llama : MTP support for DeepSeek V3.2 (#26457)

* llama : MTP support for DeepSeek V3.2

* model : no need to include MTP layers during DeepSeek V3.2 model type discovery

---------

Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>

conversion/deepseek.py

@@ -447,12 +447,43 @@ class DeepseekV2Model(TextModel):
 class DeepseekV32Model(DeepseekV2Model):
     model_arch = gguf.MODEL_ARCH.DEEPSEEK32
     skip_mtp = False
+    supports_mtp_export = True
+    _n_main_layers: int | None = None
 
     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.block_count = self.hparams["num_hidden_layers"]
+        if not self.no_mtp:
+            self.block_count += self.hparams.get("num_nextn_predict_layers", 0)
         self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
 
+    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)
+
+    @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
+
+        # DeepSeek V3.2 appends the NextN/MTP block past num_hidden_layers
+        # (model.layers.61 -> blk.61 in the 62-block file).
+        assert cls._n_main_layers is not None
+        is_mtp = (m := re.match(r"model\.layers\.(\d+)\.", name)) is not None and int(m.group(1)) >= cls._n_main_layers
+
+        # --no-mtp: drop the appended NextN block entirely.
+        if is_mtp and cls.no_mtp:
+            return None
+        # --mtp: keep ONLY NextN-block tensors plus the shared embeddings/
+        # norm/lm_head (so the resulting GGUF carries just the draft head).
+        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_vocab(self):
         from transformers import AutoTokenizer
         tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
@@ -463,7 +494,7 @@ class DeepseekV32Model(DeepseekV2Model):
         super().set_gguf_parameters()
 
         # NextN/MTP prediction layers
-        if (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
+        if not self.no_mtp and (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

src/llama-model.cpp

@@ -2071,24 +2071,8 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
             {
                 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;
         case LLM_ARCH_GLM_DSA:
+        case LLM_ARCH_DEEPSEEK32:
             {
                 if (params.ctx_type == LLAMA_CONTEXT_TYPE_MTP && hparams.n_layer_nextn > 0) {
                     // The NextN/MTP draft head runs dense MLA (no DSA indexer), so the
@@ -2313,7 +2297,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                     }
 
                     if ((arch == LLM_ARCH_STEP35 || arch == LLM_ARCH_HY_V3 || arch == LLM_ARCH_GLM_DSA ||
-                            arch == LLM_ARCH_MIMO2) &&
+                            arch == LLM_ARCH_MIMO2 || arch == LLM_ARCH_DEEPSEEK32) &&
                             hparams.n_layer_nextn > 0) {
                         if (params.ctx_type == LLAMA_CONTEXT_TYPE_MTP) {
                             filter = [&](uint32_t il) { return il >= hparams.n_layer(); };

src/models/deepseek32.cpp

@@ -44,13 +44,24 @@ void llama_model_deepseek32::load_arch_hparams(llama_model_loader & ml) {
     GGML_ASSERT(hparams.n_layer_nextn < hparams.n_layer_all && "n_layer_nextn must be < n_layer");
 
     switch (hparams.n_layer()) {
-        case 62: type = LLM_TYPE_685B_A37B; break;
+        case 61: type = LLM_TYPE_685B_A37B; break;
         default: type = LLM_TYPE_UNKNOWN;
     }
 }
 
-void llama_model_deepseek32::load_arch_tensors(llama_model_loader &) {
+void llama_model_deepseek32::load_arch_tensors(llama_model_loader & ml) {
     LLAMA_LOAD_LOCALS;
+
+    const bool mtp_only = (hparams.n_layer_nextn > 0) && (ml.get_weight("blk.0.attn_norm.weight") == nullptr);
+    const std::string mtp_probe = "blk." + std::to_string(n_layer) + ".nextn.eh_proj.weight";
+    const bool trunk_only = (hparams.n_layer_nextn > 0) && (ml.get_weight(mtp_probe.c_str()) == nullptr);
+    const int trunk_flags = mtp_only   ? TENSOR_NOT_REQUIRED : 0;
+    int       mtp_flags   = trunk_only ? TENSOR_NOT_REQUIRED : 0;
+
+    if (!ml.load_mtp) {
+        mtp_flags |= TENSOR_SKIP;
+    }
+
     const bool is_mla = hparams.is_mla();
     if (!is_mla) {
         throw std::runtime_error("DEEPSEEK32 architecture requires MLA");
@@ -80,12 +91,7 @@ void llama_model_deepseek32::load_arch_tensors(llama_model_loader &) {
     }
 
