feat: Overhaul build_recurrent_state / build_inp_s_copy to match attention pattern

https://github.com/ggml-org/llama.cpp/pull/13979/files#r2141701738

This is a big overhaul to bring consistency between how inputs and per-
layer components are created for attention layers and recurrent layers. The
main changes are:

- Rename class llm_graph_input_s_copy -> llm_graph_input_rs
- Add a corresponding llm_graph_input_rs_hybrid_recurrent
- Rename build_inp_s_copy -> build_rs_inp_recurrent
- Add a corresponding build_rs_inp_hybrid_recurrent
- Rename build_recurrent_state -> build_rs to match build_attn w/
llm_graph_input_rs android-build AUTHORS bamba-9b-2.2T.gguf bamba-9b-2.2T.q4_k_m.gguf broken.log build build-rel build-xcframework.sh build.android build.android.bak ci cmake CMakeLists.txt CMakePresets.json CODEOWNERS common common.o CONTRIBUTING.md convert_hf_to_gguf_update.py convert_hf_to_gguf.py convert_llama_ggml_to_gguf.py convert_lora_to_gguf.py debug.log docs examples flake.lock flake.nix ggml ggml-alloc.o ggml-backend.o ggml-metal.o ggml-model-BF16.gguf ggml-model-Q4_K_M.gguf ggml-quants.o ggml.o gguf-py grammar-parser.o grammars include LICENSE licenses llama.log llama.o llamacpp_trace.log main.log Makefile media models mypy.ini pocs poetry.lock prompts pyproject.toml pyrightconfig.json q4_k_m_boot.log q8_0_boot.log quant.log quant2.log README.md requirements requirements.txt sampling.o scripts SECURITY.md src test-grammar-output.tmp test-json-schema-input.tmp tests tools vendor working.log as the first input
- Add a corresponding overload of build_rs w/
llm_graph_input_rs_hybrid_recurrent android-build AUTHORS bamba-9b-2.2T.gguf bamba-9b-2.2T.q4_k_m.gguf broken.log build build-rel build-xcframework.sh build.android build.android.bak ci cmake CMakeLists.txt CMakePresets.json CODEOWNERS common common.o CONTRIBUTING.md convert_hf_to_gguf_update.py convert_hf_to_gguf.py convert_llama_ggml_to_gguf.py convert_lora_to_gguf.py debug.log docs examples flake.lock flake.nix ggml ggml-alloc.o ggml-backend.o ggml-metal.o ggml-model-BF16.gguf ggml-model-Q4_K_M.gguf ggml-quants.o ggml.o gguf-py grammar-parser.o grammars include LICENSE licenses llama.log llama.o llamacpp_trace.log main.log Makefile media models mypy.ini pocs poetry.lock prompts pyproject.toml pyrightconfig.json q4_k_m_boot.log q8_0_boot.log quant.log quant2.log README.md requirements requirements.txt sampling.o scripts SECURITY.md src test-grammar-output.tmp test-json-schema-input.tmp tests tools vendor working.log as the first input
- Add a llm_graph_input_attn_kv_hybrid_recurrent analogous to
llm_graph_input_attn_kv_unified
- Add a build_attn override that takes
llm_graph_input_attn_kv_hybrid_recurrent android-build AUTHORS bamba-9b-2.2T.gguf bamba-9b-2.2T.q4_k_m.gguf broken.log build build-rel build-xcframework.sh build.android build.android.bak ci cmake CMakeLists.txt CMakePresets.json CODEOWNERS common common.o CONTRIBUTING.md convert_hf_to_gguf_update.py convert_hf_to_gguf.py convert_llama_ggml_to_gguf.py convert_lora_to_gguf.py debug.log docs examples flake.lock flake.nix ggml ggml-alloc.o ggml-backend.o ggml-metal.o ggml-model-BF16.gguf ggml-model-Q4_K_M.gguf ggml-quants.o ggml.o gguf-py grammar-parser.o grammars include LICENSE licenses llama.log llama.o llamacpp_trace.log main.log Makefile media models mypy.ini pocs poetry.lock prompts pyproject.toml pyrightconfig.json q4_k_m_boot.log q8_0_boot.log quant.log quant2.log README.md requirements requirements.txt sampling.o scripts SECURITY.md src test-grammar-output.tmp test-json-schema-input.tmp tests tools vendor working.log as the first input

This makes the two paradigms fully consistent. The main drawback is the
code duplication in the build_attn and build_rs implementations where the
only difference between implementations is how they cast the memory state.

