Files
llama.cpp/src/models/ricursive.cpp
T
Anton SokolchenkoandJames Pickering de32b268c4 Support DFLASH2 for Recursive Ripple-65B-Preview (#31904)
* fixed support for DFlash2 for Ripple 65B

* spec: Add NoPE handling to DFlash

(cherry picked from commit 89ea8e3e77e3359d17902a2db136022d5d419d88)

* spec: Fix NoPE not being picked up by DFlash

(cherry picked from commit 43393b95fda2ce7e7d243e2650bf7b1cadf1c732)

---------

Co-authored-by: James Pickering <james_pic@hotmail.com>
2026-11-15 19:36:08 +08:00

230 lines
11 KiB
C++

#include "models.h"
void llama_model_ricursive::load_arch_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
ml.get_key(LLM_KV_EMBEDDING_SCALE, hparams.f_embedding_scale, false);
// optional: absent or 0 means no clamp (the reference tests `if clamp_threshold:`); a scalar applies to every layer
ml.get_key_or_arr(LLM_KV_SWIGLU_CLAMP, hparams.swiglu_clamp, hparams.n_layer_all, false);
for (uint32_t il = 0; il < hparams.n_layer_all; ++il) {
if (!std::isfinite(hparams.swiglu_clamp[il]) || hparams.swiglu_clamp[il] < 0.0f) {
throw std::runtime_error(format("ricursive: swiglu_clamp for layer %u must be finite and non-negative, got %g", il, hparams.swiglu_clamp[il]));
}
}
if (hparams.n_embd_head_k() != hparams.n_embd_head_v()) {
throw std::runtime_error("ricursive: key_length and value_length must be equal");
}
// RoPE is only applied on the SWA layers, so a model without SWA would have no positional encoding
ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
if (hparams.n_swa == 0) {
throw std::runtime_error("ricursive: sliding_window must be > 0");
}
hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
if (ml.get_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl, false)) {
// load_swa_pattern only bounds the array by LLAMA_MAX_LAYERS: a short one would leave the last layers dense and without RoPE
uint32_t n_entries = 0;
ml.get_arr_n(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, n_entries);
if (n_entries != hparams.n_layer()) {
throw std::runtime_error(format("ricursive: sliding_window_pattern has %u entries, expected %u", n_entries, hparams.n_layer()));
}
} else {
// default: every 5th layer is global (RicursiveConfig layer_types: full attention iff i % 5 == 4)
load_swa_pattern(ml, 5);
}
if (!hparams.is_swa_any()) {
throw std::runtime_error("ricursive: no sliding window layer, the model would have no positional encoding");
}
// RoPE exactly on the SWA layers; rope_pattern is optional because the first published GGUFs do not carry it
hparams.rope_pattern = hparams.is_swa_impl;
ml.get_key_or_arr(LLM_KV_ATTENTION_ROPE_PATTERN, hparams.rope_pattern, hparams.n_layer(), false);
for (uint32_t il = 0; il < hparams.n_layer(); ++il) {
if (hparams.has_rope(il) != hparams.is_swa(il)) {
throw std::runtime_error(format("ricursive: layer %u has %s but %s", il,
hparams.has_rope(il) ? "RoPE" : "no RoPE", hparams.is_swa(il) ? "a sliding window" : "no sliding window"));
}
}
// the SWA layers are the only RoPE layers, they use the model RoPE parameters
hparams.rope_freq_base_train_swa = hparams.rope_freq_base_train;
hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
switch (hparams.n_layer()) {
case 80: type = LLM_TYPE_65B; break;
default: type = LLM_TYPE_UNKNOWN;
}
}
void llama_model_ricursive::load_arch_tensors(llama_model_loader &) {
LLAMA_LOAD_LOCALS;
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
if (output == NULL) {
output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
}
const int64_t n_embd_q = n_embd_head_k * n_head;
for (int i = 0; i < n_layer; ++i) {
auto & layer = layers[i];
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_q}, 0);
layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}, 0);
layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}, 0);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_q, n_embd}, 0);
layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), {n_embd, n_embd_q}, TENSOR_NOT_REQUIRED);
layer.attn_vres_lambda = create_tensor(tn(LLM_TENSOR_ATTN_VRES_LAMBDA, "weight", i), {1}, TENSOR_NOT_REQUIRED);
layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
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_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
