spec: fix failed to decode mtmd chunk with DFlash (#28587)

* speculative: fix failed to decode mtmd chunk with DFlash

When using DFlash w/ vision models, the drafter memory fails to
allocate new tokens because images report a fixed offset. Stop copying
them to allow the drafter to continue.

* address PR feedback

limit M-RoPE skip to images only, allow audio to pass through. Clean up
comments to align to the updated implementation

common/speculative.cpp

@@ -1094,8 +1094,7 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
 
         // Target prefill may contain token IDs or multimodal embeddings. Both
         // produce the target-layer features used to seed the draft KV cache, so
-        // skipping the embedding batches leaves a hole in the draft's cache and
-        // the next injection fails to initialize.
+        // embeddings are injected too, except the pinned ones skipped below.
         // TODO: revisit after https://github.com/ggml-org/llama.cpp/pull/24669 is merged
         const bool has_tokens     = batch_in.token != nullptr;
         const bool has_embeddings = batch_in.embd  != nullptr;
@@ -1131,6 +1130,13 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
             }
             const int32_t n_rows = i_batch_end[seq_id] - i_batch_beg[seq_id] + 1;
 
+            // an M-RoPE image pins all its rows to one position, so a windowed draft
+            // cache cannot free cells for it - skip it, the draft can jump over the gap
+            const bool pos_pinned = batch_in.pos[i_batch_beg[seq_id]] == batch_in.pos[i_batch_end[seq_id]];
+            if (has_embeddings && n_rows > 1 && pos_pinned) {
+                continue;
+            }
+
             for (int32_t offset = 0; offset < n_rows; offset += n_ubatch) {
                 const int32_t n_chunk = std::min(n_ubatch, n_rows - offset);
 

scripts/make-release-checks.sh

@@ -22,9 +22,9 @@ for arg in "$@"; do
     esac
 done
 
-MAJOR=$(grep "set(LLAMA_VERSION_MAJOR" "$REPO_ROOT/CMakeLists.txt" | grep -oP '\d+')
-MINOR=$(grep "set(LLAMA_VERSION_MINOR" "$REPO_ROOT/CMakeLists.txt" | grep -oP '\d+')
-PATCH=$(grep "set(LLAMA_VERSION_PATCH" "$REPO_ROOT/CMakeLists.txt" | grep -oP '\d+')
+MAJOR=$(grep "set(LLAMA_VERSION_MAJOR" "$REPO_ROOT/CMakeLists.txt" | sed 's/.*MAJOR \([0-9]*\).*/\1/')
+MINOR=$(grep "set(LLAMA_VERSION_MINOR" "$REPO_ROOT/CMakeLists.txt" | sed 's/.*MINOR \([0-9]*\).*/\1/')
+PATCH=$(grep "set(LLAMA_VERSION_PATCH" "$REPO_ROOT/CMakeLists.txt" | sed 's/.*PATCH \([0-9]*\).*/\1/')
 VERSION="v${MAJOR}.${MINOR}.${PATCH}"
 echo "Determined version: ${VERSION}"
 if [[ -n "${GITHUB_OUTPUT:-}" ]]; then
@@ -91,9 +91,9 @@ else
     fi
 fi
 
-MAJOR=$(grep "set(GGML_VERSION_MAJOR" "$REPO_ROOT/ggml/CMakeLists.txt" | grep -oP '\d+')
-MINOR=$(grep "set(GGML_VERSION_MINOR" "$REPO_ROOT/ggml/CMakeLists.txt" | grep -oP '\d+')
-PATCH=$(grep "set(GGML_VERSION_PATCH" "$REPO_ROOT/ggml/CMakeLists.txt" | grep -oP '\d+')
+MAJOR=$(grep "set(GGML_VERSION_MAJOR" "$REPO_ROOT/ggml/CMakeLists.txt" | sed 's/.*MAJOR \([0-9]*\).*/\1/')
+MINOR=$(grep "set(GGML_VERSION_MINOR" "$REPO_ROOT/ggml/CMakeLists.txt" | sed 's/.*MINOR \([0-9]*\).*/\1/')
+PATCH=$(grep "set(GGML_VERSION_PATCH" "$REPO_ROOT/ggml/CMakeLists.txt" | sed 's/.*PATCH \([0-9]*\).*/\1/')
 GGML_VERSION="v${MAJOR}.${MINOR}.${PATCH}"
 echo "Local ggml version: ${GGML_VERSION}"
 

