sync : ggml

examples/common-ggml.cpp

@@ -75,6 +75,7 @@ bool ggml_common_quantize_0(
         case GGML_FTYPE_MOSTLY_MXFP4:
         case GGML_FTYPE_MOSTLY_NVFP4:
         case GGML_FTYPE_MOSTLY_Q1_0:
+        case GGML_FTYPE_MOSTLY_Q2_0:
                 {
                     fprintf(stderr, "%s: invalid model type %d\n", __func__, ftype);
                     return false;
@@ -217,6 +218,7 @@ bool ggml_common_quantize_0(
                 case GGML_TYPE_MXFP4:
                 case GGML_TYPE_NVFP4:
                 case GGML_TYPE_Q1_0:
+                case GGML_TYPE_Q2_0:
                 case GGML_TYPE_COUNT:
                     {
                         fprintf(stderr, "%s: unsupported quantization type %d (%s)\n", __func__, ttype, ggml_type_name((ggml_type) ttype));

examples/talk-llama/CMakeLists.txt

@@ -21,6 +21,7 @@ if (WHISPER_SDL2)
         llama-kv-cache.cpp
         llama-kv-cache-iswa.cpp
         llama-kv-cache-dsa.cpp
+        llama-kv-cache-dsv4.cpp
         llama-memory-recurrent.cpp
         llama-memory-hybrid.cpp
         llama-memory-hybrid-iswa.cpp

examples/talk-llama/llama-arch.cpp

@@ -77,6 +77,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_DEEPSEEK2,        "deepseek2"        },
     { LLM_ARCH_DEEPSEEK2OCR,     "deepseek2-ocr"    },
     { LLM_ARCH_DEEPSEEK32,       "deepseek32"       },
+    { LLM_ARCH_DEEPSEEK4,        "deepseek4"        },
     { LLM_ARCH_CHATGLM,          "chatglm"          },
     { LLM_ARCH_GLM4,             "glm4"             },
     { LLM_ARCH_GLM4_MOE,         "glm4moe"          },
@@ -129,6 +130,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_PANGU_EMBED,      "pangu-embedded"   },
     { LLM_ARCH_MISTRAL3,         "mistral3"         },
     { LLM_ARCH_EAGLE3,           "eagle3"           },
+    { LLM_ARCH_DFLASH,           "dflash"           },
     { LLM_ARCH_MISTRAL4,         "mistral4"         },
     { LLM_ARCH_PADDLEOCR,        "paddleocr"        },
     { LLM_ARCH_MIMO2,            "mimo2"            },
@@ -249,9 +251,19 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ATTENTION_INDEXER_HEAD_COUNT,           "%s.attention.indexer.head_count"           },
     { LLM_KV_ATTENTION_INDEXER_KEY_LENGTH,           "%s.attention.indexer.key_length"           },
     { LLM_KV_ATTENTION_INDEXER_TOP_K,                "%s.attention.indexer.top_k"                },
+    { LLM_KV_ATTENTION_OUTPUT_GROUP_COUNT,           "%s.attention.output_group_count"           },
+    { LLM_KV_ATTENTION_OUTPUT_LORA_RANK,             "%s.attention.output_lora_rank"             },
+    { LLM_KV_ATTENTION_COMPRESS_ROPE_FREQ_BASE,      "%s.attention.compress_rope_freq_base"      },
+    { LLM_KV_ATTENTION_COMPRESS_RATIOS,              "%s.attention.compress_ratios"              },
     { LLM_KV_ATTENTION_SHARED_KV_LAYERS,             "%s.attention.shared_kv_layers"             },
     { LLM_KV_ATTENTION_RECURRENT_LAYERS,             "%s.attention.recurrent_layers"             },
 