}
// the gate and the value residual are per-model switches (use_gated_attention, use_valres)
for (int i = 1; i < n_layer; ++i) {
if ((layers[i].wqkv_gate == nullptr) != (layers[0].wqkv_gate == nullptr)) {
throw std::runtime_error(format("ricursive: attn_gate must be present in all layers or in none (layers 0 and %d differ)", i));
}
if ((layers[i].attn_vres_lambda == nullptr) != (layers[0].attn_vres_lambda == nullptr)) {
throw std::runtime_error(format("ricursive: attn_vres_lambda must be present in all layers or in none (layers 0 and %d differ)", i));
}
}
}
std::unique_ptr<llm_graph_context> llama_model_ricursive::build_arch_graph(const llm_graph_params & params) const {
return std::make_unique<graph>(params);
}
llama_model_ricursive::graph::graph(const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v();
ggml_tensor * cur;
ggml_tensor * inpL;
inpL = build_inp_embd(model.tok_embd);
// value residual: x0 = token embeddings * embedding_scale, the same tensor for all layers (modeling_ricursive.py:380-381,405)
ggml_tensor * inp_vres = inpL;
ggml_tensor * inp_pos = build_inp_pos();
auto * inp_attn = build_attn_inp_kv_iswa();
ggml_tensor * inp_out_ids = build_inp_out_ids();
const float kq_scale = 1.0f / sqrtf(float(n_embd_head));
for (int il = 0; il < n_layer; ++il) {
const auto & layer = model.layers[il];
res->t_layer_inp[il] = inpL;
ggml_tensor * inpSA = inpL;
cur = build_norm(inpL, layer.attn_norm, NULL, LLM_NORM_RMS, il);
cb(cur, "attn_norm", il);
ggml_tensor * attn_in = cur;
{
ggml_tensor * Qcur = build_lora_mm(layer.wq, attn_in, layer.wq_s);
ggml_tensor * Kcur = build_lora_mm(layer.wk, attn_in, layer.wk_s);
// v_in = lerp(x0, attn_norm(x), sigmoid(lambda)), lambda is learned per layer (modeling_ricursive.py:182-185)
ggml_tensor * v_in = attn_in;
if (layer.attn_vres_lambda) {
ggml_tensor * mix = ggml_sigmoid(ctx0, layer.attn_vres_lambda);
v_in = ggml_add(ctx0, inp_vres, ggml_mul(ctx0, ggml_sub(ctx0, attn_in, inp_vres), mix));
cb(v_in, "vres_in", il);
}
ggml_tensor * Vcur = build_lora_mm(layer.wv, v_in, layer.wv_s);
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
Qcur = build_norm(Qcur, layer.attn_q_norm, NULL, LLM_NORM_RMS, il);
Kcur = build_norm(Kcur, layer.attn_k_norm, NULL, LLM_NORM_RMS, il);
cb(Qcur, "Qcur_normed", il);
cb(Kcur, "Kcur_normed", il);
// global layers use no RoPE (NoPE): the reference rotates only sliding_attention layers (modeling_ricursive.py:127-129,203-205)
Qcur = build_rope(Qcur, inp_pos, il);
Kcur = build_rope(Kcur, inp_pos, il);
cb(Qcur, "Qcur", il);
cb(Kcur, "Kcur", il);
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn, NULL, NULL, NULL, Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
cb(cur, "attn_out", il);
if (layer.wqkv_gate) {
ggml_tensor * gate = build_lora_mm(layer.wqkv_gate, attn_in, layer.wqkv_gate_s);
gate = ggml_sigmoid(ctx0, gate);
cb(gate, "attn_gate", il);
cur = ggml_mul(ctx0, cur, gate);
cb(cur, "attn_gated", il);
}
cur = build_lora_mm(layer.wo, cur, layer.wo_s);
}
if (il == n_layer - 1 && inp_out_ids) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
}
cur = build_norm(cur, layer.attn_post_norm, NULL, LLM_NORM_RMS, il);
cb(cur, "attn_post_norm", il);
ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
cb(ffn_inp, "ffn_inp", il);
cur = build_norm(ffn_inp, layer.ffn_norm, NULL, LLM_NORM_RMS, il);
cb(cur, "ffn_norm", il);
{
ggml_tensor * gate = build_lora_mm(layer.ffn_gate, cur, layer.ffn_gate_s);
ggml_tensor * up = build_lora_mm(layer.ffn_up, cur, layer.ffn_up_s);
cur = hparams.swiglu_clamp[il] > 0.0f
? ggml_swiglu_clamp(ctx0, gate, up, hparams.swiglu_clamp[il])
: ggml_swiglu_split(ctx0, gate, up);
cb(cur, "ffn_swiglu", il);
cur = build_lora_mm(layer.ffn_down, cur, layer.ffn_down_s);
cb(cur, "ffn_out", il);
}
cur = build_norm(cur, layer.ffn_post_norm, NULL, LLM_NORM_RMS, il);
cb(cur, "ffn_post_norm", il);
cur = ggml_add(ctx0, cur, ffn_inp);
cur = build_cvec(cur, il);
cb(cur, "l_out", il);
inpL = cur;
}
cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, -1);
cb(cur, "result_norm", -1);
res->t_embd = cur;
cur = build_lora_mm(model.output, cur, model.output_s);
cb(cur, "result_output", -1);
res->t_logits = cur;
}