LLM inference in C/C++
C++ ggml
model: add dots3-note (#27060)
Xuan-Son Nguyen5a32f7b66ef6
diff --git a/conversion/__init__.py b/conversion/__init__.pyindex 4b8817ead4..b4afdf0f8d 100644--- a/conversion/__init__.py+++ b/conversion/__init__.py@@ -64,6 +64,9 @@ TEXT_MODEL_MAP: dict[str, str] = {"DistilBertForSequenceClassification": "bert","DistilBertModel": "bert","Dots1ForCausalLM": "dots1",+ "Dots3NoteForCausalLM": "dots3",+ "Dots3NoteForConditionalGeneration": "dots3",+ "Dots3NoteTextForCausalLM": "dots3","DotsOCRForCausalLM": "qwen","DreamModel": "dream","Ernie4_5ForCausalLM": "ernie",diff --git a/conversion/dots3.py b/conversion/dots3.pynew file mode 100644index 0000000000..7c36b84826--- /dev/null+++ b/conversion/dots3.py@@ -0,0 +1,195 @@+from __future__ import annotations++import math+import re++from typing import TYPE_CHECKING, Callable, Iterable++if TYPE_CHECKING:+ from torch import Tensor++from .base import ModelBase, gguf++from .deepseek import DeepseekV2Model++[email protected]("Dots3NoteForCausalLM", "Dots3NoteForConditionalGeneration", "Dots3NoteTextForCausalLM")+class Dots3NoteModel(DeepseekV2Model):+ model_arch = gguf.MODEL_ARCH.DOTS3NOTE+ skip_mtp = False+ supports_mtp_export = True++ # trunk layer count, stashed before indexing for filter_tensors (mirrors DeepseekV32Model)+ _n_main_layers: int | None = None++ def index_tensors(self, remote_hf_model_id: str | None = None):+ type(self)._n_main_layers = self.hparams["num_hidden_layers"]+ return super().index_tensors(remote_hf_model_id=remote_hf_model_id)++ def __init__(self, *args, **kwargs):+ super().__init__(*args, **kwargs)++ hparams = self.hparams++ # config file doesn't specify MTP block, detect it from model weight+ self.n_nextn = 1 if "model.mtp.embed_tokens.weight" in self.model_tensors else 0+ if self.n_nextn:+ self.block_count += self.n_nextn+ self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)++ self.layer_types = hparams["layer_types"]+ if len(self.layer_types) < hparams["num_hidden_layers"]:+ raise ValueError("layer_types is shorter than num_hidden_layers")++ if hparams.get("use_dsa", True) is not True:+ raise ValueError("dots3-note conversion requires use_dsa=true")+ if hparams.get("normalization", "RMSNorm") != "RMSNorm" or hparams.get("final_norm", "RMSNorm") != "RMSNorm":+ raise ValueError("dots3-note conversion only supports RMSNorm")+ if hparams.get("k_rope_only_layernorm", True) is not True:+ raise ValueError("dots3-note conversion requires k_rope_only_layernorm=true")+ if hparams.get("topk_method", "noaux_tc") != "noaux_tc" or hparams.get("scoring_func") != "sigmoid":+ raise ValueError("dots3-note conversion only supports noaux_tc/sigmoid expert gating")+ if hparams.get("n_group", 1) != 1 or hparams.get("topk_group", 1) != 1:+ raise ValueError("dots3-note conversion does not support grouped expert routing")+ if hparams.get("use_dynamic_rsf", False) or hparams.get("moe_gating_fp32", False):+ raise ValueError("dots3-note conversion does not support use_dynamic_rsf/moe_gating_fp32")+ for key in ("attention_gate_type", "swa_attention_gate_type"):+ if hparams.get(key, "headwise") != "headwise":+ raise ValueError(f"dots3-note conversion only supports headwise attention gate, got {key}={hparams.get(key)!r}")+ if hparams["swa_qk_nope_head_dim"] + hparams["swa_qk_rope_head_dim"] != hparams.get("swa_head_dim", 256):+ raise ValueError("swa_head_dim must equal swa_qk_nope_head_dim + swa_qk_rope_head_dim")+ if hparams["swa_qk_rope_head_dim"] != hparams["qk_rope_head_dim"]:+ # both layer kinds share a single rope_dimension_count+ raise ValueError("swa_qk_rope_head_dim must match qk_rope_head_dim")++ self.apply_lora_rescale = hparams.get("apply_mla_qkv_lora_rescale", False)++ def _is_swa_layer(self, bid: int) -> bool:+ if bid >= self.hparams["num_hidden_layers"]:+ # note: the NextN/MTP block uses the sliding-attention MLA+ return True+ return self.layer_types[bid] == "sliding_attention"++ def set_vocab(self):+ from transformers import AutoTokenizer+ tokenizer = AutoTokenizer.from_pretrained(self.dir_model)+ special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)+ tokens, toktypes, tokpre = self.get_vocab_base()+ self.gguf_writer.add_tokenizer_model("gpt2")+ self.gguf_writer.add_tokenizer_pre(tokpre)+ self.gguf_writer.add_token_list(tokens)+ self.gguf_writer.add_token_types(toktypes)+ special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|endofassistant|>"]) # ty: ignore[unresolved-attribute]+ special_vocab.add_to_gguf(self.gguf_writer)++ @classmethod+ def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:+ if (titem := super().filter_tensors(item)) is None:+ return None+ name, gen = titem+ if name.startswith(("vision_encoder.", "audio_encoder.")):+ return None++ assert cls._n_main_layers is not None+ is_mtp = name.startswith("model.mtp.") or \+ ((m := re.match(r"model\.layers\.(\d+)\.", name)) is not None and int(m.group(1)) >= cls._n_main_layers)++ # --no-mtp: drop the NextN/MTP block; --mtp: keep only that block plus the shared embeddings/norm/lm_head+ if is_mtp and cls.no_mtp:+ return None+ if cls.mtp_only and not is_mtp and name not in (+ "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight",+ ):+ return None++ return name, gen++ def set_gguf_parameters(self):+ hparams = self.hparams++ # head_count is a per-layer array because the two layer kinds have different head counts+ n_layer = hparams["num_hidden_layers"]+ hparams["num_attention_heads"] = [+ hparams["swa_num_attention_heads"] if self._is_swa_layer(il) else hparams["num_attention_heads"]+ for il in range(self.block_count)+ ]++ # prevent the base class from emitting key/value_length from the unused head_dim+ hparams.pop("head_dim", None)++ super().set_gguf_parameters()++ # MLA geometry of the sliding-window layers (rope.freq_base_swa is emitted by the base class)+ swa_kv_lora_rank = hparams["swa_kv_lora_rank"]+ self.gguf_writer.add_sliding_window(hparams["sliding_window_size"])+ self.gguf_writer.add_sliding_window_pattern([self._is_swa_layer(il) for il in range(n_layer)])+ self.gguf_writer.add_kv_lora_rank_swa(swa_kv_lora_rank)+ self.gguf_writer.add_key_length_swa(swa_kv_lora_rank + hparams["swa_qk_rope_head_dim"])+ self.gguf_writer.add_value_length_swa(swa_kv_lora_rank)+ self.gguf_writer.add_key_length_mla_swa(hparams["swa_qk_nope_head_dim"] + hparams["swa_qk_rope_head_dim"])+ self.gguf_writer.add_value_length_mla_swa(hparams["swa_v_head_dim"])+ if hparams["swa_q_lora_rank"] != hparams["q_lora_rank"]:+ raise ValueError("dots3-note conversion assumes