src/models/bert.cpp

@@ -29,15 +29,13 @@ void llama_model_bert::load_arch_tensors(llama_model_loader &) {
     tok_embd     = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, 0);
     type_embd    = create_tensor(tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_token_types}, TENSOR_NOT_REQUIRED);
 
-    if (arch == LLM_ARCH_BERT) {
-        pos_embd = create_tensor(tn(LLM_TENSOR_POS_EMBD,    "weight"), {n_embd, n_ctx_train}, 0);
+    pos_embd = create_tensor(tn(LLM_TENSOR_POS_EMBD,    "weight"), {n_embd, n_ctx_train}, 0);
 
-        cls   = create_tensor(tn(LLM_TENSOR_CLS, "weight"), {n_embd, n_embd}, TENSOR_NOT_REQUIRED);
-        cls_b = create_tensor(tn(LLM_TENSOR_CLS, "bias"),   {n_embd},         TENSOR_NOT_REQUIRED);
+    cls   = create_tensor(tn(LLM_TENSOR_CLS, "weight"), {n_embd, n_embd}, TENSOR_NOT_REQUIRED);
+    cls_b = create_tensor(tn(LLM_TENSOR_CLS, "bias"),   {n_embd},         TENSOR_NOT_REQUIRED);
 
-        cls_out   = create_tensor(tn(LLM_TENSOR_CLS_OUT, "weight"), {n_embd, hparams.n_cls_out}, TENSOR_NOT_REQUIRED);
-        cls_out_b = create_tensor(tn(LLM_TENSOR_CLS_OUT, "bias"),   {hparams.n_cls_out},         TENSOR_NOT_REQUIRED);
-    }
+    cls_out   = create_tensor(tn(LLM_TENSOR_CLS_OUT, "weight"), {n_embd, hparams.n_cls_out}, TENSOR_NOT_REQUIRED);
+    cls_out_b = create_tensor(tn(LLM_TENSOR_CLS_OUT, "bias"),   {hparams.n_cls_out},         TENSOR_NOT_REQUIRED);
 
     tok_norm   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {n_embd}, 0);
     tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias",   0), {n_embd}, 0);
@@ -53,20 +51,10 @@ void llama_model_bert::load_arch_tensors(llama_model_loader &) {
         layer.attn_out_norm   = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}, 0);
         layer.attn_out_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "bias", i),   {n_embd}, 0);
 
-        if (hparams.moe_every_n_layers > 0 && i % hparams.moe_every_n_layers == 1) {
-            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff,   n_expert}, 0);
-            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff,   n_embd, n_expert}, 0);
-            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,   "weight", i), {n_embd, n_expert}, 0);
-        } else {
-            layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
-            layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, TENSOR_NOT_REQUIRED);
-            layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
-            layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, TENSOR_NOT_REQUIRED);
-
-            if (arch == LLM_ARCH_NOMIC_BERT) {
-                layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
-            }
-        }
+        layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
+        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, TENSOR_NOT_REQUIRED);
+        layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, TENSOR_NOT_REQUIRED);
 
         layer.layer_out_norm   = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}, 0);
         layer.layer_out_norm_b = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "bias", i),   {n_embd}, 0);

src/models/jina-bert-v3.cpp

@@ -19,16 +19,6 @@ void llama_model_jina_bert_v3::load_arch_tensors(llama_model_loader &) {
     tok_embd     = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, 0);
     type_embd    = create_tensor(tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_token_types}, TENSOR_NOT_REQUIRED);
 
-    if (arch == LLM_ARCH_BERT) {
-        pos_embd = create_tensor(tn(LLM_TENSOR_POS_EMBD,    "weight"), {n_embd, n_ctx_train}, 0);
-
-        cls   = create_tensor(tn(LLM_TENSOR_CLS, "weight"), {n_embd, n_embd}, TENSOR_NOT_REQUIRED);
-        cls_b = create_tensor(tn(LLM_TENSOR_CLS, "bias"),   {n_embd},         TENSOR_NOT_REQUIRED);
-
-        cls_out   = create_tensor(tn(LLM_TENSOR_CLS_OUT, "weight"), {n_embd, hparams.n_cls_out}, TENSOR_NOT_REQUIRED);
-        cls_out_b = create_tensor(tn(LLM_TENSOR_CLS_OUT, "bias"),   {hparams.n_cls_out},         TENSOR_NOT_REQUIRED);
-    }
-
     tok_norm   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {n_embd}, 0);
     tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias",   0), {n_embd}, 0);
 
@@ -43,20 +33,10 @@ void llama_model_jina_bert_v3::load_arch_tensors(llama_model_loader &) {
         layer.attn_out_norm   = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}, 0);
         layer.attn_out_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "bias", i),   {n_embd}, 0);
 
-        if (hparams.moe_every_n_layers > 0 && i % hparams.moe_every_n_layers == 1) {
-            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff,   n_expert}, 0);
-            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff,   n_embd, n_expert}, 0);
-            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,   "weight", i), {n_embd, n_expert}, 0);
-        } else {
-            layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
-            layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, TENSOR_NOT_REQUIRED);
-            layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
-            layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, TENSOR_NOT_REQUIRED);
-
-            if (arch == LLM_ARCH_NOMIC_BERT) {
-                layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
-            }
-        }
+        layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
+        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, TENSOR_NOT_REQUIRED);
+        layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, TENSOR_NOT_REQUIRED);
 
         layer.layer_out_norm   = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}, 0);
         layer.layer_out_norm_b = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "bias", i),   {n_embd}, 0);

src/models/nomic-bert-moe.cpp

@@ -4,12 +4,10 @@ void llama_model_nomic_bert_moe::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,    hparams.f_norm_eps);
     ml.get_key(LLM_KV_MOE_EVERY_N_LAYERS,         hparams.moe_every_n_layers, 0);
 
-    if (hparams.n_layer() == 12 && hparams.n_embd == 768) {
-        if (arch == LLM_ARCH_NOMIC_BERT) {
-            type = LLM_TYPE_137M;
-        } else if (arch == LLM_ARCH_NOMIC_BERT_MOE && hparams.moe_every_n_layers == 2) {
-            type = LLM_TYPE_475M;
-        }
+    switch (hparams.n_layer()) {
+        case 12:
+            type = LLM_TYPE_475M; break;
+        default: type = LLM_TYPE_UNKNOWN;
     }
 }
 
@@ -22,16 +20,6 @@ void llama_model_nomic_bert_moe::load_arch_tensors(llama_model_loader &) {
     tok_embd     = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, 0);
     type_embd    = create_tensor(tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_token_types}, TENSOR_NOT_REQUIRED);
 
-    if (arch == LLM_ARCH_BERT) {
-        pos_embd = create_tensor(tn(LLM_TENSOR_POS_EMBD,    "weight"), {n_embd, n_ctx_train}, 0);
-
-        cls   = create_tensor(tn(LLM_TENSOR_CLS, "weight"), {n_embd, n_embd}, TENSOR_NOT_REQUIRED);
-        cls_b = create_tensor(tn(LLM_TENSOR_CLS, "bias"),   {n_embd},         TENSOR_NOT_REQUIRED);
-
-        cls_out   = create_tensor(tn(LLM_TENSOR_CLS_OUT, "weight"), {n_embd, hparams.n_cls_out}, TENSOR_NOT_REQUIRED);
-        cls_out_b = create_tensor(tn(LLM_TENSOR_CLS_OUT, "bias"),   {hparams.n_cls_out},         TENSOR_NOT_REQUIRED);
-    }
-
     tok_norm   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {n_embd}, 0);
     tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias",   0), {n_embd}, 0);
 