Branch: HybridRecurrentCache

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

Author: gabe-l-hart

src/llama-graph.cpp

@@ -239,7 +239,7 @@ void llm_graph_input_cls::set_input(const llama_ubatch * ubatch) {
     }
 }
 
-void llm_graph_input_s_copy::set_input(const llama_ubatch * ubatch) {
+void llm_graph_input_rs::set_input(const llama_ubatch * ubatch) {
     GGML_UNUSED(ubatch);
 
     const int64_t n_kv = kv_state->get_n_kv();
@@ -255,6 +255,11 @@ void llm_graph_input_s_copy::set_input(const llama_ubatch * ubatch) {
     }
 }
 
+llm_graph_input_rs_hybrid_recurrent::llm_graph_input_rs_hybrid_recurrent(
+        const llama_kv_cache_hybrid_recurrent_state * kv_state) :
+    llm_graph_input_rs(kv_state->get_state_recurrent()) {
+}
+
 void llm_graph_input_cross_embd::set_input(const llama_ubatch * ubatch) {
     GGML_UNUSED(ubatch);
 
@@ -360,6 +365,13 @@ void llm_graph_input_attn_kv_unified::set_input(const llama_ubatch * ubatch) {
     }
 }
 
+llm_graph_input_attn_kv_hybrid_recurrent::llm_graph_input_attn_kv_hybrid_recurrent(
+        const llama_hparams & hparams,
+        const llama_cparams & cparams,
+        const llama_kv_cache_hybrid_recurrent_state * kv_state) :
+    llm_graph_input_attn_kv_unified(hparams, cparams, kv_state->get_state_attn()) {
+}
+
 void llm_graph_input_attn_kv_unified_iswa::set_input(const llama_ubatch * ubatch) {
     if (self_kq_mask) {
         kv_state->get_base()->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
@@ -962,25 +974,6 @@ ggml_tensor * llm_graph_context::build_inp_cls() const {
     return cur;
 }
 
-ggml_tensor * llm_graph_context::build_inp_s_copy(const llama_kv_cache_recurrent_state * kv_state) const {
-    if (kv_state == nullptr) {
-        kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
-    }
-
-    auto inp = std::make_unique<llm_graph_input_s_copy>(kv_state);
-
-    const auto n_kv = kv_state->get_n_kv();
-
-    auto & cur = inp->s_copy;
-
-    cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_kv);
-    ggml_set_input(cur);
-
-    res->add_input(std::move(inp));
-
-    return cur;
-}
-
 ggml_tensor * llm_graph_context::build_inp_cross_embd() const {
     auto inp = std::make_unique<llm_graph_input_cross_embd>(cross);
 
@@ -1262,9 +1255,7 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_build_forward_expand(gf, k_cur);
     ggml_build_forward_expand(gf, v_cur);
 
-    // NOTE: For hybrid caches, this may be a child of mstate, so we use the one
-    //  encapsulated in inp
-    const auto * kv_state = inp->kv_state;
+    const auto * kv_state = static_cast<const llama_kv_cache_unified_state *>(mstate);
 
     // store to KV cache
     {
@@ -1296,15 +1287,14 @@ ggml_tensor * llm_graph_context::build_attn(
     return cur;
 }
 
-llm_graph_input_attn_kv_unified * llm_graph_context::build_attn_inp_kv_hybrid_recurrent() const {
-    const auto * kv_state = static_cast<const llama_kv_cache_hybrid_recurrent_state *>(mstate);
-
-    auto inp = std::make_unique<llm_graph_input_attn_kv_unified>(hparams, cparams, kv_state->get_state_attn());
+llm_graph_input_attn_kv_hybrid_recurrent * llm_graph_context::build_attn_inp_kv_hybrid_recurrent() const {
+    auto inp = std::make_unique<llm_graph_input_attn_kv_hybrid_recurrent>(
+        hparams, cparams, static_cast<const llama_kv_cache_hybrid_recurrent_state *>(mstate));
 
     {
         GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Hybrid recurrent is not supported with SWA attention layers");
 
-        const auto n_kv = kv_state->get_state_attn()->get_n_kv();
+        const auto n_kv = inp->kv_state->get_n_kv();
 
         inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
         //cb(inp->self_kq_mask, "KQ_mask", -1);
@@ -1313,7 +1303,57 @@ llm_graph_input_attn_kv_unified * llm_graph_context::build_attn_inp_kv_hybrid_re
         inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
     }
 
-    return (llm_graph_input_attn_kv_unified *) res->add_input(std::move(inp));
+    return (llm_graph_input_attn_kv_hybrid_recurrent *) res->add_input(std::move(inp));
+}
+
+ggml_tensor * llm_graph_context::build_attn(
+        llm_graph_input_attn_kv_hybrid_recurrent * inp,
+        ggml_cgraph * gf,
+        ggml_tensor * wo,
+        ggml_tensor * wo_b,
+        ggml_tensor * q_cur,
+        ggml_tensor * k_cur,
+        ggml_tensor * v_cur,
+        ggml_tensor * kq_b,
+        ggml_tensor * v_mla,
+            float     kq_scale,
+            int       il) const {
+    // these nodes are added to the graph together so that they are not reordered
+    // by doing so, the number of splits in the graph is reduced
+    ggml_build_forward_expand(gf, q_cur);
+    ggml_build_forward_expand(gf, k_cur);
+    ggml_build_forward_expand(gf, v_cur);
+
+    const auto * kv_state = static_cast<const llama_kv_cache_hybrid_recurrent_state *>(mstate)->get_state_attn();
+
+    // store to KV cache
+    {
+        ggml_build_forward_expand(gf, kv_state->cpy_k(ctx0, k_cur, il));
+        ggml_build_forward_expand(gf, kv_state->cpy_v(ctx0, v_cur, il));
+    }
+
+    const auto & kq_mask = inp->get_kq_mask();
+
+    ggml_tensor * q = q_cur;
+    ggml_tensor * k = kv_state->get_k(ctx0, il);
+    ggml_tensor * v = kv_state->get_v(ctx0, il);
+
+    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale);
+    cb(cur, "kqv_out", il);
+
+    if (wo) {
+        cur = build_lora_mm(wo, cur);
+        if (arch == LLM_ARCH_GLM4) {
+            // GLM4 seems to have numerical issues with half-precision accumulators
+            ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
+        }
+    }
+
+    if (wo_b) {
+        cur = ggml_add(ctx0, cur, wo_b);
+    }
+
+    return cur;
 }
 
 llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unified_iswa() const {
@@ -1455,19 +1495,90 @@ ggml_tensor * llm_graph_context::build_attn(
     return cur;
 }
 
-ggml_tensor * llm_graph_context::build_recurrent_state(
-                             ggml_cgraph * gf,
-                             ggml_tensor * s,
-                             ggml_tensor * state_copy,
-                                 int32_t   state_size,
-                                 int32_t   n_seqs,
-                                    bool   avoid_copies,
-    const llama_kv_cache_recurrent_state * kv_state) const {
+llm_graph_input_rs * llm_graph_context::build_rs_inp_recurrent() const {
+    const auto * kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
+
+    auto inp = std::make_unique<llm_graph_input_rs>(kv_state);
+
+    const auto n_kv = kv_state->get_n_kv();
+
+    auto & cur = inp->s_copy;
+
+    cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_kv);
+    ggml_set_input(cur);
+
+    return (llm_graph_input_rs *) res->add_input(std::move(inp));
+}
+
+ggml_tensor * llm_graph_context::build_rs(
+        llm_graph_input_rs * inp,
+        ggml_cgraph * gf,
+        ggml_tensor * s,
+            int32_t   state_size,
+            int32_t   n_seqs,
+               bool   avoid_copies) const {
+
+    const auto * kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
+
+    const auto n_kv    = kv_state->get_n_kv();
+    const auto kv_head = kv_state->get_head();
+    const auto rs_zero = kv_state->get_rs_z();
+
+    ggml_tensor * states = ggml_reshape_2d(ctx0, s, state_size, kv_state->get_size());
+
+    // Clear a single state which will then be copied to the other cleared states.
+    // Note that this is a no-op when the view is zero-sized.
+    ggml_tensor * state_zero = ggml_view_1d(ctx0, states, state_size*(rs_zero >= 0), rs_zero*states->nb[1]*(rs_zero >= 0));
+    ggml_build_forward_expand(gf, ggml_scale_inplace(ctx0, state_zero, 0));
+
+    ggml_tensor * output_states;
 
-    if (kv_state == nullptr) {
-        kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
+    if (!avoid_copies) {
+        // copy states
+        // NOTE: assuming the copy destinations are ALL contained between kv_head and kv_head + n_kv
+        // {state_size, kv_size} -> {state_size, n_seqs}
+        output_states = ggml_get_rows(ctx0, states, ggml_view_1d(ctx0, inp->s_copy, n_seqs, 0));
+        ggml_build_forward_expand(gf, output_states);
+    } else {
+        // FIXME: make the gathering operation happen before the copy below
+        //        (maybe with an optional lambda function passed as a parameter instead of `avoid_copies`?)
+        output_states = states;
     }
 
+    // copy extra states which won't be changed further (between n_seqs and n_kv)
+    ggml_tensor * states_extra = ggml_get_rows(ctx0, states, ggml_view_1d(ctx0, inp->s_copy, n_kv - n_seqs, n_seqs*inp->s_copy->nb[0]));
+    ggml_build_forward_expand(gf,
+        ggml_cpy(ctx0,
+            states_extra,
+            ggml_view_1d(ctx0, s, state_size*(n_kv - n_seqs), (kv_head + n_seqs)*state_size*ggml_element_size(s))));
+
+    return output_states;
+}
+
+llm_graph_input_rs_hybrid_recurrent * llm_graph_context::build_rs_inp_hybrid_recurrent() const {
+    auto inp = std::make_unique<llm_graph_input_rs_hybrid_recurrent>(
+        static_cast<const llama_kv_cache_hybrid_recurrent_state *>(mstate));
+
+    const auto n_kv = inp->kv_state->get_n_kv();
+
+    auto & cur = inp->s_copy;
+
+    cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_kv);
+    ggml_set_input(cur);
+
+    return (llm_graph_input_rs_hybrid_recurrent *) res->add_input(std::move(inp));
+}
+
+ggml_tensor * llm_graph_context::build_rs(
+        llm_graph_input_rs_hybrid_recurrent * inp,
+        ggml_cgraph * gf,
+        ggml_tensor * s,
+            int32_t   state_size,
+            int32_t   n_seqs,
+               bool   avoid_copies) const {
+
+    const auto * kv_state = static_cast<const llama_kv_cache_hybrid_recurrent_state *>(mstate)->get_state_recurrent();
+
     const auto n_kv    = kv_state->get_n_kv();
     const auto kv_head = kv_state->get_head();
     const auto rs_zero = kv_state->get_rs_z();
@@ -1485,7 +1596,7 @@ ggml_tensor * llm_graph_context::build_recurrent_state(
         // copy states
         // NOTE: assuming the copy destinations are ALL contained between kv_head and kv_head + n_kv
         // {state_size, kv_size} -> {state_size, n_seqs}
-        output_states = ggml_get_rows(ctx0, states, ggml_view_1d(ctx0, state_copy, n_seqs, 0));
+        output_states = ggml_get_rows(ctx0, states, ggml_view_1d(ctx0, inp->s_copy, n_seqs, 0));
         ggml_build_forward_expand(gf, output_states);
     } else {
         // FIXME: make the gathering operation happen before the copy below
@@ -1494,7 +1605,7 @@ ggml_tensor * llm_graph_context::build_recurrent_state(
     }
 
     // copy extra states which won't be changed further (between n_seqs and n_kv)
-    ggml_tensor * states_extra = ggml_get_rows(ctx0, states, ggml_view_1d(ctx0, state_copy, n_kv - n_seqs, n_seqs*state_copy->nb[0]));
+    ggml_tensor * states_extra = ggml_get_rows(ctx0, states, ggml_view_1d(ctx0, inp->s_copy, n_kv - n_seqs, n_seqs*inp->s_copy->nb[0]));
     ggml_build_forward_expand(gf,
         ggml_cpy(ctx0,
             states_extra,
@@ -1504,9 +1615,9 @@ ggml_tensor * llm_graph_context::build_recurrent_state(
 }
 
 ggml_tensor * llm_graph_context::build_rwkv_token_shift_load(
-         ggml_cgraph * gf,
-         ggml_tensor * state_copy,
-  const llama_ubatch & ubatch,
+    llm_graph_input_rs * inp,
+           ggml_cgraph * gf,
+    const llama_ubatch & ubatch,
                  int   il) const {
     const auto * kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
 
@@ -1516,8 +1627,8 @@ ggml_tensor * llm_graph_context::build_rwkv_token_shift_load(
 
     ggml_tensor * token_shift_all = kv_state->get_k_l(il);
 
-    ggml_tensor * token_shift = build_recurrent_state(
-            gf, token_shift_all, state_copy,
+    ggml_tensor * token_shift = build_rs(
+            inp, gf, token_shift_all,
             hparams.n_embd_k_s(), n_seqs);
 
     token_shift = ggml_reshape_3d(ctx0, token_shift, hparams.n_embd, token_shift_count, n_seqs);