llama : move the Engram pad id to llama_hparams and the token map to llama_model

The saver no longer needs the deepseek41 model class.

src/llama-hparams.h

@@ -358,6 +358,8 @@ struct llama_hparams {
     llama_token ple_embd_token   = LLAMA_TOKEN_NULL;
     // hash constant sets, stored back to back: 1 shared by all PLE layers (qwen4exp), or 1 per PLE layer in layer order (deepseek41)
     uint32_t ple_n_modules       = 1;
+    // deepseek41 Engram: the token that pads the n-gram history, an index into llama_model::engram_token_map
+    uint32_t engram_pad_id       = 0;
     // the file lists PLE layer indices, so this is never a per-layer gguf array and can hold one bit per layer
     std::bitset<LLAMA_MAX_LAYERS> is_ple_impl;
     // the hash multipliers reach ~2e13 and have to stay 64-bit

src/llama-model-saver.cpp

@@ -8,7 +8,6 @@
 #include "llama-hparams.h"
 #include "llama-model.h"
 #include "llama-vocab.h"
-#include "models/models.h"
 
 #include <cstdint>
 #include <string>
@@ -342,8 +341,6 @@ void llama_model_saver::add_kv_from_model() {
     add_kv(LLM_KV_HYPER_CONNECTION_LOW_RANK,             hparams.hc_low_rank);
 
     if (model->arch == LLM_ARCH_DEEPSEEK41) {
-        const auto * m41 = dynamic_cast<const llama_model_deepseek41 *>(model);
-        GGML_ASSERT(m41 != nullptr);
         std::vector<uint32_t> kv_src;
         std::vector<uint32_t> idx_src;
         for (uint32_t il = 0; il < hparams.n_layer(); ++il) {
@@ -362,8 +359,8 @@ void llama_model_saver::add_kv_from_model() {
             add_kv(LLM_KV_ATTENTION_INDEXER_CANDIDATE_BLOCK_SIZE,   hparams.dsv41_cand_block_size);
         }
         if (hparams.ple_n_heads > 0) {
-            add_kv(LLM_KV_ENGRAM_PAD_TOKEN_ID, m41->engram_pad_id);
-            add_kv(LLM_KV_ENGRAM_TOKEN_MAP,    m41->engram_token_map);
+            add_kv(LLM_KV_ENGRAM_PAD_TOKEN_ID, hparams.engram_pad_id);
+            add_kv(LLM_KV_ENGRAM_TOKEN_MAP,    model->engram_token_map);
         }
     }
 

src/llama-model.h

@@ -730,6 +730,9 @@ struct llama_model {
     std::vector<std::string> classifier_labels;
     std::vector<std::string> decision_labels;
 
+    // deepseek41 Engram: token id -> id of its normalized form, hashed into the n-grams
+    std::vector<int32_t> engram_token_map;
+
     struct ggml_tensor * tok_embd   = nullptr;
     struct ggml_tensor * type_embd  = nullptr;
     struct ggml_tensor * pos_embd   = nullptr;

src/models/deepseek41.cpp

@@ -96,7 +96,7 @@ void llama_model_deepseek41::load_arch_hparams(llama_model_loader & ml) {
         ml.get_arr(LLM_KV_PLE_LAYER_MULTIPLIERS,      mult);
         ml.get_arr(LLM_KV_PLE_HEAD_OFFSETS,           offsets);
         ml.get_arr(LLM_KV_PLE_HEAD_VOCAB_SIZES,       primes);
-        ml.get_key(LLM_KV_ENGRAM_PAD_TOKEN_ID,        engram_pad_id);
+        ml.get_key(LLM_KV_ENGRAM_PAD_TOKEN_ID,        hparams.engram_pad_id);
         ml.get_arr(LLM_KV_ENGRAM_TOKEN_MAP,           engram_token_map);
 
         const size_t n_e    = e_layers.size();
@@ -105,7 +105,7 @@ void llama_model_deepseek41::load_arch_hparams(llama_model_loader & ml) {
         if (n_cols == 0 || hparams.ple_head_dim == 0 ||
                 mult.size() != n_e*n_gram || primes.size() != n_e*n_cols || offsets.size() != n_e*n_cols ||
                 mult.size() > LLAMA_MAX_PLE_NGRAM || primes.size() > LLAMA_MAX_PLE_HEADS ||
-                engram_pad_id >= engram_token_map.size()) {
+                hparams.engram_pad_id >= engram_token_map.size()) {
             throw std::runtime_error("DeepSeek-V4.1 engram metadata is inconsistent");
         }
         for (size_t e = 0; e < n_e; ++e) {
@@ -247,7 +247,7 @@ void llama_model_deepseek41::engram_hash(int e, const llama_ubatch & ubatch, con
 
     const uint64_t * mult   = hparams.ple_layer_multipliers.data() + (size_t) e*n_gram;
     const uint32_t * primes = hparams.ple_head_vocab_sizes.data()  + (size_t) e*n_cols;
-    const int64_t    pad    = engram_token_map[engram_pad_id];
+    const int64_t    pad    = engram_token_map[hparams.engram_pad_id];
     const uint32_t * offs   = hparams.ple_head_offsets.data()      + (size_t) e*n_cols;
 
     const auto cid = [&](llama_token t) -> int64_t {

src/models/models.h

@@ -1498,11 +1498,7 @@ struct llama_model_deepseek41 : public llama_model_deepseek4 {
     void load_arch_hparams(llama_model_loader & ml) override;
     void load_arch_tensors(llama_model_loader & ml) override;
 
-    // Engram: the hash constants are hparams.ple_* (one module per Engram layer), hashed on the host (ggml has no
-    // int64 mul/xor/mod); the token map and its pad id are V4.1 only
-    uint32_t engram_pad_id = 0;
-    std::vector<int32_t> engram_token_map;
-
+    // Engram: the hash constants are hparams.ple_* (one module per Engram layer), hashed on the host (ggml has no int64 mul/xor/mod)
     int engram_index(int il) const;
 
     void engram_hash(int e, const llama_ubatch & ubatch, const std::vector<llama_token> & prev, std::vector<int32_t> & out) const;