"DeepseekForCausalLM": "deepseek",
"DeepseekV2ForCausalLM": "deepseek",
"DeepseekV3ForCausalLM": "deepseek",
+ "DeepseekV32ForCausalLM": "deepseek",
"DistilBertForMaskedLM": "bert",
"DistilBertForSequenceClassification": "bert",
"DistilBertModel": "bert",
gguf.MODEL_TENSOR.SSM_CONV1D_Q,
gguf.MODEL_TENSOR.SSM_CONV1D_K,
gguf.MODEL_TENSOR.SSM_CONV1D_V,
+ # DSA indexer weights should be F32
+ gguf.MODEL_TENSOR.INDEXER_PROJ,
)
)
or new_name[-7:] not in (".weight", ".lora_a", ".lora_b")
experts = [k for d in self._experts for k in d.keys()]
if len(experts) > 0:
raise ValueError(f"Unprocessed experts: {experts}")
+
+
+@ModelBase.register("DeepseekV32ForCausalLM")
+class DeepseekV32Model(DeepseekV2Model):
+ model_arch = gguf.MODEL_ARCH.DEEPSEEK32
+ skip_mtp = False
+
+ def __init__(self, *args, **kwargs):
+ super().__init__(*args, **kwargs)
+ self.block_count = self.hparams["num_hidden_layers"] + self.hparams.get("num_nextn_predict_layers", 0)
+ self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+ def set_vocab(self):
+ from transformers import AutoTokenizer
+ tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
+ assert getattr(tokenizer, "add_bos_token", False), "Change value of add_bos_token to true in tokenizer_config.json file."
+ self._set_vocab_gpt2()
+
+ def set_gguf_parameters(self):
+ super().set_gguf_parameters()
+
+ # NextN/MTP prediction layers
+ if (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
+ self.gguf_writer.add_nextn_predict_layers(num_nextn_predict_layers)
+
+ # DSA indexer parameters
+ self.gguf_writer.add_indexer_head_count(self.hparams["index_n_heads"])
+ self.gguf_writer.add_indexer_key_length(self.hparams["index_head_dim"])
+ self.gguf_writer.add_indexer_top_k(self.hparams["index_topk"])
}
}
+static void ggml_compute_forward_fill_f16(const ggml_compute_params * params, ggml_tensor * dst) {
+ const ggml_fp16_t c = GGML_CPU_FP32_TO_FP16(ggml_get_op_params_f32(dst, 0));
+
+ GGML_TENSOR_LOCALS(int64_t, ne, dst, ne);
+ GGML_TENSOR_LOCALS(size_t, nb, dst, nb);
+
+ const auto [ir0, ir1] = get_thread_range(params, dst);
+
+ for (int64_t ir = ir0; ir < ir1; ++ir) {
+ const int64_t i03 = ir/(ne2*ne1);
+ const int64_t i02 = (ir - i03*ne2*ne1)/ne1;
+ const int64_t i01 = (ir - i03*ne2*ne1 - i02*ne1);
+
+ ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1);
+
+ ggml_vec_set_f16(ne0, dst_ptr, c);
+ }
+}
+
void ggml_compute_forward_fill(const ggml_compute_params * params, ggml_tensor * dst) {
- ggml_compute_forward_fill_f32(params, dst);
+ const ggml_tensor * src0 = dst->src[0];
+
+ switch (src0->type) {
+ case GGML_TYPE_F32:
+ {
+ ggml_compute_forward_fill_f32(params, dst);
+ } break;
+ case GGML_TYPE_F16:
+ {
+ ggml_compute_forward_fill_f16(params, dst);
+ } break;
+ default:
+ {
+ GGML_ABORT("unsupported type for ggml_compute_forward_fill: %s", ggml_type_name(src0->type));
+ }
+ }
}
// ggml_compute_tri
struct ggml_tensor * a,
float c,
bool inplace) {
- GGML_ASSERT(a->type == GGML_TYPE_F32);
+ GGML_ASSERT(a->type == GGML_TYPE_F32 || a->type == GGML_TYPE_F16);
GGML_ASSERT(ggml_is_contiguous(a));
struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
DEEPSEEK = auto()
DEEPSEEK2 = auto()
DEEPSEEK2OCR = auto()
+ DEEPSEEK32 = auto()
CHATGLM = auto()
GLM4 = auto()
GLM4_MOE = auto()
MODEL_ARCH.DEEPSEEK: "deepseek",
MODEL_ARCH.DEEPSEEK2: "deepseek2",
MODEL_ARCH.DEEPSEEK2OCR: "deepseek2-ocr",
+ MODEL_ARCH.DEEPSEEK32: "deepseek32",
MODEL_ARCH.CHATGLM: "chatglm",
MODEL_ARCH.GLM4: "glm4",
MODEL_ARCH.GLM4_MOE: "glm4moe",
MODEL_TENSOR.FFN_UP_SHEXP,
MODEL_TENSOR.FFN_EXP_PROBS_B,
],
+ MODEL_ARCH.DEEPSEEK32: [
+ MODEL_TENSOR.TOKEN_EMBD,
+ MODEL_TENSOR.OUTPUT_NORM,
+ MODEL_TENSOR.OUTPUT,
+ MODEL_TENSOR.ROPE_FREQS,
+ MODEL_TENSOR.ATTN_NORM,
+ MODEL_TENSOR.ATTN_Q,
+ MODEL_TENSOR.ATTN_Q_A,
+ MODEL_TENSOR.ATTN_Q_B,
+ MODEL_TENSOR.ATTN_KV_A_MQA,
+ MODEL_TENSOR.ATTN_K_B,
+ MODEL_TENSOR.ATTN_V_B,
+ MODEL_TENSOR.ATTN_Q_A_NORM,
+ MODEL_TENSOR.ATTN_KV_A_NORM,
+ MODEL_TENSOR.ATTN_OUT,
+ MODEL_TENSOR.ATTN_ROT_EMBD,
+ MODEL_TENSOR.FFN_GATE_INP,
+ MODEL_TENSOR.FFN_NORM,
+ MODEL_TENSOR.FFN_GATE,
+ MODEL_TENSOR.FFN_DOWN,
+ MODEL_TENSOR.FFN_UP,
+ MODEL_TENSOR.FFN_GATE_EXP,
+ MODEL_TENSOR.FFN_DOWN_EXP,
+ MODEL_TENSOR.FFN_UP_EXP,
+ MODEL_TENSOR.FFN_GATE_SHEXP,
+ MODEL_TENSOR.FFN_DOWN_SHEXP,
+ MODEL_TENSOR.FFN_UP_SHEXP,
+ MODEL_TENSOR.FFN_EXP_PROBS_B,
+ MODEL_TENSOR.INDEXER_K_NORM,
+ MODEL_TENSOR.INDEXER_PROJ,
+ MODEL_TENSOR.INDEXER_ATTN_K,
+ MODEL_TENSOR.INDEXER_ATTN_Q_B,
+ # NextN/MTP tensors - preserved but unused
+ MODEL_TENSOR.NEXTN_EH_PROJ,
+ MODEL_TENSOR.NEXTN_EMBED_TOKENS,
+ MODEL_TENSOR.NEXTN_ENORM,
+ MODEL_TENSOR.NEXTN_HNORM,
+ MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD,
+ MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM,
+ ],
MODEL_ARCH.ERNIE4_5_MOE: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_ROT_EMBD,
],
+ MODEL_ARCH.DEEPSEEK32: [
+ MODEL_TENSOR.ROPE_FREQS,
+ MODEL_TENSOR.ATTN_ROT_EMBD,
+ ],
MODEL_ARCH.CHATGLM: [
MODEL_TENSOR.ROPE_FREQS,
],
llama-io.cpp
llama-kv-cache.cpp
llama-kv-cache-iswa.cpp
+ llama-kv-cache-dsa.cpp
llama-memory.cpp
llama-memory-hybrid.cpp
llama-memory-hybrid-iswa.cpp
{ LLM_ARCH_DEEPSEEK, "deepseek" },
{ LLM_ARCH_DEEPSEEK2, "deepseek2" },
{ LLM_ARCH_DEEPSEEK2OCR, "deepseek2-ocr" },
+ { LLM_ARCH_DEEPSEEK32, "deepseek32" },
{ LLM_ARCH_CHATGLM, "chatglm" },
{ LLM_ARCH_GLM4, "glm4" },
{ LLM_ARCH_GLM4_MOE, "glm4moe" },
case LLM_ARCH_OLMO2:
case LLM_ARCH_OLMOE:
case LLM_ARCH_DEEPSEEK2:
+ case LLM_ARCH_DEEPSEEK32:
case LLM_ARCH_GLM_DSA:
case LLM_ARCH_BITNET:
case LLM_ARCH_T5:
LLM_ARCH_DEEPSEEK,
LLM_ARCH_DEEPSEEK2,
LLM_ARCH_DEEPSEEK2OCR,
+ LLM_ARCH_DEEPSEEK32,
LLM_ARCH_CHATGLM,
LLM_ARCH_GLM4,
LLM_ARCH_GLM4_MOE,
#include "llama-kv-cache.h"
#include "llama-kv-cache-iswa.h"
+#include "llama-kv-cache-dsa.h"
#include "llama-memory-hybrid.h"
#include "llama-memory-hybrid-iswa.h"
#include "llama-memory-recurrent.h"
return res;
}
+void llm_graph_input_attn_k_dsa::set_input(const llama_ubatch * ubatch) {
+ mctx->get_mla()->set_input_k_idxs(self_k_idxs_mla, ubatch);
+
+ mctx->get_mla()->set_input_kq_mask(self_kq_mask_mla, ubatch, cparams.causal_attn);
+
+ mctx->get_lid()->set_input_k_idxs(self_k_idxs_lid, ubatch);
+
+ mctx->get_lid()->set_input_kq_mask(self_kq_mask_lid, ubatch, cparams.causal_attn);
+
+ mctx->get_lid()->set_input_k_rot(self_k_rot_lid);
+}
+
+bool llm_graph_input_attn_k_dsa::can_reuse(const llm_graph_params & params) {
+ const auto * mctx = static_cast<const llama_kv_cache_dsa_context *>(params.mctx);
+
+ this->mctx = mctx;
+
+ bool res = true;
+
+ res &= self_k_idxs_mla->ne[0] == params.ubatch.n_tokens;
+ res &= self_k_idxs_lid->ne[0] == params.ubatch.n_tokens;
+
+ res &= can_reuse_kq_mask(self_kq_mask_mla, mctx->get_mla(), params.ubatch, params.cparams);
+ res &= can_reuse_kq_mask(self_kq_mask_lid, mctx->get_lid(), params.ubatch, params.cparams);
+
+ return res;
+}
+
void llm_graph_input_attn_kv_iswa::set_input(const llama_ubatch * ubatch) {
// base tensors may not be allocated if there are no non-SWA attention layers
if (self_k_idxs && self_k_idxs->buffer) {
return cur;
}
+ggml_tensor * llm_graph_context::build_attn(
+ llm_graph_input_attn_k_dsa * inp,
+ ggml_tensor * wo,
+ ggml_tensor * wo_b,
+ ggml_tensor * wo_s,
+ ggml_tensor * q_cur,
+ ggml_tensor * k_cur,
+ ggml_tensor * v_cur,
+ ggml_tensor * kq_b,
+ ggml_tensor * sinks,
+ ggml_tensor * v_mla,
+ ggml_tensor * top_k,
+ 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
+ // expand k later to enable rope fusion which directly writes into k-v cache
+ ggml_build_forward_expand(gf, q_cur);
+ ggml_build_forward_expand(gf, v_cur);
+ ggml_build_forward_expand(gf, k_cur);
+
+ const auto * mctx_cur = inp->mctx->get_mla();
+
+ // store to KV cache
+ {
+ const auto & k_idxs = inp->get_k_idxs_mla();
+
+ ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, k_idxs, il));
+ }
+
+ const auto & kq_mask = inp->get_kq_mask_mla();
+
+ // prepare new kq mask - starts filled with -INFINITY
+ ggml_tensor * kq_mask_all = ggml_fill(ctx0, kq_mask, -INFINITY);
+
+ // reshape KQ mask into tensor with rows of size 1:
+ // [n_kv, n_batch, 1, n_stream] -> [1, n_kv, n_batch, n_stream]
+ kq_mask_all = ggml_view_4d(ctx0, kq_mask_all, 1, kq_mask_all->ne[0], kq_mask_all->ne[1], kq_mask_all->ne[3], kq_mask_all->nb[0], kq_mask_all->nb[1], kq_mask_all->nb[2], 0);
+
+ // reshape top_k indices: [n_top_k, n_batch, 1, n_stream] -> [n_top_k, n_batch, n_stream, 1]
+ ggml_tensor * top_k_3d = ggml_view_4d(ctx0, top_k, top_k->ne[0], top_k->ne[1], top_k->ne[3], 1, top_k->nb[1], top_k->nb[2], top_k->ne[3]*top_k->nb[3], 0);
+
+ // prepare zero-filled tensor with rows of size 1: [1, n_top_k, n_batch, n_stream]
+ // this will be our source of zero values for unmasking top k mask elements
+ ggml_tensor * zeros = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, 1, top_k_3d->ne[0], top_k_3d->ne[1], top_k_3d->ne[2]);
+ zeros = ggml_fill(ctx0, zeros, 0.0f);
+
+ // modify KQ mask by unmasking elements that are in top_k indices
+ // ggml_set_rows([1, n_kv, n_batch, n_stream], [1, n_top_k, n_batch, n_stream], [n_top_k, n_batch, n_stream, 1])
+ ggml_tensor * kq_mask_top_k = ggml_set_rows(ctx0, kq_mask_all, zeros, top_k_3d);
+
+ // reshape to restore the original shape of KQ mask:
+ // [1, n_kv, n_batch, n_stream] -> [n_kv, n_batch, 1, n_stream]
+ kq_mask_top_k = ggml_view_4d(ctx0, kq_mask_top_k, kq_mask_top_k->ne[1], kq_mask_top_k->ne[2], 1, kq_mask_top_k->ne[3], kq_mask_top_k->nb[2], kq_mask_top_k->nb[3], kq_mask_top_k->nb[3], 0);
+
+ // combine with the original kq mask
+ kq_mask_top_k = ggml_add(ctx0, kq_mask_top_k, kq_mask);
+
+ ggml_tensor * q = q_cur;
+ ggml_tensor * k = mctx_cur->get_k(ctx0, il);
+ ggml_tensor * v = ggml_view_4d(ctx0, k, v_cur->ne[0], k->ne[1], k->ne[2], k->ne[3], k->nb[1], k->nb[2], k->nb[3], 0);
+
+ ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask_top_k, sinks, v_mla, kq_scale, il);
+ cb(cur, "kqv_out", il);
+
+ if (wo) {
+ cur = build_lora_mm(wo, cur, wo_s);
+ }
+
+ if (wo_b) {
+ cur = ggml_add(ctx0, cur, wo_b);
+ }
+
+ return cur;
+}
+
ggml_tensor * llm_graph_context::build_attn(
llm_graph_input_attn_kv_iswa * inp,
ggml_tensor * wo,
return cur;
}
+llm_graph_input_attn_k_dsa * llm_graph_context::build_attn_inp_k_dsa() const {
+ const auto * mctx_cur = static_cast<const llama_kv_cache_dsa_context *>(mctx);
+
+ auto inp = std::make_unique<llm_graph_input_attn_k_dsa>(hparams, cparams, mctx_cur);
+
+ {
+ inp->self_k_idxs_mla = mctx_cur->get_mla()->build_input_k_idxs(ctx0, ubatch);
+
+ inp->self_kq_mask_mla = build_attn_inp_kq_mask(ctx0, mctx_cur->get_mla(), ubatch, cparams);
+ inp->self_kq_mask_mla_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_mla, GGML_TYPE_F16) : inp->self_kq_mask_mla;
+ }
+
+ {
+ inp->self_k_idxs_lid = mctx_cur->get_lid()->build_input_k_idxs(ctx0, ubatch);
+
+ inp->self_kq_mask_lid = build_attn_inp_kq_mask(ctx0, mctx_cur->get_lid(), ubatch, cparams);
+ inp->self_kq_mask_lid_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_lid, GGML_TYPE_F16) : inp->self_kq_mask_lid;
+
+ inp->self_k_rot_lid = mctx_cur->get_lid()->build_input_k_rot(ctx0);
+ }
+
+ return (llm_graph_input_attn_k_dsa *) res->add_input(std::move(inp));
+}
+
// TODO: maybe separate the inner implementation into a separate function
// like with the non-sliding window equivalent
// once sliding-window hybrid caches are a thing.
struct llama_memory_context_i;
class llama_kv_cache_context;
+class llama_kv_cache_dsa_context;
class llama_kv_cache_iswa_context;
class llama_memory_recurrent_context;
class llama_memory_hybrid_context;
const llama_kv_cache_context * mctx;
};
+class llm_graph_input_attn_k_dsa : public llm_graph_input_i {
+public:
+ llm_graph_input_attn_k_dsa(
+ const llama_hparams & hparams,
+ const llama_cparams & cparams,
+ const llama_kv_cache_dsa_context * mctx) :
+ hparams(hparams),
+ cparams(cparams),
+ mctx(mctx) {
+ }
+ ~llm_graph_input_attn_k_dsa() = default;
+
+ void set_input(const llama_ubatch * ubatch) override;
+
+ bool can_reuse(const llm_graph_params & params) override;
+
+ ggml_tensor * get_k_idxs_mla() const { return self_k_idxs_mla; }
+ ggml_tensor * get_k_idxs_lid() const { return self_k_idxs_lid; }
+
+ ggml_tensor * get_kq_mask_mla() const { return self_kq_mask_mla_cnv; }
+ ggml_tensor * get_kq_mask_lid() const { return self_kq_mask_lid; }
+
+ ggml_tensor * self_k_idxs_mla = nullptr; // I64 [n_batch]
+ ggml_tensor * self_k_idxs_lid = nullptr; // I64 [n_batch]
+
+ ggml_tensor * self_kq_mask_mla = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
+ ggml_tensor * self_kq_mask_mla_cnv = nullptr; // [n_kv, n_batch/n_stream, 1, n_stream]
+ ggml_tensor * self_kq_mask_lid = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
+ ggml_tensor * self_kq_mask_lid_cnv = nullptr; // [n_kv, n_batch/n_stream, 1, n_stream]
+
+ ggml_tensor * self_k_rot_lid = nullptr;
+
+ const llama_hparams hparams;
+ const llama_cparams cparams;
+
+ const llama_kv_cache_dsa_context * mctx;
+};
+
class llm_graph_input_attn_kv_iswa : public llm_graph_input_i {
public:
llm_graph_input_attn_kv_iswa(
float kq_scale,
int il) const;
+ llm_graph_input_attn_k_dsa * build_attn_inp_k_dsa() const;
+
+ ggml_tensor * build_attn(
+ llm_graph_input_attn_k_dsa * inp,
+ ggml_tensor * wo,
+ ggml_tensor * wo_b,
+ ggml_tensor * wo_s,
+ ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
+ ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
+ ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
+ ggml_tensor * kq_b,
+ ggml_tensor * sinks, // [n_head_q]
+ ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
+ ggml_tensor * top_k, // [n_indexer_top_k, n_tokens]
+ float kq_scale,
+ int il) const;
+
llm_graph_input_attn_kv_iswa * build_attn_inp_kv_iswa() const;
// note: if k_cur or v_cur are not provided, they will not be stored in the memory
--- /dev/null
+#include "llama-kv-cache-dsa.h"
+
+#include "llama-impl.h"
+#include "llama-batch.h"
+#include "llama-model.h"
+
+#include <algorithm>
+#include <cassert>
+
+//
+// llama_kv_cache_dsa
+//
+
+llama_kv_cache_dsa::llama_kv_cache_dsa(
+ const llama_model & model,
+ ggml_type type_k,
+ ggml_type type_v,
+ bool v_trans,
+ bool offload,
+ bool unified,
+ uint32_t kv_size,
+ uint32_t n_seq_max,
+ uint32_t n_pad,
+ uint32_t n_swa,
+ llama_swa_type swa_type,
+ const layer_filter_cb & filter,
+ const layer_reuse_cb & reuse) :
+ hparams_lid(model.hparams), n_stream(unified ? 1 : n_seq_max) {
+
+ LLAMA_LOG_INFO("%s: creating main KV cache, size = %u cells\n", __func__, kv_size);
+
+ kv_mla = std::make_unique<llama_kv_cache>(
+ model, model.hparams, type_k, type_v,
+ v_trans, offload, unified, kv_size, n_seq_max, n_pad,
+ n_swa, swa_type, filter, reuse);
+
+ // we use llama_kv_cache for caching indexer keys
+ // by hand-tweaking some hparams we fool it to create
+ // indexer key cache tensors with correct dimensions
+ // https://github.com/ggml-org/llama.cpp/pull/21149#discussion_r3015940823
+
+ // DSA lightning indexer uses MQA with single key head
+ std::fill(hparams_lid.n_head_kv_arr.begin(), hparams_lid.n_head_kv_arr.end(), 1);
+ hparams_lid.n_embd_head_k_full = model.hparams.indexer_head_size;
+ hparams_lid.rope_type = LLAMA_ROPE_TYPE_NEOX;
+
+ LLAMA_LOG_INFO("%s: creating indexer KV cache, size = %u cells\n", __func__, kv_size);
+
+ kv_lid = std::make_unique<llama_kv_cache>(
+ model, hparams_lid, type_k, type_v,
+ v_trans, offload, unified, kv_size, n_seq_max, n_pad,
+ n_swa, swa_type, filter, reuse);
+}
+
+void llama_kv_cache_dsa::clear(bool data) {
+ kv_mla->clear(data);
+ kv_lid->clear(data);
+}
+
+bool llama_kv_cache_dsa::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
+ bool res = true;
+
+ res = res & kv_mla->seq_rm(seq_id, p0, p1);
+ res = res & kv_lid->seq_rm(seq_id, p0, p1);
+
+ return res;
+}
+
+void llama_kv_cache_dsa::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
+ kv_mla->seq_cp(seq_id_src, seq_id_dst, p0, p1);
+ kv_lid->seq_cp(seq_id_src, seq_id_dst, p0, p1);
+}
+
+void llama_kv_cache_dsa::seq_keep(llama_seq_id seq_id) {
+ kv_mla->seq_keep(seq_id);
+ kv_lid->seq_keep(seq_id);
+}
+
+void llama_kv_cache_dsa::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
+ kv_mla->seq_add(seq_id, p0, p1, shift);
+ kv_lid->seq_add(seq_id, p0, p1, shift);
+}
+
+void llama_kv_cache_dsa::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
+ kv_mla->seq_div(seq_id, p0, p1, d);
+ kv_lid->seq_div(seq_id, p0, p1, d);
+}
+
+llama_pos llama_kv_cache_dsa::seq_pos_min(llama_seq_id seq_id) const {
+ return kv_mla->seq_pos_min(seq_id);
+}
+
+llama_pos llama_kv_cache_dsa::seq_pos_max(llama_seq_id seq_id) const {
+ return kv_mla->seq_pos_max(seq_id);
+}
+
+std::map<ggml_backend_buffer_type_t, size_t> llama_kv_cache_dsa::memory_breakdown() const {
+ std::map<ggml_backend_buffer_type_t, size_t> mb = kv_mla->memory_breakdown();
+ for (const auto & buft_size : kv_lid->memory_breakdown()) {
+ mb[buft_size.first] += buft_size.second;
+ }
+ return mb;
+}
+
+llama_memory_context_ptr llama_kv_cache_dsa::init_batch(
+ llama_batch_allocr & balloc,
+ uint32_t n_ubatch,
+ bool embd_all) {
+ GGML_UNUSED(embd_all);
+
+ do {
+ balloc.split_reset();
+
+ std::vector<llama_ubatch> ubatches;
+ while (true) {
+ auto ubatch = n_stream == 1 ? balloc.split_simple(n_ubatch) : balloc.split_equal(n_ubatch, true);
+
+ if (ubatch.n_tokens == 0) {
+ break;
+ }
+
+ ubatches.push_back(std::move(ubatch)); // NOLINT
+ }
+
+ if (balloc.get_n_used() < balloc.get_n_tokens()) {
+ // failed to find a suitable split
+ break;
+ }
+
+ auto sinfos_mla = kv_mla->prepare(ubatches);
+ if (sinfos_mla.empty()) {
+ break;
+ }
+
+ auto sinfos_lid = kv_lid->prepare(ubatches);
+ if (sinfos_lid.empty()) {
+ break;
+ }
+
+ assert(sinfos_mla.size() == sinfos_lid.size());
+
+ return std::make_unique<llama_kv_cache_dsa_context>(
+ this, std::move(sinfos_mla), std::move(sinfos_lid), std::move(ubatches));
+ } while (false);
+
+ return std::make_unique<llama_kv_cache_dsa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
+}
+
+llama_memory_context_ptr llama_kv_cache_dsa::init_full() {
+ return std::make_unique<llama_kv_cache_dsa_context>(this);
+}
+
+llama_memory_context_ptr llama_kv_cache_dsa::init_update(llama_context * lctx, bool optimize) {
+ return std::make_unique<llama_kv_cache_dsa_context>(this, lctx, optimize);
+}
+
+bool llama_kv_cache_dsa::get_can_shift() const {
+ return kv_mla->get_can_shift() &&
+ kv_lid->get_can_shift() &&
+ kv_mla->get_size() == kv_lid->get_size();
+}
+
+void llama_kv_cache_dsa::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
+ kv_mla->state_write(io, seq_id, flags);
+ kv_lid->state_write(io, seq_id, flags);
+}
+
+void llama_kv_cache_dsa::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
+ kv_mla->state_read(io, seq_id, flags);
+ kv_lid->state_read(io, seq_id, flags);
+}
+
+llama_kv_cache * llama_kv_cache_dsa::get_mla() const {
+ return kv_mla.get();
+}
+
+llama_kv_cache * llama_kv_cache_dsa::get_lid() const {
+ return kv_lid.get();
+}
+
+//
+// llama_kv_cache_dsa_context
+//
+
+llama_kv_cache_dsa_context::llama_kv_cache_dsa_context(llama_memory_status status) : status(status) {}
+
+llama_kv_cache_dsa_context::llama_kv_cache_dsa_context(
+ llama_kv_cache_dsa * kv) :
+ ctx_mla(kv->get_mla()->init_full()),
+ ctx_lid(kv->get_lid()->init_full()),
+ status(llama_memory_status_combine(ctx_mla->get_status(), ctx_lid->get_status())) {
+}
+
+llama_kv_cache_dsa_context::llama_kv_cache_dsa_context(
+ llama_kv_cache_dsa * kv,
+ llama_context * lctx,
+ bool optimize) :
+ ctx_mla(kv->get_mla()->init_update(lctx, optimize)),
+ ctx_lid(kv->get_lid()->init_update(lctx, optimize)),
+ status(llama_memory_status_combine(ctx_mla->get_status(), ctx_lid->get_status())) {
+}
+
+llama_kv_cache_dsa_context::llama_kv_cache_dsa_context(
+ llama_kv_cache_dsa * kv,
+ slot_info_vec_t sinfos_mla,
+ slot_info_vec_t sinfos_lid,
+ std::vector<llama_ubatch> ubatches) :
+ ubatches(std::move(ubatches)),
+ // note: here we copy the ubatches. not sure if this is ideal
+ ctx_mla(new llama_kv_cache_context(kv->get_mla(), std::move(sinfos_mla), this->ubatches)),
+ ctx_lid(new llama_kv_cache_context(kv->get_lid(), std::move(sinfos_lid), this->ubatches)),
+ status(llama_memory_status_combine(ctx_mla->get_status(), ctx_lid->get_status())) {
+}
+
+llama_kv_cache_dsa_context:: ~llama_kv_cache_dsa_context() = default;
+
+bool llama_kv_cache_dsa_context::next() {
+ assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+
+ ctx_mla->next();
+ ctx_lid->next();
+
+ if (++i_next >= ubatches.size()) {
+ return false;
+ }
+
+ return true;
+}
+
+bool llama_kv_cache_dsa_context::apply() {
+ assert(!llama_memory_status_is_fail(status));
+
+ bool res = true;
+
+ res = res & ctx_mla->apply();
+ res = res & ctx_lid->apply();
+
+ return res;
+}
+
+llama_memory_status llama_kv_cache_dsa_context::get_status() const {
+ return status;
+}
+
+const llama_ubatch & llama_kv_cache_dsa_context::get_ubatch() const {
+ assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+
+ return ubatches[i_next];
+}
+
+const llama_kv_cache_context * llama_kv_cache_dsa_context::get_mla() const {
+ assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+
+ return static_cast<const llama_kv_cache_context *>(ctx_mla.get());
+}
+
+const llama_kv_cache_context * llama_kv_cache_dsa_context::get_lid() const {
+ assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+
+ return static_cast<const llama_kv_cache_context *>(ctx_lid.get());
+}
--- /dev/null
+#pragma once
+
+#include "llama-kv-cache.h"
+
+#include <vector>
+
+//
+// llama_kv_cache_dsa
+//
+
+// utilizes two instances of llama_kv_cache:
+// - the first instance is for caching key tensors of the model,
+// - the second instance is for caching lightning indexer key tensors
+
+class llama_kv_cache_dsa : public llama_memory_i {
+public:
+ llama_kv_cache_dsa(
+ const llama_model & model,
+ ggml_type type_k,
+ ggml_type type_v,
+ bool v_trans,
+ bool offload,
+ bool unified,
+ uint32_t kv_size,
+ uint32_t n_seq_max,
+ uint32_t n_pad,
+ uint32_t n_swa,
+ llama_swa_type swa_type,
+ const layer_filter_cb & filter,
+ const layer_reuse_cb & reuse);
+
+ ~llama_kv_cache_dsa() = default;
+
+ //
+ // llama_memory_i
+ //
+
+ llama_memory_context_ptr init_batch(
+ llama_batch_allocr & balloc,
+ uint32_t n_ubatch,
+ bool embd_all) override;
+
+ llama_memory_context_ptr init_full() override;
+
+ llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) override;
+
+ bool get_can_shift() const override;
+
+ void clear(bool data) override;
+
+ bool seq_rm (llama_seq_id seq_id, llama_pos p0, llama_pos p1) override;
+ void seq_cp (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) override;
+ void seq_keep(llama_seq_id seq_id) override;
+ void seq_add (llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) override;
+ void seq_div (llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) override;
+
+ llama_pos seq_pos_min(llama_seq_id seq_id) const override;
+ llama_pos seq_pos_max(llama_seq_id seq_id) const override;
+
+ std::map<ggml_backend_buffer_type_t, size_t> memory_breakdown() const override;
+
+ // state write/load
+
+ void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const override;
+ void state_read (llama_io_read_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) override;
+
+ //
+ // llama_kv_cache_dsa specific API
+ //
+
+ llama_kv_cache * get_mla() const;
+ llama_kv_cache * get_lid() const;
+
+private:
+ // we keep indexer KV cache hparams instance here as llama_kv_cache stores only reference to it
+ llama_hparams hparams_lid;
+ const uint32_t n_stream = 1;
+
+ std::unique_ptr<llama_kv_cache> kv_mla;
+ std::unique_ptr<llama_kv_cache> kv_lid;
+};
+
+class llama_kv_cache_dsa_context : public llama_memory_context_i {
+public:
+ using slot_info_vec_t = llama_kv_cache::slot_info_vec_t;
+
+ // used for errors
+ llama_kv_cache_dsa_context(llama_memory_status status);
+
+ // used to create a full-cache context
+ llama_kv_cache_dsa_context(
+ llama_kv_cache_dsa * kv);
+
+ // used to create an update context
+ llama_kv_cache_dsa_context(
+ llama_kv_cache_dsa * kv,
+ llama_context * lctx,
+ bool optimize);
+
+ // used to create a batch processing context from a batch
+ llama_kv_cache_dsa_context(
+ llama_kv_cache_dsa * kv,
+ slot_info_vec_t sinfos_base,
+ slot_info_vec_t sinfos_ik,
+ std::vector<llama_ubatch> ubatches);
+
+ virtual ~llama_kv_cache_dsa_context();
+
+ //
+ // llama_memory_context_i
+ //
+
+ bool next() override;
+ bool apply() override;
+
+ llama_memory_status get_status() const override;
+ const llama_ubatch & get_ubatch() const override;
+
+ //
+ // llama_kv_cache_dsa_context specific API
+ //
+
+ const llama_kv_cache_context * get_mla() const;
+ const llama_kv_cache_context * get_lid() const;
+
+private:
+ //llama_kv_cache_dsa * kv;
+
+ // the index of the next ubatch to process
