model : support classifier_pooling for rerankers (#29627)

* model : support classifier_pooling for ModernBERT rerankers

Assisted-by: Claude Opus 5.5

* model : read classifier pooling type in load_hparams

Write classifier.pooling_type from _try_set_pooling_type whenever the
config has classifier_pooling, and read it in
llama_model_base::load_hparams. ModernBERT falls back to mean when it
is unspecified.

Assisted-by: Claude Opus 5.5

* conversion : only accept cls and mean for classifier_pooling

Assisted-by: Claude Opus 5.5

* model : rename classifier_pooling_type to pooling_type_cls

Assisted-by: Claude Opus 5.5

src/llama-arch.cpp

@@ -361,6 +361,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_CONVNEXT_BLOCK_COUNT,      "%s.convnext.block_count"      },
 
     { LLM_KV_CLASSIFIER_OUTPUT_LABELS, "%s.classifier.output_labels" },
+    { LLM_KV_CLASSIFIER_POOLING_TYPE,  "%s.classifier.pooling_type"  },
 
     { LLM_KV_TARGET_LAYERS,         "%s.target_layers"        },
     { LLM_KV_TARGET_HIDDEN_SIZE,    "%s.target_hidden_size"   },

src/llama-arch.h

@@ -407,6 +407,7 @@ enum llm_kv {
     LLM_KV_CONVNEXT_BLOCK_COUNT,
 
     LLM_KV_CLASSIFIER_OUTPUT_LABELS,
+    LLM_KV_CLASSIFIER_POOLING_TYPE,
 
     LLM_KV_TARGET_LAYERS,
     LLM_KV_TARGET_HIDDEN_SIZE,

src/llama-graph.cpp

@@ -3727,8 +3727,8 @@ void llm_graph_context::build_pooling(
             } break;
         case LLAMA_POOLING_TYPE_RANK:
             {
-                if (arch == LLM_ARCH_MODERN_BERT) {
-                    // modern bert gte reranker builds mean first then applies prediction head and classifier
+                if (hparams.pooling_type_cls == LLAMA_POOLING_TYPE_MEAN) {
+                    // modern bert with classifier_pooling = "mean" builds mean first then applies prediction head and classifier
                     // https://github.com/huggingface/transformers/blob/main/src/transformers/models/modernbert/modular_modernbert.py#L1404-1411
                     ggml_tensor * inp_mean = build_inp_mean();
                     cur = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, inp)), inp_mean);

src/llama-hparams.h

@@ -350,6 +350,7 @@ struct llama_hparams {
     uint32_t    dec_n_layer        = 0;
 
     enum llama_pooling_type      pooling_type            = LLAMA_POOLING_TYPE_NONE;
+    enum llama_pooling_type      pooling_type_cls        = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling before the classifier head (RANK)
     enum llama_rope_type         rope_type               = LLAMA_ROPE_TYPE_NONE;
     enum llama_rope_scaling_type rope_scaling_type_train = LLAMA_ROPE_SCALING_TYPE_NONE;
 

src/llama-model.cpp

@@ -1320,6 +1320,7 @@ void llama_model_base::load_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_EMBEDDING_LENGTH_OUT,    hparams.n_embd_out_impl, false);
     ml.get_key(LLM_KV_ATTENTION_CAUSAL,        hparams.causal_attn,     false);
     ml.get_key(LLM_KV_POOLING_TYPE,            hparams.pooling_type,    false);
+    ml.get_key(LLM_KV_CLASSIFIER_POOLING_TYPE, hparams.pooling_type_cls, false);
     ml.get_key(LLM_KV_BLOCK_COUNT,             hparams.n_layer_all);
     GGML_ASSERT(hparams.n_layer_all > 0 && hparams.n_layer_all <= LLAMA_MAX_LAYERS);
     ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS,    hparams.n_layer_nextn,   false);

src/models/modern-bert.cpp

@@ -20,6 +20,11 @@ void llama_model_modern_bert::load_arch_hparams(llama_model_loader & ml) {
         hparams.llm_ffn_op = llm_ffn_op_type_from_string(hidden_act, LLM_FFN_GEGLU);
     }
 
+    // GGUFs without a classifier pooling type use mean (gte-reranker-modernbert-base)
+    if (hparams.pooling_type_cls == LLAMA_POOLING_TYPE_UNSPECIFIED) {
+        hparams.pooling_type_cls = LLAMA_POOLING_TYPE_MEAN;
+    }
+
     switch (hparams.n_layer()) {
         case 12:
             type = LLM_TYPE_47M; break; // granite-embedding-small