Skip to content

LLM inference in C/C++

C++ ggml

model : support MTP in GLM-4.5-Air (#26534)

jacekpoplawskic060ca974c77
diff --git a/conversion/glm.py b/conversion/glm.py
index abd7f279f9..23fdbca881 100644
--- a/conversion/glm.py
+++ b/conversion/glm.py
@@ -112,12 +112,36 @@ class GlmOCRModel(Glm4Model):
@ModelBase.example("zai-org/GLM-4.5-Air")
class Glm4MoeModel(TextModel):
model_arch = gguf.MODEL_ARCH.GLM4_MOE
+ supports_mtp_export = True
+ _n_main_layers: int | None = None
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
- # GLM4_MOE has num_hidden_layers + 1 actual layers (including NextN layer)
- self.block_count = self.hparams["num_hidden_layers"] + self.hparams.get("num_nextn_predict_layers", 0)
- self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+ 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
+
+ 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
+
+ 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_vocab(self):
return self._set_vocab_glm()
@@ -153,10 +177,22 @@ class Glm4MoeModel(TextModel):
if (norm_topk_prob := self.hparams.get("norm_topk_prob")) is not None:
self.gguf_writer.add_expert_weights_norm(norm_topk_prob)
- # 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)
+ 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"
+
_experts: list[dict[str, Tensor]] | None = None
# note: unlike GLM4V non-MoE, we don't need to permute Q/K here since GLM4V_MOE uses Neox ordering already
@@ -348,6 +384,7 @@ class GlmMoeDsaModel(DeepseekV2Model):
@ModelBase.example("upstage/Solar-Open-100B")
class SolarOpenModel(Glm4MoeModel):
model_arch = gguf.MODEL_ARCH.GLM4_MOE
+ supports_mtp_export = False
def set_vocab(self):
from transformers import AutoTokenizer
diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py
index 8f6f55519c..f236a5d2c9 100644
--- a/gguf-py/gguf/constants.py
+++ b/gguf-py/gguf/constants.py
@@ -3822,7 +3822,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_DOWN_SHEXP,
MODEL_TENSOR.FFN_UP_SHEXP,
MODEL_TENSOR.FFN_EXP_PROBS_B,
- # NextN/MTP tensors - preserved but unused
+ # NextN/MTP tensors
MODEL_TENSOR.NEXTN_EH_PROJ,
MODEL_TENSOR.NEXTN_EMBED_TOKENS,
MODEL_TENSOR.NEXTN_ENORM,
diff --git a/src/models/glm4-moe.cpp b/src/models/glm4-moe.cpp
index 8cde66978c..83ea7f8ac6 100644
--- a/src/models/glm4-moe.cpp
+++ b/src/models/glm4-moe.cpp
@@ -29,10 +29,19 @@ void llama_model_glm4_moe::load_arch_hparams(llama_model_loader & ml) {
}
}
-void llama_model_glm4_moe::load_arch_tensors(llama_model_loader &) {
+void llama_model_glm4_moe::load_arch_tensors(llama_model_loader & ml) {
LLAMA_LOAD_LOCALS;
const int64_t n_expert_shared = hparams.n_expert_shared;
+ 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;
+ }
GGML_ASSERT(hparams.n_expert > 0 && "n_expert must be > 0 for GLM4_MOE MoE layers");
GGML_ASSERT(hparams.n_expert_used > 0 && "n_expert_used must be > 0 for GLM4_MOE MoE layers");
@@ -47,16 +56,9 @@ void llama_model_glm4_moe::load_arch_tensors(llama_model_loader &) {
output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, TENSOR_DUPLICATED);
}
- // Load ALL tensors including NextN layer to satisfy total tensor count
- // but only PROCESS up to last layer (skipping final NextN layer) in forward pass
for (int i = 0; i < n_layer_all; ++i) {
- int flags = 0;
- if (i >= n_layer) {
- // skip all tensors in the NextN layers
- flags |= TENSOR_SKIP;
- }
-
auto & layer = layers[i];
+ const int flags = i < n_layer ? trunk_flags : mtp_flags;
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, flags);
@@ -110,24 +112,186 @@ void llama_model_glm4_moe::load_arch_tensors(llama_model_loader &) {
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), { n_embd, n_ff }, flags);
}
- // NextN/MTP tensors (preserved but unused) - conditionally load for last nextn_predict_layers
+ // NextN/MTP tensors
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);
layer.nextn.hnorm = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), { n_embd }, flags);
// Optional tensors
- layer.nextn.embed_tokens = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED);
- layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED);
- layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, flags | TENSOR_NOT_REQUIRED);
+ layer.nextn.embed_tokens = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED | flags);
+ layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED | flags);
+ layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, TENSOR_NOT_REQUIRED | flags);
}
}
}
std::unique_ptr<llm_graph_context> llama_model_glm4_moe::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);
}
+llama_model_glm4_moe::graph_mtp::graph_mtp(const llama_model & model, const llm_graph_params & params)
+ : llm_graph_context(params) {
+ GGML_ASSERT(hparams.n_layer_nextn > 0 && "GLM4_MOE MTP requires n_layer_nextn > 0");
+ GGML_ASSERT(hparams.n_layer_nextn == 1 && "GLM4_MOE MTP currently only supports a single MTP block");
+
+ const int64_t n_embd_head = hparams.n_embd_head_v();
+ GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+
+ 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");
+
+ 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();
+
+ auto * inp_attn = build_attn_inp_kv();
+
+ 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, 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);
+
+ auto [Qcur, Kcur, Vcur] = build_qkv(layer, cur,
+ n_embd_head, n_head, n_head_kv, il);
+
+ if (layer.attn_q_norm) {
+ Qcur = build_norm(Qcur, layer.attn_q_norm, nullptr, LLM_NORM_RMS, il);
+ cb(Qcur, "mtp_Qcur_normed", il);
+ }
+ if (layer.attn_k_norm) {
+ Kcur = build_norm(Kcur, layer.attn_k_norm, nullptr, LLM_NORM_RMS, il);
+ cb(Kcur, "mtp_Kcur_normed", il);
+ }
+
+ Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot,
+ rope_type, n_ctx_orig, freq_base, freq_scale,
+ ext_factor, attn_factor, beta_fast, beta_slow);
+
+ Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot,
+ rope_type, n_ctx_orig, freq_base, freq_scale,
+ ext_factor, attn_factor, beta_fast, beta_slow);
+
+ cb(Qcur, "mtp_Qcur", il);
+ cb(Kcur, "mtp_Kcur", il);
+ cb(Vcur, "mtp_Vcur", il);
+
+ cur = build_attn(inp_attn,
+ layer.wo, nullptr, layer.wo_s,
+ Qcur, Kcur, Vcur, nullptr, nullptr, nullptr,
+ 1.0f / sqrtf(float(n_embd_head)), 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.attn_post_norm, nullptr, LLM_NORM_RMS, il);
+ cb(cur, "mtp_post_attn_norm", il);
+
+ ggml_tensor * routed_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);
+ cb(routed_out, "mtp_ffn_moe_out", il);
+
+ ggml_tensor * shared_out = build_ffn(cur,
+ layer.ffn_up_shexp, nullptr, nullptr,
+ layer.ffn_gate_shexp, nullptr, nullptr,
+ layer.ffn_down_shexp, nullptr, nullptr,
+ nullptr,
+ LLM_FFN_SILU, LLM_FFN_PAR, il);
+ cb(shared_out, "mtp_ffn_shexp_out", il);
+
+ cur = ggml_add(ctx0, routed_out, shared_out);
+ cb(cur, "mtp_ffn_out", il);
+
+ cur = ggml_add(ctx0, cur, ffn_inp);
+ cb(cur, "mtp_post_ffn", il);
+
+ ggml_tensor * head_norm_w = layer.nextn.shared_head_norm
+ ? layer.nextn.shared_head_norm
+ : model.output_norm;
+ GGML_ASSERT(head_norm_w && "GLM4_MOE 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;
+
+ if (inp_out_ids) {
+ 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 && "GLM4_MOE 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);
+}
+
llama_model_glm4_moe::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v();
@@ -154,8 +318,7 @@ llama_model_glm4_moe::graph::graph(const llama_model & model, const llm_graph_pa
ggml_tensor * inp_out_ids = build_inp_out_ids();
- // Only process up to last layer (skip final NextN layer)
- // Final layer tensors are loaded but not processed in forward pass
+ // NextN layers are processed by graph_mtp.
for (int il = 0; il < n_layer; ++il) {
ggml_tensor * inpSA = inpL;
@@ -205,7 +368,7 @@ llama_model_glm4_moe::graph::graph(const llama_model & model, const llm_graph_pa
model.layers[il].wo, NULL, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
- if (il == n_layer - 1 && inp_out_ids) {
+ 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);
}
@@ -265,6 +428,13 @@ llama_model_glm4_moe::graph::graph(const llama_model & model, const llm_graph_pa
cur = inpL;
cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+ 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;
diff --git a/src/models/models.h b/src/models/models.h
index 157b05dc05..969429e3b6 100644
--- a/src/models/models.h
+++ b/src/models/models.h
@@ -1412,6 +1412,10 @@ struct llama_model_glm4_moe : 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;
};