+ size_t i_next = 0;
+
+ std::vector<llama_ubatch> ubatches;
+
+ const llama_memory_context_ptr ctx_mla;
+ const llama_memory_context_ptr ctx_lid;
+
+ const llama_memory_status status;
+};
LLAMA_LOG_INFO("%s: creating non-SWA KV cache, size = %u cells\n", __func__, size_base);
kv_base = std::make_unique<llama_kv_cache>(
- model, type_k, type_v,
+ model, hparams, type_k, type_v,
v_trans, offload, unified, size_base, n_seq_max, n_pad,
0, LLAMA_SWA_TYPE_NONE, filter_base, reuse);
LLAMA_LOG_INFO("%s: creating SWA KV cache, size = %u cells\n", __func__, size_swa);
kv_swa = std::make_unique<llama_kv_cache>(
- model, type_k, type_v,
+ model, hparams, type_k, type_v,
v_trans, offload, unified, size_swa, n_seq_max, n_pad,
hparams.n_swa, hparams.swa_type, filter_swa, reuse);
}
llama_kv_cache::llama_kv_cache(
const llama_model & model,
+ const llama_hparams & hparams,
ggml_type type_k,
ggml_type type_v,
bool v_trans,
llama_swa_type swa_type,
const layer_filter_cb & filter,
const layer_reuse_cb & reuse) :
- model(model), hparams(model.hparams), v_trans(v_trans),
+ model(model), hparams(hparams), v_trans(v_trans),
n_seq_max(n_seq_max), n_stream(unified ? 1 : n_seq_max), n_pad(n_pad), n_swa(n_swa), swa_type(swa_type) {
GGML_ASSERT(kv_size % n_pad == 0);
// allocate tensors and initialize the buffers to avoid NaNs in the padding
for (auto & [buft, ctx] : ctx_map) {
ggml_backend_buffer_t buf;
- if (model.hparams.no_alloc) {
+ if (hparams.no_alloc) {
buf = ggml_backend_buft_alloc_buffer(buft, /*size =*/ 0); // dummy buffer
for (ggml_tensor * t = ggml_get_first_tensor(ctx.get()); t != nullptr; t = ggml_get_next_tensor(ctx.get(), t)) {
t->buffer = buf; // set dummy buffer for KV cache so that the backend scheduler won't try to allocate it
ggml_is_quantized(type_k) &&
hparams.n_embd_head_k() % 64 == 0;
+ // always create Hadamard rotation tensors for DeepSeek V3.2 DSA lightning indexer
+ if (model.arch == LLM_ARCH_DEEPSEEK32 && hparams.n_embd_head_k_full == hparams.indexer_head_size) {
+ attn_rot_k = true;
+ }
+
attn_rot_v =
!attn_rot_disable &&
n_embd_head_v_all > 0 &&
using slot_info_vec_t = std::vector<slot_info>;
+ // TODO: refactor the memory instances to not depend on `llama_model`
+ // instead pass all necessary info (e.g. hparams, dev layers, arch, etc.) directly
+ // likely through `struct llama_memory_params`
llama_kv_cache(
const llama_model & model,
+ const llama_hparams & hparams,
ggml_type type_k,
ggml_type type_v,
bool v_trans,
hparams(model.hparams),
mem_attn(new llama_kv_cache(
model,
+ model.hparams,
type_k,
type_v,
v_trans,
#include "llama-kv-cache.h"
#include "llama-kv-cache-iswa.h"
+#include "llama-kv-cache-dsa.h"
#include "llama-memory-hybrid.h"
#include "llama-memory-hybrid-iswa.h"
#include "llama-memory-recurrent.h"
return new llama_model_deepseek2(params);
case LLM_ARCH_DEEPSEEK2OCR:
return new llama_model_deepseek2ocr(params);
+ case LLM_ARCH_DEEPSEEK32:
+ return new llama_model_deepseek32(params);
case LLM_ARCH_GLM_DSA:
return new llama_model_glm_dsa(params);
case LLM_ARCH_MISTRAL4:
case LLM_TYPE_310B_A15B: return "310B.A15B";
case LLM_TYPE_355B_A32B: return "355B.A32B";
case LLM_TYPE_397B_A17B: return "397B.A17B";
+ case LLM_TYPE_685B_A37B: return "685B.A37B";
case LLM_TYPE_744B_A40B: return "744B.A40B";
case LLM_TYPE_E2B: return "E2B";
case LLM_TYPE_E4B: return "E4B";
LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n", __func__, hparams.expert_weights_scale);
}
- if (arch == LLM_ARCH_DEEPSEEK2 || arch == LLM_ARCH_DEEPSEEK2OCR || arch == LLM_ARCH_GLM_DSA || arch == LLM_ARCH_MISTRAL4) {
+ if (arch == LLM_ARCH_DEEPSEEK2 || arch == LLM_ARCH_DEEPSEEK2OCR || arch == LLM_ARCH_DEEPSEEK32 || arch == LLM_ARCH_GLM_DSA || arch == LLM_ARCH_MISTRAL4) {
LLAMA_LOG_INFO("%s: n_layer_dense_lead = %d\n", __func__, hparams.n_layer_dense_lead);
LLAMA_LOG_INFO("%s: n_lora_q = %d\n", __func__, hparams.n_lora_q);
LLAMA_LOG_INFO("%s: n_lora_kv = %d\n", __func__, hparams.n_lora_kv);
{
res = nullptr;
} break;
+ case LLM_ARCH_DEEPSEEK32:
+ {
+ res = new llama_kv_cache_dsa(
+ *this,
+ params.type_k,
+ params.type_v,
+ !cparams.flash_attn,
+ cparams.offload_kqv,
+ cparams.kv_unified,
+ cparams.n_ctx_seq,
+ cparams.n_seq_max,
+ 1,
+ hparams.n_swa,
+ hparams.swa_type,
+ nullptr,
+ nullptr);
+ } break;
// Models that need standard caching should rely on recurrent/hybrid
// checks
default:
res = new llama_kv_cache(
*this,
+ hparams,
params.type_k,
params.type_v,
!cparams.flash_attn,
case LLM_ARCH_DEEPSEEK:
case LLM_ARCH_DEEPSEEK2:
case LLM_ARCH_DEEPSEEK2OCR:
+ case LLM_ARCH_DEEPSEEK32:
case LLM_ARCH_PLM:
case LLM_ARCH_CHATGLM:
case LLM_ARCH_GRANITE:
LLM_TYPE_310B_A15B, // /MiMo-V2-Flash
LLM_TYPE_355B_A32B, // GLM-4.5
LLM_TYPE_397B_A17B, // Qwen3.5
+ LLM_TYPE_685B_A37B, // DeepSeek V3.2
LLM_TYPE_744B_A40B, // GLM-5
LLM_TYPE_E2B,
LLM_TYPE_E4B,
--- /dev/null
+#include "models.h"
+
+#include "llama-kv-cache.h"
+#include "llama-kv-cache-dsa.h"
+
+void llama_model_deepseek32::load_arch_hparams(llama_model_loader & ml) {
+ ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
+ ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+ hparams.f_norm_eps = 1e-6; // eps for layer norm
+ ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false);
+
+ // MoE parameters
+ ml.get_key(LLM_KV_EXPERT_COUNT, hparams.n_expert);
+ ml.get_key(LLM_KV_EXPERT_USED_COUNT, hparams.n_expert_used);
+ ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared);
+ ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead, false);
+ ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE, hparams.expert_weights_scale, false);
+ ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM, hparams.expert_weights_norm, false);
+
+ // deepseek MLA parameters
+ ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK, hparams.n_lora_q);
+ ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK, hparams.n_lora_kv);
+ ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA, hparams.n_embd_head_k_mla_impl, false);
+ ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla_impl, false);
+ ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
+ ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared);
+
+ // DSA parameters
+ ml.get_key(LLM_KV_ATTENTION_INDEXER_HEAD_COUNT, hparams.indexer_n_head);
+ ml.get_key(LLM_KV_ATTENTION_INDEXER_KEY_LENGTH, hparams.indexer_head_size);
+ ml.get_key(LLM_KV_ATTENTION_INDEXER_TOP_K, hparams.indexer_top_k);
+
+ // Expert gating function
+ ml.get_key(LLM_KV_EXPERT_GATING_FUNC, hparams.expert_gating_func);
+
+ if (ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL, hparams.rope_yarn_log_mul, 0.0f)) {
+ // [TAG_DEEPSEEK2_YARN_LOG_MUL_FIX]
+ // cancel the factor from the convert script
+ hparams.rope_yarn_log_mul /= 0.1f;
+ }
+
+ // NextN/MTP parameters
+ ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.nextn_predict_layers, false);
+ GGML_ASSERT(hparams.nextn_predict_layers < hparams.n_layer && "nextn_predict_layers must be < n_layer");
+
+ // TODO: when MTP is implemented, this should probably be updated if needed
+ hparams.n_layer_kv_from_start = hparams.n_layer - hparams.nextn_predict_layers;
+
+ switch (hparams.n_layer) {
+ case 62: type = LLM_TYPE_685B_A37B; break;
+ default: type = LLM_TYPE_UNKNOWN;
+ }
+}
+
+void llama_model_deepseek32::load_arch_tensors(llama_model_loader &) {
+ LLAMA_LOAD_LOCALS;
+ const bool is_mla = hparams.is_mla();
+ if (!is_mla) {
+ throw std::runtime_error("DEEPSEEK32 architecture requires MLA");
+ }
+
+ // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
+ const int64_t n_embd_head_k_mla = hparams.n_embd_head_k_mla();
+ const int64_t n_embd_head_v_mla = hparams.n_embd_head_v_mla();
+
+ const int64_t n_embd_head_qk_rope = hparams.n_rot();
+ const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;
+
+ const int64_t q_lora_rank = hparams.n_lora_q;
+ const int64_t kv_lora_rank = hparams.n_lora_kv;
+
+ const int64_t n_ff_exp = hparams.n_ff_exp;
+ const int64_t n_expert_shared = hparams.n_expert_shared;
+
+ tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+ // output
+ output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+ // try to load output.weight, if not found, use token_embd (tied embeddings)
+ output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+ if (!output) {
+ output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+ }
+
+ for (int i = 0; i < n_layer; ++i) {
+ int flags = 0;
+ if (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers) {
+ // skip all tensors in the NextN layers
+ // TODO @ngxson : TENSOR_NOT_REQUIRED was a hack, need to remove it later
+ flags |= TENSOR_SKIP | TENSOR_NOT_REQUIRED;
+ }
+
+ auto & layer = layers[i];
+
+ layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, flags);
+ layer.attn_q_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_A_NORM, "weight", i), {q_lora_rank}, flags);
+ layer.attn_kv_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank}, flags);
+
+ layer.wq_a = create_tensor(tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}, flags);
+ layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k_mla}, flags);
+
+ layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + n_embd_head_qk_rope}, flags);
+
+ // note: only old legacy GGUF files will have the unsplit wkv_b tensor in
+ layer.wk_b = create_tensor(tn(LLM_TENSOR_ATTN_K_B, "weight", i), {n_embd_head_qk_nope, kv_lora_rank, n_head}, flags);
+ layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V_B, "weight", i), {kv_lora_rank, n_embd_head_v_mla, n_head}, flags);
+
+ layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head * n_embd_head_v_mla, n_embd}, flags);
+
+ layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags);
+
+ // DSA indexer
+ layer.indexer_k_norm = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM, "weight", i), {hparams.indexer_head_size}, flags);
+ layer.indexer_k_norm_b = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM, "bias", i), {hparams.indexer_head_size}, flags);
+ layer.indexer_proj = create_tensor(tn(LLM_TENSOR_INDEXER_PROJ, "weight", i), {n_embd, hparams.indexer_n_head}, flags);
+ layer.indexer_attn_k = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_K, "weight", i), {n_embd, hparams.indexer_head_size}, flags);
+ layer.indexer_attn_q_b = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_Q_B, "weight", i), {q_lora_rank, hparams.indexer_n_head * hparams.indexer_head_size}, flags);
+ if (i < (int) hparams.n_layer_dense_lead) {
+ layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, flags);
+ layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, flags);
+ layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, flags);
+ } else {
+ layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, flags);
+ layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED);
+
+ if (n_expert == 0) {
+ throw std::runtime_error("n_expert must be > 0");
+ }
+ if (n_expert_used == 0) {
+ throw std::runtime_error("n_expert_used must be > 0");
+ }
+
+ // MoE branch
+ layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert}, flags);
+ layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, flags);
+ layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert}, flags);
+
+ // Shared expert branch
+ layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, flags);
+ layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { n_ff_exp * n_expert_shared, n_embd}, flags);
+ layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, flags);
+ }
+
+ // NextN/MTP tensors (preserved but unused) - conditionally load for last nextn_predict_layers
+ if (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers) {
+ layer.nextn.eh_proj = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, flags);
+ layer.nextn.enorm = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), { n_embd }, flags);
+ layer.nextn.hnorm = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), { n_embd }, flags);
+
+ // Optional tensors
+ layer.nextn.embed_tokens = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED);
+ layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED);
+ layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, flags | TENSOR_NOT_REQUIRED);
+ }
+ }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_deepseek32::build_arch_graph(const llm_graph_params & params) const {
+ return std::make_unique<graph>(*this, params);
+}
+
+llama_model_deepseek32::graph::graph(const llama_model & model, const llm_graph_params & params) :
+ llm_graph_context(params) {
+ const bool is_mla = hparams.is_mla();
+ GGML_ASSERT(is_mla);
+
+ // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
+ const int64_t n_embd_head_k = hparams.n_embd_head_k_mla();
+ const int64_t n_embd_head_v = hparams.n_embd_head_v_mla();
+ GGML_UNUSED(n_embd_head_v);
+
+ const int64_t n_embd_head_qk_rope = hparams.n_rot();
+ const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope;
+
+ const int64_t n_indexer_head = hparams.indexer_n_head;
+ const int64_t n_embd_indexer_head = hparams.indexer_head_size;
+ const int64_t n_embd_indexer_head_rope = hparams.n_rot();
+ const int64_t n_embd_indexer_head_nope = n_embd_indexer_head - n_embd_indexer_head_rope;
+ const uint32_t n_indexer_top_k = hparams.indexer_top_k;
+
+ const uint32_t kv_lora_rank = hparams.n_lora_kv;
+
+ // We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly.
+ // See https://github.com/ggml-org/llama.cpp/discussions/7416 for detailed explanation.
+ // And also: https://github.com/ggml-org/llama.cpp/pull/17945 [TAG_DEEPSEEK2_YARN_LOG_MUL_FIX]
+
+ // first cancel the adjustment from llama_hparams::yarn_attn_factor_adjust to get the original attn_factor
+ GGML_ASSERT(ext_factor >= 0.0f);
+ const float attn_factor_org = attn_factor * (1.0f + 0.1f * logf(1.0f / freq_scale));
+
+ // use the original attn_factor to pre-scale the kq_scale
+ const float mscale = attn_factor_org * (1.0f + 0.1f * hparams.rope_yarn_log_mul * logf(1.0f / freq_scale));
+ const float kq_scale = 1.0f * mscale * mscale / sqrtf(float(n_embd_head_k));
+
+ ggml_tensor * cur;
+ ggml_tensor * inpL;
+
+ // {n_embd, n_tokens}
+ inpL = build_inp_embd(model.tok_embd);
+
+ // inp_pos - contains the positions
+ ggml_tensor * inp_pos = build_inp_pos();
+
+ llm_graph_input_attn_k_dsa * inp_attn_dsa = build_attn_inp_k_dsa();
+
+ ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+ int effective_n_layers = hparams.n_layer - hparams.nextn_predict_layers;
+ for (int il = 0; il < effective_n_layers; ++il) {
+ ggml_tensor * inpSA = inpL;
+
+ // norm
+ cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+ cb(cur, "attn_norm", il);
+
+ // self_attention
+ {
+ ggml_tensor * qr = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur);
+ cb(qr, "qr", il);
+
+ qr = build_norm(qr, model.layers[il].attn_q_a_norm, nullptr, LLM_NORM_RMS, il);
+ cb(qr, "qr", il);
+
+ ggml_tensor * top_k = nullptr;
+
+ // lightning indexer
+ {
+ ggml_tensor * indexer_q = ggml_mul_mat(ctx0, model.layers[il].indexer_attn_q_b, qr);
+ cb(indexer_q, "indexer_q", il);
+
+ // split into {n_embd_indexer_head_rope, n_indexer_head, n_tokens}
+ ggml_tensor * indexer_q_pe =
+ ggml_view_3d(ctx0, indexer_q, n_embd_indexer_head_rope, n_indexer_head, n_tokens,
+ ggml_row_size(indexer_q->type, n_embd_indexer_head),
+ ggml_row_size(indexer_q->type, n_embd_indexer_head) * n_indexer_head, 0);
+ cb(indexer_q_pe, "indexer_q_pe", il);
+
+ // and {n_embd_indexer_head_nope, n_indexer_head, n_tokens}
+ ggml_tensor * indexer_q_nope =
+ ggml_view_3d(ctx0, indexer_q, n_embd_indexer_head_nope, n_indexer_head, n_tokens,
+ ggml_row_size(indexer_q->type, n_embd_indexer_head),
+ ggml_row_size(indexer_q->type, n_embd_indexer_head) * n_indexer_head,
+ ggml_row_size(indexer_q->type, n_embd_indexer_head_nope));
+ cb(indexer_q_nope, "indexer_q_nope", il);
+
+ indexer_q_pe = ggml_rope_ext(ctx0, indexer_q_pe, inp_pos, nullptr, n_rot,
+ LLAMA_ROPE_TYPE_NEOX, n_ctx_orig, freq_base, freq_scale,
+ ext_factor, attn_factor, beta_fast, beta_slow);
+ cb(indexer_q_pe, "indexer_q_pe", il);
+
+ // {n_embd_indexer_head_rope + n_embd_indexer_head_nope, n_head, n_tokens}
+ indexer_q = ggml_concat(ctx0, indexer_q_pe, indexer_q_nope, 0);
+ cb(indexer_q, "indexer_q", il);
+
+ ggml_tensor * indexer_k = ggml_mul_mat(ctx0, model.layers[il].indexer_attn_k, cur);
+ cb(indexer_k, "indexer_k", il);
+
+ indexer_k = build_norm(indexer_k, model.layers[il].indexer_k_norm, model.layers[il].indexer_k_norm_b, LLM_NORM, il);
+ cb(indexer_k, "indexer_k", il);
+
+ // split into {n_embd_indexer_head_rope, 1, n_tokens}
+ ggml_tensor * indexer_k_pe =
+ ggml_view_3d(ctx0, indexer_k, n_embd_indexer_head_rope, 1, n_tokens,
+ ggml_row_size(indexer_k->type, n_embd_indexer_head),
+ ggml_row_size(indexer_k->type, n_embd_indexer_head) * 1, 0);
+ cb(indexer_k_pe, "indexer_k_pe", il);
+
+ // and {n_embd_indexer_head_nope, 1, n_tokens}
+ ggml_tensor * indexer_k_nope =
+ ggml_view_3d(ctx0, indexer_k, n_embd_indexer_head_nope, 1, n_tokens,
+ ggml_row_size(indexer_k->type, n_embd_indexer_head),
+ ggml_row_size(indexer_k->type, n_embd_indexer_head) * 1,
+ ggml_row_size(indexer_k->type, n_embd_indexer_head_nope));
+ cb(indexer_k_nope, "indexer_k_nope", il);
+
+ indexer_k_pe = ggml_rope_ext(ctx0, indexer_k_pe, inp_pos, nullptr, n_rot,
+ LLAMA_ROPE_TYPE_NEOX, n_ctx_orig, freq_base, freq_scale,
+ ext_factor, attn_factor, beta_fast, beta_slow);
+ cb(indexer_k_pe, "indexer_k_pe", il);
+
+ // {n_embd_indexer_head_rope + n_embd_indexer_head_nope, 1, n_tokens}
+ indexer_k = ggml_concat(ctx0, indexer_k_pe, indexer_k_nope, 0);
+ cb(indexer_k, "indexer_k", il);
+
+ // perform Hadamard transform on indexer q and k
+ indexer_q = ggml_mul_mat(ctx0, inp_attn_dsa->self_k_rot_lid, indexer_q);
+ cb(indexer_q, "indexer_q", il);
+ indexer_k = ggml_mul_mat(ctx0, inp_attn_dsa->self_k_rot_lid, indexer_k);
+ cb(indexer_k, "indexer_k", il);
+
+ // store indexer keys to KV cache
+ const auto * mctx_lid = inp_attn_dsa->mctx->get_lid();
+ const auto & k_idxs_lid = inp_attn_dsa->get_k_idxs_lid();
+ ggml_build_forward_expand(gf, mctx_lid->cpy_k(ctx0, indexer_k, k_idxs_lid, il));
+
+ // prepare indexer weights
+ ggml_tensor * indexer_weights = ggml_mul_mat(ctx0, model.layers[il].indexer_proj, cur);
+ cb(indexer_weights, "indexer_weights", il);
+
+ // get cached indexer keys
+ indexer_k = mctx_lid->get_k(ctx0, il);
+
+ // split the batch into streams if needed
+ const auto n_stream = indexer_k->ne[3];
+ indexer_q = ggml_view_4d(ctx0, indexer_q, indexer_q->ne[0], indexer_q->ne[1], indexer_q->ne[2]/n_stream, n_stream, indexer_q->nb[1], indexer_q->nb[2], indexer_q->nb[3]/n_stream, 0);
+ indexer_weights = ggml_view_4d(ctx0, indexer_weights, indexer_weights->ne[0], indexer_weights->ne[1]/n_stream, indexer_weights->ne[2], n_stream, indexer_weights->nb[1], indexer_weights->nb[2]/n_stream, indexer_weights->nb[3]/n_stream, 0);
+
+ // calculate indexer kq
+ indexer_q = ggml_permute(ctx0, indexer_q, 0, 2, 1, 3);
+ cb(indexer_q, "indexer_q", il);
+ indexer_k = ggml_permute(ctx0, indexer_k, 0, 2, 1, 3);
+ cb(indexer_k, "indexer_k", il);
+
+ ggml_tensor * indexer_kq = ggml_mul_mat(ctx0, indexer_k, indexer_q);
+ cb(indexer_kq, "indexer_kq", il);
+
+ // ReLU requires contiguous tensors
+ indexer_kq = ggml_cont(ctx0, ggml_permute(ctx0, indexer_kq, 2, 1, 0, 3));
+ cb(indexer_kq, "indexer_kq", il);
+
+ // apply ReLU
+ ggml_tensor * indexer_score = ggml_relu(ctx0, indexer_kq);
+ cb(indexer_score, "indexer_score", il);
+
+ // pre-scale weights to avoid scaling operations on huge indexer_score tensor
+ indexer_weights = ggml_scale(ctx0, indexer_weights, 1.0f / sqrtf(float(n_embd_indexer_head * n_indexer_head)));
+ cb(indexer_weights, "indexer_weights", il);
+
+ // multiply scores by indexer weights
+ indexer_score = ggml_mul(ctx0, indexer_score, indexer_weights);
+ cb(indexer_score, "indexer_score", il);
+
+ // sum by q n_indexer_head dimension
+ indexer_score = ggml_sum_rows(ctx0, indexer_score);
+ cb(indexer_score, "indexer_score", il);
+
+ // permute result to match KQ mask
+ indexer_score = ggml_cont(ctx0, ggml_permute(ctx0, indexer_score, 2, 1, 0, 3));
+ cb(indexer_score, "indexer_score", il);
+
+ // mask indexer scores
+ ggml_tensor * indexer_kq_mask = inp_attn_dsa->get_kq_mask_lid();
+ indexer_score = ggml_add(ctx0, indexer_score, indexer_kq_mask);
+ cb(indexer_score, "indexer_score", il);
+
+ // get indices of top k indexer scores
+ uint32_t n_top_k = indexer_score->ne[0] < n_indexer_top_k ? indexer_score->ne[0] : n_indexer_top_k;
+ top_k = ggml_cont(ctx0, ggml_top_k(ctx0, indexer_score, n_top_k));
+ cb(top_k, "top_k", il);
+ }
+
+ ggml_tensor * q = ggml_mul_mat(ctx0, model.layers[il].wq_b, qr);
+ cb(q, "q", il);
+
+ // split into {n_embd_head_qk_nope, n_head, n_tokens}
+ ggml_tensor * q_nope =
+ ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens, ggml_row_size(q->type, n_embd_head_k),
+ ggml_row_size(q->type, n_embd_head_k) * n_head, 0);
+ cb(q_nope, "q_nope", il);
+
+ // and {n_embd_head_qk_rope, n_head, n_tokens}
+ ggml_tensor * q_pe = ggml_view_3d(
+ ctx0, q, n_embd_head_qk_rope, n_head, n_tokens, ggml_row_size(q->type, n_embd_head_k),
+ ggml_row_size(q->type, n_embd_head_k) * n_head, ggml_row_size(q->type, n_embd_head_qk_nope));
+ cb(q_pe, "q_pe", il);
+
+ ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
+ cb(kv_cmpr_pe, "kv_cmpr_pe", il);
+
+ // split into {kv_lora_rank, n_tokens}
+ ggml_tensor * kv_cmpr =
+ ggml_view_2d(ctx0, kv_cmpr_pe, kv_lora_rank, n_tokens,
+ ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), 0);
+ cb(kv_cmpr, "kv_cmpr", il);
+
+ // and {n_embd_head_qk_rope, 1, n_tokens}
+ ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_cmpr_pe, n_embd_head_qk_rope, 1, n_tokens,
+ ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+ ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+ ggml_row_size(kv_cmpr_pe->type, kv_lora_rank));
+ cb(k_pe, "k_pe", il);
+
+ q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+ ext_factor, attn_factor, beta_fast, beta_slow);
+ cb(q_pe, "q_pe", il);
+
+ k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+ ext_factor, attn_factor, beta_fast, beta_slow);
+ cb(k_pe, "k_pe", il);
+
+ kv_cmpr = build_norm(kv_cmpr, model.layers[il].attn_kv_a_norm, nullptr, LLM_NORM_RMS, il);
+ cb(kv_cmpr, "kv_cmpr", il);
+
+ // MLA attention
+ {
+ // {n_embd_head_qk_nope, n_tokens, n_head}
+ q_nope = ggml_permute(ctx0, q_nope, 0, 2, 1, 3);
+ cb(q_nope, "q_nope_perm", il);
+
+ // {n_embd_head_qk_nope, kv_lora_rank, n_head} x {n_embd_head_qk_nope, n_tokens, n_head}
+ ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, model.layers[il].wk_b, q_nope);
+ cb(q_nope_absorbed, "q_nope_absorbed", il);
+
+ // {kv_lora_rank, n_head, n_tokens}
+ q_nope_absorbed = ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3);
+ cb(q_nope_absorbed, "q_nope_absorbed_perm", il);
+
+ // {n_embd_head_qk_rope + kv_lora_rank, n_head, n_tokens}
+ // note: rope must go first for in-place context shifting in build_rope_shift()
+ ggml_tensor * Qcur = ggml_concat(ctx0, q_nope_absorbed, q_pe, 0);
+ cb(Qcur, "Qcur", il);
+
+ kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens);
+ cb(kv_cmpr, "kv_cmpr_reshape", il);
+
+ // {n_embd_head_qk_rope + kv_lora_rank, 1, n_tokens}
+ ggml_tensor * Kcur = ggml_concat(ctx0, kv_cmpr, k_pe, 0);
+ cb(Kcur, "Kcur", il);
+
+ // {kv_lora_rank, 1, n_tokens}
+ ggml_tensor * Vcur = kv_cmpr;
+ cb(Vcur, "Vcur", il);
+
+ // note: MLA with the absorption optimization converts into MQA (ie: GQA with 1 group)
+ cur = build_attn(inp_attn_dsa,
+ model.layers[il].wo, NULL, model.layers[il].wo_s,
+ Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, top_k, kq_scale, il);
+ }
+ }
+ if (il == effective_n_layers - 1 && inp_out_ids) {
+ cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+ inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+ }
+ ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+ cb(ffn_inp, "ffn_inp", il);
+
+ cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+ cb(cur, "ffn_norm", il);
+
+ if ((uint32_t) il < hparams.n_layer_dense_lead) {
+ cur = build_ffn(cur,
+ model.layers[il].ffn_up, NULL, model.layers[il].ffn_up_s,
+ model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_s,
+ model.layers[il].ffn_down, NULL, model.layers[il].ffn_down_s,
+ NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+ cb(cur, "ffn_out", il);
+ } else {
+ // MoE branch
+ ggml_tensor * moe_out = build_moe_ffn(cur,
+ model.layers[il].ffn_gate_inp,
+ model.layers[il].ffn_up_exps,
+ model.layers[il].ffn_gate_exps,
+ model.layers[il].ffn_down_exps,
+ model.layers[il].ffn_exp_probs_b,
+ n_expert, n_expert_used,
+ LLM_FFN_SILU, hparams.expert_weights_norm,
+ hparams.expert_weights_scale,
+ (llama_expert_gating_func_type) hparams.expert_gating_func,
+ il,
+ nullptr,
+ model.layers[il].ffn_gate_up_exps,
+ model.layers[il].ffn_up_exps_s,
+ model.layers[il].ffn_gate_exps_s,
+ model.layers[il].ffn_down_exps_s);
+ cb(moe_out, "ffn_moe_out", il);
+
+ // FFN shared expert
+ {
+ ggml_tensor * ffn_shexp =
+ build_ffn(cur,
+ model.layers[il].ffn_up_shexp, NULL, model.layers[il].ffn_up_shexp_s,
+ model.layers[il].ffn_gate_shexp, NULL, model.layers[il].ffn_gate_shexp_s,
+ model.layers[il].ffn_down_shexp, NULL, model.layers[il].ffn_down_shexp_s,
+ NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+ cb(ffn_shexp, "ffn_shexp", il);
+
+ cur = ggml_add(ctx0, moe_out, ffn_shexp);
+ cb(cur, "ffn_out", il);
+ }
+ }
+ cur = ggml_add(ctx0, cur, ffn_inp);
+
+ cur = build_cvec(cur, il);
+ cb(cur, "l_out", il);
+
+ // input for next layer
+ inpL = cur;
+ }
+ cur = inpL;
+
+ cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+ cb(cur, "result_norm", -1);
+ res->t_embd = cur;
+
+ // lm_head
+ cur = ggml_mul_mat(ctx0, model.output, cur);
+
+ cb(cur, "result_output", -1);
+ res->t_logits = cur;
+
+ ggml_build_forward_expand(gf, cur);
+}
};
+struct llama_model_deepseek32 : public llama_model_base {
+ llama_model_deepseek32(const struct llama_model_params & params) : llama_model_base(params) {}
+ void load_arch_hparams(llama_model_loader & ml) override;
+ void load_arch_tensors(llama_model_loader & ml) override;
+
+ struct graph : public llm_graph_context {
+ graph(const llama_model & model, const llm_graph_params & params);
+ };
+
+ std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
+};
+
+
struct llama_model_deepseek2ocr : public llama_model_base {
llama_model_deepseek2ocr(const struct llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
n_ff = 96;
n_layer = 22; // hparams.n_layer_kv_from_start = 20 is hardcoded
} else if (arch == LLM_ARCH_DEEPSEEK2
+ || arch == LLM_ARCH_DEEPSEEK32
|| arch == LLM_ARCH_GLM_DSA
|| arch == LLM_ARCH_KIMI_LINEAR
|| arch == LLM_ARCH_MISTRAL4) {
ms.add_kv(LLM_KV_ATTENTION_MAX_ALIBI_BIAS, 8.0f);
if (arch == LLM_ARCH_DEEPSEEK2
+ || arch == LLM_ARCH_DEEPSEEK32
|| arch == LLM_ARCH_GLM_DSA
|| arch == LLM_ARCH_KIMI_LINEAR
|| arch == LLM_ARCH_MISTRAL4) {
case LLM_ARCH_ARCTIC:
case LLM_ARCH_DEEPSEEK:
case LLM_ARCH_DEEPSEEK2:
+ case LLM_ARCH_DEEPSEEK32:
case LLM_ARCH_GLM4_MOE:
case LLM_ARCH_GLM_DSA:
case LLM_ARCH_EXAONE_MOE:
overview / pkg / ggml / sources / llama.cpp / commitdiff