+    { LLM_KV_HYPER_CONNECTION_COUNT,                 "%s.hyper_connection.count"                 },
+    { LLM_KV_HYPER_CONNECTION_SINKHORN_ITERATIONS,   "%s.hyper_connection.sinkhorn_iterations"   },
+    { LLM_KV_HYPER_CONNECTION_EPSILON,               "%s.hyper_connection.epsilon"               },
+
+    { LLM_KV_HASH_LAYER_COUNT,                       "%s.hash_layer_count"                       },
+
     { LLM_KV_ROPE_DIMENSION_COUNT,           "%s.rope.dimension_count"                 },
     { LLM_KV_ROPE_DIMENSION_COUNT_SWA,       "%s.rope.dimension_count_swa"             },
     { LLM_KV_ROPE_DIMENSION_SECTIONS,        "%s.rope.dimension_sections"              },
@@ -439,6 +451,23 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
     { LLM_TENSOR_ATTN_Q_B,                               "blk.%d.attn_q_b" },
     { LLM_TENSOR_ATTN_KV_A_MQA,                          "blk.%d.attn_kv_a_mqa" },
     { LLM_TENSOR_ATTN_KV_B,                              "blk.%d.attn_kv_b" },
+    { LLM_TENSOR_ATTN_KV,                                "blk.%d.attn_kv" },
+    { LLM_TENSOR_ATTN_KV_NORM,                           "blk.%d.attn_kv_a_norm" },
+    { LLM_TENSOR_ATTN_OUT_A,                             "blk.%d.attn_output_a" },
+    { LLM_TENSOR_ATTN_OUT_B,                             "blk.%d.attn_output_b" },
+    { LLM_TENSOR_HC_HEAD_FN,                             "output_hc_fn" },
+    { LLM_TENSOR_HC_HEAD_BASE,                           "output_hc_base" },
+    { LLM_TENSOR_HC_HEAD_SCALE,                          "output_hc_scale" },
+    { LLM_TENSOR_HC_ATTN_FN,                             "blk.%d.hc_attn_fn" },
+    { LLM_TENSOR_HC_ATTN_BASE,                           "blk.%d.hc_attn_base" },
+    { LLM_TENSOR_HC_ATTN_SCALE,                          "blk.%d.hc_attn_scale" },
+    { LLM_TENSOR_HC_FFN_FN,                              "blk.%d.hc_ffn_fn" },
+    { LLM_TENSOR_HC_FFN_BASE,                            "blk.%d.hc_ffn_base" },
+    { LLM_TENSOR_HC_FFN_SCALE,                           "blk.%d.hc_ffn_scale" },
+    { LLM_TENSOR_ATTN_COMPRESSOR_WKV,                    "blk.%d.attn_compressor_kv" },
+    { LLM_TENSOR_ATTN_COMPRESSOR_WGATE,                  "blk.%d.attn_compressor_gate" },
+    { LLM_TENSOR_ATTN_COMPRESSOR_APE,                    "blk.%d.attn_compressor_ape" },
+    { LLM_TENSOR_ATTN_COMPRESSOR_NORM,                   "blk.%d.attn_compressor_norm" },
     { LLM_TENSOR_PER_LAYER_TOKEN_EMBD,                   "per_layer_token_embd" },
     { LLM_TENSOR_PER_LAYER_MODEL_PROJ,                   "per_layer_model_proj" },
     { LLM_TENSOR_PER_LAYER_PROJ_NORM,                    "per_layer_proj_norm" },
@@ -565,6 +594,11 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
     { LLM_TENSOR_INDEXER_PROJ,                           "blk.%d.indexer.proj" },
     { LLM_TENSOR_INDEXER_ATTN_K,                         "blk.%d.indexer.attn_k" },
     { LLM_TENSOR_INDEXER_ATTN_Q_B,                       "blk.%d.indexer.attn_q_b" },
+    { LLM_TENSOR_INDEXER_COMPRESSOR_WKV,                 "blk.%d.indexer_compressor_kv" },
+    { LLM_TENSOR_INDEXER_COMPRESSOR_WGATE,               "blk.%d.indexer_compressor_gate" },
+    { LLM_TENSOR_INDEXER_COMPRESSOR_APE,                 "blk.%d.indexer_compressor_ape" },
+    { LLM_TENSOR_INDEXER_COMPRESSOR_NORM,                "blk.%d.indexer_compressor_norm" },
+    { LLM_TENSOR_FFN_GATE_TID2EID,                       "blk.%d.ffn_gate_tid2eid" },
     { LLM_TENSOR_MASKED_EMBD_CENTROIDS,                  "masked_embd_centroids" },
     { LLM_TENSOR_MASKED_EMBD_ORDERING,                   "masked_embd_ordering" },
     { LLM_TENSOR_FC,                                     "fc" },
@@ -615,6 +649,23 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_ATTN_Q_B,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_ATTN_KV_A_MQA,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_ATTN_KV_B,                  {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ATTN_KV,                    {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ATTN_KV_NORM,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_ATTN_OUT_A,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ATTN_OUT_B,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_HC_HEAD_FN,                 {LLM_TENSOR_LAYER_OUTPUT,    GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_HC_HEAD_BASE,               {LLM_TENSOR_LAYER_OUTPUT,    GGML_OP_ADD}},
+    {LLM_TENSOR_HC_HEAD_SCALE,              {LLM_TENSOR_LAYER_OUTPUT,    GGML_OP_MUL}},
+    {LLM_TENSOR_HC_ATTN_FN,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_HC_ATTN_BASE,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
+    {LLM_TENSOR_HC_ATTN_SCALE,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_HC_FFN_FN,                  {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_HC_FFN_BASE,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
+    {LLM_TENSOR_HC_FFN_SCALE,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_ATTN_COMPRESSOR_WKV,        {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ATTN_COMPRESSOR_WGATE,      {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ATTN_COMPRESSOR_APE,        {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
+    {LLM_TENSOR_ATTN_COMPRESSOR_NORM,       {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_ATTN_K_B,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_ATTN_V_B,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_ATTN_SINKS,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SCALE}},
@@ -778,6 +829,11 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_INDEXER_PROJ,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_INDEXER_ATTN_K,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_INDEXER_ATTN_Q_B,           {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_INDEXER_COMPRESSOR_WKV,     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_INDEXER_COMPRESSOR_WGATE,   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_INDEXER_COMPRESSOR_APE,     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
+    {LLM_TENSOR_INDEXER_COMPRESSOR_NORM,    {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_FFN_GATE_TID2EID,           {LLM_TENSOR_LAYER_REPEATING, GGML_OP_GET_ROWS}},
     {LLM_TENSOR_NEXTN_PROJ_PRE,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_NEXTN_PROJ_POST,            {LLM_TENSOR_LAYER_OUTPUT,    GGML_OP_MUL_MAT}},
     // NextN/MTP tensors are stored per-block (blk.%d.nextn.*) even though only the
@@ -932,6 +988,7 @@ bool llm_arch_supports_sm_tensor(const llm_arch & arch) {
         case LLM_ARCH_OLMOE:
         case LLM_ARCH_DEEPSEEK2:
         case LLM_ARCH_DEEPSEEK32:
+        case LLM_ARCH_DEEPSEEK4:
         case LLM_ARCH_GLM_DSA:
         case LLM_ARCH_BITNET:
         case LLM_ARCH_T5:

examples/talk-llama/llama-arch.h

@@ -82,6 +82,7 @@ enum llm_arch {
     LLM_ARCH_DEEPSEEK2,
     LLM_ARCH_DEEPSEEK2OCR,
     LLM_ARCH_DEEPSEEK32,
+    LLM_ARCH_DEEPSEEK4,
     LLM_ARCH_CHATGLM,
     LLM_ARCH_GLM4,
     LLM_ARCH_GLM4_MOE,
@@ -143,6 +144,7 @@ enum llm_arch {
     LLM_ARCH_TALKIE,
     LLM_ARCH_MELLUM,
     LLM_ARCH_EAGLE3,
+    LLM_ARCH_DFLASH,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -254,9 +256,19 @@ enum llm_kv {
     LLM_KV_ATTENTION_INDEXER_HEAD_COUNT,
     LLM_KV_ATTENTION_INDEXER_KEY_LENGTH,
     LLM_KV_ATTENTION_INDEXER_TOP_K,
+    LLM_KV_ATTENTION_OUTPUT_GROUP_COUNT,
+    LLM_KV_ATTENTION_OUTPUT_LORA_RANK,
+    LLM_KV_ATTENTION_COMPRESS_ROPE_FREQ_BASE,
+    LLM_KV_ATTENTION_COMPRESS_RATIOS,
     LLM_KV_ATTENTION_SHARED_KV_LAYERS,
     LLM_KV_ATTENTION_RECURRENT_LAYERS,
 
+    LLM_KV_HYPER_CONNECTION_COUNT,
+    LLM_KV_HYPER_CONNECTION_SINKHORN_ITERATIONS,
+    LLM_KV_HYPER_CONNECTION_EPSILON,
+
+    LLM_KV_HASH_LAYER_COUNT,
+
     LLM_KV_ROPE_DIMENSION_COUNT,
     LLM_KV_ROPE_DIMENSION_COUNT_SWA,
     LLM_KV_ROPE_DIMENSION_SECTIONS,
@@ -500,10 +512,27 @@ enum llm_tensor {
     LLM_TENSOR_ATTN_Q_B,
     LLM_TENSOR_ATTN_KV_A_MQA,
     LLM_TENSOR_ATTN_KV_B,
+    LLM_TENSOR_ATTN_KV,
+    LLM_TENSOR_ATTN_KV_NORM,
+    LLM_TENSOR_ATTN_OUT_A,
+    LLM_TENSOR_ATTN_OUT_B,
     LLM_TENSOR_ATTN_K_B,
     LLM_TENSOR_ATTN_V_B,
     LLM_TENSOR_ATTN_Q_A_NORM,
     LLM_TENSOR_ATTN_KV_A_NORM,
+    LLM_TENSOR_HC_HEAD_FN,
+    LLM_TENSOR_HC_HEAD_BASE,
+    LLM_TENSOR_HC_HEAD_SCALE,
+    LLM_TENSOR_HC_ATTN_FN,
+    LLM_TENSOR_HC_ATTN_BASE,
+    LLM_TENSOR_HC_ATTN_SCALE,
+    LLM_TENSOR_HC_FFN_FN,
+    LLM_TENSOR_HC_FFN_BASE,
+    LLM_TENSOR_HC_FFN_SCALE,
+    LLM_TENSOR_ATTN_COMPRESSOR_WKV,
+    LLM_TENSOR_ATTN_COMPRESSOR_WGATE,
+    LLM_TENSOR_ATTN_COMPRESSOR_APE,
+    LLM_TENSOR_ATTN_COMPRESSOR_NORM,
     LLM_TENSOR_ATTN_SUB_NORM,
     LLM_TENSOR_FFN_SUB_NORM,
     LLM_TENSOR_DEC_ATTN_NORM,
@@ -565,6 +594,11 @@ enum llm_tensor {
     LLM_TENSOR_INDEXER_PROJ,
     LLM_TENSOR_INDEXER_ATTN_K,
     LLM_TENSOR_INDEXER_ATTN_Q_B,
+    LLM_TENSOR_INDEXER_COMPRESSOR_WKV,
+    LLM_TENSOR_INDEXER_COMPRESSOR_WGATE,
+    LLM_TENSOR_INDEXER_COMPRESSOR_APE,
+    LLM_TENSOR_INDEXER_COMPRESSOR_NORM,
+    LLM_TENSOR_FFN_GATE_TID2EID,
     LLM_TENSOR_NEXTN_PROJ_PRE,
     LLM_TENSOR_NEXTN_PROJ_POST,
     LLM_TENSOR_NEXTN_EH_PROJ,

examples/talk-llama/llama-batch.cpp

@@ -379,6 +379,8 @@ bool llama_batch_allocr::init(
                     LLAMA_LOG_ERROR("%s: sequence %d positions are decreasing (not allowed)\n", __func__, seq_id);
                     return false;
                 }
+
+                cur_seq_pos[seq_id] = pos;
             }
         }
     }
@@ -505,7 +507,7 @@ 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, bool sequential) {
+llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch, bool sequential, uint32_t n_keep_tail) {
     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__);
 
@@ -548,7 +550,7 @@ llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch, bool sequential)
         }
     }
 
-    const uint32_t n_seqs = cur_seq_set.size();
+    uint32_t n_seqs = cur_seq_set.size();
 
     // we are done
     if (n_seqs == 0) {
@@ -569,7 +571,7 @@ llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch, bool sequential)
     std::vector<idx_vec_t> idxs_per_seq(n_seqs);
 
     while (true) {
-        // we can only add new n_seq_tokens tokens if all the sequence sets have at least one more unused token and
+        // we can only add new n_seq_tokens tokens if all the sequence sets have at least 1 more unused tokens and
         //   if we haven't reached n_ubatch
         bool can_expand = true;
 
@@ -600,6 +602,72 @@ llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch, bool sequential)
         }
     }
 
+    // if n_keep_tail > 0, keep only the seqs that either finish in this ubatch or have at least
+    //   n_keep_tail tokens remaining for a future ubatch, so that the trailing n_keep_tail tokens
+    //   of each seq are never split across ubatches
+    if (n_keep_tail > 0) {
+        GGML_ASSERT(n_ubatch > n_keep_tail);
+
+        auto n_remaining = [&](uint32_t s) {
+            return (uint32_t) (seq_set_map[cur_seq_set[s]].size() - cur_idx[s]);
+        };
+
+        // keep the longest prefix of seqs that satisfy the constraint, to preserve sequential seq ids
+        uint32_t n_keep = 0;
+        while (n_keep < n_seqs) {
+            const uint32_t remaining = n_remaining(n_keep);
+
+            if (remaining != 0 && remaining < n_keep_tail) {
+                break;
+            }
+
+            n_keep++;
+        }
+
+        // all seqs violate the constraint - resolve the first one directly and emit it alone
+        if (n_keep == 0) {
+            auto & idxs = idxs_per_seq[0];
+
+            const auto & seq_idxs = seq_set_map[cur_seq_set[0]];
+
+            if (idxs.size() + n_remaining(0) <= n_ubatch) {
+                // extend the seq to completion
+                while (n_remaining(0) > 0) {
+                    const int32_t idx = seq_idxs[cur_idx[0]];
+
+                    idxs.push_back(idx);
+
+                    used[idx] = true;
+                    ++n_used;
+
+                    ++cur_idx[0];
+                }
+            } else {
+                // truncate the seq so that at least n_keep_tail tokens remain
+                while (n_remaining(0) < n_keep_tail) {
+                    used[idxs.back()] = false;
+                    --n_used;
+
+                    idxs.pop_back();
+
+                    --cur_idx[0];
+                }
+            }
+
+            n_keep = 1;
+        }
+
+        // return the tokens of the deferred seqs back to the pool
+        for (uint32_t s = n_keep; s < n_seqs; ++s) {
+            for (const int32_t idx : idxs_per_seq[s]) {
+                used[idx] = false;
+                --n_used;
+            }
+        }
+
+        n_seqs = n_keep;
+    }
+
     // concat the per-sequence-set lists
     std::vector<int32_t> idxs;
 
@@ -814,7 +882,7 @@ void llama_batch_allocr::ubatch_print(const llama_ubatch & ubatch, int debug) {
         LLAMA_LOG_DEBUG("%s:   output     = %p\n", __func__, (void *) ubatch.output);
         LLAMA_LOG_DEBUG("%s:   n_outputs  = %d\n", __func__, n_outputs);
 
-        if (debug > 1) {
+        if (debug > 0) {
             int seq_id_max = 0;
             for (uint32_t i = 0; i < ubatch.n_tokens; ++i) {
                 for (int s = 0; s < ubatch.n_seq_id[i]; ++s) {

examples/talk-llama/llama-batch.h

@@ -104,7 +104,8 @@ class llama_batch_allocr {
 
     // make ubatches of equal-length sequences sets
     // if sequential == true, the tokens in the ubatch will have increasing sequential sequence ids
-    llama_ubatch split_equal(uint32_t n_ubatch, bool sequential);
+    // n_keep_tail = minimum trailing tokens of a seq that must land in the same ubatch
+    llama_ubatch split_equal(uint32_t n_ubatch, bool sequential, uint32_t n_keep_tail);
 
     // sequence-set-wise split - each ubatch contains a single sequence-set
     llama_ubatch split_seq(uint32_t n_ubatch);