@@ -55,10 +43,6 @@ void llama_model_nomic_bert_moe::load_arch_tensors(llama_model_loader &) {
             layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, TENSOR_NOT_REQUIRED);
             layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
             layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, TENSOR_NOT_REQUIRED);
-
-            if (arch == LLM_ARCH_NOMIC_BERT) {
-                layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
-            }
         }
 
         layer.layer_out_norm   = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}, 0);

src/models/nomic-bert.cpp

@@ -1,15 +1,12 @@
 #include "models.h"
 
 void llama_model_nomic_bert::load_arch_hparams(llama_model_loader & ml) {
-    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,    hparams.f_norm_eps);
-    ml.get_key(LLM_KV_MOE_EVERY_N_LAYERS,         hparams.moe_every_n_layers, 0);
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
 
-    if (hparams.n_layer() == 12 && hparams.n_embd == 768) {
-        if (arch == LLM_ARCH_NOMIC_BERT) {
-            type = LLM_TYPE_137M;
-        } else if (arch == LLM_ARCH_NOMIC_BERT_MOE && hparams.moe_every_n_layers == 2) {
-            type = LLM_TYPE_475M;
-        }
+    switch (hparams.n_layer()) {
+        case 12:
+            type = LLM_TYPE_137M; break;
+        default: type = LLM_TYPE_UNKNOWN;
     }
 }
 
@@ -22,16 +19,6 @@ void llama_model_nomic_bert::load_arch_tensors(llama_model_loader &) {
     tok_embd     = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, 0);
     type_embd    = create_tensor(tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_token_types}, TENSOR_NOT_REQUIRED);
 
-    if (arch == LLM_ARCH_BERT) {
-        pos_embd = create_tensor(tn(LLM_TENSOR_POS_EMBD,    "weight"), {n_embd, n_ctx_train}, 0);
-
-        cls   = create_tensor(tn(LLM_TENSOR_CLS, "weight"), {n_embd, n_embd}, TENSOR_NOT_REQUIRED);
-        cls_b = create_tensor(tn(LLM_TENSOR_CLS, "bias"),   {n_embd},         TENSOR_NOT_REQUIRED);
-
-        cls_out   = create_tensor(tn(LLM_TENSOR_CLS_OUT, "weight"), {n_embd, hparams.n_cls_out}, TENSOR_NOT_REQUIRED);
-        cls_out_b = create_tensor(tn(LLM_TENSOR_CLS_OUT, "bias"),   {hparams.n_cls_out},         TENSOR_NOT_REQUIRED);
-    }
-
     tok_norm   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {n_embd}, 0);
     tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias",   0), {n_embd}, 0);
 
@@ -46,20 +33,12 @@ void llama_model_nomic_bert::load_arch_tensors(llama_model_loader &) {
         layer.attn_out_norm   = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}, 0);
         layer.attn_out_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "bias", i),   {n_embd}, 0);
 
-        if (hparams.moe_every_n_layers > 0 && i % hparams.moe_every_n_layers == 1) {
-            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff,   n_expert}, 0);
-            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff,   n_embd, n_expert}, 0);
-            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,   "weight", i), {n_embd, n_expert}, 0);
-        } else {
-            layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
-            layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, TENSOR_NOT_REQUIRED);
-            layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
-            layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, TENSOR_NOT_REQUIRED);
+        layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
+        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, TENSOR_NOT_REQUIRED);
+        layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, TENSOR_NOT_REQUIRED);
 
-            if (arch == LLM_ARCH_NOMIC_BERT) {
-                layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
-            }
-        }
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
 
         layer.layer_out_norm   = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}, 0);
         layer.layer_out_norm_b = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "bias", i),   {n_embd}, 0);