Skip to content

LLM inference in C/C++

C++ ggml

model: Add support for Nanbeige4.2 (#25994)

zqlb77d646751d0
diff --git a/conversion/__init__.py b/conversion/__init__.py
index b2bb7e5161..45c001b78f 100644
--- a/conversion/__init__.py
+++ b/conversion/__init__.py
@@ -167,6 +167,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
"ModernBertForMaskedLM": "bert",
"ModernBertForSequenceClassification": "bert",
"ModernBertModel": "bert",
+ "NanbeigeForCausalLM": "nanbeige",
"NemotronForCausalLM": "nemotron",
"NemotronHForCausalLM": "nemotron",
"NeoBERT": "bert",
diff --git a/conversion/nanbeige.py b/conversion/nanbeige.py
new file mode 100644
index 0000000000..f1fc425b3a
--- /dev/null
+++ b/conversion/nanbeige.py
@@ -0,0 +1,24 @@
+from __future__ import annotations
+
+from .base import ModelBase, gguf, logger
+from .llama import LlamaModel
+
+
[email protected]("NanbeigeForCausalLM")
+class NanbeigeModel(LlamaModel):
+ model_arch = gguf.MODEL_ARCH.NANBEIGE
+ undo_permute = True
+
+ def set_gguf_parameters(self):
+ super().set_gguf_parameters()
+ hparams = self.hparams
+
+ n_loops = int(hparams.get("num_loops", 1) or 1)
+ if n_loops < 1:
+ n_loops = 1
+ self.gguf_writer.add_num_loops(n_loops)
+ logger.info(f"gguf: num_loops = {n_loops}")
+
+ skip_loop_final_norm = bool(hparams.get("skip_loop_final_norm", False))
+ self.gguf_writer.add_skip_loop_final_norm(skip_loop_final_norm)
+ logger.info(f"gguf: skip_loop_final_norm = {skip_loop_final_norm}")
diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py
index 2071e3eaa8..78e3c29a0d 100644
--- a/gguf-py/gguf/constants.py
+++ b/gguf-py/gguf/constants.py
@@ -145,6 +145,8 @@ class Keys:
TOKEN_SHIFT_COUNT = "{arch}.token_shift_count"
INTERLEAVE_MOE_LAYER_STEP = "{arch}.interleave_moe_layer_step"
FULL_ATTENTION_INTERVAL = "{arch}.full_attention_interval"
+ NUM_LOOPS = "{arch}.num_loops"
+ SKIP_LOOP_FINAL_NORM = "{arch}.skip_loop_final_norm"
HASH_LAYER_COUNT = "{arch}.hash_layer_count"
ACTIVATION_SPARSITY_SCALE = "{arch}.activation_sparsity_scale"
ALTUP_ACTIVE_IDX = "{arch}.altup.active_idx"
@@ -545,6 +547,7 @@ class MODEL_ARCH(IntEnum):
KIMI_LINEAR = auto()
TALKIE = auto()
MELLUM = auto()
+ NANBEIGE = auto()
class VISION_PROJECTOR_TYPE(IntEnum):
@@ -1134,6 +1137,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.KIMI_LINEAR: "kimi-linear",
MODEL_ARCH.TALKIE: "talkie",
MODEL_ARCH.MELLUM: "mellum",
+ MODEL_ARCH.NANBEIGE: "nanbeige",
}
VISION_PROJECTOR_TYPE_NAMES: dict[VISION_PROJECTOR_TYPE, str] = {
@@ -4505,7 +4509,22 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_DOWN_EXP,
MODEL_TENSOR.FFN_UP_EXP,
],
- # TODO
+ MODEL_ARCH.NANBEIGE: [
+ MODEL_TENSOR.TOKEN_EMBD,
+ MODEL_TENSOR.OUTPUT_NORM,
+ MODEL_TENSOR.OUTPUT,
+ MODEL_TENSOR.ROPE_FREQS,
+ MODEL_TENSOR.ATTN_NORM,
+ MODEL_TENSOR.ATTN_Q,
+ MODEL_TENSOR.ATTN_K,
+ MODEL_TENSOR.ATTN_V,
+ MODEL_TENSOR.ATTN_OUT,
+ MODEL_TENSOR.ATTN_ROT_EMBD,
+ MODEL_TENSOR.FFN_NORM,
+ MODEL_TENSOR.FFN_GATE,
+ MODEL_TENSOR.FFN_DOWN,
+ MODEL_TENSOR.FFN_UP,
+ ],
}
# tensors that will not be serialized
@@ -4572,6 +4591,10 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_ROT_EMBD,
],
+ MODEL_ARCH.NANBEIGE: [
+ MODEL_TENSOR.ROPE_FREQS,
+ MODEL_TENSOR.ATTN_ROT_EMBD,
+ ],
}
#
diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py
index ba08f8d650..ecf1f17ee8 100644
--- a/gguf-py/gguf/gguf_writer.py
+++ b/gguf-py/gguf/gguf_writer.py
@@ -908,6 +908,12 @@ class GGUFWriter:
def add_token_shift_count(self, count: int) -> None:
self.add_uint32(Keys.LLM.TOKEN_SHIFT_COUNT.format(arch=self.arch), count)
+ def add_num_loops(self, count: int) -> None:
+ self.add_uint32(Keys.LLM.NUM_LOOPS.format(arch=self.arch), count)
+
+ def add_skip_loop_final_norm(self, value: bool) -> None:
+ self.add_bool(Keys.LLM.SKIP_LOOP_FINAL_NORM.format(arch=self.arch), value)
+
def add_interleave_moe_layer_step(self, value: int) -> None:
self.add_uint32(Keys.LLM.INTERLEAVE_MOE_LAYER_STEP.format(arch=self.arch), value)
diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
index 39bf2c7959..c017067850 100644
--- a/src/llama-arch.cpp
+++ b/src/llama-arch.cpp
@@ -143,6 +143,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_KIMI_LINEAR, "kimi-linear" },
{ LLM_ARCH_TALKIE, "talkie" },
{ LLM_ARCH_MELLUM, "mellum" },
+ { LLM_ARCH_NANBEIGE, "nanbeige" },
{ LLM_ARCH_UNKNOWN, "(unknown)" },
};
@@ -221,6 +222,8 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_TOKEN_SHIFT_COUNT, "%s.token_shift_count" },
{ LLM_KV_INTERLEAVE_MOE_LAYER_STEP, "%s.interleave_moe_layer_step" },
{ LLM_KV_FULL_ATTENTION_INTERVAL, "%s.full_attention_interval" },
+ { LLM_KV_NUM_LOOPS, "%s.num_loops" },
+ { LLM_KV_SKIP_LOOP_FINAL_NORM, "%s.skip_loop_final_norm" },
{ LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" },
{ LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" },
diff --git a/src/llama-arch.h b/src/llama-arch.h
index 2e3916a0be..1c9aebb0bb 100644
--- a/src/llama-arch.h
+++ b/src/llama-arch.h
@@ -148,6 +148,7 @@ enum llm_arch {
LLM_ARCH_EAGLE3,
LLM_ARCH_MINIMAX_M3,
LLM_ARCH_DFLASH,
+ LLM_ARCH_NANBEIGE,
LLM_ARCH_UNKNOWN,
};
@@ -226,6 +227,8 @@ enum llm_kv {
LLM_KV_TOKEN_SHIFT_COUNT,
LLM_KV_INTERLEAVE_MOE_LAYER_STEP,
LLM_KV_FULL_ATTENTION_INTERVAL,
+ LLM_KV_NUM_LOOPS,
+ LLM_KV_SKIP_LOOP_FINAL_NORM,
LLM_KV_ATTENTION_HEAD_COUNT,
LLM_KV_ATTENTION_HEAD_COUNT_KV,
diff --git a/src/llama-context.cpp b/src/llama-context.cpp
index 012894e13f..9b399d6096 100644
--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
@@ -2339,6 +2339,7 @@ uint32_t llama_context::graph_max_nodes(uint32_t n_tokens) const {
model.arch == LLM_ARCH_QWEN35 ||
model.arch == LLM_ARCH_QWEN35MOE ||
model.arch == LLM_ARCH_DEEPSEEK4 ||
+ model.arch == LLM_ARCH_NANBEIGE ||
model.arch == LLM_ARCH_MINIMAX_M3) {
return std::max<uint32_t>(n_tokens * 40, 32u * model.n_tensors());
}
diff --git a/src/llama-model.cpp b/src/llama-model.cpp
index 074acbe1fa..be0a0df55d 100644
--- a/src/llama-model.cpp
+++ b/src/llama-model.cpp
@@ -85,6 +85,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params
return new llama_model_stablelm(params);
case LLM_ARCH_MELLUM:
return new llama_model_mellum(params);
+ case LLM_ARCH_NANBEIGE:
+ return new llama_model_nanbeige(params);
case LLM_ARCH_QWEN:
return new llama_model_qwen(params);
case LLM_ARCH_QWEN2:
@@ -2491,6 +2493,7 @@ 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_NANBEIGE:
return LLAMA_ROPE_TYPE_NORM;
// the pairs of head values are offset by n_rot/2
diff --git a/src/models/models.h b/src/models/models.h
index 916459e127..92ebfafa1e 100644
--- a/src/models/models.h
+++ b/src/models/models.h
@@ -424,6 +424,22 @@ struct llama_model_mellum : public llama_model_base {
std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
};
+struct llama_model_nanbeige : public llama_model_base {
+ llama_model_nanbeige(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;
+
+ int n_loops = 1;
+ int n_layer_phys = 0;
+ bool skip_loop_final_norm = false;
+
+ 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_qwen : public llama_model_base {
llama_model_qwen(const struct llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
diff --git a/src/models/nanbeige.cpp b/src/models/nanbeige.cpp
new file mode 100644
index 0000000000..3a546600fa
--- /dev/null
+++ b/src/models/nanbeige.cpp
@@ -0,0 +1,184 @@
+#include "models.h"
+
+void llama_model_nanbeige::load_arch_hparams(llama_model_loader & ml) {
+ ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+ uint32_t n_loops_u = 1;
+ ml.get_key(LLM_KV_NUM_LOOPS, n_loops_u, false);
+ GGML_ASSERT(n_loops_u >= 1);
+
+ skip_loop_final_norm = false;
+ ml.get_key(LLM_KV_SKIP_LOOP_FINAL_NORM, skip_loop_final_norm, false);
+
+ n_layer_phys = (int) hparams.n_layer();
+
+ // Bound-check before casting: signed int mul can overflow and bypass the guard.
+ GGML_ASSERT((size_t) n_layer_phys * (size_t) n_loops_u <= (size_t) LLAMA_MAX_LAYERS);
+ n_loops = (int) n_loops_u;
+
+ // Expand logical layer count before load_tensors() allocates layers / KV.
+ if (n_loops > 1) {
+ for (int j = 1; j < n_loops; ++j) {
+ for (int i = 0; i < n_layer_phys; ++i) {
+ const int dst = i + j * n_layer_phys;
+ hparams.n_head_arr[dst] = hparams.n_head_arr[i];
+ hparams.n_head_kv_arr[dst] = hparams.n_head_kv_arr[i];
+ hparams.n_ff_arr[dst] = hparams.n_ff_arr[i];
+ hparams.is_swa_impl[dst] = hparams.is_swa_impl[i];
+ hparams.is_recr_impl[dst] = hparams.is_recr_impl[i];
+ }
+ }
+ hparams.n_layer_all = (uint32_t) ((size_t) n_layer_phys * (size_t) n_loops);
+ }
+
+ type = LLM_TYPE_UNKNOWN;
+}
+
+void llama_model_nanbeige::load_arch_tensors(llama_model_loader &) {
+ LLAMA_LOAD_LOCALS;
+
+ 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 == NULL) {
+ output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+ }
+
+ const int n_phys = n_layer_phys > 0 ? n_layer_phys : n_layer;
+ for (int i = 0; i < n_phys; ++i) {
+ auto & layer = layers[i];
+
+ layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+ create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
+ layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+ layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2},
+ TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+
+ layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+ layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
+ layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0);
+ layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);
+ }
+
+ // Share physical weights across loops; each slot still has its own KV index.
+ if (n_loops > 1) {
+ for (int j = 1; j < n_loops; ++j) {
+ for (int i = 0; i < n_phys; ++i) {
+ layers[i + j * n_phys] = layers[i];
+ }
+ }
+ }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_nanbeige::build_arch_graph(const llm_graph_params & params) const {
+ return std::make_unique<graph>(*this, params);
+}
+
+llama_model_nanbeige::graph::graph(const llama_model & model, const llm_graph_params & params) :
+ llm_graph_context(params) {
+ const auto & nb = static_cast<const llama_model_nanbeige &>(model);
+
+ const int64_t n_embd_head = hparams.n_embd_head_v();
+ GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+
+ const int n_phys = nb.n_layer_phys > 0 ? nb.n_layer_phys : (int) n_layer;
+ const int n_loops = nb.n_loops > 0 ? nb.n_loops : 1;
+
+ ggml_tensor * cur;
+ ggml_tensor * inpL;
+
+ inpL = build_inp_embd(model.tok_embd);
+
+ ggml_tensor * inp_pos = build_inp_pos();
+
+ auto * inp_attn = build_attn_inp_kv();
+
+ const float kq_scale = hparams.f_attention_scale == 0.0f
+ ? 1.0f / sqrtf(float(n_embd_head))
+ : hparams.f_attention_scale;
+
+ ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+ for (int il = 0; il < n_layer; ++il) {
+ ggml_tensor * inpSA = inpL;
+
+ cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+ cb(cur, "attn_norm", il);
+
+ {
+ ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+ auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
+ n_embd_head, n_head, n_head_kv, il);
+
+ Qcur = ggml_rope_ext(
+ ctx0, Qcur, inp_pos, rope_factors,
+ 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, rope_factors,
+ n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+ ext_factor, attn_factor, beta_fast, beta_slow);
+
+ cb(Qcur, "Qcur", il);
+ cb(Kcur, "Kcur", il);
+ cb(Vcur, "Vcur", il);
+
+ cur = build_attn(inp_attn,
+ model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
+ Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+ cb(cur, "attn_out", 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);
+
+ cur = build_ffn(cur,
+ model.layers[il].ffn_up, model.layers[il].ffn_up_b, model.layers[il].ffn_up_s,
+ model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, model.layers[il].ffn_gate_s,
+ model.layers[il].ffn_down, model.layers[il].ffn_down_b, model.layers[il].ffn_down_s,
+ NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+ cb(cur, "ffn_out", il);
+
+ cur = ggml_add(ctx0, cur, ffn_inp);
+ cb(cur, "ffn_out", il);
+
+ cur = build_cvec(cur, il);
+ cb(cur, "l_out", il);
+
+ inpL = cur;
+
+ if (n_loops > 1 &&
+ ((il + 1) % n_phys) == 0 &&
+ (il + 1) < n_layer &&
+ !nb.skip_loop_final_norm) {
+ cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, il);
+ cb(cur, "loop_norm", 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 = build_lora_mm(model.output, cur, model.output_s);
+ cb(cur, "result_output", -1);
+ res->t_logits = cur;
+
+ ggml_build_forward_expand(gf, cur);
+}