Add arch support for cohere2-MoE

conversion/__init__.py

@@ -40,6 +40,7 @@
     "ChatGLMModel": "chatglm",
     "CodeShellForCausalLM": "codeshell",
     "CogVLMForCausalLM": "cogvlm",
+    "Cohere2MoeForCausalLM": "command_r",
     "Cohere2ForCausalLM": "command_r",
     "CohereForCausalLM": "command_r",
     "DbrxForCausalLM": "dbrx",

conversion/base.py

@@ -1192,7 +1192,7 @@
             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}")
 
@@ -1277,7 +1277,7 @@
             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":
@@ -1335,15 +1335,15 @@
         vocab_size = self.hparams.get("vocab_size", len(tokenizer.vocab))  # ty: ignore[unresolved-attribute]
         assert max(tokenizer.vocab.values()) < vocab_size  # ty: ignore[unresolved-attribute]
 
         tokpre = self.get_vocab_base_pre(tokenizer)
 
         reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}  # ty: ignore[unresolved-attribute]
         added_vocab = tokenizer.get_added_vocab()  # ty: ignore[unresolved-attribute]
 
         added_tokens_decoder = tokenizer.added_tokens_decoder  # ty: ignore[unresolved-attribute]
 
         for i in range(vocab_size):
             if i not in reverse_vocab:
                 tokens.append(f"[PAD{i}]")
                 toktypes.append(gguf.TokenType.UNUSED)
             else:
@@ -1356,7 +1356,7 @@
                         token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False))  # ty: ignore[unresolved-attribute, invalid-assignment]
                         if previous_token != token:
                             logger.info(f"{repr(previous_token)} is encoded and decoded back to {repr(token)} using AutoTokenizer")
 
                     if added_tokens_decoder[i].special or self.does_token_look_special(token):
                         toktypes.append(gguf.TokenType.CONTROL)
                     else:
@@ -1492,6 +1492,9 @@
         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"
@@ -1717,14 +1720,14 @@
         vocab_size = self.hparams.get("vocab_size", len(tokenizer.vocab))  # ty: ignore[unresolved-attribute]
         assert max(tokenizer.vocab.values()) < vocab_size  # ty: ignore[unresolved-attribute]
 
         reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}  # ty: ignore[unresolved-attribute]
         # k-mers can share text with a base-vocab BPE token (e.g. CCCCCC) and get
         # dropped by get_vocab(); a reserved marker suffix (U+E000) keeps each
         # k-mer's own id (llama.cpp strips it on detokenization)
         for kmer in tokenizer.kmers:  # ty: ignore[unresolved-attribute]
             reverse_vocab[tokenizer.dna_token_to_id[kmer]] = kmer + "\ue000"  # ty: ignore[unresolved-attribute]
         added_vocab = tokenizer.get_added_vocab()  # ty: ignore[unresolved-attribute]
         added_tokens_decoder = tokenizer.added_tokens_decoder  # ty: ignore[unresolved-attribute]
 
         tokens: list[str] = []
         toktypes: list[int] = []

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 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
             return
 
         yield from super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("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}")

convert_hf_to_gguf_update.py

@@ -100,6 +100,7 @@ class TOKENIZER_TYPE(IntEnum):
     {"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", },

gguf-py/gguf/constants.py

@@ -452,6 +452,7 @@ class MODEL_ARCH(IntEnum):
     XVERSE           = auto()
     COMMAND_R        = auto()
     COHERE2          = auto()
+    COHERE2MOE       = auto()
     DBRX             = auto()
     OLMO             = auto()
     OLMO2            = auto()
@@ -998,6 +999,7 @@ class MODEL_TENSOR(IntEnum):
     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",
@@ -2831,6 +2833,33 @@ class MODEL_TENSOR(IntEnum):
         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,

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"            },

src/llama-arch.h

@@ -70,6 +70,7 @@ 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,

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:

src/llama-model.cpp

@@ -155,6 +155,8 @@ 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:
@@ -1463,9 +1465,12 @@ 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)) {
@@ -1838,6 +1843,7 @@ 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 ||
@@ -2347,6 +2353,7 @@ 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:

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);
         }