     for (int i = 0; i < n_layer_all; ++i) {
-        int flags = 0;
+        const int flags = (i >= n_layer) ? mtp_flags : trunk_flags;
-        if (i >= n_layer) {
-            // 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];
 
@@ -138,7 +144,7 @@ void llama_model_deepseek32::load_arch_tensors(llama_model_loader &) {
             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
+        // NextN/MTP tensors - conditionally load for last nextn_predict_layers
         if (i >= n_layer) {
             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);
@@ -153,6 +159,9 @@ void llama_model_deepseek32::load_arch_tensors(llama_model_loader &) {
 }
 
 std::unique_ptr<llm_graph_context> llama_model_deepseek32::build_arch_graph(const llm_graph_params & params) const {
+    if (params.gtype == LLM_GRAPH_TYPE_DECODER_MTP) {
+        return std::make_unique<graph_mtp>(*this, params);
+    }
     return std::make_unique<graph>(*this, params);
 }
 
@@ -430,7 +439,9 @@ llama_model_deepseek32::graph::graph(const llama_model & model, const llm_graph_
                         Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, top_k, kq_scale, il);
             }
         }
-        if (il == n_layer - 1 && inp_out_ids) {
+        // when unmasked nextn embeddings are requested, t_h_nextn must keep all rows,
+        // so the early output masking has to be skipped (it is applied after the final norm instead)
+        if (il == n_layer - 1 && inp_out_ids && (!cparams.embeddings_nextn || cparams.embeddings_nextn_masked)) {
             cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
             inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
         }
@@ -493,6 +504,14 @@ llama_model_deepseek32::graph::graph(const llama_model & model, const llm_graph_
 
     cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
 
+    // post-norm hidden state feeds the NextN/MTP draft head
+    cb(cur, "h_nextn", -1);
+    res->t_h_nextn = cur;
+
+    if (cparams.embeddings_nextn && !cparams.embeddings_nextn_masked && inp_out_ids) {
+        cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+    }
+
     cb(cur, "result_norm", -1);
     res->t_embd = cur;
 
@@ -504,3 +523,243 @@ llama_model_deepseek32::graph::graph(const llama_model & model, const llm_graph_
 
     ggml_build_forward_expand(gf, cur);
 }
+
+// LLM_GRAPH_TYPE_DECODER_MTP draft head for DeepSeek V3.2 (DEEPSEEK32).
+// Semantics mirror the deepseek-family NextN/MTP layer:
+//   enorm(embed) + hnorm(prev_hidden) -> concat(e, h) -> eh_proj ->
+//   full deepseek32 decoder block (dense MLA attention + sigmoid-gated MoE FFN
+//   with shared expert, exactly as the trunk deepseek2 graph builds it) ->
+//   shared_head_norm (fallback output_norm) -> shared LM head.
+// The DSA indexer is not used at runtime.
+llama_model_deepseek32::graph_mtp::graph_mtp(const llama_model & model, const llm_graph_params & params)
+    : llm_graph_context(params) {
+    GGML_ASSERT(hparams.n_layer_nextn > 0 && "DEEPSEEK32 MTP requires n_layer_nextn > 0");
+    GGML_ASSERT(hparams.n_layer_nextn == 1 && "DEEPSEEK32 MTP currently only supports a single MTP block");
+    GGML_ASSERT(hparams.is_mla() && "DEEPSEEK32 MTP requires MLA");
+
+    const int il = hparams.n_layer() + cparams.nextn_layer_offset;
+    GGML_ASSERT(cparams.nextn_layer_offset >= 0 &&
+                cparams.nextn_layer_offset < (int) hparams.n_layer_nextn &&
+                "nextn_layer_offset out of range [0, n_layer_nextn)");
+    const auto & layer = model.layers[il];
+
+    GGML_ASSERT(layer.nextn.eh_proj && "MTP block missing nextn.eh_proj");
+    GGML_ASSERT(layer.nextn.enorm   && "MTP block missing nextn.enorm");
+    GGML_ASSERT(layer.nextn.hnorm   && "MTP block missing nextn.hnorm");
+    GGML_ASSERT(layer.ffn_gate_inp  && "MTP block missing ffn_gate_inp");
+
+    // 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_qk_rope = hparams.n_rot();
+    const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope;
+
+    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 the deepseek2 trunk graph for the detailed explanation - this must match it EXACTLY.
+    GGML_ASSERT(ext_factor >= 0.0f);
+    const float attn_factor_org = attn_factor * (1.0f + 0.1f * logf(1.0f / freq_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));
+
+    // TODO: extract in a common llm_graph_context::build_inp_embd_h()
+    auto inp = std::make_unique<llm_graph_input_embd_h>(hparams.n_embd);
+
+    inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+    ggml_set_input(inp->tokens);
+
+    inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd_inp(), n_tokens);
+    ggml_set_input(inp->embd);
+
+    ggml_tensor * tok_embd;
+    if (ubatch.token) {
+        ggml_tensor * tok_embd_w = layer.nextn.embed_tokens ? layer.nextn.embed_tokens : model.tok_embd;
+
+        tok_embd = ggml_get_rows(ctx0, tok_embd_w, inp->tokens);
+    } else {
+        tok_embd = inp->embd;
+    }
+    cb(tok_embd, "mtp_tok_embd", il);
+
+    inp->h = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd, n_tokens);
+    ggml_set_input(inp->h);
+    ggml_set_name(inp->h, "mtp_h_input");
+
+    ggml_tensor * h_embd = inp->h;
+
+    res->add_input(std::move(inp));
+
+    ggml_tensor * inp_pos     = build_inp_pos();
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    // MLA with the absorption optimization uses a K-only cache (V is a view of K)
+    auto * inp_attn = build_attn_inp_k();
+
+    ggml_tensor * h_norm = build_norm(h_embd, layer.nextn.hnorm, nullptr, LLM_NORM_RMS, il);
+    cb(h_norm, "mtp_hnorm", il);
+
+    ggml_tensor * e_norm = build_norm(tok_embd, layer.nextn.enorm, nullptr, LLM_NORM_RMS, il);
+    cb(e_norm, "mtp_enorm", il);
+
+    ggml_tensor * concat = ggml_concat(ctx0, e_norm, h_norm, /*dim=*/ 0);
+    cb(concat, "mtp_concat", il);
+
+    ggml_tensor * cur = build_lora_mm(layer.nextn.eh_proj, concat, layer.nextn.eh_proj_s);
+    cb(cur, "mtp_eh_proj", il);
+
+    ggml_tensor * inpSA = cur;
+
+    cur = build_norm(cur, layer.attn_norm, nullptr, LLM_NORM_RMS, il);
+    cb(cur, "mtp_attn_norm", il);
+
+    // self-attention: dense MLA, same construction as the deepseek2 trunk graph
+    {
+        ggml_tensor * q = ggml_mul_mat(ctx0, layer.wq_a, cur);
+        cb(q, "mtp_q", il);
+
+        q = build_norm(q, layer.attn_q_a_norm, nullptr, LLM_NORM_RMS, il);
+        cb(q, "mtp_q", il);
+
+        q = ggml_mul_mat(ctx0, layer.wq_b, q);
+        cb(q, "mtp_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, "mtp_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, "mtp_q_pe", il);
+
+        ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, layer.wkv_a_mqa, cur);
+        cb(kv_cmpr_pe, "mtp_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, "mtp_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, "mtp_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, "mtp_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, "mtp_k_pe", il);
+
+        kv_cmpr = build_norm(kv_cmpr, layer.attn_kv_a_norm, nullptr, LLM_NORM_RMS, il);
+        cb(kv_cmpr, "mtp_kv_cmpr", il);
+
+        // {n_embd_head_qk_nope, n_tokens, n_head}
+        q_nope = ggml_permute(ctx0, q_nope, 0, 2, 1, 3);
+        cb(q_nope, "mtp_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, layer.wk_b, q_nope);
+        cb(q_nope_absorbed, "mtp_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, "mtp_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, "mtp_Qcur", il);
+
+        kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens);
+        cb(kv_cmpr, "mtp_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, "mtp_Kcur", il);
+
+        // {kv_lora_rank, 1, n_tokens}
+        ggml_tensor * Vcur = kv_cmpr;
+        cb(Vcur, "mtp_Vcur", il);
+
+        // note: MLA with the absorption optimization converts into MQA (ie: GQA with 1 group)
+        cur = build_attn(inp_attn,
+                layer.wo, NULL, layer.wo_s,
+                Qcur, Kcur, Vcur, nullptr, nullptr, layer.wv_b, kq_scale, il);
+        cb(cur, "mtp_attn_out", il);
+    }
+
+    ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+    cb(ffn_inp, "mtp_ffn_inp", il);
+
+    cur = build_norm(ffn_inp, layer.ffn_norm, NULL, LLM_NORM_RMS, il);
+    cb(cur, "mtp_ffn_norm", il);
+
+    // MoE FFN with shared expert - same construction as the deepseek2 trunk graph
+    ggml_tensor * moe_out = build_moe_ffn(cur,
+        layer.ffn_gate_inp,
+        layer.ffn_up_exps,
+        layer.ffn_gate_exps,
+        layer.ffn_down_exps,
+        layer.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,
+        layer.ffn_gate_up_exps,
+        layer.ffn_up_exps_s,
+        layer.ffn_gate_exps_s,
+        layer.ffn_down_exps_s);
+    cb(moe_out, "mtp_ffn_moe_out", il);
+
+    // FFN shared expert
+    ggml_tensor * ffn_shexp =
+        build_ffn(cur,
+            layer.ffn_up_shexp, NULL, layer.ffn_up_shexp_s,
+            layer.ffn_gate_shexp, NULL, layer.ffn_gate_shexp_s,
+            layer.ffn_down_shexp, NULL, layer.ffn_down_shexp_s,
+            NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+    cb(ffn_shexp, "mtp_ffn_shexp", il);
+
+    cur = ggml_add(ctx0, moe_out, ffn_shexp);
+    cb(cur, "mtp_ffn_out", il);
+
+    cur = ggml_add(ctx0, cur, ffn_inp);
+    cb(cur, "mtp_post_ffn", il);
+
+    // shared_head_norm applied after the decoder block, before the shared LM head.
+    // The post-norm hidden state seeds the next MTP step.
+    ggml_tensor * head_norm_w = layer.nextn.shared_head_norm
+            ? layer.nextn.shared_head_norm
+            : model.output_norm;
+    GGML_ASSERT(head_norm_w && "DEEPSEEK32 MTP: missing both nextn.shared_head_norm and output_norm");
+    cur = build_norm(cur, head_norm_w, nullptr, LLM_NORM_RMS, -1);
+
+    cb(cur, "h_nextn", -1);
+    res->t_h_nextn = cur;
+
+    cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+    cb(cur, "mtp_shared_head_norm", -1);
+
+    ggml_tensor * head_w = layer.nextn.shared_head_head ? layer.nextn.shared_head_head : model.output;
+    ggml_tensor * head_s = layer.nextn.shared_head_head ? layer.nextn.shared_head_head_s : model.output_s;
+    GGML_ASSERT(head_w && "DEEPSEEK32 MTP: missing LM head (nextn.shared_head_head or model.output)");
+    cur = build_lora_mm(head_w, cur, head_s);
+    cb(cur, "result_output", -1);
+
+    res->t_logits = cur;
+    ggml_build_forward_expand(gf, cur);
+}
+

src/models/models.h

@@ -1097,6 +1097,10 @@ struct llama_model_deepseek32 : public llama_model_base {
         graph(const llama_model & model, const llm_graph_params & params);
     };
 
+    struct graph_mtp : public llm_graph_context {
+        graph_mtp(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;
 };