a shared q_lora_rank for both layer kinds")++ if self.n_nextn:+ self.gguf_writer.add_nextn_predict_layers(self.n_nextn)++ # DSA indexer (full-attention layers only)+ self.gguf_writer.add_indexer_head_count(hparams["index_n_heads"])+ self.gguf_writer.add_indexer_key_length(hparams["index_head_dim"])+ self.gguf_writer.add_indexer_top_k(hparams["index_topk"])+ self.gguf_writer.add_indexer_types([not self._is_swa_layer(il) for il in range(n_layer)])++ def prepare_metadata(self, vocab_only: bool):+ from_dir = self.fname_out.is_dir()+ super().prepare_metadata(vocab_only=vocab_only)++ if not self.mtp_only or not from_dir:+ return++ output_type: str = self.ftype.name.partition("_")[2]+ fname_default: str = gguf.naming_convention(+ self.metadata.name, self.metadata.basename, self.metadata.finetune,+ self.metadata.version, size_label=None, output_type=output_type, model_type=None)+ self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf"++ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:+ # move the MTP token embedding into the NextN block so the standard nextn mapping picks it up+ if name == "model.mtp.embed_tokens.weight":+ name = f"model.layers.{self.hparams['num_hidden_layers']}.embed_tokens.weight"+ bid = self.hparams["num_hidden_layers"]++ # fold the activation rescale sqrt(n_embd/lora_rank) into the preceding RMSNorm weight+ # this also covers the indexer wq_b, which reads the same rescaled q_lora activation+ if self.apply_lora_rescale and bid is not None:+ if name.endswith("q_a_layernorm.weight"):+ data_torch = data_torch * math.sqrt(self.hparams["hidden_size"] / self.hparams["q_lora_rank"])+ elif name.endswith("kv_a_layernorm.weight"):+ rank = self.hparams["swa_kv_lora_rank"] if self._is_swa_layer(bid) else self.hparams["kv_lora_rank"]+ data_torch = data_torch * math.sqrt(self.hparams["hidden_size"] / rank)++ # MLA absorption: split kv_b_proj into k_b (transposed) and v_b, per-layer-kind geometry+ if name.endswith("kv_b_proj.weight"):+ assert bid is not None+ if self._is_swa_layer(bid):+ n_head = self.hparams["swa_num_attention_heads"]+ qk_nope_head_dim = self.hparams["swa_qk_nope_head_dim"]+ v_head_dim = self.hparams["swa_v_head_dim"]+ else:+ n_head = self.hparams["num_attention_heads"]+ qk_nope_head_dim = self.hparams["qk_nope_head_dim"]+ v_head_dim = self.hparams["v_head_dim"]+ if isinstance(n_head, list): # set_gguf_parameters turns this into a per-layer array+ n_head = n_head[bid]++ assert data_torch.shape[0] == n_head * (qk_nope_head_dim + v_head_dim)++ kv_b = data_torch.view(n_head, qk_nope_head_dim + v_head_dim, data_torch.shape[-1])+ k_b, v_b = kv_b.split([qk_nope_head_dim, v_head_dim], dim=1)+ k_b = k_b.transpose(1, 2)++ yield from ModelBase.modify_tensors(self, k_b, name.replace("kv_b_proj", "k_b_proj"), bid)+ yield from ModelBase.modify_tensors(self, v_b, name.replace("kv_b_proj", "v_b_proj"), bid)+ return++ yield from super().modify_tensors(data_torch, name, bid)diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.pyindex fad8d1fd8c..886253d889 100644--- a/gguf-py/gguf/constants.py+++ b/gguf-py/gguf/constants.py@@ -205,6 +205,9 @@ class Keys:VALUE_LENGTH_MLA = "{arch}.attention.value_length_mla"KEY_LENGTH_SWA = "{arch}.attention.key_length_swa"VALUE_LENGTH_SWA = "{arch}.attention.value_length_swa"+ KEY_LENGTH_MLA_SWA = "{arch}.attention.key_length_mla_swa"+ VALUE_LENGTH_MLA_SWA = "{arch}.attention.value_length_mla_swa"+ KV_LORA_RANK_SWA = "{arch}.attention.kv_lora_rank_swa"SHARED_KV_LAYERS = "{arch}.attention.shared_kv_layers"SLIDING_WINDOW_PATTERN = "{arch}.attention.sliding_window_pattern"TEMPERATURE_SCALE = "{arch}.attention.temperature_scale"@@ -558,6 +561,7 @@ class MODEL_ARCH(IntEnum):BAILINGMOE2 = auto()BAILINGMOE3 = auto()DOTS1 = auto()+ DOTS3NOTE = auto()ARCEE = auto()AFMOE = auto()LAGUNA = auto()@@ -1275,6 +1279,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {MODEL_ARCH.BAILINGMOE2: "bailingmoe2",MODEL_ARCH.BAILINGMOE3: "bailingmoe3",MODEL_ARCH.DOTS1: "dots1",+ MODEL_ARCH.DOTS3NOTE: "dots3note",MODEL_ARCH.ARCEE: "arcee",MODEL_ARCH.AFMOE: "afmoe",MODEL_ARCH.LAGUNA: "laguna",@@ -4334,6 +4339,44 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {MODEL_TENSOR.FFN_UP_EXP,MODEL_TENSOR.FFN_UP_SHEXP,],+ MODEL_ARCH.DOTS3NOTE: [+ MODEL_TENSOR.TOKEN_EMBD,+ MODEL_TENSOR.OUTPUT_NORM,+ MODEL_TENSOR.OUTPUT,+ MODEL_TENSOR.ATTN_NORM,+ 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_K_NORM,+ MODEL_TENSOR.ATTN_GATE,+ MODEL_TENSOR.ATTN_OUT,+ MODEL_TENSOR.FFN_NORM,+ MODEL_TENSOR.FFN_GATE,+ MODEL_TENSOR.FFN_DOWN,+ MODEL_TENSOR.FFN_UP,+ MODEL_TENSOR.FFN_GATE_INP,+ MODEL_TENSOR.FFN_EXP_PROBS_B,+ 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.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_NORM,+ ],MODEL_ARCH.ARCEE: [MODEL_TENSOR.TOKEN_EMBD,MODEL_TENSOR.OUTPUT_NORM,diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.pyindex 16ae9f999d..4968820251 100644--- a/gguf-py/gguf/gguf_writer.py+++ b/gguf-py/gguf/gguf_writer.py@@ -785,6 +785,15 @@ class GGUFWriter:def add_key_length_swa(self, length: int) -> None:self.add_uint32(Keys.Attention.KEY_LENGTH_SWA.format(arch=self.arch), length)+ def add_key_length_mla_swa(self, length: int) -> None:+ self.add_uint32(Keys.Attention.KEY_LENGTH_MLA_SWA.format(arch=self.arch), length)++ def add_value_length_mla_swa(self, length: int) -> None:+ self.add_uint32(Keys.Attention.VALUE_LENGTH_MLA_SWA.format(arch=self.arch), length)++ def add_kv_lora_rank_swa(self, length: int) -> None:+ self.add_uint32(Keys.Attention.KV_LORA_RANK_SWA.format(arch=self.arch), length)+def add_value_length_swa(self, length: int) -> None:self.add_uint32(Keys.Attention.VALUE_LENGTH_SWA.format(arch=self.arch), length)diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.pyindex a0571ccd32..1ff4b61d96 100644--- a/gguf-py/gguf/tensor_mapping.py+++ b/gguf-py/gguf/tensor_mapping.py@@ -723,6 +723,7 @@ class TensorNameMap:"model.layers.layers.{bid}.mixer.k", # plamo2"model.layers.layers.{bid}.mixer.k_norm", # plamo3"layers.{bid}.self_attn.k_norm", # qwen3-embedding+ "model.layers.{bid}.self_attn.k_rope_only_layernorm", # dots3note"model.layers.