llama : add Gemma4 MTP

common/speculative.cpp

@@ -3,13 +3,14 @@
 #include "common.h"
 #include "ggml.h"
 #include "llama.h"
-#include "../src/llama-ext.h" // staging API: llama_set_embeddings_nextn / llama_get_embeddings_nextn_ith (used by MTP)
 #include "log.h"
 #include "ngram-cache.h"
 #include "ngram-map.h"
 #include "ngram-mod.h"
 #include "sampling.h"
 
+#include "../src/llama-ext.h" // staging API: llama_set_embeddings_nextn / llama_get_embeddings_nextn_ith (used by MTP)
+
 #include <algorithm>
 #include <cassert>
 #include <cstring>
@@ -418,6 +419,8 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
 
     int32_t n_embd = 0;
 
+    bool is_mem_shared = false;
+
     // Per-sequence cross-batch carryover: pair (h_p, x_{p+1}) at MTP pos p+1.
     // The last h-row of one process() call needs the first token of the NEXT
     // call to pair with, so it's stashed here until that next call fires.
@@ -444,7 +447,9 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
         auto * ctx_dft = this->params.ctx_dft;
         GGML_ASSERT(ctx_tgt && ctx_dft && "MTP requires ctx_tgt and ctx_dft to be set");
 
-        n_embd = llama_model_n_embd(llama_get_model(ctx_dft));
+        n_embd = llama_model_n_embd_out(llama_get_model(ctx_dft));
+        GGML_ASSERT(n_embd == llama_model_n_embd(llama_get_model(ctx_tgt)) &&
+                "MTP input row width must match the target h_nextn width");
 
         LOG_INF("%s: adding speculative implementation 'draft-mtp'\n", __func__);
         LOG_INF("%s: - n_max=%d, n_min=%d, p_min=%.2f, n_embd=%d, backend_sampling=%d\n", __func__, this->params.n_max, this->params.n_min, this->params.p_min, n_embd, (int) this->params.backend_sampling);
@@ -490,6 +495,8 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
         llama_set_embeddings_nextn(ctx_tgt, true, /*masked*/ false);
         llama_set_embeddings_nextn(ctx_dft, true, /*masked*/ true);
 
+        is_mem_shared = llama_get_ctx_src(ctx_dft) == ctx_tgt;
+
         pending_h.assign(n_seq, std::vector<float>(n_embd, 0.0f));
 
         i_batch_beg.assign(n_seq, -1);
@@ -526,9 +533,11 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
         if (N <= 0) {
             return;
         }
+
         auto * ctx_dft = this->params.ctx_dft;
         const llama_pos pos_max = llama_memory_seq_pos_max(llama_get_memory(ctx_dft), seq_id);
-        if (pos_max < N - 1) {
+
+        if (pos_max < N - 1 && !is_mem_shared) {
             LOG_WRN("%s: ctx_dft pos_max=%d < N-1=%d - "
                     "process() hook may not have run on every prefill ubatch "
                     "(need_embd / logits=1 on every prompt position?). "
@@ -571,48 +580,42 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
 
         const size_t row_bytes = (size_t) n_embd * sizeof(float);
 
-        common_batch_clear(batch);
+        // if kv is shared with target (e.g Gemma4), then we can skip this catch-up decode
+        if (!is_mem_shared) {
+            common_batch_clear(batch);
 
-        for (int k = 0; k < n_tokens; ++k) {
-            common_batch_add(batch, batch_in.token[k], batch_in.pos[k], { batch_in.seq_id[k][0] }, 0);
-        }
+            for (int k = 0; k < n_tokens; ++k) {
+                common_batch_add(batch, batch_in.token[k], batch_in.pos[k], { batch_in.seq_id[k][0] }, 0);
+            }
 
-        // shift the tgt embeddings to the right by one position
-        // assumes that the tokens in the batch are sequential for each sequence
-        // i.e. we cannot have seq_id like this: [0, 0, 0, 1, 1, 0, 1, 1]
-        //                                                       ^--- this is a problem
-        // TODO:this is generally true, but would be nice to assert it
-        {
-            const float * h_tgt = llama_get_embeddings_nextn(ctx_tgt);
-            std::memcpy(batch.embd + (size_t) 1 * n_embd, h_tgt, row_bytes * (n_tokens-1));
+            // shift the tgt embeddings to the right by one position
+            // assumes that the tokens in the batch are sequential for each sequence
+            // i.e. we cannot have seq_id like this: [0, 0, 0, 1, 1, 0, 1, 1]
+            //                                                       ^--- this is a problem
+            // TODO:this is generally true, but would be nice to assert it
+            {
+                const float * h_tgt = llama_get_embeddings_nextn(ctx_tgt);
+                std::memcpy(batch.embd + (size_t) 1 * n_embd, h_tgt, row_bytes * (n_tokens-1));
+            }
 
-            //{
-            //    // string with seq_ids in the batch
-            //    std::stringstream ss;
-            //    for (int i = 0; i < n_tokens; ++i) {
-            //        ss << batch_in.seq_id[i][0] << ",";
-            //    }
-            //    LOG_WRN("%s: batch_in.seq_id = %s\n", __func__, ss.str().c_str());
-            //}
-        }
+            // fill the pending embeddings from a previous run
+            auto set_h = [&](int idx, const float * h_row) {
+                std::memcpy(batch.embd + (size_t) idx * n_embd, h_row, row_bytes);
+            };
 
-        // fill the pending embeddings from a previous run
-        auto set_h = [&](int idx, const float * h_row) {
-            std::memcpy(batch.embd + (size_t) idx * n_embd, h_row, row_bytes);
-        };
+            for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
+                if (i_batch_beg[seq_id] < 0) {
+                    continue;
+                }
 
-        for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
-            if (i_batch_beg[seq_id] < 0) {
-                continue;
+                set_h(i_batch_beg[seq_id], pending_h[seq_id].data());
             }
 
-            set_h(i_batch_beg[seq_id], pending_h[seq_id].data());
-        }
-
-        const int32_t rc = llama_decode(ctx_dft, batch);
-        if (rc != 0) {
-            LOG_ERR("%s: llama_decode(ctx_dft) failed rc=%d (pos=%d)\n", __func__, (int) rc, (int) batch_in.pos[0]);
-            return false;
+            const int32_t rc = llama_decode(ctx_dft, batch);
+            if (rc != 0) {
+                LOG_ERR("%s: llama_decode(ctx_dft) failed rc=%d (pos=%d)\n", __func__, (int) rc, (int) batch_in.pos[0]);
+                return false;
+            }
         }
 
         for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
@@ -721,7 +724,11 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
                     continue;
                 }
 
-                common_batch_add(batch, id, dp.n_past + i + 1, { seq_id }, true);
+                if (is_mem_shared) {
+                    common_batch_add(batch, id, dp.n_past, { seq_id }, true);
+                } else {
+                    common_batch_add(batch, id, dp.n_past + i + 1, { seq_id }, true);
+                }
                 std::memcpy(batch.embd + n_embd*(batch.n_tokens - 1), h_row, row_bytes);
             }
 

conversion/__init__.py

@@ -75,9 +75,11 @@
     "Gemma3TextModel": "gemma",
     "Gemma3nForCausalLM": "gemma",
     "Gemma3nForConditionalGeneration": "gemma",
+    "Gemma4AssistantForCausalLM": "gemma",
     "Gemma4ForConditionalGeneration": "gemma",
     "Gemma4ForCausalLM": "gemma",
     "Gemma4UnifiedForConditionalGeneration": "gemma",
+    "Gemma4UnifiedAssistantForCausalLM": "gemma",
     "GemmaForCausalLM": "gemma",
     "Glm4ForCausalLM": "glm",
     "Glm4MoeForCausalLM": "glm",

conversion/gemma.py

@@ -785,6 +785,16 @@ def set_gguf_parameters(self):
             self.gguf_writer.add_suppress_tokens(suppress_tokens)
 
