Added support for the ArcticForCausalLM.

convert-hf-to-gguf.py

@@ -2277,6 +2277,156 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
         return [(self.map_tensor_name(name), data_torch)]
 
 
+@Model.register("ArcticForCausalLM")
+class ArcticModel(Model):
+    model_arch = gguf.MODEL_ARCH.ARCTIC
+
+    def set_vocab(self):
+        # The reason for using a custom implementation here is that the
+        # snowflake-arctic-instruct model redefined tokens 31998 and 31999 from
+        # tokenizer.model and used them as BOS and EOS instead of adding new tokens.
+        from sentencepiece import SentencePieceProcessor
+
+        tokenizer_path = self.dir_model / 'tokenizer.model'
+
+        if not tokenizer_path.is_file():
+            logger.error(f'Error: Missing {tokenizer_path}')
+            sys.exit(1)
+
+        # Read the whole vocabulary from the tokenizer.model file
+        tokenizer = SentencePieceProcessor(str(tokenizer_path))
+
+        vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
+
+        tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
+        scores: list[float] = [-10000.0] * vocab_size
+        toktypes: list[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size
+
+        for token_id in range(tokenizer.vocab_size()):
+
+            piece = tokenizer.id_to_piece(token_id)
+            text = piece.encode("utf-8")
+            score = tokenizer.get_score(token_id)
+
+            toktype = SentencePieceTokenTypes.NORMAL
+            if tokenizer.is_unknown(token_id):
+                toktype = SentencePieceTokenTypes.UNKNOWN
+            elif tokenizer.is_control(token_id):
+                toktype = SentencePieceTokenTypes.CONTROL
+            elif tokenizer.is_unused(token_id):
+                toktype = SentencePieceTokenTypes.UNUSED
+            elif tokenizer.is_byte(token_id):
+                toktype = SentencePieceTokenTypes.BYTE
+
+            tokens[token_id] = text
+            scores[token_id] = score
+            toktypes[token_id] = toktype
+
+        # Use the added_tokens_decoder field from tokeniser_config.json as the source
+        # of information about added/redefined tokens and modify them accordingly.
+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        if tokenizer_config_file.is_file():
+            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+                tokenizer_config_json = json.load(f)
+
+                if "added_tokens_decoder" in tokenizer_config_json:
+                    added_tokens_decoder = tokenizer_config_json["added_tokens_decoder"]
+                    for token_id, token_json in added_tokens_decoder.items():
+                        token_id = int(token_id)
+                        if (token_id >= vocab_size):
+                            logger.debug(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
+                            continue
+
+                        token_content = token_json["content"]
+                        token_type = SentencePieceTokenTypes.USER_DEFINED
+                        token_score = -10000.0
+
+                        # Map unk_token to UNKNOWN, other special tokens to CONTROL
+                        # Set the score to 0.0 as in the original tokenizer.model
+                        if ("special" in token_json) and token_json["special"]:
+                            if token_content == tokenizer_config_json["unk_token"]:
+                                token_type = SentencePieceTokenTypes.UNKNOWN
+                            else:
+                                token_type = SentencePieceTokenTypes.CONTROL
+                            token_score = 0.0
+
+                        logger.info(f"Setting added token {token_id} to '{token_content}' (type: {token_type}, score: {token_score:.2f})")
+                        tokens[token_id] = token_content.encode("utf-8")
+                        toktypes[token_id] = token_type
+                        scores[token_id] = token_score
+
+        self.gguf_writer.add_tokenizer_model("llama")
+        self.gguf_writer.add_tokenizer_pre("default")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+        self.gguf_writer.add_rope_dimension_count(hparams["hidden_size"] // hparams["num_attention_heads"])
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        n_head = self.hparams["num_attention_heads"]
+        n_kv_head = self.hparams.get("num_key_value_heads")
+
+        if name.endswith("q_proj.weight"):
+            data_torch = LlamaModel.permute(data_torch, n_head, n_head)
+        if name.endswith("k_proj.weight"):
+            data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
+
+        # process the experts separately
+        if name.find("block_sparse_moe.experts") != -1:
+            n_experts = self.hparams["num_local_experts"]
+
+            assert bid is not None
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+                tensors: list[tuple[str, Tensor]] = []
+
+                # merge the experts into a single 3d tensor
+                for wid in ["w1", "w2", "w3"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename = f"model.layers.{bid}.block_sparse_moe.experts.{xid}.{wid}.weight"
+                        datas.append(self._experts[bid][ename])
+                        del self._experts[bid][ename]
+
+                    data_torch = torch.stack(datas, dim=0)
+
+                    merged_name = f"layers.{bid}.feed_forward.experts.{wid}.weight"
+
+                    new_name = self.map_tensor_name(merged_name)
+
+                    tensors.append((new_name, data_torch))
+                return tensors
+            else:
+                return []
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def write_tensors(self):
+        super().write_tensors()
+
+        if self._experts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
+
+
 ###### CONVERSION LOGIC ######
 
 

gguf-py/gguf/constants.py

@@ -137,6 +137,7 @@ class MODEL_ARCH(IntEnum):
     COMMAND_R  = auto()
     DBRX       = auto()
     OLMO       = auto()
+    ARCTIC     = auto()
 
 
 class MODEL_TENSOR(IntEnum):
@@ -179,6 +180,7 @@ class MODEL_TENSOR(IntEnum):
     SSM_A              = auto()
     SSM_D              = auto()
     SSM_OUT            = auto()
+    FFN_NORM_EXP       = auto()
 
 
 MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
@@ -214,6 +216,7 @@ class MODEL_TENSOR(IntEnum):
     MODEL_ARCH.COMMAND_R:      "command-r",
     MODEL_ARCH.DBRX:           "dbrx",
     MODEL_ARCH.OLMO:           "olmo",
+    MODEL_ARCH.ARCTIC:         "arctic",
 }
 
 TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@@ -256,6 +259,7 @@ class MODEL_TENSOR(IntEnum):
     MODEL_TENSOR.SSM_A:              "blk.{bid}.ssm_a",
     MODEL_TENSOR.SSM_D:              "blk.{bid}.ssm_d",
     MODEL_TENSOR.SSM_OUT:            "blk.{bid}.ssm_out",
+    MODEL_TENSOR.FFN_NORM_EXP:       "blk.{bid}.ffn_norm_exps",
 }
 
 MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
@@ -724,6 +728,27 @@ class MODEL_TENSOR(IntEnum):
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
+    MODEL_ARCH.ARCTIC: [
+        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_GATE_INP,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_NORM_EXP,
+    ],
     # TODO
 }
 

gguf-py/gguf/tensor_mapping.py

@@ -370,6 +370,64 @@ class TensorNameMap:
             "model.layers.{bid}.out_proj",
             "backbone.layers.{bid}.mixer.out_proj",
         ),
+
+    }
+
+    # architecture-specific block mappings
+    arch_block_mappings_cfg: dict[MODEL_ARCH, dict[MODEL_TENSOR, tuple[str, ...]]] = {
+        MODEL_ARCH.ARCTIC: {
+            MODEL_TENSOR.TOKEN_EMBD: (
+                "model.embed_tokens",
+            ),
+            MODEL_TENSOR.OUTPUT_NORM: (
+                "model.norm",
+            ),
+            MODEL_TENSOR.OUTPUT: (
+                "lm_head",
+            ),
+            MODEL_TENSOR.ATTN_NORM: (
+                "model.layers.{bid}.input_layernorm",
+            ),
+            MODEL_TENSOR.ATTN_Q: (
+                "model.layers.{bid}.self_attn.q_proj",
+            ),
+            MODEL_TENSOR.ATTN_K: (
+                "model.layers.{bid}.self_attn.k_proj",
+            ),
+            MODEL_TENSOR.ATTN_V: (
+                "model.layers.{bid}.self_attn.v_proj",
+            ),
+            MODEL_TENSOR.ATTN_OUT: (
+                "model.layers.{bid}.self_attn.o_proj",
+            ),
+            MODEL_TENSOR.FFN_GATE_INP: (
+                "model.layers.{bid}.block_sparse_moe.gate",
+            ),
+            MODEL_TENSOR.FFN_NORM: (
+                "model.layers.{bid}.residual_layernorm",
+            ),
+            MODEL_TENSOR.FFN_GATE: (
+                "model.layers.{bid}.residual_mlp.w1",
+            ),
+            MODEL_TENSOR.FFN_DOWN: (
+                "model.layers.{bid}.residual_mlp.w2",
+            ),
+            MODEL_TENSOR.FFN_UP: (
+                "model.layers.{bid}.residual_mlp.w3",
+            ),
+            MODEL_TENSOR.FFN_GATE_EXP: (
+                "layers.{bid}.feed_forward.experts.w1",
+            ),
+            MODEL_TENSOR.FFN_DOWN_EXP: (
+                "layers.{bid}.feed_forward.experts.w2",
+            ),
+            MODEL_TENSOR.FFN_UP_EXP: (
+                "layers.{bid}.feed_forward.experts.w3",
+            ),
+            MODEL_TENSOR.FFN_NORM_EXP: (
+                "model.layers.{bid}.post_attention_layernorm",
+            ),
+        },
     }
 
     mapping: dict[str, tuple[MODEL_TENSOR, str]]
@@ -383,12 +441,16 @@ def __init__(self, arch: MODEL_ARCH, n_blocks: int):
             self.mapping[tensor_name] = (tensor, tensor_name)
             for key in keys:
                 self.mapping[key] = (tensor, tensor_name)
+        if arch in self.arch_block_mappings_cfg:
+            block_mappings = self.arch_block_mappings_cfg[arch]
+        else:
+            block_mappings = self.block_mappings_cfg
         for bid in range(n_blocks):
-            for tensor, keys in self.block_mappings_cfg.items():
+            for tensor, keys in block_mappings.items():
                 if tensor not in MODEL_TENSORS[arch]:
                     continue
                 # TODO: make this configurable
-                n_experts = 60
+                n_experts = 128
                 for xid in range(n_experts):
                     tensor_name = TENSOR_NAMES[tensor].format(bid = bid, xid = xid)
                     self.mapping[tensor_name] = (tensor, tensor_name)