{bid}.attention.key_layernorm", # apertus),diff --git a/src/CMakeLists.txt b/src/CMakeLists.txtindex 39ba3061f7..c6df19f2ec 100644--- a/src/CMakeLists.txt+++ b/src/CMakeLists.txt@@ -25,6 +25,7 @@ add_library(llamallama-kv-cache.cppllama-kv-cache-iswa.cppllama-kv-cache-dsa.cpp+ llama-kv-cache-dsa-iswa.cppllama-kv-cache-msa.cppllama-kv-cache-dsv4.cppllama-memory.cppdiff --git a/src/llama-arch.cpp b/src/llama-arch.cppindex c9b504c333..60b8e461c2 100644--- a/src/llama-arch.cpp+++ b/src/llama-arch.cpp@@ -110,6 +110,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {{ LLM_ARCH_BAILINGMOE2, "bailingmoe2" },{ LLM_ARCH_BAILINGMOE3, "bailingmoe3" },{ LLM_ARCH_DOTS1, "dots1" },+ { LLM_ARCH_DOTS3NOTE, "dots3note" },{ LLM_ARCH_ARCEE, "arcee" },{ LLM_ARCH_AFMOE, "afmoe" },{ LLM_ARCH_LAGUNA, "laguna" },@@ -273,6 +274,9 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {{ LLM_KV_ATTENTION_VALUE_LENGTH_MLA, "%s.attention.value_length_mla" },{ LLM_KV_ATTENTION_KEY_LENGTH_SWA, "%s.attention.key_length_swa" },{ LLM_KV_ATTENTION_VALUE_LENGTH_SWA, "%s.attention.value_length_swa" },+ { LLM_KV_ATTENTION_KEY_LENGTH_MLA_SWA, "%s.attention.key_length_mla_swa" },+ { LLM_KV_ATTENTION_VALUE_LENGTH_MLA_SWA, "%s.attention.value_length_mla_swa" },+ { LLM_KV_ATTENTION_KV_LORA_RANK_SWA, "%s.attention.kv_lora_rank_swa" },{ LLM_KV_ATTENTION_INDEXER_HEAD_COUNT, "%s.attention.indexer.head_count" },{ LLM_KV_ATTENTION_INDEXER_KEY_LENGTH, "%s.attention.indexer.key_length" },{ LLM_KV_ATTENTION_INDEXER_TOP_K, "%s.attention.indexer.top_k" },@@ -1056,6 +1060,7 @@ bool llm_arch_supports_sm_tensor(const llm_arch & arch) {case LLM_ARCH_DEEPSEEK2:case LLM_ARCH_DEEPSEEK32:case LLM_ARCH_DEEPSEEK4:+ case LLM_ARCH_DOTS3NOTE:case LLM_ARCH_GLM_DSA:case LLM_ARCH_BITNET:case LLM_ARCH_T5:diff --git a/src/llama-arch.h b/src/llama-arch.hindex 48fe051a99..7159e23bf7 100644--- a/src/llama-arch.h+++ b/src/llama-arch.h@@ -115,6 +115,7 @@ enum llm_arch {LLM_ARCH_BAILINGMOE2,LLM_ARCH_BAILINGMOE3,LLM_ARCH_DOTS1,+ LLM_ARCH_DOTS3NOTE,LLM_ARCH_ARCEE,LLM_ARCH_AFMOE,LLM_ARCH_LAGUNA,@@ -278,6 +279,9 @@ enum llm_kv {LLM_KV_ATTENTION_VALUE_LENGTH_MLA,LLM_KV_ATTENTION_KEY_LENGTH_SWA,LLM_KV_ATTENTION_VALUE_LENGTH_SWA,+ LLM_KV_ATTENTION_KEY_LENGTH_MLA_SWA,+ LLM_KV_ATTENTION_VALUE_LENGTH_MLA_SWA,+ LLM_KV_ATTENTION_KV_LORA_RANK_SWA,LLM_KV_ATTENTION_INDEXER_HEAD_COUNT,LLM_KV_ATTENTION_INDEXER_KEY_LENGTH,LLM_KV_ATTENTION_INDEXER_TOP_K,diff --git a/src/llama-graph.cpp b/src/llama-graph.cppindex 5212e19a25..8fca8e1bc0 100644--- a/src/llama-graph.cpp+++ b/src/llama-graph.cpp@@ -9,6 +9,7 @@#include "llama-kv-cache.h"#include "llama-kv-cache-iswa.h"#include "llama-kv-cache-dsa.h"+#include "llama-kv-cache-dsa-iswa.h"#include "llama-kv-cache-msa.h"#include "llama-kv-cache-dsv4.h"#include "llama-memory-hybrid.h"@@ -507,10 +508,12 @@ void llm_graph_input_attn_k::set_input(const llama_ubatch * ubatch) {}bool llm_graph_input_attn_k::can_reuse(const llm_graph_params & params) {- const auto * mctx = static_cast<const llama_kv_cache_context *>(params.mctx);+ mctx = static_cast<const llama_kv_cache_context *>(params.mctx);- this->mctx = mctx;+ return can_reuse_impl(params);+}+bool llm_graph_input_attn_k::can_reuse_impl(const llm_graph_params & params) {bool res = true;res &= self_k_idxs->ne[0] == params.ubatch.n_tokens;@@ -567,10 +570,12 @@ void llm_graph_input_attn_k_dsa::set_input(const llama_ubatch * ubatch) {}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);+ mctx = static_cast<const llama_kv_cache_dsa_context *>(params.mctx);- this->mctx = mctx;+ return can_reuse_impl(params);+}+bool llm_graph_input_attn_k_dsa::can_reuse_impl(const llm_graph_params & params) {bool res = true;res &= self_k_idxs_mla->ne[0] == params.ubatch.n_tokens;@@ -582,6 +587,25 @@ bool llm_graph_input_attn_k_dsa::can_reuse(const llm_graph_params & params) {return res;}+void llm_graph_input_attn_k_dsa_iswa::set_input(const llama_ubatch * ubatch) {+ inp_dsa->set_input(ubatch);+ inp_swa->set_input(ubatch);+}++bool llm_graph_input_attn_k_dsa_iswa::can_reuse(const llm_graph_params & params) {+ mctx = static_cast<const llama_kv_cache_dsa_iswa_context *>(params.mctx);++ inp_dsa->mctx = mctx->get_dsa();+ inp_swa->mctx = mctx->get_swa();++ bool res = true;++ res &= inp_dsa->can_reuse_impl(params);+ res &= inp_swa->can_reuse_impl(params);++ 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 layersif (self_k_idxs && self_k_idxs->buffer) {@@ -3210,8 +3234,12 @@ ggml_tensor * llm_graph_context::build_attn(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);+static std::unique_ptr<llm_graph_input_attn_k_dsa> build_attn_inp_k_dsa_impl(+ ggml_context * ctx0,+ const llama_ubatch & ubatch,+ const llama_hparams & hparams,+ const llama_cparams & cparams,+ const llama_kv_cache_dsa_context * mctx_cur) {auto inp = std::make_unique<llm_graph_input_attn_k_dsa>(hparams, cparams, mctx_cur);@@ -3235,9 +3263,35 @@ llm_graph_input_attn_k_dsa * llm_graph_context::build_attn_inp_k_dsa() const {inp->self_k_rot_lid = mctx_cur->get_lid()->build_input_k_rot(ctx0);}+ return inp;+}++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 = build_attn_inp_k_dsa_impl(ctx0, ubatch, hparams, cparams, mctx_cur);+return (llm_graph_input_attn_k_dsa *) res->add_input(std::move(inp));}+llm_graph_input_attn_k_dsa_iswa * llm_graph_context::build_attn_inp_k_dsa_iswa() const {+ const auto * mctx_cur = static_cast<const llama_kv_cache_dsa_iswa_context *>(mctx);++ auto inp_dsa = build_attn_inp_k_dsa_impl(ctx0, ubatch, hparams, cparams, mctx_cur->get_dsa());++ // build_attn_inp_k_impl rejects SWA caches, so construct the input directly+ auto inp_swa = std::make_unique<llm_graph_input_attn_k>(hparams, cparams, mctx_cur->get_swa());++ inp_swa->self_k_idxs = mctx_cur->get_swa()->build_input_k_idxs(ctx0, ubatch);++ inp_swa->self_kq_mask = build_attn_inp_kq_mask(ctx0, mctx_cur->get_swa(), ubatch, cparams);+ inp_swa->self_kq_mask_cnv = inp_swa->self_kq_mask;++ auto inp = std::make_unique<llm_graph_input_attn_k_dsa_iswa>(std::move(inp_dsa), std::move(inp_swa), mctx_cur);++ return (llm_graph_input_attn_k_dsa_iswa *) res->add_input(std::move(inp));+}+llm_graph_input_attn_kv_msa * llm_graph_context::build_attn_inp_kv_msa(bool msa_enabled) const {const auto * mctx_cur = static_cast<const llama_kv_cache_msa_context *>(mctx);diff --git a/src/llama-graph.h b/src/llama-graph.hindex 94324c7457..b388e028cb 100644--- a/src/llama-graph.h+++ b/src/llama-graph.h@@ -23,6 +23,7 @@ struct llama_memory_context_i;class llama_kv_cache_context;class llama_kv_cache_dsa_context;+class llama_kv_cache_dsa_iswa_context;class llama_kv_cache_msa_context;class llama_kv_cache_dsv4_raw_context;class llama_kv_cache_dsv4_context;@@ -374,6 +375,9 @@ public:bool can_reuse(const llm_graph_params & params) override;+ // like can_reuse, but does not re-bind mctx+ bool can_reuse_impl(const llm_graph_params & params);+ggml_tensor * get_k_idxs() const { return self_k_idxs; }ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; }@@ -405,6 +409,9 @@ public:bool can_reuse(const llm_graph_params & params) override;+ // like can_reuse, but does not re-bind mctx+ bool can_reuse_impl(const llm_graph_params & params);+ggml_tensor * get_k_idxs_mla() const { return self_k_idxs_mla; }ggml_tensor * get_k_idxs_lid() const { return self_k_idxs_lid; }@@ -427,6 +434,32 @@ public:const llama_kv_cache_dsa_context * mctx;};+// DSA input (full-attention layers + indexer) with K-only input for the SWA layers+class llm_graph_input_attn_k_dsa_iswa : public llm_graph_input_i {+public:+ llm_graph_input_attn_k_dsa_iswa(+ std::unique_ptr<llm_graph_input_attn_k_dsa> inp_dsa,+ std::unique_ptr<llm_graph_input_attn_k> inp_swa,+ const llama_kv_cache_dsa_iswa_context * mctx) :+ inp_dsa(std::move(inp_dsa)),+ inp_swa(std::move(inp_swa)),+ mctx(mctx) {+ }+ ~llm_graph_input_attn_k_dsa_iswa() = default;++ void set_input(const llama_ubatch * ubatch) override;++ bool can_reuse(const llm_graph_params & params) override;++ llm_graph_input_attn_k_dsa * get_dsa() const { return inp_dsa.get(); }+ llm_graph_input_attn_k * get_swa() const { return inp_swa.get(); }++ std::unique_ptr<llm_graph_input_attn_k_dsa> inp_dsa;+ std::unique_ptr<llm_graph_input_attn_k> inp_swa;++ const llama_kv_cache_dsa_iswa_context * mctx;+};+// standard K/V attention input against the base cache, plus destination indices for the indexer key cacheclass llm_graph_input_attn_kv_msa : public llm_graph_input_attn_kv {public:@@ -1191,6 +1224,8 @@ struct llm_graph_context {llm_graph_input_attn_k_dsa * build_attn_inp_k_dsa() const;+ llm_graph_input_attn_k_dsa_iswa * build_attn_inp_k_dsa_iswa() const;+llm_graph_input_attn_kv_msa * build_attn_inp_kv_msa(bool msa_enabled) const;ggml_tensor * build_attn(diff --git a/src/llama-hparams.h b/src/llama-hparams.hindex f6af36436b..c3c14292c3 100644--- a/src/llama-hparams.h+++ b/src/llama-hparams.h@@ -101,6 +101,11 @@ struct llama_hparams {uint32_t n_group_used = 0;uint32_t n_group_experts = 0;+ // MLA + SWA (i.e. dots3note)+ uint32_t n_lora_kv_swa = 0;+ uint32_t n_embd_head_k_mla_swa = 0;+ uint32_t n_embd_head_v_mla_swa = 0;+float expert_group_scale = 0.05f;float expert_weights_scale = 0.0f;bool expert_weights_norm = false;diff --git a/src/llama-kv-cache-dsa-iswa.cpp b/src/llama-kv-cache-dsa-iswa.cppnew file mode 100644index 0000000000..dc10342a19--- /dev/null+++ b/src/llama-kv-cache-dsa-iswa.cpp@@ -0,0 +1,341 @@+#include "llama-kv-cache-dsa-iswa.h"++#include "llama-impl.h"+#include "llama-batch.h"+#include "llama-model.h"++#include <algorithm>+#include <cassert>++//+// llama_kv_cache_dsa_iswa+//++llama_kv_cache_dsa_iswa::llama_kv_cache_dsa_iswa(+ const llama_model & model,+ ggml_type type_k,+ ggml_type type_v,+ bool v_trans,+ bool offload,+ bool swa_full,+ bool unified,+ uint32_t kv_size,+ uint32_t n_seq_max,+ uint32_t n_ubatch,+ uint32_t n_pad,+ const layer_filter_cb & filter_mla,+ const layer_filter_cb & filter_lid,+ const layer_reuse_cb & reuse) : unified(unified) {++ const auto & hparams = model.hparams;++ // chain filters+ const layer_filter_cb filter_dsa = [&](int32_t il) {+ if (filter_mla && !filter_mla(il)) {+ return false;+ }++ return !hparams.is_swa(il);+ };++ const layer_filter_cb filter_swa = [&](int32_t il) {+ if (filter_mla && !filter_mla(il)) {+ return false;+ }++ return hparams.is_swa(il);+ };++ const uint32_t size_dsa = kv_size;++ // note: the SWA cache is always padded to 256 for performance+ // https://github.com/ggml-org/llama.cpp/issues/17037+ uint32_t size_swa = GGML_PAD(std::min(size_dsa, hparams.n_swa*(unified ? n_seq_max : 1) + n_ubatch), 256);++ // when using full-size SWA cache, we set the SWA cache size to be equal to the base cache size+ if (swa_full) {+ LLAMA_LOG_WARN("%s: using full-size SWA cache (ref: %s)\n",+ __func__, "https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055");++ size_swa = size_dsa;+ }++ LLAMA_LOG_INFO("%s: creating DSA KV cache, size = %u cells\n", __func__, size_dsa);++ kv_dsa = std::make_unique<llama_kv_cache_dsa>(+ model, type_k, type_v,+ v_trans, offload, unified, size_dsa, n_seq_max, n_pad,+ 0, LLAMA_SWA_TYPE_NONE, filter_dsa, filter_lid, 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, hparams, type_k, type_v,+ v_trans, offload, unified, size_swa, n_seq_max, n_pad,+ hparams.n_swa, hparams.swa_type, nullptr, filter_swa, reuse, nullptr);+}++void llama_kv_cache_dsa_iswa::clear(bool data) {+ kv_dsa->clear(data);+ kv_swa->clear(data);+}++bool llama_kv_cache_dsa_iswa::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {+ bool res = true;++ res = res & kv_dsa->seq_rm(seq_id, p0, p1);+ res = res & kv_swa->seq_rm(seq_id, p0, p1);++ return res;+}++void llama_kv_cache_dsa_iswa::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {+ kv_dsa->seq_cp(seq_id_src, seq_id_dst, p0, p1);+ kv_swa->seq_cp(seq_id_src, seq_id_dst, p0, p1);+}++void llama_kv_cache_dsa_iswa::seq_keep(llama_seq_id seq_id) {+ kv_dsa->seq_keep(seq_id);+ kv_swa->seq_keep(seq_id);+}++void llama_kv_cache_dsa_iswa::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {+ kv_dsa->seq_add(seq_id, p0, p1, shift);+ kv_swa->seq_add(seq_id, p0, p1, shift);+}++void llama_kv_cache_dsa_iswa::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {+ kv_dsa->seq_div(seq_id, p0, p1, d);+ kv_swa->seq_div(seq_id, p0, p1, d);+}++llama_pos llama_kv_cache_dsa_iswa::seq_pos_min(llama_seq_id seq_id) const {+ // the DSA cache is a superset of the SWA cache, so we can just check the SWA cache+ return kv_swa->seq_pos_min(seq_id);+}++llama_pos llama_kv_cache_dsa_iswa::seq_pos_max(llama_seq_id seq_id) const {+ return kv_swa->seq_pos_max(seq_id);+}++std::map<ggml_backend_buffer_type_t, size_t> llama_kv_cache_dsa_iswa::memory_breakdown() const {+ std::map<ggml_backend_buffer_type_t, size_t> mb = kv_dsa->memory_breakdown();+ for (const auto & buft_size : kv_swa->memory_breakdown()) {+ mb[buft_size.first] += buft_size.second;+ }+ return mb;+}++llama_memory_context_ptr llama_kv_cache_dsa_iswa::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) {+ GGML_UNUSED(embd_all);++ // first try simple split+ do {+ if (!unified) {+ // requires equal splits, so we skip the simple split+ break;+ }++ balloc.split_reset();++ std::vector<llama_ubatch> ubatches;+ while (true) {+ auto ubatch = balloc.split_simple(n_ubatch);++ 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_dsa->get_mla()->prepare(ubatches);+ if (sinfos_mla.empty()) {+ break;+ }++ auto sinfos_lid = kv_dsa->get_lid()->prepare(ubatches);+ if (sinfos_lid.empty()) {+ break;+ }++ auto sinfos_swa = kv_swa->prepare(ubatches);+ if (sinfos_swa.empty()) {+ break;+ }++ assert(sinfos_mla.size() == sinfos_swa.size());++ return std::make_unique<llama_kv_cache_dsa_iswa_context>(+ this, std::move(sinfos_mla), std::move(sinfos_lid), std::move(sinfos_swa), std::move(ubatches));+ } while (false);++ // if it fails, try equal split+ do {+ balloc.split_reset();++ std::vector<llama_ubatch> ubatches;+ while (true) {+ auto ubatch = balloc.split_equal(n_ubatch, !unified, 0);++ 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_dsa->get_mla()->prepare(ubatches);+ if (sinfos_mla.empty()) {+ break;+ }++ auto sinfos_lid = kv_dsa->get_lid()->prepare(ubatches);+ if (sinfos_lid.empty()) {+ break;+ }++ auto sinfos_swa = kv_swa->prepare(ubatches);+ if (sinfos_swa.empty()) {+ break;+ }++ assert(sinfos_mla.size() == sinfos_swa.size());++ return std::make_unique<llama_kv_cache_dsa_iswa_context>(+ this, std::move(sinfos_mla), std::move(sinfos_lid), std::move(sinfos_swa), std::move(ubatches));+ } while (false);++ return std::make_unique<llama_kv_cache_dsa_iswa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);+}++llama_memory_context_ptr llama_kv_cache_dsa_iswa::init_full() {+ return std::make_unique<llama_kv_cache_dsa_iswa_context>(this);+}++llama_memory_context_ptr llama_kv_cache_dsa_iswa::init_update(llama_context * lctx, bool optimize) {+ return std::make_unique<llama_kv_cache_dsa_iswa_context>(this, lctx, optimize);+}++bool llama_kv_cache_dsa_iswa::get_can_shift() const {+ return kv_dsa->get_can_shift() &&+ kv_swa->get_can_shift() &&+ kv_dsa->get_mla()->get_size() == kv_swa->get_size();+}++void llama_kv_cache_dsa_iswa::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {+ if ((flags & LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY) == 0) {+ kv_dsa->state_write(io, seq_id, flags);+ }++ kv_swa->state_write(io, seq_id, flags);+}++void llama_kv_cache_dsa_iswa::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {+ if ((flags & LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY) == 0) {+ kv_dsa->state_read(io, seq_id, flags);+ }++ kv_swa->state_read(io, seq_id, flags);+}++llama_kv_cache_dsa * llama_kv_cache_dsa_iswa::get_dsa() const {+ return kv_dsa.get();+}++llama_kv_cache * llama_kv_cache_dsa_iswa::get_swa() const {+ return kv_swa.get();+}++//+// llama_kv_cache_dsa_iswa_context+//++llama_kv_cache_dsa_iswa_context::llama_kv_cache_dsa_iswa_context(llama_memory_status status) : status(status) {}++llama_kv_cache_dsa_iswa_context::llama_kv_cache_dsa_iswa_context(+ llama_kv_cache_dsa_iswa * kv) :+ ctx_dsa(kv->get_dsa()->init_full()),+ ctx_swa(kv->get_swa()->init_full()),+ status(llama_memory_status_combine(ctx_dsa->get_status(), ctx_swa->get_status())) {+}++llama_kv_cache_dsa_iswa_context::llama_kv_cache_dsa_iswa_context(+ llama_kv_cache_dsa_iswa * kv,+ llama_context * lctx,+ bool optimize) :+ ctx_dsa(kv->get_dsa()->init_update(lctx, optimize)),+ ctx_swa(kv->get_swa()->init_update(lctx, optimize)),+ status(llama_memory_status_combine(ctx_dsa->get_status(), ctx_swa->get_status())) {+}++llama_kv_cache_dsa_iswa_context::llama_kv_cache_dsa_iswa_context(+ llama_kv_cache_dsa_iswa * kv,+ slot_info_vec_t sinfos_mla,+ slot_info_vec_t sinfos_lid,+ slot_info_vec_t sinfos_swa,+ std::vector<llama_ubatch> ubatches) :+ ubatches(std::move(ubatches)),+ // note: here we copy the ubatches. not sure if this is ideal+ ctx_dsa(new llama_kv_cache_dsa_context(kv->get_dsa(), std::move(sinfos_mla), std::move(sinfos_lid), this->ubatches)),+ ctx_swa(new llama_kv_cache_context(kv->get_swa(), std::move(sinfos_swa), this->ubatches)),+ status(llama_memory_status_combine(ctx_dsa->get_status(), ctx_swa->get_status())) {+}++llama_kv_cache_dsa_iswa_context:: ~llama_kv_cache_dsa_iswa_context() = default;++bool llama_kv_cache_dsa_iswa_context::next() {+ assert(status == LLAMA_MEMORY_STATUS_SUCCESS);++ ctx_dsa->next();+ ctx_swa->next();++ if (++i_next >= ubatches.size()) {+ return false;+ }++ return true;+}++bool llama_kv_cache_dsa_iswa_context::apply() {+ assert(!llama_memory_status_is_fail(status));++ bool res = true;++ res = res & ctx_dsa->apply();+ res = res & ctx_swa->apply();++ return res;+}++llama_memory_status llama_kv_cache_dsa_iswa_context::get_status() const {+ return status;+}++const llama_ubatch & llama_kv_cache_dsa_iswa_context::get_ubatch() const {+ assert(status == LLAMA_MEMORY_STATUS_SUCCESS);++ return ubatches[i_next];+}++const llama_kv_cache_dsa_context * llama_kv_cache_dsa_iswa_context::get_dsa() const {+ assert(status == LLAMA_MEMORY_STATUS_SUCCESS);++ return static_cast<const llama_kv_cache_dsa_context *>(ctx_dsa.get());+}++const llama_kv_cache_context * llama_kv_cache_dsa_iswa_context::get_swa() const {+ assert(status == LLAMA_MEMORY_STATUS_SUCCESS);++ return static_cast<const llama_kv_cache_context *>(ctx_swa.get());+}diff --git a/src/llama-kv-cache-dsa-iswa.h b/src/llama-kv-cache-dsa-iswa.hnew file mode 100644index 0000000000..28cf95bf05--- /dev/null+++ b/src/llama-kv-cache-dsa-iswa.h@@ -0,0 +1,134 @@+#pragma once++#include "llama-kv-cache-dsa.h"++#include <vector>++//+// llama_kv_cache_dsa_iswa+//++// utilizes two child memories: llama_kv_cache_dsa for the full-attention (DSA) layers and llama_kv_cache for the SWA layers++class llama_kv_cache_dsa_iswa : public llama_memory_i {+public:+ llama_kv_cache_dsa_iswa(+ const llama_model & model,+ ggml_type type_k,+ ggml_type type_v,+ bool v_trans,+ bool offload,+ bool swa_full,+ bool unified,+ uint32_t kv_size,+ uint32_t n_seq_max,+ uint32_t n_ubatch,+ uint32_t n_pad,+ const layer_filter_cb & filter_mla,+ const layer_filter_cb & filter_lid,+ const layer_reuse_cb & reuse);++ ~llama_kv_cache_dsa_iswa() = 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_iswa specific API+ //++ llama_kv_cache_dsa * get_dsa() const;+ llama_kv_cache * get_swa() const;++private:+ const bool unified;++ std::unique_ptr<llama_kv_cache_dsa> kv_dsa;+ std::unique_ptr<llama_kv_cache> kv_swa;+};++class llama_kv_cache_dsa_iswa_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_iswa_context(llama_memory_status status);++ // used to create a full-cache context+ llama_kv_cache_dsa_iswa_context(+ llama_kv_cache_dsa_iswa * kv);++ // used to create an update context+ llama_kv_cache_dsa_iswa_context(+ llama_kv_cache_dsa_iswa * kv,+ llama_context * lctx,+ bool optimize);++ // used to create a batch processing context from a batch+ llama_kv_cache_dsa_iswa_context(+ llama_kv_cache_dsa_iswa * kv,+ slot_info_vec_t sinfos_mla,+ slot_info_vec_t sinfos_lid,+ slot_info_vec_t sinfos_swa,+ std::vector<llama_ubatch> ubatches);++ virtual ~llama_kv_cache_dsa_iswa_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_iswa_context specific API+ //++ const llama_kv_cache_dsa_context * get_dsa() const;+ const llama_kv_cache_context * get_swa() const;++private:+ // the index of the next ubatch to process+ size_t i_next = 0;++ std::vector<llama_ubatch> ubatches;++ const llama_memory_context_ptr ctx_dsa;+ const llama_memory_context_ptr ctx_swa;++ const llama_memory_status status;+};diff --git a/src/llama-kv-cache.cpp b/src/llama-kv-cache.cppindex 5382cd7266..2e2bd7dc6d 100644--- a/src/llama-kv-cache.cpp+++ b/src/llama-kv-cache.cpp@@ -323,7 +323,8 @@ llama_kv_cache::llama_kv_cache(hparams.n_embd_head_k() % 64 == 0;// always create Hadamard rotation tensors for DeepSeek lightning indexers- if ((model.arch == LLM_ARCH_DEEPSEEK32 || model.arch == LLM_ARCH_DEEPSEEK4 || model.arch == LLM_ARCH_GLM_DSA) &&+ if ((model.arch == LLM_ARCH_DEEPSEEK32 || model.arch == LLM_ARCH_DEEPSEEK4 ||+ model.arch == LLM_ARCH_GLM_DSA || model.arch == LLM_ARCH_DOTS3NOTE) &&hparams.n_embd_head_k_full == hparams.indexer_head_size) {attn_rot_k = true;}diff --git a/src/llama-model-saver.cpp b/src/llama-model-saver.cppindex b9e0a60094..0d39e6de89 100644--- a/src/llama-model-saver.cpp+++ b/src/llama-model-saver.cpp@@ -31,6 +31,7 @@ bool llama_model_saver_supports_arch(llm_arch arch) {case LLM_ARCH_MELLUM:case LLM_ARCH_LAGUNA:case LLM_ARCH_GRANITE_SWA:+ case LLM_ARCH_DOTS3NOTE: // TODO: need to handle SWA pattern and MLA+SWA configreturn false;default:return true;diff --git a/src/llama-model.cpp b/src/llama-model.cppindex d7874e0a92..de0d3c1a6f 100644--- a/src/llama-model.cpp+++ b/src/llama-model.cpp@@ -11,6 +11,7 @@#include "llama-kv-cache.h"#include "llama-kv-cache-iswa.h"#include "llama-kv-cache-dsa.h"+#include "llama-kv-cache-dsa-iswa.h"#include "llama-kv-cache-msa.h"#include "llama-kv-cache-dsv4.h"#include "llama-memory-hybrid.h"@@ -194,6 +195,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_paramsreturn new llama_model_deepseek2ocr(params);case LLM_ARCH_DEEPSEEK32:return new llama_model_deepseek32(params);+ case LLM_ARCH_DOTS3NOTE:+ return new llama_model_dots3note(params);case LLM_ARCH_DEEPSEEK4:return new llama_model_deepseek4(params);case LLM_ARCH_GLM_DSA:@@ -851,6 +854,7 @@ const char * llm_type_name(llm_type type) {case LLM_TYPE_230B_A10B: return "230B.A10B";case LLM_TYPE_428B_A23B: return "428B.A23B";case LLM_TYPE_235B_A22B: return "235B.A22B";+ case LLM_TYPE_288B_A19B: return "288B.A19B";case LLM_TYPE_300B_A47B: return "300B.A47B";case LLM_TYPE_310B_A15B: return "310B.A15B";case LLM_TYPE_355B_A32B: return "355B.A32B";@@ -1924,7 +1928,9 @@ void llama_model::print_info() const {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_DEEPSEEK32 || 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_DOTS3NOTE || 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);@@ -2193,6 +2199,57 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,nullptr);}} break;+ case LLM_ARCH_DOTS3NOTE:+ {+ GGML_ASSERT(hparams.swa_type != LLAMA_SWA_TYPE_NONE);++ if (params.ctx_type == LLAMA_CONTEXT_TYPE_MTP && hparams.n_layer_nextn > 0) {+ // MTP draft context: plain attention KV cache holding only the nextn layer+ llama_kv_cache::layer_filter_cb filter =+ [&](uint32_t il) { return il >= hparams.n_layer(); };++ res = new llama_kv_cache(+ *this,+ hparams,+ 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,+ filter,+ nullptr,+ nullptr);+ } else {+ // main context: DSA cache for the trunk full-attention layers plus a window-sized SWA cache+ llama_kv_cache::layer_filter_cb filter_mla = nullptr;+ if (hparams.n_layer_nextn > 0) {+ filter_mla = [&](uint32_t il) { return il < hparams.n_layer(); };+ }+ llama_kv_cache::layer_filter_cb filter_lid = [&](uint32_t il) { return il < hparams.n_layer() && hparams.is_indexer_full(il); };++ res = new llama_kv_cache_dsa_iswa(+ *this,+ params.type_k,+ params.type_v,+ !cparams.flash_attn,+ cparams.offload_kqv,+ params.swa_full,+ cparams.kv_unified,+ cparams.n_ctx_seq,+ cparams.n_seq_max,+ cparams.n_ubatch,+ 1,+ filter_mla,+ filter_lid,+ nullptr);+ }+ } break;case LLM_ARCH_DEEPSEEK4:{GGML_ASSERT(hparams.swa_type != LLAMA_SWA_TYPE_NONE);@@ -2661,6 +2718,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {case LLM_ARCH_LLAMA_EMBED:case LLM_ARCH_MAINCODER:case LLM_ARCH_GLM_DSA:+ case LLM_ARCH_DOTS3NOTE:case LLM_ARCH_NANBEIGE:case LLM_ARCH_POCKETTTS:return LLAMA_ROPE_TYPE_NORM;diff --git a/src/llama-model.h b/src/llama-model.hindex 4412ef08e7..44bd967575 100644--- a/src/llama-model.h+++ b/src/llama-model.h@@ -140,6 +140,7 @@ enum llm_type {LLM_TYPE_230B_A10B, // Minimax M2LLM_TYPE_428B_A23B, // Minimax M3LLM_TYPE_235B_A22B,+ LLM_TYPE_288B_A19B, // dots3-noteLLM_TYPE_300B_A47B, // Ernie MoE bigLLM_TYPE_310B_A15B, // /MiMo-V2-FlashLLM_TYPE_355B_A32B, // GLM-4.5diff --git a/src/models/dots3note.cpp b/src/models/dots3note.cppnew file mode 100644index 0000000000..00a008c2c9--- /dev/null+++ b/src/models/dots3note.cpp@@ -0,0 +1,480 @@+#include "models.h"++#include "llama-kv-cache.h"+#include "llama-kv-cache-dsa.h"++// note: code adapted from deepseek32.cpp (DSA indexer + absorbed MLA) and step35.cpp (head-wise output gate)++void llama_model_dots3note::load_arch_hparams(llama_model_loader & ml) {+ ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);+ hparams.f_norm_eps = 1e-6; // eps for the indexer k_norm layer norm++ // TODO: use MTP layer+ ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.n_layer_nextn, false);+ GGML_ASSERT(hparams.n_layer_nextn < hparams.n_layer_all && "n_layer_nextn must be < n_layer_all");++ // MoE parameters+ ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared);+ ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);+ ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead);+ 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);+ ml.get_key(LLM_KV_EXPERT_GATING_FUNC, hparams.expert_gating_func);++ // MLA parameters of the full-attention layers+ 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);+ ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla_impl);++ // MLA parameters of the sliding-window layers+ ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK_SWA, hparams.n_lora_kv_swa);+ ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA_SWA, hparams.n_embd_head_k_mla_swa);+ ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA_SWA, hparams.n_embd_head_v_mla_swa);++ hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;+ ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);+ ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa);+ ml.get_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl);++ // 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);+ ml.get_arr(LLM_KV_ATTENTION_INDEXER_TYPES, hparams.is_indexer_full_impl);++ switch (hparams.n_layer()) {+ case 46: type = LLM_TYPE_288B_A19B; break;+ default: type = LLM_TYPE_UNKNOWN;+ }+}++void llama_model_dots3note::load_arch_tensors(llama_model_loader & ml) {+ LLAMA_LOAD_LOCALS;+ GGML_UNUSED(ml);++ if (!hparams.is_mla()) {+ throw std::runtime_error("DOTS3NOTE architecture requires MLA");+ }++ const int64_t n_embd_head_qk_rope = hparams.n_rot();++ const int64_t q_lora_rank = hparams.n_lora_q;+ 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_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) {+ output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);+ }++ for (int i = 0; i < n_layer_all; ++i) {+ auto & layer = layers[i];++ const bool is_mtp = i >= n_layer;+ // the NextN/MTP block uses the sliding-attention geometry+ const bool is_swa = is_mtp || hparams.is_swa(i);++ // MTP tensors are preserved in the GGUF but there is no MTP graph yet+ const int flags = is_mtp ? TENSOR_SKIP | TENSOR_NOT_REQUIRED : 0;++ const int64_t n_head_l = hparams.n_head(i);++ const int64_t kv_lora_rank = is_swa ? hparams.n_lora_kv_swa : hparams.n_lora_kv;+ const int64_t n_embd_head_k_mla = is_swa ? hparams.n_embd_head_k_mla_swa : hparams.n_embd_head_k_mla();+ const int64_t n_embd_head_v_mla = is_swa ? hparams.n_embd_head_v_mla_swa : hparams.n_embd_head_v_mla();+ const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;++ 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);+ // norm applied on the shared rope key before rope+ layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_qk_rope}, 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_l * 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);++ layer.wk_b = create_tensor(tn(LLM_TENSOR_ATTN_K_B, "weight", i), {n_embd_head_qk_nope, kv_lora_rank, n_head_l}, flags);+ layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V_B, "weight", i), {kv_lora_rank, n_embd_head_v_mla, n_head_l}, flags);++ layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head_l * n_embd_head_v_mla, n_embd}, flags);++ // head-wise sigmoid output gate+ layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), {n_embd, n_head_l}, flags);++ layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags);++ // DSA indexer+ if (!is_mtp && hparams.is_indexer_full(i)) {+ 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 (is_mtp || 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 {+ if (n_expert == 0 || n_expert_used == 0) {+ throw std::runtime_error("n_expert and n_expert_used must be > 0");+ }++ 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}, flags);++ 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);++ 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);+ }++ if (is_mtp) {+ 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);+ layer.nextn.embed_tokens = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, flags);+ layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, flags);+ }+ }+}++std::unique_ptr<llm_graph_context> llama_model_dots3note::build_arch_graph(const llm_graph_params & params) const {+ return std::make_unique<graph>(*this, params);+}++llama_model_dots3note::graph::graph(const llama_model & model, const llm_graph_params & params) :+ llm_graph_context(params) {+ GGML_ASSERT(hparams.is_mla());++ const int64_t n_embd_head_qk_rope = hparams.n_rot();++ const int64_t n_indexer_head = hparams.indexer_n_head;+ const int64_t n_embd_indexer_head = hparams.indexer_head_size;+ const uint32_t n_indexer_top_k = hparams.indexer_top_k;++ // the indexer head layout is [rope | nope]+ GGML_ASSERT(hparams.n_rot() <= n_embd_indexer_head);++ ggml_tensor * cur;+ ggml_tensor * inpL;++ inpL = build_inp_embd(model.tok_embd);++ ggml_tensor * inp_pos = build_inp_pos();++ llm_graph_input_attn_k_dsa_iswa * inp_attn = build_attn_inp_k_dsa_iswa();++ ggml_tensor * inp_out_ids = build_inp_out_ids();++ for (int il = 0; il < n_layer; ++il) {+ ggml_tensor * inpSA = inpL;++ const bool is_swa = hparams.is_swa(il);++ const int64_t n_head_l = hparams.n_head(il);++ const int64_t kv_lora_rank = is_swa ? hparams.n_lora_kv_swa : hparams.n_lora_kv;+ const int64_t n_embd_head_k_mla = is_swa ? hparams.n_embd_head_k_mla_swa : hparams.n_embd_head_k_mla();+ const int64_t n_embd_head_v_mla = is_swa ? hparams.n_embd_head_v_mla_swa : hparams.n_embd_head_v_mla();+ const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;++ const float kq_scale = 1.0f/sqrtf(float(n_embd_head_k_mla));+ const float freq_base_l = model.get_rope_freq_base(cparams, il);++ // norm+ cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);+ cb(cur, "attn_norm", il);++ // self_attention+ {+ ggml_tensor * attn_inp = cur;++ 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 (full-attention layers only)+ if (!is_swa) {+ ggml_tensor * indexer_q = ggml_mul_mat(ctx0, model.layers[il].indexer_attn_q_b, qr);+ cb(indexer_q, "indexer_q", il);++ // {n_embd_indexer_head, n_indexer_head, n_tokens}+ indexer_q = ggml_reshape_3d(ctx0, indexer_q, n_embd_indexer_head, n_indexer_head, n_tokens);+ indexer_q = ggml_rope_ext(ctx0, indexer_q, 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, "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);++ // {n_embd_indexer_head, 1, n_tokens}+ indexer_k = ggml_reshape_3d(ctx0, indexer_k, n_embd_indexer_head, 1, n_tokens);+ indexer_k = ggml_rope_ext(ctx0, indexer_k, 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, "indexer_k", il);++ // perform Hadamard transform on indexer q and k+ indexer_q = ggml_mul_mat(ctx0, inp_attn->get_dsa()->self_k_rot_lid, indexer_q);+ cb(indexer_q, "indexer_q", il);+ indexer_k = ggml_mul_mat(ctx0, inp_attn->get_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->get_dsa()->mctx->get_lid();+ const auto & k_idxs_lid = inp_attn->get_dsa()->get_k_idxs_lid();+ ggml_build_forward_expand(gf, mctx_lid->cpy_k(ctx0, indexer_k, k_idxs_lid, il));++ ggml_tensor * indexer_weights = ggml_mul_mat(ctx0, model.layers[il].indexer_proj, cur);+ cb(indexer_weights, "indexer_weights", il);++ 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);++ // 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);++ ggml_tensor * indexer_score = nullptr;+ if (cparams.fused_lid) {+ indexer_score = ggml_lightning_indexer(ctx0, indexer_q, indexer_k, indexer_weights, inp_attn->get_dsa()->get_kq_mask_lid());+ cb(indexer_score, "indexer_score", il);+ res->add_fused_node({LLM_FUSED_OP_LIGHTNING_INDEXER, indexer_score, il});+ } else {+ 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);++ indexer_score = ggml_relu(ctx0, indexer_kq);+ cb(indexer_score, "indexer_score", il);++ 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);++ ggml_tensor * indexer_kq_mask = inp_attn->get_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_l, n_tokens}+ ggml_tensor * q_nope =+ ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head_l, n_tokens, ggml_row_size(q->type, n_embd_head_k_mla),+ ggml_row_size(q->type, n_embd_head_k_mla) * n_head_l, 0);+ cb(q_nope, "q_nope", il);++ // and {n_embd_head_qk_rope, n_head_l, n_tokens}+ ggml_tensor * q_pe = ggml_view_3d(+ ctx0, q, n_embd_head_qk_rope, n_head_l, n_tokens, ggml_row_size(q->type, n_embd_head_k_mla),+ ggml_row_size(q->type, n_embd_head_k_mla) * n_head_l, 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);++ // norm on the shared rope key, applied before rope+ k_pe = build_norm(k_pe, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il);+ 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_l, 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_l, 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 with the absorption optimization+ {+ // {n_embd_head_qk_nope, n_tokens, n_head_l}+ 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_l} x {n_embd_head_qk_nope, n_tokens, n_head_l}+ 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_l, 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_l, n_tokens}+ 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);++ // apply the head-wise output gate before o_proj, so wo stays out of build_attn+ if (is_swa) {+ cur = build_attn(inp_attn->get_swa(),+ nullptr, nullptr, nullptr,+ Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, kq_scale, il);+ } else {+ cur = build_attn(inp_attn->get_dsa(),+ nullptr, nullptr, nullptr,+ Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, top_k, kq_scale, il);+ }+ cb(cur, "attn_out", il);++ ggml_tensor * gate = build_lora_mm(model.layers[il].wqkv_gate, attn_inp);+ cb(gate, "attn_gate", il);++ gate = ggml_sigmoid(ctx0, gate);+ cb(gate, "attn_gate_sigmoid", il);++ // broadcast the per-head gate over the head dimension+ ggml_tensor * attn_3d = ggml_reshape_3d(ctx0, cur, n_embd_head_v_mla, n_head_l, n_tokens);+ ggml_tensor * gate_3d = ggml_reshape_3d(ctx0, gate, 1, n_head_l, n_tokens);+ attn_3d = ggml_mul(ctx0, attn_3d, gate_3d);+ cb(attn_3d, "attn_gated", il);++ cur = ggml_reshape_2d(ctx0, attn_3d, n_embd_head_v_mla * n_head_l, n_tokens);++ cur = build_lora_mm(model.layers[il].wo, cur, model.layers[il].wo_s);+ cb(cur, "attn_output", il);+ }+ }++ 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);+ }++ 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 {+ 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);++ 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);++ 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;++ cur = ggml_mul_mat(ctx0, model.output, cur);++ cb(cur, "result_output", -1);+ res->t_logits = cur;++ ggml_build_forward_expand(gf, cur);+}diff --git a/src/models/models.h b/src/models/models.hindex 1dd30dfd16..157b05dc05 100644--- a/src/models/models.h+++ b/src/models/models.h@@ -1156,6 +1156,18 @@ struct llama_model_deepseek32 : public llama_model_base {};+struct llama_model_dots3note : public llama_model_base {+ llama_model_dots3note(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_deepseek4 : public llama_model_base {llama_model_deepseek4(const struct llama_model_params & params) : llama_model_base(params) {}void load_arch_hparams(llama_model_loader & ml) override;diff --git a/tests/test-llama-archs.cpp b/tests/test-llama-archs.cppindex 4eb3763cbe..07e3a7a116 100644--- a/tests/test-llama-archs.cpp+++ b/tests/test-llama-archs.cpp@@ -104,6 +104,7 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {} else if (arch == LLM_ARCH_DEEPSEEK2|| arch == LLM_ARCH_DEEPSEEK32|| arch == LLM_ARCH_GLM_DSA+ || arch == LLM_ARCH_DOTS3NOTE|| arch == LLM_ARCH_KIMI_LINEAR|| arch == LLM_ARCH_BAILINGMOE3|| arch == LLM_ARCH_KIMI_K3@@ -166,6 +167,7 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {if (arch == LLM_ARCH_DEEPSEEK2|| arch == LLM_ARCH_DEEPSEEK32|| arch == LLM_ARCH_GLM_DSA+ || arch == LLM_ARCH_DOTS3NOTE|| arch == LLM_ARCH_KIMI_LINEAR|| arch == LLM_ARCH_BAILINGMOE3|| arch == LLM_ARCH_KIMI_K3@@ -175,6 +177,22 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {ms.add_kv(LLM_KV_ROPE_DIMENSION_COUNT, uint32_t(64));ms.add_kv(LLM_KV_ATTENTION_KEY_LENGTH_MLA, uint32_t(192));ms.add_kv(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, uint32_t(128));+ if (arch == LLM_ARCH_DOTS3NOTE) {+ // SWA layers reuse the same MLA geometry as the full layers in this fixture+ ms.add_kv(LLM_KV_ATTENTION_KV_LORA_RANK_SWA, uint32_t(512));+ ms.add_kv(LLM_KV_ATTENTION_KEY_LENGTH_SWA, uint32_t(576));+ ms.add_kv(LLM_KV_ATTENTION_VALUE_LENGTH_SWA, uint32_t(512));+ ms.add_kv(LLM_KV_ATTENTION_KEY_LENGTH_MLA_SWA, uint32_t(192));+ ms.add_kv(LLM_KV_ATTENTION_VALUE_LENGTH_MLA_SWA, uint32_t(128));+ ms.add_kv(LLM_KV_ROPE_FREQ_BASE_SWA, 10000.0f);+ // indexer on the full-attention layers (inverse of the swa pattern)+ std::vector<uint32_t> indexer_types;+ indexer_types.reserve(n_layer);+ for (uint32_t il = 0; il < n_layer; il++) {+ indexer_types.push_back(il % 2 ? 0 : 1);+ }+ ms.add_kv(LLM_KV_ATTENTION_INDEXER_TYPES, indexer_types);+ }} else if (arch == LLM_ARCH_MINIMAX_M3) {// partial rotary: n_rot must not exceed the indexer key length (64)ms.add_kv(LLM_KV_ROPE_DIMENSION_COUNT, uint32_t(64));@@ -197,7 +215,8 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {ms.add_kv(LLM_KV_ROPE_FREQ_BASE_SWA, 10000.0f);// SWA pattern: every 5th layer is full attention (matches E2B layer_types)ms.add_kv(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, uint32_t(5));- } else if (arch == LLM_ARCH_COHERE2MOE || arch == LLM_ARCH_MIMO2 || arch == LLM_ARCH_STEP35 || arch == LLM_ARCH_MUSE_GLIMMER || arch == LLM_ARCH_GRANITE_SWA) {+ } else if (arch == LLM_ARCH_COHERE2MOE || arch == LLM_ARCH_MIMO2 || arch == LLM_ARCH_STEP35 ||+ arch == LLM_ARCH_MUSE_GLIMMER || arch == LLM_ARCH_GRANITE_SWA || arch == LLM_ARCH_DOTS3NOTE) {std::vector<uint32_t> pattern;pattern.reserve(n_layer);for (uint32_t il = 0; il < n_layer; il++) {@@ -365,6 +384,7 @@ static bool moe_mandatory(const llm_arch arch) {case LLM_ARCH_DEEPSEEK:case LLM_ARCH_DEEPSEEK2:case LLM_ARCH_DEEPSEEK32:+ case LLM_ARCH_DOTS3NOTE:case LLM_ARCH_GLM4_MOE:case LLM_ARCH_GLM_DSA:case LLM_ARCH_EXAONE_MOE: