Skip to content

LLM inference in C/C++

C++ ggml

Add arch support for cohere2-MoE (#24260)

Michael Wand4988f6e86605
diff --git a/conversion/__init__.py b/conversion/__init__.py
index cd6f8e6b93..00192cf33a 100644
--- a/conversion/__init__.py
+++ b/conversion/__init__.py
@@ -40,6 +40,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
"ChatGLMModel": "chatglm",
"CodeShellForCausalLM": "codeshell",
"CogVLMForCausalLM": "cogvlm",
+ "Cohere2MoeForCausalLM": "command_r",
"Cohere2ForCausalLM": "command_r",
"CohereForCausalLM": "command_r",
"DbrxForCausalLM": "dbrx",
diff --git a/conversion/base.py b/conversion/base.py
index 9d81c19b46..c872bcbb3c 100644
--- a/conversion/base.py
+++ b/conversion/base.py
@@ -1195,7 +1195,7 @@ class TextModel(ModelBase):
self.gguf_writer.add_embedding_length(n_embd)
logger.info(f"gguf: embedding length = {n_embd}")
- if (n_ff := self.find_hparam(["intermediate_size", "n_inner", "hidden_dim"], optional=True)) is not None:
+ if (n_ff := self.find_hparam(["prefix_dense_intermediate_size", "intermediate_size", "n_inner", "hidden_dim"], optional=True)) is not None:
self.gguf_writer.add_feed_forward_length(n_ff)
logger.info(f"gguf: feed forward length = {n_ff}")
@@ -1280,7 +1280,7 @@ class TextModel(ModelBase):
self.gguf_writer.add_expert_group_used_count(n_group_used)
logger.info(f"gguf: expert groups used count = {n_group_used}")
- if (score_func := self.find_hparam(["score_function", "scoring_func", "score_func", "moe_router_activation", "moe_router_activation_func"], optional=True)) is not None:
+ if (score_func := self.find_hparam(["score_function", "scoring_func", "score_func", "moe_router_activation", "moe_router_activation_func", "expert_selection_fn"], optional=True)) is not None:
if score_func == "sigmoid":
self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
elif score_func == "softmax":
@@ -1495,6 +1495,9 @@ class TextModel(ModelBase):
if chkhsh == "d772b220ace2baec124bed8cfafce0ead7d6c38a4b65ef11261cf9d5d62246d1":
# ref: https://huggingface.co/CohereLabs/tiny-aya-base
res = "tiny_aya"
+ if chkhsh == "52df12b4c8d4176e7481aab4b6e8454d1fd0a210a04a574f6d4e067d10e23c3e":
+ # ref: https://huggingface.co/CohereLabs/North-Mini-Code-1.0
+ res = "cohere2moe"
if chkhsh == "e636dc30a262dcc0d8c323492e32ae2b70728f4df7dfe9737d9f920a282b8aea":
# ref: https://huggingface.co/Qwen/Qwen1.5-7B
res = "qwen2"
diff --git a/conversion/command_r.py b/conversion/command_r.py
index 603288d165..118565c669 100644
--- a/conversion/command_r.py
+++ b/conversion/command_r.py
@@ -1,5 +1,6 @@
from __future__ import annotations
+import re
from typing import Iterable, TYPE_CHECKING
import torch
@@ -55,3 +56,122 @@ class Cohere2Model(TextModel):
return
yield from super().modify_tensors(data_torch, name, bid)
+
+
[email protected]("Cohere2MoeForCausalLM")
+class Cohere2MoeModel(TextModel):
+ model_arch = gguf.MODEL_ARCH.COHERE2MOE
+ _n_main_layers: int | None = None
+ _expert_tensor_re = re.compile(
+ r"model\.layers\.(\d+)\.mlp\.experts\.(\d+)\.(down_proj|gate_proj|up_proj)\.weight"
+ )
+
+ def __init__(self, *args, **kwargs):
+ super().__init__(*args, **kwargs)
+ if (n_nextn := int(self.hparams.get("num_nextn_predict_layers", 0) or 0)) > 0 and not self.no_mtp:
+ self.block_count += n_nextn
+ self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+ self._experts: list[dict[str, Tensor]] = [{} for _ in range(self.block_count)]
+
+ def _set_vocab_gpt2(self) -> None:
+ 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 = gguf.SpecialVocab(self.dir_model, load_merges=True)
+ special_vocab.add_to_gguf(self.gguf_writer)
+
+ def set_gguf_parameters(self):
+ hparams = self.hparams
+ expert_intermediate_size = hparams["intermediate_size"]
+ mlp_layer_types = hparams.get("mlp_layer_types")
+ n_dense_lead = hparams.get("first_k_dense_replace", 0)
+ if mlp_layer_types is not None:
+ n_dense_lead = next((i for i, t in enumerate(mlp_layer_types) if t != "dense"), len(mlp_layer_types))
+
+ super().set_gguf_parameters()
+
+ self.gguf_writer.add_logit_scale(hparams["logit_scale"])
+ self.gguf_writer.add_sliding_window(hparams["sliding_window"])
+ self.gguf_writer.add_sliding_window_pattern([t == "sliding_attention" for t in hparams["layer_types"]])
+ self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+ self.gguf_writer.add_expert_feed_forward_length(expert_intermediate_size)
+ self.gguf_writer.add_leading_dense_block_count(n_dense_lead)
+ self.gguf_writer.add_expert_weights_norm(hparams.get("norm_topk_prob", False))
+ if (num_shared_experts := hparams.get("num_shared_experts", 0)) > 0:
+ if hparams.get("shared_expert_combination_strategy", "average") != "average":
+ raise ValueError("Cohere2 MoE only supports average shared expert combination")
+ self.gguf_writer.add_expert_shared_count(num_shared_experts)
+ self.gguf_writer.add_expert_shared_feed_forward_length(expert_intermediate_size * num_shared_experts)
+ if (n_nextn := hparams.get("num_nextn_predict_layers", 0)) > 0 and not self.no_mtp:
+ self.gguf_writer.add_nextn_predict_layers(n_nextn)
+ self.gguf_writer.add_rope_dimension_count(hparams["head_dim"])
+ self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
+
+ def index_tensors(self, remote_hf_model_id: str | None = None):
+ hparams = {**self.hparams, **self.hparams.get("text_config", {})}
+ self._n_main_layers = hparams.get("num_hidden_layers")
+ type(self)._n_main_layers = self._n_main_layers
+ return super().index_tensors(remote_hf_model_id=remote_hf_model_id)
+
+ @classmethod
+ def filter_tensors(cls, item):
+ if (titem := super().filter_tensors(item)) is None:
+ return None
+ name, gen = titem
+
+ if 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 modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+ if name.endswith(".bias"):
+ if torch.any(data_torch != 0):
+ raise ValueError(f"Bias tensor {name!r} is not zero.")
+ logger.debug(f"Skipping bias tensor {name!r}.")
+ return
+
+ if (m := self._expert_tensor_re.fullmatch(name)) is not None:
+ n_experts = self.hparams["num_experts"]
+ layer_idx = int(m.group(1))
+ assert bid is None or bid == layer_idx
+
+ self._experts[layer_idx][name] = data_torch
+
+ expected = {
+ f"model.layers.{layer_idx}.mlp.experts.{xid}.{w_name}.weight"
+ for xid in range(n_experts)
+ for w_name in ("down_proj", "gate_proj", "up_proj")
+ }
+ if expected.issubset(self._experts[layer_idx]):
+ for w_name in ["down_proj", "gate_proj", "up_proj"]:
+ datas: list[Tensor] = []
+
+ for xid in range(n_experts):
+ ename = f"model.layers.{layer_idx}.mlp.experts.{xid}.{w_name}.weight"
+ datas.append(self._experts[layer_idx][ename])
+ del self._experts[layer_idx][ename]
+
+ data_torch = torch.stack(datas, dim=0)
+ merged_name = f"model.layers.{layer_idx}.mlp.experts.{w_name}.weight"
+
+ yield from super().modify_tensors(data_torch, merged_name, layer_idx)
+ return
+
+ yield from super().modify_tensors(data_torch, name, bid)
+
+ def prepare_tensors(self):
+ super().prepare_tensors()
+
+ experts = [k for d in self._experts for k in d.keys()]
+ if len(experts) > 0:
+ raise ValueError(f"Unprocessed experts: {experts}")
diff --git a/convert_hf_to_gguf_update.py b/convert_hf_to_gguf_update.py
index b4c8a7cf00..91c006278f 100755
--- a/convert_hf_to_gguf_update.py
+++ b/convert_hf_to_gguf_update.py
@@ -100,6 +100,7 @@ models = [
{"name": "refact", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/smallcloudai/Refact-1_6-base", },
{"name": "command-r", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/CohereForAI/c4ai-command-r-v01", },
{"name": "tiny_aya", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/CohereLabs/tiny-aya-base", },
+ {"name": "cohere2moe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/CohereLabs/North-Mini-Code-1.0", },
{"name": "qwen2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen1.5-7B", },
{"name": "olmo", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/allenai/OLMo-1.7-7B-hf", },
{"name": "dbrx", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/databricks/dbrx-base", },
diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py
index 4b6dfea64d..463963f2ac 100644
--- a/gguf-py/gguf/constants.py
+++ b/gguf-py/gguf/constants.py
@@ -457,6 +457,7 @@ class MODEL_ARCH(IntEnum):
XVERSE = auto()
COMMAND_R = auto()
COHERE2 = auto()
+ COHERE2MOE = auto()
DBRX = auto()
OLMO = auto()
OLMO2 = auto()
@@ -1012,6 +1013,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.XVERSE: "xverse",
MODEL_ARCH.COMMAND_R: "command-r",
MODEL_ARCH.COHERE2: "cohere2",
+ MODEL_ARCH.COHERE2MOE: "cohere2moe",
MODEL_ARCH.DBRX: "dbrx",
MODEL_ARCH.OLMO: "olmo",
MODEL_ARCH.OLMO2: "olmo2",
@@ -2872,6 +2874,33 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
+ MODEL_ARCH.COHERE2MOE: [
+ MODEL_TENSOR.TOKEN_EMBD,
+ MODEL_TENSOR.OUTPUT_NORM,
+ MODEL_TENSOR.OUTPUT,
+ MODEL_TENSOR.ATTN_NORM,
+ MODEL_TENSOR.ATTN_Q,
+ MODEL_TENSOR.ATTN_K,
+ MODEL_TENSOR.ATTN_V,
+ MODEL_TENSOR.ATTN_OUT,
+ MODEL_TENSOR.FFN_GATE,
+ MODEL_TENSOR.FFN_DOWN,
+ MODEL_TENSOR.FFN_UP,
+ MODEL_TENSOR.FFN_GATE_INP,
+ MODEL_TENSOR.FFN_GATE_EXP,
+ MODEL_TENSOR.FFN_GATE_UP_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.NEXTN_EH_PROJ,
+ MODEL_TENSOR.NEXTN_EMBED_TOKENS,
+ MODEL_TENSOR.NEXTN_ENORM,
+ MODEL_TENSOR.NEXTN_HNORM,
+ MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD,
+ MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM,
+ ],
MODEL_ARCH.DBRX: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
index 9f93d5bc7c..4a52d97729 100644
--- a/src/llama-arch.cpp
+++ b/src/llama-arch.cpp
@@ -66,6 +66,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_XVERSE, "xverse" },
{ LLM_ARCH_COMMAND_R, "command-r" },
{ LLM_ARCH_COHERE2, "cohere2" },
+ { LLM_ARCH_COHERE2MOE, "cohere2moe" },
{ LLM_ARCH_DBRX, "dbrx" },
{ LLM_ARCH_OLMO, "olmo" },
{ LLM_ARCH_OLMO2, "olmo2" },
diff --git a/src/llama-arch.h b/src/llama-arch.h
index c5245fb589..989da06d8d 100644
--- a/src/llama-arch.h
+++ b/src/llama-arch.h
@@ -71,6 +71,7 @@ enum llm_arch {
LLM_ARCH_XVERSE,
LLM_ARCH_COMMAND_R,
LLM_ARCH_COHERE2,
+ LLM_ARCH_COHERE2MOE,
LLM_ARCH_DBRX,
LLM_ARCH_OLMO,
LLM_ARCH_OLMO2,
diff --git a/src/llama-model-saver.cpp b/src/llama-model-saver.cpp
index 67d4a9df0f..a3928523ba 100644
--- a/src/llama-model-saver.cpp
+++ b/src/llama-model-saver.cpp
@@ -18,6 +18,7 @@ bool llama_model_saver_supports_arch(llm_arch arch) {
case LLM_ARCH_GEMMA3:
case LLM_ARCH_GEMMA3N:
case LLM_ARCH_COHERE2:
+ case LLM_ARCH_COHERE2MOE:
case LLM_ARCH_OLMO2:
case LLM_ARCH_BITNET:
case LLM_ARCH_T5:
diff --git a/src/llama-model.cpp b/src/llama-model.cpp
index 7281ed79f1..c528755339 100644
--- a/src/llama-model.cpp
+++ b/src/llama-model.cpp
@@ -157,6 +157,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params
return new llama_model_command_r(params);
case LLM_ARCH_COHERE2:
return new llama_model_cohere2(params);
+ case LLM_ARCH_COHERE2MOE:
+ return new llama_model_cohere2moe(params);
case LLM_ARCH_DBRX:
return new llama_model_dbrx(params);
case LLM_ARCH_OLMO:
@@ -1467,9 +1469,12 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
}
ml.done_getting_tensors();
+ // Tied NVFP4 output is valid when no separate LM-head scale tensors are present.
+ // If sidecar scales exist, the output weight must be an actual output tensor.
GGML_ASSERT(!(output && tok_embd &&
strcmp(output->name, tok_embd->name) == 0 &&
- output->type == GGML_TYPE_NVFP4));
+ output->type == GGML_TYPE_NVFP4 &&
+ (output_s || output_in_s)));
// populate tensors_by_name
for (auto & [_, ctx_ptr] : ml.ctx_map) {
for (auto * cur = ggml_get_first_tensor(ctx_ptr.get()); cur != NULL; cur = ggml_get_next_tensor(ctx_ptr.get(), cur)) {
@@ -1844,6 +1849,7 @@ void llama_model::print_info() const {
}
if (arch == LLM_ARCH_MELLUM ||
+ arch == LLM_ARCH_COHERE2MOE ||
arch == LLM_ARCH_QWEN3MOE ||
arch == LLM_ARCH_OPENAI_MOE ||
arch == LLM_ARCH_QWEN3VLMOE ||
@@ -2389,6 +2395,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
case LLM_ARCH_XVERSE:
case LLM_ARCH_COMMAND_R:
case LLM_ARCH_COHERE2:
+ case LLM_ARCH_COHERE2MOE:
case LLM_ARCH_OLMO:
case LLM_ARCH_ARCTIC:
case LLM_ARCH_DEEPSEEK:
diff --git a/src/models/cohere2.cpp b/src/models/cohere2.cpp
index 61a5945a19..e2b3662560 100644
--- a/src/models/cohere2.cpp
+++ b/src/models/cohere2.cpp
@@ -122,9 +122,9 @@ llama_model_cohere2::graph::graph(const llama_model & model, const llm_graph_par
// feed-forward network
{
cur = build_ffn(ffn_inp,
- model.layers[il].ffn_up, NULL, NULL,
- model.layers[il].ffn_gate, NULL, NULL,
- model.layers[il].ffn_down, NULL, NULL,
+ 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);
}
diff --git a/src/models/cohere2moe.cpp b/src/models/cohere2moe.cpp
new file mode 100644
index 0000000000..499c73a1c4
--- /dev/null
+++ b/src/models/cohere2moe.cpp
@@ -0,0 +1,443 @@
+#include "models.h"
+
+void llama_model_cohere2moe::load_arch_hparams(llama_model_loader & ml) {
+ const bool found_norm = ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps, false);
+ const bool found_norm_rms = ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps, false);
+ if (!found_norm && !found_norm_rms) {
+ throw std::runtime_error("missing Cohere2 MoE norm epsilon");
+ }
+ if (!found_norm_rms) {
+ hparams.f_norm_rms_eps = 0.0f;
+ }
+
+ ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
+ ml.get_key(LLM_KV_LOGIT_SCALE, hparams.f_logit_scale);
+ ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead);
+ ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
+ ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
+ ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared, false);
+ ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM, hparams.expert_weights_norm, false);
+ ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE, hparams.expert_weights_scale, false);
+ ml.get_key(LLM_KV_EXPERT_GATING_FUNC, hparams.expert_gating_func, false);
+
+ 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");
+
+ if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) {
+ hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID;
+ }
+
+ hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
+ uint32_t swa_period = 4;
+ if (ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false)) {
+ hparams.set_swa_pattern(swa_period, true);
+ } else {
+ ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl, hparams.n_layer());
+ }
+
+ hparams.rope_freq_base_train_swa = hparams.rope_freq_base_train;
+ hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
+ ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
+
+ switch (hparams.n_layer()) {
+ case 49: type = LLM_TYPE_30B_A3B; break;
+ default: type = LLM_TYPE_UNKNOWN;
+ }
+}
+
+void llama_model_cohere2moe::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);
+ // Trunk-only: the GGUF declares MTP layers in metadata but the actual MTP
+ // tensors live in a separate file. Mark MTP tensors NOT_REQUIRED so the
+ // trunk loads cleanly.
+ 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;
+ const int mtp_flags = trunk_only ? TENSOR_NOT_REQUIRED : 0;
+
+ tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);
+
+ // output
+ 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 is NULL, init from the input tok embed
+ if (output == NULL) {
+ output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, TENSOR_DUPLICATED);
+ }
+
+ if (n_expert == 0) {
+ throw std::runtime_error("n_expert must be > 0 for Cohere2Moe");
+ }
+ if (n_expert_used == 0) {
+ throw std::runtime_error("n_expert_used must be > 0 for Cohere2Moe");
+ }
+
+ auto load_block_trunk = [&](int i, int flags) {
+ auto & layer = layers[i];
+
+ layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, flags);
+
+ create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_gqa, n_embd_gqa, flags);
+ layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, flags);
+
+ if (static_cast<uint32_t>(i) < 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 {
+ const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff;
+
+ layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), { n_embd, 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);
+ create_tensor_gate_up_exps(layer, i, n_embd, n_ff_exp, n_expert, flags);
+
+ if (hparams.n_expert_shared > 0) {
+ const int64_t n_ff_shexp = hparams.n_ff_shexp ? hparams.n_ff_shexp : n_ff_exp * hparams.n_expert_shared;
+ layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), { n_embd, n_ff_shexp }, flags);
+ layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { n_ff_shexp, n_embd }, flags);
+ layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), { n_embd, n_ff_shexp }, flags);
+ }
+ }
+ };
+
+ auto load_block_mtp = [&](int i, int flags) {
+ auto & layer = layers[i];
+
+ // MTP block looks like a full-attention Cohere2 MoE decoder block.
+ layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, flags);
+
+ create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_gqa, n_embd_gqa, flags);
+ layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, flags);
+
+ const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff;
+
+ // Routed experts
+ layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), { n_embd, 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);
+ create_tensor_gate_up_exps(layer, i, n_embd, n_ff_exp, n_expert, flags);
+
+ if (hparams.n_expert_shared > 0) {
+ const int64_t n_ff_shexp = hparams.n_ff_shexp ? hparams.n_ff_shexp : n_ff_exp * hparams.n_expert_shared;
+
+ // Shared experts
+ layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), { n_embd, n_ff_shexp }, flags);
+ layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { n_ff_shexp, n_embd }, flags);
+ layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), { n_embd, n_ff_shexp }, flags);
+ }
+
+ // NextN-specific tensors that define the MTP block.
+ 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 }, TENSOR_NOT_REQUIRED);
+ layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED);
+ layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, TENSOR_NOT_REQUIRED);
+ };
+
+ for (int i = 0; i < n_layer; ++i) {
+ load_block_trunk(i, trunk_flags);
+ }
+ // MTP/NextN layers are loaded as extra decoder blocks.
+ for (int i = n_layer; i < n_layer_all; ++i) {
+ load_block_mtp(i, mtp_flags);
+ }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_cohere2moe::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_cohere2moe::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();
+
+ GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+ GGML_ASSERT(n_embd_head == n_rot);
+
+ const llm_norm_type cohere2moe_norm_type = hparams.f_norm_rms_eps == 0.0f ? LLM_NORM : LLM_NORM_RMS;
+ const float f_logit_scale = hparams.f_logit_scale;
+ ggml_tensor * cur;
+ ggml_tensor * inpL = build_inp_embd(model.tok_embd);
+ ggml_tensor * inp_pos = build_inp_pos();
+
+ auto * inp_attn = build_attn_inp_kv_iswa();
+ ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+ // MTP/NextN layers are loaded as extra decoder blocks but not executed in the main pass.
+ for (int il = 0; il < n_layer; ++il) {
+ const bool is_swa = hparams.is_swa(il);
+ // Dense-prefix full-attention layers use RoPE; later layers follow the SWA pattern.
+ const bool force_rope = static_cast<uint32_t>(il) < hparams.n_layer_dense_lead;
+
+ cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, cohere2moe_norm_type, il);
+ cb(cur, "attn_norm", il);
+
+ ggml_tensor * ffn_inp = cur;
+
+ {
+ const auto & layer = model.layers[il];
+
+ auto [Qcur, Kcur, Vcur] = build_qkv(layer, cur,
+ n_embd_head, n_head, n_head_kv, il);
+
+ if (is_swa || force_rope) {
+ ggml_tensor * rope_factors = model.get_rope_factors(cparams, 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,
+ layer.wo, layer.wo_b, layer.wo_s,
+ Qcur, Kcur, Vcur, nullptr, nullptr, nullptr,
+ 1.0f / sqrtf(float(n_embd_head)), il);
+ }
+
+ if (il == n_layer - 1 && inp_out_ids && cparams.embeddings_nextn_masked) {
+ cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+ inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+ ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
+ }
+
+ ggml_tensor * attn_out = cur;
+
+ const auto & layer = model.layers[il];
+
+ if (layer.ffn_gate_inp == nullptr) {
+ cur = build_ffn(ffn_inp,
+ layer.ffn_up, nullptr, layer.ffn_up_s,
+ layer.ffn_gate, nullptr, layer.ffn_gate_s,
+ layer.ffn_down, nullptr, layer.ffn_down_s,
+ nullptr, LLM_FFN_SILU, LLM_FFN_PAR, il);
+ cb(cur, "ffn_out", il);
+ } else {
+ cur = build_moe_ffn(ffn_inp,
+ layer.ffn_gate_inp,
+ layer.ffn_up_exps,
+ layer.ffn_gate_exps,
+ layer.ffn_down_exps,
+ nullptr,
+ 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(cur, "ffn_moe_out", il);
+
+ if (layer.ffn_up_shexp) {
+ ggml_tensor * ffn_shexp = build_ffn(ffn_inp,
+ layer.ffn_up_shexp, nullptr, layer.ffn_up_shexp_s,
+ layer.ffn_gate_shexp, nullptr, layer.ffn_gate_shexp_s,
+ layer.ffn_down_shexp, nullptr, layer.ffn_down_shexp_s,
+ nullptr, LLM_FFN_SILU, LLM_FFN_PAR, il);
+ cb(ffn_shexp, "ffn_shexp", il);
+
+ cur = ggml_add(ctx0, cur, ffn_shexp);
+ cur = ggml_scale(ctx0, cur, 0.5f);
+ cb(cur, "ffn_out", il);
+ }
+ }
+
+ cur = ggml_add(ctx0, cur, inpL);
+ cur = ggml_add(ctx0, cur, attn_out);
+
+ cur = build_cvec(cur, il);
+ cb(cur, "l_out", il);
+
+ inpL = cur;
+ }
+
+ cur = inpL;
+ cur = build_norm(cur, model.output_norm, nullptr, cohere2moe_norm_type, -1);
+
+ cb(cur, "h_nextn", -1);
+ res->t_h_nextn = cur;
+
+ if (!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;
+
+ cur = build_lora_mm(model.output, cur);
+
+ if (f_logit_scale) {
+ cur = ggml_scale(ctx0, cur, f_logit_scale);
+ }
+
+ cb(cur, "result_output", -1);
+ res->t_logits = cur;
+
+ ggml_build_forward_expand(gf, cur);
+}
+
+llama_model_cohere2moe::graph_mtp::graph_mtp(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+ GGML_ASSERT(hparams.n_layer_nextn > 0 && "COHERE2MOE MTP requires n_layer_nextn > 0");
+ GGML_ASSERT(hparams.n_layer_nextn == 1 && "COHERE2MOE 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());
+ GGML_ASSERT(n_embd_head == n_rot);
+
+ const int il = hparams.n_layer();
+ 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");
+
+ const llm_norm_type cohere2moe_norm_type = hparams.f_norm_rms_eps == 0.0f ? LLM_NORM : LLM_NORM_RMS;
+
+ // 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);
+
+ // TODO: make static using `ggml_build_forward_select()`
+ // see llm_graph_context::build_inp_embd() for reference
+ 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_iswa();
+
+ ggml_tensor * h_norm = build_norm(h_embd, layer.nextn.hnorm, nullptr, cohere2moe_norm_type, il);
+ cb(h_norm, "mtp_hnorm", il);
+
+ ggml_tensor * e_norm = build_norm(tok_embd, layer.nextn.enorm, nullptr, cohere2moe_norm_type, 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 * inpL = cur;
+
+ cur = build_norm(cur, layer.attn_norm, nullptr, cohere2moe_norm_type, il);
+ cb(cur, "mtp_attn_norm", il);
+ ggml_tensor * ffn_inp = cur;
+
+ auto [Qcur, Kcur, Vcur] = build_qkv(layer, cur, n_embd_head, n_head, n_head_kv, il);
+ ggml_tensor * rope_factors = model.get_rope_factors(cparams, 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, "mtp_Qcur", il);
+ cb(Kcur, "mtp_Kcur", il);
+ cb(Vcur, "mtp_Vcur", il);
+
+ cur = build_attn(inp_attn,
+ layer.wo, layer.wo_b, 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 * attn_out = cur;
+
+ cur = build_moe_ffn(ffn_inp,
+ layer.ffn_gate_inp,
+ layer.ffn_up_exps,
+ layer.ffn_gate_exps,
+ layer.ffn_down_exps,
+ nullptr,
+ 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(cur, "mtp_ffn_moe_out", il);
+
+ if (layer.ffn_up_shexp) {
+ ggml_tensor * ffn_shexp = build_ffn(ffn_inp,
+ layer.ffn_up_shexp, nullptr, layer.ffn_up_shexp_s,
+ layer.ffn_gate_shexp, nullptr, layer.ffn_gate_shexp_s,
+ layer.ffn_down_shexp, nullptr, layer.ffn_down_shexp_s,
+ nullptr, LLM_FFN_SILU, LLM_FFN_PAR, il);
+ cb(ffn_shexp, "mtp_ffn_shexp", il);
+
+ cur = ggml_add(ctx0, cur, ffn_shexp);
+ cur = ggml_scale(ctx0, cur, 0.5f);
+ cb(cur, "mtp_ffn_out", il);
+ }
+
+ cur = ggml_add(ctx0, cur, inpL);
+ cur = ggml_add(ctx0, cur, attn_out);
+ 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 && "COHERE2MOE MTP: missing both nextn.shared_head_norm and output_norm");
+ cur = build_norm(cur, head_norm_w, nullptr, cohere2moe_norm_type, -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_ASSERT(head_w && "COHERE2MOE MTP: missing LM head (nextn.shared_head_head or model.output)");
+ cur = build_lora_mm(head_w, cur, layer.nextn.shared_head_head ? layer.nextn.shared_head_head_s : nullptr);
+
+ if (hparams.f_logit_scale) {
+ cur = ggml_scale(ctx0, cur, hparams.f_logit_scale);
+ }
+
+ 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.h
index ee3aff07b9..2ac8415a36 100644
--- a/src/models/models.h
+++ b/src/models/models.h
@@ -937,6 +937,23 @@ struct llama_model_cohere2 : public llama_model_base {
};
+struct llama_model_cohere2moe : public llama_model_base {
+ llama_model_cohere2moe(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);
+ };
+
+ 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;
+};
+
+
struct llama_model_dbrx : public llama_model_base {
llama_model_dbrx(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.cpp
index 4d06274ef1..524971ae4b 100644
--- a/tests/test-llama-archs.cpp
+++ b/tests/test-llama-archs.cpp
@@ -185,7 +185,7 @@ 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_MIMO2 || arch == LLM_ARCH_STEP35) {
+ } else if (arch == LLM_ARCH_COHERE2MOE || arch == LLM_ARCH_MIMO2 || arch == LLM_ARCH_STEP35) {
std::vector<uint32_t> pattern;
pattern.reserve(n_layer);
for (uint32_t il = 0; il < n_layer; il++) {
@@ -322,6 +322,7 @@ static std::vector<float> get_logits(
static bool moe_mandatory(const llm_arch arch) {
switch (arch) {
case LLM_ARCH_LLAMA4:
+ case LLM_ARCH_COHERE2MOE:
case LLM_ARCH_GROK:
case LLM_ARCH_QWEN2MOE:
case LLM_ARCH_QWEN3MOE: