feat: Hybrid unified/recurrent cache

include/llama.h

@@ -552,6 +552,9 @@ extern "C" {
     // Returns true if the model is recurrent (like Mamba, RWKV, etc.)
     LLAMA_API bool llama_model_is_recurrent(const struct llama_model * model);
 
+    // Returns true if the model is hybrid-recurrent (like Jamba, Bamba, etc.)
+    LLAMA_API bool llama_model_is_hybrid(const struct llama_model * model);
+
     // Returns 0 on success
     LLAMA_API uint32_t llama_model_quantize(
             const char * fname_inp,

src/llama-arch.cpp

@@ -144,6 +144,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ATTENTION_SCALE,                        "%s.attention.scale"                        },
     { LLM_KV_ATTENTION_KEY_LENGTH_MLA,               "%s.attention.key_length_mla"               },
     { LLM_KV_ATTENTION_VALUE_LENGTH_MLA,             "%s.attention.value_length_mla"             },
+    { LLM_KV_ATTENTION_LAYER_INDICES,                "%s.attention.layer_indices"                },
 
     { LLM_KV_ROPE_DIMENSION_COUNT,      "%s.rope.dimension_count"                 },
     { LLM_KV_ROPE_DIMENSION_SECTIONS,   "%s.rope.dimension_sections"              },
@@ -1744,3 +1745,25 @@ llm_arch llm_arch_from_string(const std::string & name) {
 const llm_tensor_info & llm_tensor_info_for(llm_tensor tensor) {
     return LLM_TENSOR_INFOS.at(tensor);
 }
+
+bool llm_arch_is_recurrent(const llm_arch & arch) {
+    switch (arch) {
+        case LLM_ARCH_MAMBA:
+        case LLM_ARCH_RWKV6:
+        case LLM_ARCH_RWKV6QWEN2:
+        case LLM_ARCH_RWKV7:
+        case LLM_ARCH_ARWKV7:
+            return true;
+        default:
+            return false;
+    }
+}
+
+bool llm_arch_is_hybrid(const llm_arch & arch) {
+    // TODO: There are currently no hybrid models! Once there are, this will be
+    //  the place to identify them
+    switch (arch) {
+        default:
+            return false;
+    }
+}

src/llama-arch.h

@@ -148,6 +148,7 @@ enum llm_kv {
     LLM_KV_ATTENTION_SCALE,
     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,
@@ -435,3 +436,6 @@ const char * llm_arch_name(llm_arch arch);
 llm_arch llm_arch_from_string(const std::string & name);
 
 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);

src/llama-hparams.cpp

@@ -49,7 +49,10 @@ uint32_t llama_hparams::n_embd_v_gqa(uint32_t il) const {
     return n_embd_head_v * n_head_kv;
 }
 
-uint32_t llama_hparams::n_embd_k_s() const {
+uint32_t llama_hparams::n_embd_k_s(uint32_t il) const {
+    if (!recurrent_layer(il)) {
+        return 0;
+    }
     if (wkv_head_size != 0) {
         // for RWKV models
         return token_shift_count * n_embd;
@@ -60,7 +63,10 @@ uint32_t llama_hparams::n_embd_k_s() const {
     return (ssm_d_conv > 0 ? ssm_d_conv - 1 : 0) * ssm_d_inner;
 }
 
-uint32_t llama_hparams::n_embd_v_s() const {
+uint32_t llama_hparams::n_embd_v_s(uint32_t il) const {
+    if (!recurrent_layer(il)) {
+        return 0;
+    }
     if (wkv_head_size != 0) {
         // corresponds to RWKV's wkv_states size
         return n_embd * wkv_head_size;
@@ -70,6 +76,10 @@ uint32_t llama_hparams::n_embd_v_s() const {
     return ssm_d_state * ssm_d_inner;
 }
 
+bool llama_hparams::recurrent_layer(uint32_t il) const {
+    return recurrent_layer_arr[il];
+}
+
 bool llama_hparams::is_swa(uint32_t il) const {
     if (il < n_layer) {
         return n_swa > 0 && n_swa_pattern > 0 && il % n_swa_pattern < (n_swa_pattern - 1);

src/llama-hparams.h

@@ -102,6 +102,9 @@ struct llama_hparams {
     uint32_t ssm_d_state = 0;
     uint32_t ssm_dt_rank = 0;
 
+    // for hybrid state space models
+    std::array<bool, LLAMA_MAX_LAYERS> recurrent_layer_arr;
+
     bool ssm_dt_b_c_rms = false;
 
     float f_clamp_kqv      = 0.0f;
@@ -149,10 +152,13 @@ struct llama_hparams {
 
     // dimension of the rolling state embeddings
     // corresponds to Mamba's conv_states size or RWKV's token_shift states size
-    uint32_t n_embd_k_s() const;
+    uint32_t n_embd_k_s(uint32_t il = 0) const;
 
     // dimension of the recurrent state embeddings
-    uint32_t n_embd_v_s() const;
+    uint32_t n_embd_v_s(uint32_t il = 0) const;
+
+    // whether or not the given layer is recurrent (for hybrid models)
+    bool recurrent_layer(uint32_t il) const;
 
     bool is_swa(uint32_t il) const;
 };