapping sequence sets participating in this ubatch
     for (int32_t i = 0; i < batch.n_tokens; ++i) {
         if (used[i]) {
@@ -478,9 +502,16 @@ llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch) {
             }
         }
 
+        // accept only increasing sequence ids
+        if (sequential) {
+            add = add && (cur_seq_set.empty() || batch.seq_id[i][0] == last_seq_id + 1);
+        }
+
         if (add) {
             cur_seq_set.push_back(seq_set[i]);
 
+            last_seq_id = batch.seq_id[i][0];
+
             if (cur_seq_set.size() > n_ubatch) {
                 break;
             }
@@ -529,6 +560,7 @@ llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch) {
             idxs_per_seq[s].push_back(idx);
 
             used[idx] = true;
+            ++n_used;
 
             ++cur_idx[s];
         }
@@ -570,6 +602,7 @@ llama_ubatch llama_batch_allocr::split_seq(uint32_t n_ubatch) {
         idxs.push_back(cur_idx);
 
         used[cur_idx] = true;
+        ++n_used;
 
         if (idxs.size() >= n_ubatch) {
             break;
@@ -620,78 +653,77 @@ llama_ubatch llama_batch_allocr::ubatch_add(const std::vector<int32_t> & idxs, u
 
     assert(n_tokens%n_seqs == 0);
 
-    ubatches.emplace_back();
-
-    auto & ubatch = ubatches.back();
+    auto udata = std::make_shared<llama_ubatch::data_t>();
 
     const int32_t n_pos_cur = batch.embd ? n_pos_per_embd : 1;
 
     const int64_t n_embd_all = batch.embd ? (int64_t) n_tokens*n_embd : 0;
     const int64_t n_pos_all  =              (int64_t) n_tokens*n_pos_cur;
 
-    ubatch.token     .resize(n_tokens);
-    ubatch.embd      .resize(n_embd_all);
-    ubatch.pos       .resize(n_pos_all);
-    ubatch.n_seq_id  .resize(n_tokens);
-    ubatch.seq_id    .resize(n_tokens);
-    ubatch.seq_id_unq.resize(0);
-    ubatch.seq_idx   .resize(LLAMA_MAX_SEQ, -1);
-    ubatch.output    .resize(n_tokens);
+    udata->token     .resize(n_tokens);
+    udata->embd      .resize(n_embd_all);
+    udata->pos       .resize(n_pos_all);
+    udata->n_seq_id  .resize(n_tokens);
+    udata->seq_id    .resize(n_tokens);
+    udata->seq_id_unq.resize(0);
+    udata->seq_idx   .resize(LLAMA_MAX_SEQ, -1);
+    udata->output    .resize(n_tokens);
 
     seq_set_t seq_set_unq;
 
     for (size_t i = 0; i < idxs.size(); ++i) {
         if (batch.token) {
-            ubatch.token[i] = batch.token[idxs[i]];
+            udata->token[i] = batch.token[idxs[i]];
         }
 
         if (batch.embd) {
-            memcpy(ubatch.embd.data() + i*n_embd, batch.embd + (int64_t) idxs[i]*n_embd, n_embd*sizeof(float));
+            memcpy(udata->embd.data() + i*n_embd, batch.embd + (int64_t) idxs[i]*n_embd, n_embd*sizeof(float));
         }
 
         for (int j = 0; j < n_pos_cur; ++j) {
-            ubatch.pos[j*n_tokens + i] = batch.pos[j*batch.n_tokens + idxs[i]];
+            udata->pos[j*n_tokens + i] = batch.pos[j*batch.n_tokens + idxs[i]];
         }
 
-        ubatch.n_seq_id[i] = batch.n_seq_id[idxs[i]];
-        ubatch.seq_id[i]   = batch.seq_id[idxs[i]];
-        ubatch.output[i]   = batch.logits[idxs[i]];
+        udata->n_seq_id[i] = batch.n_seq_id[idxs[i]];
+        udata->seq_id[i]   = batch.seq_id[idxs[i]];
+        udata->output[i]   = batch.logits[idxs[i]];
 
-        for (int s = 0; s < ubatch.n_seq_id[i]; ++s) {
-            seq_set_unq.set(ubatch.seq_id[i][s]);
+        for (int s = 0; s < udata->n_seq_id[i]; ++s) {
+            seq_set_unq.set(udata->seq_id[i][s]);
         }
 
-        if (ubatch.output[i]) {
+        if (udata->output[i]) {
             out_ids.push_back(idxs[i]);
         }
     }
 
-    for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+    for (uint32_t s = 0; s < n_seq_max; ++s) {
         if (seq_set_unq.test(s)) {
-            ubatch.seq_idx[s] = ubatch.seq_id_unq.size();
-            ubatch.seq_id_unq.push_back(s);
+            udata->seq_idx[s] = udata->seq_id_unq.size();
+            udata->seq_id_unq.push_back(s);
         }
     }
 
     llama_ubatch res {
-        /*.equal_seqs   =*/ equal_seqs,
+        /*.b_equal_seqs =*/ equal_seqs,
         /*.n_tokens     =*/ n_tokens,
         /*.n_seq_tokens =*/ n_tokens/n_seqs,
         /*.n_seqs       =*/ n_seqs,
-        /*.n_seqs_unq   =*/ (uint32_t) ubatch.seq_id_unq.size(),
-
-        /*.token        =*/ batch.token ? ubatch.token.data() : nullptr,
-        /*.embd         =*/ batch.embd ? ubatch.embd.data() : nullptr,
-        /*.pos          =*/ ubatch.pos.data(),
-        /*.n_seq_id     =*/ ubatch.n_seq_id.data(),
-        /*.seq_id       =*/ ubatch.seq_id.data(),
-        /*.seq_id_unq   =*/ ubatch.seq_id_unq.data(),
-        /*.seq_idx      =*/ ubatch.seq_idx.data(),
-        /*.output       =*/ ubatch.output.data(),
+        /*.n_seqs_unq   =*/ (uint32_t) udata->seq_id_unq.size(),
+
+        /*.token        =*/ batch.token ? udata->token.data() : nullptr,
+        /*.embd         =*/ batch.embd ? udata->embd.data() : nullptr,
+        /*.pos          =*/ udata->pos.data(),
+        /*.n_seq_id     =*/ udata->n_seq_id.data(),
+        /*.seq_id       =*/ udata->seq_id.data(),
+        /*.seq_id_unq   =*/ udata->seq_id_unq.data(),
+        /*.seq_idx      =*/ udata->seq_idx.data(),
+        /*.output       =*/ udata->output.data(),
+        /*.data         =*/ std::move(udata),
     };
 
     if (debug > 0) {
-        LLAMA_LOG_DEBUG("%s: added ubatch %d to split:\n", __func__, (int) ubatches.size() - 1);
+        LLAMA_LOG_DEBUG("%s: added ubatch to split:\n", __func__);
 
         ubatch_print(res, debug);
     }
@@ -701,7 +733,7 @@ llama_ubatch llama_batch_allocr::ubatch_add(const std::vector<int32_t> & idxs, u
 
 void llama_batch_allocr::ubatch_print(const llama_ubatch & ubatch, int debug) {
     if (debug > 0) {
-        LLAMA_LOG_DEBUG("%s:   equal_seqs   = %d\n", __func__, ubatch.equal_seqs);
+        LLAMA_LOG_DEBUG("%s:   equal_seqs   = %d\n", __func__, ubatch.equal_seqs());
         LLAMA_LOG_DEBUG("%s:   n_tokens     = %d\n", __func__, ubatch.n_tokens);
         LLAMA_LOG_DEBUG("%s:   n_seq_tokens = %d\n", __func__, ubatch.n_seq_tokens);
         LLAMA_LOG_DEBUG("%s:   n_seqs       = %d\n", __func__, ubatch.n_seqs);

data/ext/sources/examples/talk-llama/llama-batch.h

@@ -8,12 +8,17 @@
 #include <vector>
 #include <set>
 #include <bitset>
+#include <memory>
 #include <unordered_map>
 
 // keep this struct lightweight
-// it points to data in `llama_batch_allocr`
 struct llama_ubatch {
-    bool equal_seqs;
+    bool equal_seqs() const {
+        return b_equal_seqs != 0;
+    }
+
+    uint32_t b_equal_seqs; // note: this is a boolean, but we use an int32_t for alignment
+                           //       otherwise address sanitizer complains
     // TODO: whole_seqs for embeddings?
 
     uint32_t n_tokens;     // total tokens (n_seq_tokens * n_seqs)
@@ -34,6 +39,20 @@ struct llama_ubatch {
     llama_seq_id *  seq_id_unq; // [n_seqs_unq]       | s   | seq_id
     int32_t      *  seq_idx;    // [LLAMA_MAX_SEQ]    | -   | seq_idx
     int8_t       *  output;     // [n_tokens]         | i   | -
+
+    struct data_t {
+        std::vector<llama_token>    token;
+        std::vector<float>          embd;
+        std::vector<llama_pos>      pos;
+        std::vector<int32_t>        n_seq_id;
+        std::vector<llama_seq_id *> seq_id;
+        std::vector<llama_seq_id>   seq_id_unq;
+        std::vector<int32_t>        seq_idx;
+        std::vector<int8_t>         output;
+    };
+
+    // the llama_ubatch pointers above point to this data if set. otherwise - points to non-owning data
+    std::shared_ptr<data_t> data;
 };
 
 // a helper for sanitizing, fulfilling and splitting a batch
@@ -48,12 +67,14 @@ public:
             const llama_vocab & vocab,
             const llama_memory_i * memory,
             uint32_t n_embd,
+            uint32_t n_seq_max,
             bool output_all);
 
     const llama_batch & get_batch() const;
 
     uint32_t get_n_tokens()  const;
     uint32_t get_n_outputs() const;
+    uint32_t get_n_used()    const;
 
     // the array of output indices in the order they were encountered during the ubatch splitting
     std::vector<int32_t> & get_out_ids();
@@ -69,7 +90,8 @@ public:
     llama_ubatch split_simple(uint32_t n_ubatch);
 
     // make ubatches of equal-length sequences sets
-    llama_ubatch split_equal(uint32_t n_ubatch);
+    // if sequential == true, the tokens in the ubatch will have increasing sequential sequence ids
+    llama_ubatch split_equal(uint32_t n_ubatch, bool sequential);
 
     // sequence-set-wise split - each ubatch contains a single sequence-set
     llama_ubatch split_seq(uint32_t n_ubatch);
@@ -98,6 +120,7 @@ private:
     const uint32_t n_pos_per_embd;
 
     uint32_t n_embd;
+    uint32_t n_seq_max;
     uint32_t n_outputs;
 
     std::array<llama_seq_id, 1> seq_id_0 = { 0 }; // default sequence id
@@ -112,6 +135,9 @@ private:
     using pos_set_t = std::set<llama_pos>;
     using seq_cpl_t = std::vector<bool>;
 
+    // helper flag to quickly determine if there are any coupled sequences in the batch
+    bool has_cpl = false;
+
     std::vector<pos_set_t> seq_pos; // seq_pos[s]: the set of positions in sequence s
     std::vector<seq_cpl_t> seq_cpl; // seq_cpl[s0][s1]: if sequence s0 is coupled to sequence s1
 
@@ -125,23 +151,10 @@ private:
     // batch indices of the output
     std::vector<int32_t> out_ids;
 
+    uint32_t n_used;
+
     // used[i] indicates if token i has already been used in a previous ubatch
     std::vector<bool> used;
 
-    // llama_ubatch points to this data:
-    struct ubatch {
-        std::vector<llama_token>    token;
-        std::vector<float>          embd;
-        std::vector<llama_pos>      pos;
-        std::vector<int32_t>        n_seq_id;
-        std::vector<llama_seq_id *> seq_id;
-        std::vector<llama_seq_id>   seq_id_unq;
-        std::vector<int32_t>        seq_idx;
-        std::vector<int8_t>         output;
-    };
-
-    // current splitting state:
-    std::vector<ubatch> ubatches;
-
     int debug;
 };