 
+@ModelBase.register("Gemma4AssistantForCausalLM", "Gemma4UnifiedAssistantForCausalLM")
+class Gemma4AssistantModel(Gemma4Model):
+    model_arch = gguf.MODEL_ARCH.GEMMA4_ASSISTANT
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_embedding_length_out(self.hparams["backbone_hidden_size"])
+        self.gguf_writer.add_nextn_predict_layers(self.block_count)
+
+
 @ModelBase.register("Gemma4ForConditionalGeneration")
 class Gemma4VisionAudioModel(MmprojModel):
     has_audio_encoder = True

gguf-py/gguf/constants.py

@@ -440,6 +440,7 @@ class MODEL_ARCH(IntEnum):
     GEMMA3           = auto()
     GEMMA3N          = auto()
     GEMMA4           = auto()
+    GEMMA4_ASSISTANT = auto()
     GEMMA_EMBEDDING  = auto()
     STARCODER2       = auto()
     RWKV6            = auto()
@@ -897,6 +898,8 @@ class MODEL_TENSOR(IntEnum):
     A_PER_DIM_K_SCALE     = auto() # gemma4
     A_PER_DIM_SCALE       = auto() # gemma4
     # nextn/mtp
+    NEXTN_PRE_PROJ       = auto()
+    NEXTN_POST_PROJ      = auto()
     NEXTN_EH_PROJ        = auto()
     NEXTN_EMBED_TOKENS   = auto()
     NEXTN_ENORM          = auto()
@@ -986,6 +989,7 @@ class MODEL_TENSOR(IntEnum):
     MODEL_ARCH.GEMMA3:           "gemma3",
     MODEL_ARCH.GEMMA3N:          "gemma3n",
     MODEL_ARCH.GEMMA4:           "gemma4",
+    MODEL_ARCH.GEMMA4_ASSISTANT: "gemma4-assistant",
     MODEL_ARCH.GEMMA_EMBEDDING:  "gemma-embedding",
     MODEL_ARCH.STARCODER2:       "starcoder2",
     MODEL_ARCH.RWKV6:            "rwkv6",
@@ -1471,6 +1475,8 @@ class MODEL_TENSOR(IntEnum):
     MODEL_TENSOR.A_QF_FFN_DOWN:             "a.proj_blk.{bid}.ffn_down",
     MODEL_TENSOR.A_QF_FFN_NORM:             "a.proj_blk.{bid}.ffn_norm",
     # NextN/MTP
+    MODEL_TENSOR.NEXTN_PRE_PROJ:            "nextn.pre_projection",
+    MODEL_TENSOR.NEXTN_POST_PROJ:           "nextn.post_projection",
     MODEL_TENSOR.NEXTN_EH_PROJ:             "blk.{bid}.nextn.eh_proj",
     MODEL_TENSOR.NEXTN_EMBED_TOKENS:        "blk.{bid}.nextn.embed_tokens",
     MODEL_TENSOR.NEXTN_ENORM:               "blk.{bid}.nextn.enorm",
@@ -2577,6 +2583,24 @@ class MODEL_TENSOR(IntEnum):
         MODEL_TENSOR.PER_LAYER_PROJ_NORM,
         MODEL_TENSOR.PER_LAYER_POST_NORM,
     ],
+    MODEL_ARCH.GEMMA4_ASSISTANT: [
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.NEXTN_PRE_PROJ,
+        MODEL_TENSOR.NEXTN_POST_PROJ,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_POST_NORM,
+        MODEL_TENSOR.FFN_PRE_NORM,
+        MODEL_TENSOR.FFN_POST_NORM,
+        MODEL_TENSOR.LAYER_OUT_SCALE,
+    ],
     MODEL_ARCH.GEMMA_EMBEDDING: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT,

gguf-py/gguf/tensor_mapping.py

@@ -2367,6 +2367,14 @@ class TensorNameMap:
         ),
 
         # NextN/MTP tensors
+        MODEL_TENSOR.NEXTN_PRE_PROJ: (
+            "pre_projection",
+        ),
+
+        MODEL_TENSOR.NEXTN_POST_PROJ: (
+            "post_projection",
+        ),
+
         MODEL_TENSOR.NEXTN_EH_PROJ: (
             "model.layers.{bid}.eh_proj",
         ),

include/llama.h

@@ -388,6 +388,8 @@ extern "C" {
         // note: the samplers must be sampler chains (i.e. use llama_sampler_chain_init)
         struct llama_sampler_seq_config * samplers;
         size_t                            n_samplers;
+
+        struct llama_context * ctx_src;
     };
 
     struct llama_model_tensor_override {

src/llama-arch.cpp

@@ -57,6 +57,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_GEMMA3,           "gemma3"           },
     { LLM_ARCH_GEMMA3N,          "gemma3n"          },
     { LLM_ARCH_GEMMA4,           "gemma4"           },
+    { LLM_ARCH_GEMMA4_ASSISTANT, "gemma4-assistant" },
     { LLM_ARCH_GEMMA_EMBEDDING,  "gemma-embedding"  },
     { LLM_ARCH_STARCODER2,       "starcoder2"       },
     { LLM_ARCH_MAMBA,            "mamba"            },
@@ -453,6 +454,8 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
     { LLM_TENSOR_FFN_NORM_EXPS,                          "blk.%d.ffn_norm_exps" },
     { LLM_TENSOR_ATTN_K_B,                               "blk.%d.attn_k_b" },
     { LLM_TENSOR_ATTN_V_B,                               "blk.%d.attn_v_b" },
+    { LLM_TENSOR_NEXTN_PRE_PROJ,                         "nextn.pre_projection" },
+    { LLM_TENSOR_NEXTN_POST_PROJ,                        "nextn.post_projection" },
     { LLM_TENSOR_NEXTN_EH_PROJ,                          "blk.%d.nextn.eh_proj" },
     { LLM_TENSOR_NEXTN_EMBED_TOKENS,                     "blk.%d.nextn.embed_tokens" },
     { LLM_TENSOR_NEXTN_ENORM,                            "blk.%d.nextn.enorm" },
@@ -765,6 +768,8 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_INDEXER_PROJ,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_INDEXER_ATTN_K,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_INDEXER_ATTN_Q_B,           {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_NEXTN_PRE_PROJ,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_NEXTN_POST_PROJ,            {LLM_TENSOR_LAYER_OUTPUT,    GGML_OP_MUL_MAT}},
     // NextN/MTP tensors are stored per-block (blk.%d.nextn.*) even though only the
     // last nextn_predict_layers blocks carry them. Classify as LAYER_REPEATING so
     // the model loader doesn't fault on the block index.

src/llama-arch.h

@@ -61,6 +61,7 @@ enum llm_arch {
     LLM_ARCH_GEMMA3,
     LLM_ARCH_GEMMA3N,
     LLM_ARCH_GEMMA4,
+    LLM_ARCH_GEMMA4_ASSISTANT,
     LLM_ARCH_GEMMA_EMBEDDING,
     LLM_ARCH_STARCODER2,
     LLM_ARCH_MAMBA,
@@ -557,6 +558,8 @@ enum llm_tensor {
     LLM_TENSOR_INDEXER_PROJ,
     LLM_TENSOR_INDEXER_ATTN_K,
     LLM_TENSOR_INDEXER_ATTN_Q_B,
+    LLM_TENSOR_NEXTN_PRE_PROJ,  // TODO: rename to PROJ_PRE
+    LLM_TENSOR_NEXTN_POST_PROJ, // TODO: rename to PROJ_POST
     LLM_TENSOR_NEXTN_EH_PROJ,
     LLM_TENSOR_NEXTN_EMBED_TOKENS,
     LLM_TENSOR_NEXTN_ENORM,