bigdl-core-cpp 2.5.0b20240620__py3-none-manylinux2010_x86_64.whl → 2.5.0b20240621__py3-none-manylinux2010_x86_64.whl

bigdl/cpp/convert-hf-to-gguf.py

@@ -49,9 +49,6 @@ AnyModel = TypeVar("AnyModel", bound="type[Model]")
 class Model:
     _model_classes: dict[str, type[Model]] = {}
 
-<<<<<<< HEAD
-    def __init__(self, dir_model: Path, ftype: int, fname_out: Path, is_big_endian: bool, use_temp_file: bool):
-=======
     dir_model: Path
     ftype: int
     is_big_endian: bool
@@ -73,26 +70,17 @@ class Model:
     def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool, use_temp_file: bool, eager: bool):
         if type(self) is Model:
             raise TypeError(f"{type(self).__name__!r} should not be directly instantiated")
->>>>>>> uupstream/master
         self.dir_model = dir_model
         self.ftype = ftype
         self.is_big_endian = is_big_endian
         self.endianess = gguf.GGUFEndian.BIG if is_big_endian else gguf.GGUFEndian.LITTLE
         self.use_temp_file = use_temp_file
-<<<<<<< HEAD
-        self.is_safetensors = self._is_model_safetensors()
-        self.num_parts = Model.count_model_parts(self.dir_model, ".safetensors" if self.is_safetensors else ".bin")
-        self.part_names = self._get_part_names()
-        self.hparams = Model.load_hparams(self.dir_model)
-        self.gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=self.use_temp_file)
-=======
         self.lazy = not eager
         self.part_names = Model.get_model_part_names(self.dir_model, ".safetensors")
         self.is_safetensors = len(self.part_names) > 0
         if not self.is_safetensors:
             self.part_names = Model.get_model_part_names(self.dir_model, ".bin")
         self.hparams = Model.load_hparams(self.dir_model)
->>>>>>> uupstream/master
         self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer"])
         self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
         self.tensor_names = None
@@ -238,8 +226,6 @@ class Model:
 
         self.gguf_writer.add_file_type(self.ftype)
         logger.info(f"gguf: file type = {self.ftype}")
-<<<<<<< HEAD
-=======
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused
@@ -255,7 +241,6 @@ class Model:
         del name, new_name, bid, n_dims  # unused
 
         return False
->>>>>>> uupstream/master
 
     def write_tensors(self):
         max_name_len = max(len(s) for _, s in self.tensor_map.mapping.values()) + len(".weight,")
@@ -278,18 +263,11 @@ class Model:
                     bid = int(part)
                     break
 
-<<<<<<< HEAD
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                raise ValueError(f"Can not map tensor {name!r}")
-=======
             for new_name, data in ((n, d.squeeze().numpy()) for n, d in self.modify_tensors(data_torch, name, bid)):
                 data: np.ndarray = data  # type hint
                 n_dims = len(data.shape)
                 data_dtype = data.dtype
                 data_qtype: gguf.GGMLQuantizationType | None = None
->>>>>>> uupstream/master
 
                 # when both are True, f32 should win
                 extra_f32 = self.extra_f32_tensors(name, new_name, bid, n_dims)
@@ -303,18 +281,12 @@ class Model:
                     new_name.endswith("_norm.weight"),
                 ))
 
-<<<<<<< HEAD
-            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-            if self.ftype == 1 and data_dtype == np.float16 and (n_dims == 1 or new_name.endswith("_norm.weight")):
-                data = data.astype(np.float32)
-=======
                 # Some tensor types are always in float32
                 extra_f32 = extra_f32 or any(self.match_model_tensor_name(new_name, key, bid) for key in (
                     gguf.MODEL_TENSOR.FFN_GATE_INP,
                     gguf.MODEL_TENSOR.POS_EMBD,
                     gguf.MODEL_TENSOR.TOKEN_TYPES,
                 ))
->>>>>>> uupstream/master
 
                 # if f16 desired, convert any float32 2-dim weight tensors to float16
                 extra_f16 = any(cond for cond in (
@@ -322,15 +294,11 @@ class Model:
                     (name.endswith(".weight") and n_dims >= 2),
                 ))
 
-<<<<<<< HEAD
-            logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-=======
                 if self.ftype != gguf.LlamaFileType.ALL_F32 and extra_f16 and not extra_f32:
                     if self.ftype == gguf.LlamaFileType.MOSTLY_BF16:
                         data = gguf.quantize_bf16(data)
                         assert data.dtype == np.int16
                         data_qtype = gguf.GGMLQuantizationType.BF16
->>>>>>> uupstream/master
 
                     elif self.ftype == gguf.LlamaFileType.MOSTLY_Q8_0 and gguf.can_quantize_to_q8_0(data):
                         data = gguf.quantize_q8_0(data)
@@ -402,22 +370,6 @@ class Model:
         except KeyError:
             raise NotImplementedError(f'Architecture {arch!r} not supported!') from None
 
-<<<<<<< HEAD
-    def _is_model_safetensors(self) -> bool:
-        return Model.count_model_parts(self.dir_model, ".safetensors") > 0
-
-    def _get_part_names(self):
-        if self.is_safetensors:
-            if self.num_parts == 1:  # there's only one .safetensors file
-                return ("model.safetensors",)
-            return (f"model-{n:05}-of-{self.num_parts:05}.safetensors" for n in range(1, self.num_parts + 1))
-
-        if self.num_parts == 1:  # there's only one .bin file
-            return ("pytorch_model.bin",)
-        return (f"pytorch_model-{n:05}-of-{self.num_parts:05}.bin" for n in range(1, self.num_parts + 1))
-
-=======
->>>>>>> uupstream/master
     # used for GPT-2 BPE and WordPiece vocabs
     def get_vocab_base(self) -> tuple[list[str], list[int], str]:
         tokens: list[str] = []
@@ -452,10 +404,7 @@ class Model:
     # NOTE: this function is generated by convert-hf-to-gguf-update.py
     #       do not modify it manually!
     # ref:  https://github.com/ggerganov/llama.cpp/pull/6920
-<<<<<<< HEAD
-=======
     # Marker: Start get_vocab_base_pre
->>>>>>> uupstream/master
     def get_vocab_base_pre(self, tokenizer) -> str:
         # encoding this string and hashing the resulting tokens would (hopefully) give us a unique identifier that
         # is specific for the BPE pre-tokenizer used by the model
@@ -499,23 +448,15 @@ class Model:
         if chkhsh == "3ce83efda5659b07b1ad37ca97ca5797ea4285d9b9ab0dc679e4a720c9da7454":
             # ref: https://huggingface.co/openai-community/gpt2
             res = "gpt-2"
-<<<<<<< HEAD
-=======
         if chkhsh == "32d85c31273f8019248f2559fed492d929ea28b17e51d81d3bb36fff23ca72b3":
             # ref: https://huggingface.co/stabilityai/stablelm-2-zephyr-1_6b
             res = "stablelm2"
->>>>>>> uupstream/master
         if chkhsh == "6221ad2852e85ce96f791f476e0b390cf9b474c9e3d1362f53a24a06dc8220ff":
             # ref: https://huggingface.co/smallcloudai/Refact-1_6-base
             res = "refact"
         if chkhsh == "9c2227e4dd922002fb81bde4fc02b0483ca4f12911410dee2255e4987644e3f8":
             # ref: https://huggingface.co/CohereForAI/c4ai-command-r-v01
             res = "command-r"
-<<<<<<< HEAD
-        if chkhsh == "b6dc8df998e1cfbdc4eac8243701a65afe638679230920b50d6f17d81c098166":
-            # ref: https://huggingface.co/allenai/OLMo-1.7-7B-hf
-            res = "olmo"
-=======
         if chkhsh == "e636dc30a262dcc0d8c323492e32ae2b70728f4df7dfe9737d9f920a282b8aea":
             # ref: https://huggingface.co/Qwen/Qwen1.5-7B
             res = "qwen2"
@@ -537,7 +478,6 @@ class Model:
         if chkhsh == "c136ed14d01c2745d4f60a9596ae66800e2b61fa45643e72436041855ad4089d":
             # ref: https://huggingface.co/abacusai/Smaug-Llama-3-70B-Instruct
             res = "smaug-bpe"
->>>>>>> uupstream/master
 
         if res is None:
             logger.warning("\n")
@@ -558,10 +498,7 @@ class Model:
         logger.debug(f"chkhsh: {chkhsh}")
 
         return res
-<<<<<<< HEAD
-=======
         # Marker: End get_vocab_base_pre
->>>>>>> uupstream/master
 
     def _set_vocab_gpt2(self) -> None:
         tokens, toktypes, tokpre = self.get_vocab_base()
@@ -677,28 +614,17 @@ class Model:
                         logger.warning(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
                         continue
 
-<<<<<<< HEAD
-=======
                     tokens[token_id] = key.encode("utf-8")
                     scores[token_id] = -1000.0
                     toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
 
->>>>>>> uupstream/master
         if vocab_size > len(tokens):
             pad_count = vocab_size - len(tokens)
             logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]")
             for i in range(1, pad_count + 1):
-<<<<<<< HEAD
-                tokens.append(f"[PAD{i}]")
-                scores.append(-1000.0)
-                toktypes.append(SentencePieceTokenTypes.UNUSED)
-
-        assert len(tokens) == vocab_size
-=======
                 tokens.append(bytes(f"[PAD{i}]", encoding="utf-8"))
                 scores.append(-1000.0)
                 toktypes.append(SentencePieceTokenTypes.UNUSED)
->>>>>>> uupstream/master
 
         self.gguf_writer.add_tokenizer_model("llama")
         self.gguf_writer.add_tokenizer_pre("default")
@@ -848,67 +774,11 @@ class BloomModel(Model):
         if name == "word_embeddings.weight":
             assert self.tensor_names is not None
 
-<<<<<<< HEAD
-            if re.match(r"h\.\d+\.self_attention\.query_key_value\.weight", name):
-                # Map bloom-style qkv_linear to gpt-style qkv_linear
-                # bloom: https://github.com/huggingface/transformers/blob/main/src/transformers/models/bloom/modeling_bloom.py#L238-L252  # noqa
-                # gpt-2: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt2/modeling_gpt2.py#L312  # noqa
-                qkv_weights = data.reshape((n_head, 3, n_embed // n_head, n_embed))
-                data = np.concatenate(
-                    (
-                        qkv_weights[:, 0, :, :].reshape((-1, n_embed)),
-                        qkv_weights[:, 1, :, :].reshape((-1, n_embed)),
-                        qkv_weights[:, 2, :, :].reshape((-1, n_embed)),
-                    ),
-                    axis=0,
-                )
-                logger.info("re-format attention.linear_qkv.weight")
-            elif re.match(r"h\.\d+\.self_attention\.query_key_value\.bias", name):
-                qkv_bias = data.reshape((n_head, 3, n_embed // n_head))
-                data = np.concatenate(
-                    (
-                        qkv_bias[:, 0, :].reshape((n_embed,)),
-                        qkv_bias[:, 1, :].reshape((n_embed,)),
-                        qkv_bias[:, 2, :].reshape((n_embed,)),
-                    ),
-                    axis=0,
-                )
-                logger.info("re-format attention.linear_qkv.bias")
-
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                raise ValueError(f"Can not map tensor {name!r}")
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            # if f32 desired, convert any float16 to float32
-            if self.ftype == 0 and data_dtype == np.float16:
-                data = data.astype(np.float32)
-
-            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-                data = data.astype(np.float16)
-
-            logger.info(f"=> {new_name}, shape = {data.shape}, {old_dtype} --> {data.dtype}")
-
-            self.gguf_writer.add_tensor(new_name, data)
-
-            if not has_lm_head and name == "word_embeddings.weight":
-                self.gguf_writer.add_tensor("output.weight", data)
-                logger.info(name, f"=> output.weight, shape = {data.shape}, {old_dtype} --> {data.dtype}")
-=======
             # TODO: tie them at runtime, don't duplicate in the model file
             if all(s not in self.tensor_names for s in ("lm_head.weight", "output.weight")):
                 tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch))
 
         return tensors
->>>>>>> uupstream/master
 
 
 @Model.register("MPTForCausalLM")
@@ -953,44 +823,7 @@ class MPTModel(Model):
         else:
             new_name = self.map_tensor_name(name, try_suffixes=(".weight", ".bias"))
 
-<<<<<<< HEAD
-            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
-                data_torch = data_torch.to(torch.float32)
-
-            data = data_torch.squeeze().numpy()
-
-            # map tensor names
-            if "scales" in name:
-                new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias", ".scales"))
-                if new_name is not None:
-                    new_name = new_name.replace("scales", "act.scales")
-            else:
-                new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                raise ValueError(f"Can not map tensor {name!r}")
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            # if f32 desired, convert any float16 to float32
-            if self.ftype == 0 and data_dtype == np.float16:
-                data = data.astype(np.float32)
-
-            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-                data = data.astype(np.float16)
-
-            logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-
-            self.gguf_writer.add_tensor(new_name, data)
-=======
         return [(new_name, data_torch)]
->>>>>>> uupstream/master
 
 
 @Model.register("OrionForCausalLM")
@@ -1030,51 +863,6 @@ class OrionModel(Model):
         # ref:  https://huggingface.co/OrionStarAI/Orion-14B-Chat/blob/276a17221ce42beb45f66fac657a41540e71f4f5/modeling_orion.py#L570-L571
         self.gguf_writer.add_layer_norm_eps(self.hparams["rms_norm_eps"])
 
-<<<<<<< HEAD
-    def write_tensors(self):
-        # Collect tensors from generator object
-        model_kv = dict(self.get_tensors())
-        block_count = self.hparams["num_hidden_layers"]
-        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
-
-        for name, data_torch in model_kv.items():
-            # we don't need these
-            if name.endswith(".rotary_emb.inv_freq"):
-                continue
-
-            old_dtype = data_torch.dtype
-
-            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
-                data_torch = data_torch.to(torch.float32)
-
-            data = data_torch.squeeze().numpy()
-
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                raise ValueError(f"Can not map tensor {name!r}")
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            # if f32 desired, convert any float16 to float32
-            if self.ftype == 0 and data_dtype == np.float16:
-                data = data.astype(np.float32)
-
-            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-                data = data.astype(np.float16)
-
-            logger.info(f"{name} -> {new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-            self.gguf_writer.add_tensor(new_name, data)
-
-=======
->>>>>>> uupstream/master
 
 @Model.register("BaichuanForCausalLM", "BaiChuanForCausalLM")
 class BaichuanModel(Model):
@@ -1121,20 +909,7 @@ class BaichuanModel(Model):
         head_count = self.hparams["num_attention_heads"]
         head_count_kv = self.hparams.get("num_key_value_heads", head_count)
 
-<<<<<<< HEAD
-        for i in range(block_count):
-            if (w := model_kv.get(f"model.layers.{i}.self_attn.W_pack.weight")) is not None:
-                logger.info(f"Unpacking and permuting layer {i}")
-                model_kv[f"model.layers.{i}.self_attn.q_proj.weight"] = \
-                    self._reverse_hf_permute_part(w, 0, head_count, head_count)
-                model_kv[f"model.layers.{i}.self_attn.k_proj.weight"] = \
-                    self._reverse_hf_permute_part(w, 1, head_count, head_count_kv)
-                model_kv[f"model.layers.{i}.self_attn.v_proj.weight"] = \
-                    self._reverse_hf_part(w, 2)
-                del model_kv[f"model.layers.{i}.self_attn.W_pack.weight"]
-=======
         tensors: list[tuple[str, Tensor]] = []
->>>>>>> uupstream/master
 
         if bid is not None and name == f"model.layers.{bid}.self_attn.W_pack.weight":
             logger.info(f"Unpacking and permuting layer {bid}")
@@ -1149,40 +924,7 @@ class BaichuanModel(Model):
         else:
             tensors = [(self.map_tensor_name(name), data_torch)]
 
-<<<<<<< HEAD
-            old_dtype = data_torch.dtype
-
-            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
-                data_torch = data_torch.to(torch.float32)
-
-            data = data_torch.squeeze().numpy()
-
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                raise ValueError(f"Can not map tensor {name!r}")
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            # if f32 desired, convert any float16 to float32
-            if self.ftype == 0 and data_dtype == np.float16:
-                data = data.astype(np.float32)
-
-            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-                data = data.astype(np.float16)
-
-            logger.info(f"{name} -> {new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-            self.gguf_writer.add_tensor(new_name, data)
-=======
         return tensors
->>>>>>> uupstream/master
 
     def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor:
         if n_kv_head is not None and n_head != n_kv_head:
@@ -1298,46 +1040,7 @@ class XverseModel(Model):
         if name.endswith("k_proj.weight"):
             data_torch = self._reverse_hf_permute(data_torch, head_count, head_count_kv)
 
-<<<<<<< HEAD
-            old_dtype = data_torch.dtype
-
-            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
-                data_torch = data_torch.to(torch.float32)
-
-            # HF models permute some of the tensors, so we need to undo that
-            if name.endswith(("q_proj.weight")):
-                data_torch = self._reverse_hf_permute(data_torch, head_count, head_count)
-            if name.endswith(("k_proj.weight")):
-                data_torch = self._reverse_hf_permute(data_torch, head_count, head_count_kv)
-
-            data = data_torch.squeeze().numpy()
-
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                raise ValueError(f"Can not map tensor {name!r}")
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            # if f32 desired, convert any float16 to float32
-            if self.ftype == 0 and data_dtype == np.float16:
-                data = data.astype(np.float32)
-
-            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-                data = data.astype(np.float16)
-
-            logger.info(f"{name} -> {new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-            self.gguf_writer.add_tensor(new_name, data)
-=======
         return [(self.map_tensor_name(name), data_torch)]
->>>>>>> uupstream/master
 
     def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor:
         if n_kv_head is not None and n_head != n_kv_head:
@@ -1402,59 +1105,7 @@ class FalconModel(Model):
             v = qkv[:, [-1]].reshape(n_head_kv * head_dim, head_dim * n_head)
             data_torch = torch.cat((q, k, v)).reshape_as(data_torch)
 
-<<<<<<< HEAD
-        for name, data_torch in self.get_tensors():
-            old_dtype = data_torch.dtype
-
-            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
-                data_torch = data_torch.to(torch.float32)
-
-            # QKV tensor transform
-            # The original query_key_value tensor contains n_head_kv "kv groups",
-            # each consisting of n_head/n_head_kv query weights followed by one key
-            # and one value weight (shared by all query heads in the kv group).
-            # This layout makes it a big pain to work with in GGML.
-            # So we rearrange them here,, so that we have n_head query weights
-            # followed by n_head_kv key weights followed by n_head_kv value weights,
-            # in contiguous fashion.
-            # ref: https://github.com/jploski/ggml/blob/falcon40b/examples/falcon/convert-hf-to-ggml.py
-
-            if "query_key_value" in name:
-                qkv = data_torch.view(n_head_kv, n_head // n_head_kv + 2, head_dim, head_dim * n_head)
-                q = qkv[:, :-2].reshape(n_head * head_dim, head_dim * n_head)
-                k = qkv[:, [-2]].reshape(n_head_kv * head_dim, head_dim * n_head)
-                v = qkv[:, [-1]].reshape(n_head_kv * head_dim, head_dim * n_head)
-                data_torch = torch.cat((q, k, v)).reshape_as(data_torch)
-
-            data = data_torch.squeeze().numpy()
-
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                raise ValueError(f"Can not map tensor {name!r}")
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            # if f32 desired, convert any float16 to float32
-            if self.ftype == 0 and data_dtype == np.float16:
-                data = data.astype(np.float32)
-
-            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-                data = data.astype(np.float16)
-
-            logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-
-            self.gguf_writer.add_tensor(new_name, data)
-=======
         return [(self.map_tensor_name(name), data_torch)]
->>>>>>> uupstream/master
 
 
 @Model.register("GPTBigCodeForCausalLM")
@@ -1537,89 +1188,7 @@ class RefactModel(Model):
         if len(tensors) == 0:
             tensors.append((self.map_tensor_name(name), data_torch))
 
-<<<<<<< HEAD
-            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
-                data_torch = data_torch.to(torch.float32)
-
-            data = data_torch.squeeze().numpy()
-
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight",))
-            if new_name is None:
-                raise ValueError(f"Can not map tensor {name!r}")
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            # if f32 desired, convert any float16 to float32
-            if self.ftype == 0 and data_dtype == np.float16:
-                data = data.astype(np.float32)
-
-            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-                data = data.astype(np.float16)
-
-            logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-
-            self.gguf_writer.add_tensor(new_name, data)
-
-
-@Model.register("PersimmonForCausalLM")
-class PersimmonModel(Model):
-    model_arch = gguf.MODEL_ARCH.PERSIMMON
-
-    def set_gguf_parameters(self):
-        block_count = self.hparams.get("num_layers", self.hparams.get("num_hidden_layers"))
-        head_count = self.hparams["num_attention_heads"]
-        head_count_kv = head_count
-        hidden_size = self.hparams["hidden_size"]
-
-        self.gguf_writer.add_name('persimmon-8b-chat')
-        self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
-        self.gguf_writer.add_embedding_length(hidden_size)
-        self.gguf_writer.add_block_count(block_count)
-        self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
-
-        # NOTE: not sure about this change - why does the model not have a rope dimension count when it is smaller
-        #       than the head size?
-        #       ref: https://github.com/ggerganov/llama.cpp/pull/4889
-        # self.gguf_writer.add_rope_dimension_count(hidden_size // head_count)
-        self.gguf_writer.add_rope_dimension_count(hidden_size // head_count // 2)
-
-        self.gguf_writer.add_head_count(head_count)
-        self.gguf_writer.add_head_count_kv(head_count_kv)
-        self.gguf_writer.add_rope_freq_base(self.hparams["rope_theta"])
-        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"])
-
-    def set_vocab(self):
-        self._set_vocab_sentencepiece()
-        # self.gguf_writer.add_bos_token_id(71013)
-        # self.gguf_writer.add_eos_token_id(71013)
-
-    def write_tensors(self):
-        block_count = self.hparams.get("num_layers", self.hparams.get("num_hidden_layers"))
-        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
-
-        for name, data_torch in self.get_tensors():
-            if name.endswith(".self_attention.rotary_emb.inv_freq"):
-                continue
-            old_dtype = data_torch.dtype
-            # TODO: FP16 conversion produces garbage outputs. (Q8_0 does not, so..?)
-            data = data_torch.to(torch.float32).squeeze().numpy()
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                raise ValueError(f"Can not map tensor {name!r}")
-            n_dims = len(data.shape)
-            logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-            self.gguf_writer.add_tensor(new_name, data)
-=======
         return tensors
->>>>>>> uupstream/master
 
 
 @Model.register("StableLmForCausalLM", "StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM")
@@ -1829,20 +1398,6 @@ class LlamaModel(Model):
                 self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
                 self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
 
-<<<<<<< HEAD
-    # Same as super class, but permuting q_proj, k_proj
-    def write_tensors(self):
-        block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
-        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
-        n_head = self.hparams.get("num_attention_heads")
-        n_kv_head = self.hparams.get("num_key_value_heads")
-        n_experts = self.hparams.get("num_local_experts")
-        experts = dict()
-        for name, data_torch in self.get_tensors():
-            # we don't need these
-            if name.endswith((".attention.masked_bias", ".attention.bias", ".rotary_emb.inv_freq")):
-                continue
-=======
     @staticmethod
     def permute(weights: Tensor, n_head: int, n_head_kv: int | None):
         if n_head_kv is not None and n_head != n_head_kv:
@@ -1856,7 +1411,6 @@ class LlamaModel(Model):
     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")
->>>>>>> uupstream/master
 
         if name.endswith("q_proj.weight"):
             data_torch = LlamaModel.permute(data_torch, n_head, n_head)
@@ -1892,58 +1446,21 @@ class LlamaModel(Model):
 
                     new_name = self.map_tensor_name(merged_name)
 
-<<<<<<< HEAD
-                            new_name = tensor_map.get_name(merged_name, try_suffixes=(".weight", ".bias"))
-                            if new_name is None:
-                                raise ValueError(f"Can not map tensor {name!r}")
-
-                            logger.info(f"{new_name}, n_dims = {len(data.shape)}, shape = {data.shape} --> {data.dtype}")
-=======
                     tensors.append((new_name, data_torch))
                 return tensors
             else:
                 return []
 
         return [(self.map_tensor_name(name), data_torch)]
->>>>>>> uupstream/master
 
     def write_tensors(self):
         super().write_tensors()
 
-<<<<<<< HEAD
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                raise ValueError(f"Can not map tensor {name!r}")
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            # if f32 desired, convert any float16 to float32
-            if self.ftype == 0 and data_dtype == np.float16:
-                data = data.astype(np.float32)
-
-            # 1d tensors need to be converted to float32
-            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-                data = data.astype(np.float16)
-
-            logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-
-            self.gguf_writer.add_tensor(new_name, data)
-
-        if len(experts) > 0:
-            raise ValueError(f"Unprocessed experts: {experts.keys()}")
-=======
         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}")
->>>>>>> uupstream/master
 
 
 @Model.register("GrokForCausalLM")
@@ -1990,30 +1507,15 @@ class GrokModel(Model):
 
                     merged_name = f"transformer.decoder_layer.{bid}.moe.{wid}.weight"
 
-<<<<<<< HEAD
-                            new_name = tensor_map.get_name(merged_name, try_suffixes=(".weight", ".bias"))
-                            if new_name is None:
-                                raise ValueError(f"Can not map tensor {name!r}")
-
-                            logger.info(f"{new_name}, n_dims = {len(data.shape)}, shape = {data.shape} --> {data.dtype}")
-=======
                     new_name = self.map_tensor_name(merged_name)
 
                     tensors.append((new_name, data_torch))
                 return tensors
             else:
                 return []
->>>>>>> uupstream/master
 
         return [(self.map_tensor_name(name), data_torch)]
 
-<<<<<<< HEAD
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                raise ValueError(f"Can not map tensor {name!r}")
-=======
->>>>>>> uupstream/master
 
 @Model.register("DbrxForCausalLM")
 class DbrxModel(Model):
@@ -2032,201 +1534,7 @@ class DbrxModel(Model):
         self.gguf_writer.add_head_count(self.hparams["n_heads"])
         self.gguf_writer.add_head_count_kv(attn_config["kv_n_heads"])
 
-<<<<<<< HEAD
-            logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-
-            self.gguf_writer.add_tensor(new_name, data)
-
-
-@Model.register("DbrxForCausalLM")
-class DbrxModel(Model):
-    model_arch = gguf.MODEL_ARCH.DBRX
-
-    def set_gguf_parameters(self):
-        ffn_config = self.hparams["ffn_config"]
-        attn_config = self.hparams["attn_config"]
-        self.gguf_writer.add_name(self.hparams["model_type"])
-        self.gguf_writer.add_block_count(self.hparams["n_layers"])
-
-        self.gguf_writer.add_context_length(self.hparams["max_seq_len"])
-        self.gguf_writer.add_embedding_length(self.hparams["d_model"])
-        self.gguf_writer.add_feed_forward_length(ffn_config["ffn_hidden_size"])
-
-        self.gguf_writer.add_head_count(self.hparams["n_heads"])
-        self.gguf_writer.add_head_count_kv(attn_config["kv_n_heads"])
-
-        self.gguf_writer.add_rope_freq_base(attn_config["rope_theta"])
-
-        self.gguf_writer.add_clamp_kqv(attn_config["clip_qkv"])
-        self.gguf_writer.add_file_type(self.ftype)
-
-        self.gguf_writer.add_expert_count(ffn_config["moe_num_experts"])
-        self.gguf_writer.add_expert_used_count(ffn_config["moe_top_k"])
-
-        self.gguf_writer.add_layer_norm_eps(1e-5)
-
-        self.gguf_writer.add_file_type(self.ftype)
-        logger.info(f"gguf: file type = {self.ftype}")
-
-    def write_tensors(self):
-        block_count = self.hparams.get("n_layers")
-        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
-        for name, data_torch in self.get_tensors():
-            n_expert = self.hparams["ffn_config"]["moe_num_experts"]
-            n_ff = self.hparams["ffn_config"]["ffn_hidden_size"]
-            n_embd = self.hparams["d_model"]
-
-            # Specific behavior for experts tensors: suffix .weight, view as 3D and transpose
-            # original implementation expects (n_expert, n_ff, n_embd) for all experts weights
-            # But llama.cpp moe graph works differently
-            # AND the dimensions in ggml are typically in the reverse order of the pytorch dimensions
-            # so (n_expert, n_ff, n_embd) in pytorch is {n_embd, n_ff, n_expert} in ggml_tensor
-            exp_tensor_names = {"ffn.experts.mlp.w1": None,       # LLM_TENSOR_FFN_GATE_EXPS ggml_tensor->ne{n_embd, n_ff,   n_expert}
-                                "ffn.experts.mlp.w2": (0, 2, 1),  # LLM_TENSOR_FFN_DOWN_EXPS ggml_tensor->ne{n_ff,   n_embd, n_expert}
-                                "ffn.experts.mlp.v1": None}       # LLM_TENSOR_FFN_UP_EXPS   ggml_tensor->ne{n_embd, n_ff,   n_expert}
-            experts = False
-            for exp_tensor_name in exp_tensor_names.keys():
-                if name.find(exp_tensor_name) != -1 and name.find(".weight") == -1:
-                    experts = True
-                    data_torch = data_torch.view(n_expert, n_ff, n_embd)
-                    if (permute_tensor := exp_tensor_names[exp_tensor_name]) is not None:
-                        data_torch = data_torch.permute(*permute_tensor)
-                    break
-
-            old_dtype = data_torch.dtype
-
-            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
-                data_torch = data_torch.to(torch.float32)
-
-            data = data_torch.squeeze().numpy()
-
-            # map tensor names
-            # In MoE models the ffn tensors are typically most of the model weights,
-            # and need to be quantizable. Quantize expects tensor names to be suffixed by .weight.
-            # Every other model has the weight names ending in .weight,
-            # let's assume that is the convention which is not the case for dbrx:
-            # https://huggingface.co/databricks/dbrx-instruct/blob/main/model.safetensors.index.json#L15
-            new_name = tensor_map.get_name(name if not experts else name + ".weight", try_suffixes=(".weight",))
-            if new_name is None:
-                raise ValueError(f"Can not map tensor {name!r}")
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            # Most of the codebase that takes in 1D tensors only handles F32 tensors
-            # and most of the outputs tensors are F32.
-            if data_dtype != np.float32 and n_dims == 1:
-                raise ValueError(f"Can not map tensor {name!r}: all 1D tensors must be F32")
-
-            # if f32 desired, convert any float16 to float32
-            if self.ftype == 0 and data_dtype == np.float16:
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and n_dims > 1:
-                data = data.astype(np.float16)
-
-            logger.debug(f"{new_name}, n_dims = {n_dims}, shape = {data.shape}, {old_dtype} --> {data.dtype}")
-
-            self.gguf_writer.add_tensor(new_name, data)
-
-
-@Model.register("DbrxForCausalLM")
-class DbrxModel(Model):
-    model_arch = gguf.MODEL_ARCH.DBRX
-
-    def set_gguf_parameters(self):
-        ffn_config = self.hparams["ffn_config"]
-        attn_config = self.hparams["attn_config"]
-        self.gguf_writer.add_name(self.hparams["model_type"])
-        self.gguf_writer.add_block_count(self.hparams["n_layers"])
-
-        self.gguf_writer.add_context_length(self.hparams["max_seq_len"])
-        self.gguf_writer.add_embedding_length(self.hparams["d_model"])
-        self.gguf_writer.add_feed_forward_length(ffn_config["ffn_hidden_size"])
-
-        self.gguf_writer.add_head_count(self.hparams["n_heads"])
-        self.gguf_writer.add_head_count_kv(attn_config["kv_n_heads"])
-
-        self.gguf_writer.add_rope_freq_base(attn_config["rope_theta"])
-
-        self.gguf_writer.add_clamp_kqv(attn_config["clip_qkv"])
-        self.gguf_writer.add_file_type(self.ftype)
-
-        self.gguf_writer.add_expert_count(ffn_config["moe_num_experts"])
-        self.gguf_writer.add_expert_used_count(ffn_config["moe_top_k"])
-
-        self.gguf_writer.add_layer_norm_eps(1e-5)
-
-        self.gguf_writer.add_file_type(self.ftype)
-        print(f"gguf: file type = {self.ftype}")
-
-    def write_tensors(self):
-        block_count = self.hparams.get("n_layers")
-        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
-        for name, data_torch in self.get_tensors():
-            n_expert = self.hparams["ffn_config"]["moe_num_experts"]
-            n_ff = self.hparams["ffn_config"]["ffn_hidden_size"]
-            n_embd = self.hparams["d_model"]
-
-            # Specific behavior for experts tensors: suffix .weight, view as 3D and transpose
-            # original implementation expects (n_expert, n_ff, n_embd) for all experts weights
-            # But llama.cpp moe graph works differently
-            # AND the dimensions in ggml are typically in the reverse order of the pytorch dimensions
-            # so (n_expert, n_ff, n_embd) in pytorch is {n_embd, n_ff, n_expert} in ggml_tensor
-            exp_tensor_names = {"ffn.experts.mlp.w1": None,       # LLM_TENSOR_FFN_GATE_EXPS ggml_tensor->ne{n_embd, n_ff,   n_expert}
-                                "ffn.experts.mlp.w2": (0, 2, 1),  # LLM_TENSOR_FFN_DOWN_EXPS ggml_tensor->ne{n_ff,   n_embd, n_expert}
-                                "ffn.experts.mlp.v1": None}       # LLM_TENSOR_FFN_UP_EXPS   ggml_tensor->ne{n_embd, n_ff,   n_expert}
-            experts = False
-            for exp_tensor_name in exp_tensor_names.keys():
-                if name.find(exp_tensor_name) != -1 and name.find(".weight") == -1:
-                    experts = True
-                    data_torch = data_torch.view(n_expert, n_ff, n_embd)
-                    if (permute_tensor := exp_tensor_names[exp_tensor_name]) is not None:
-                        data_torch = data_torch.permute(*permute_tensor)
-                    break
-
-            old_dtype = data_torch.dtype
-
-            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
-                data_torch = data_torch.to(torch.float32)
-
-            data = data_torch.squeeze().numpy()
-
-            # map tensor names
-            # In MoE models the ffn tensors are typically most of the model weights,
-            # and need to be quantizable. Quantize expects tensor names to be suffixed by .weight.
-            # Every other model has the weight names ending in .weight,
-            # let's assume that is the convention which is not the case for dbrx:
-            # https://huggingface.co/databricks/dbrx-instruct/blob/main/model.safetensors.index.json#L15
-            new_name = tensor_map.get_name(name if not experts else name + ".weight", try_suffixes=(".weight",))
-            if new_name is None:
-                print(f"Can not map tensor {name!r}")
-                sys.exit()
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            # Most of the codebase that takes in 1D tensors only handles F32 tensors
-            # and most of the outputs tensors are F32.
-            if data_dtype != np.float32 and n_dims == 1:
-                print(f"Can not map tensor {name!r}: all 1D tensors must be F32")
-                sys.exit()
-
-            # if f32 desired, convert any float16 to float32
-            if self.ftype == 0 and data_dtype == np.float16:
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and n_dims > 1:
-                data = data.astype(np.float16)
-
-            print(f"{new_name}, n_dims = {n_dims}, shape = {data.shape}, {old_dtype} --> {data.dtype}")
-=======
         self.gguf_writer.add_rope_freq_base(attn_config["rope_theta"])
->>>>>>> uupstream/master
 
         self.gguf_writer.add_clamp_kqv(attn_config["clip_qkv"])
         self.gguf_writer.add_file_type(self.ftype)
@@ -2322,45 +1630,7 @@ class MiniCPMModel(Model):
         if name.endswith(("k_proj.weight")):
             data_torch = self._reverse_hf_permute(data_torch, n_head, n_kv_head)
 
-<<<<<<< HEAD
-            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
-                data_torch = data_torch.to(torch.float32)
-
-            # HF models permute some of the tensors, so we need to undo that
-            if name.endswith(("q_proj.weight")):
-                data_torch = self._reverse_hf_permute(data_torch, n_head, n_head)
-            if name.endswith(("k_proj.weight")):
-                data_torch = self._reverse_hf_permute(data_torch, n_head, n_kv_head)
-
-            data = data_torch.squeeze().numpy()
-
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                raise ValueError(f"Can not map tensor {name!r}")
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            # if f32 desired, convert any float16 to float32
-            if self.ftype == 0 and data_dtype == np.float16:
-                data = data.astype(np.float32)
-
-            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-                data = data.astype(np.float16)
-
-            logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-
-            self.gguf_writer.add_tensor(new_name, data)
-=======
         return [(self.map_tensor_name(name), data_torch)]
->>>>>>> uupstream/master
 
 
 @Model.register("QWenLMHeadModel")
@@ -2403,50 +1673,7 @@ class QwenModel(Model):
         self.gguf_writer.add_rope_dimension_count(self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
         self.gguf_writer.add_head_count(self.hparams["num_attention_heads"])
         self.gguf_writer.add_layer_norm_rms_eps(self.hparams["layer_norm_epsilon"])
-<<<<<<< HEAD
-
-    def write_tensors(self):
-        block_count = self.hparams["num_hidden_layers"]
-        model_kv = dict(self.get_tensors())
-        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
-        for name, data_torch in model_kv.items():
-            # we don't need these
-            if name.endswith(".rotary_emb.inv_freq"):
-                continue
-
-            old_dtype = data_torch.dtype
-
-            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
-                data_torch = data_torch.to(torch.float32)
-
-            data = data_torch.squeeze().numpy()
-
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                raise ValueError(f"Can not map tensor {name!r}")
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            # if f32 desired, convert any float16 to float32
-            if self.ftype == 0 and data_dtype == np.float16:
-                data = data.astype(np.float32)
-
-            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-                data = data.astype(np.float16)
-
-            logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-            self.gguf_writer.add_tensor(new_name, data)
-=======
         self.gguf_writer.add_file_type(self.ftype)
->>>>>>> uupstream/master
 
 
 @Model.register("Qwen2ForCausalLM")
@@ -2469,193 +1696,6 @@ class Qwen2MoeModel(Model):
         if (n_experts := self.hparams.get("num_experts")) is not None:
             self.gguf_writer.add_expert_count(n_experts)
 
-<<<<<<< HEAD
-    def write_tensors(self):
-        block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
-        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
-        n_experts = self.hparams.get("num_experts")
-        experts = dict()
-        for name, data_torch in self.get_tensors():
-            # we don't need these
-            if name.endswith((".attention.masked_bias", ".attention.bias", ".attention.rotary_emb.inv_freq")):
-                continue
-
-            old_dtype = data_torch.dtype
-
-            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
-                data_torch = data_torch.to(torch.float32)
-
-            data = data_torch.squeeze().numpy()
-
-            # process the experts separately
-            if name.find("experts") != -1:
-                experts[name] = data
-                if len(experts) >= n_experts * 3:
-                    # merge the experts into a single 3d tensor
-                    for bid in range(block_count):
-                        for w_name in ["down_proj", "gate_proj", "up_proj"]:
-                            full = True
-                            for xid in range(n_experts):
-                                ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
-                                if ename not in experts:
-                                    full = False
-                                    break
-                            if not full:
-                                continue
-
-                            datas = []
-                            for xid in range(n_experts):
-                                ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
-                                datas.append(experts[ename])
-                                del experts[ename]
-
-                            data = np.stack(datas, axis=0)
-                            data_dtype = data.dtype
-
-                            if self.ftype == 0 and data_dtype == np.float16:
-                                data = data.astype(np.float32)
-
-                            if self.ftype == 1 and data_dtype == np.float32:
-                                data = data.astype(np.float16)
-
-                            merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
-
-                            new_name = tensor_map.get_name(merged_name, try_suffixes=(".weight", ".bias"))
-                            if new_name is None:
-                                raise ValueError(f"Can not map tensor {name!r}")
-
-                            logger.debug(f"{new_name}, n_dims = {len(data.shape)}, shape = {data.shape} --> {data.dtype}")
-
-                            self.gguf_writer.add_tensor(new_name, data)
-                continue
-
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                raise ValueError(f"Can not map tensor {name!r}")
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            # if f32 desired, convert any float16 to float32
-            if self.ftype == 0 and data_dtype == np.float16:
-                data = data.astype(np.float32)
-
-            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-            if self.ftype == 1 and data_dtype == np.float16 and (n_dims == 1 or new_name.endswith("_norm.weight")):
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-                data = data.astype(np.float16)
-
-            logger.debug(f"{new_name}, n_dims = {n_dims}, shape = {data.shape}, {old_dtype} --> {data.dtype}")
-
-            self.gguf_writer.add_tensor(new_name, data)
-
-        if len(experts) > 0:
-            raise ValueError(f"Unprocessed experts: {experts.keys()}")
-
-
-@Model.register("Qwen2MoeForCausalLM")
-class Qwen2MoeModel(Model):
-    model_arch = gguf.MODEL_ARCH.QWEN2MOE
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        if (n_experts := self.hparams.get("num_experts")) is not None:
-            self.gguf_writer.add_expert_count(n_experts)
-
-    def write_tensors(self):
-        block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
-        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
-        n_experts = self.hparams.get("num_experts")
-        experts = dict()
-        for name, data_torch in self.get_tensors():
-            # we don't need these
-            if name.endswith((".attention.masked_bias", ".attention.bias", ".attention.rotary_emb.inv_freq")):
-                continue
-
-            old_dtype = data_torch.dtype
-
-            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
-                data_torch = data_torch.to(torch.float32)
-
-            data = data_torch.squeeze().numpy()
-
-            # process the experts separately
-            if name.find("experts") != -1:
-                experts[name] = data
-                if len(experts) >= n_experts * 3:
-                    # merge the experts into a single 3d tensor
-                    for bid in range(block_count):
-                        for w_name in ["down_proj", "gate_proj", "up_proj"]:
-                            full = True
-                            for xid in range(n_experts):
-                                ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
-                                if ename not in experts:
-                                    full = False
-                                    break
-                            if not full:
-                                continue
-
-                            datas = []
-                            for xid in range(n_experts):
-                                ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
-                                datas.append(experts[ename])
-                                del experts[ename]
-
-                            data = np.stack(datas, axis=0)
-                            data_dtype = data.dtype
-
-                            if self.ftype == 0 and data_dtype == np.float16:
-                                data = data.astype(np.float32)
-
-                            if self.ftype == 1 and data_dtype == np.float32:
-                                data = data.astype(np.float16)
-
-                            merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
-
-                            new_name = tensor_map.get_name(merged_name, try_suffixes=(".weight", ".bias"))
-                            if new_name is None:
-                                print(f"Can not map tensor {name!r}")
-                                sys.exit()
-
-                            print(f"{new_name}, n_dims = {len(data.shape)}, shape = {data.shape} --> {data.dtype}")
-
-                            self.gguf_writer.add_tensor(new_name, data)
-                continue
-
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                print(f"Can not map tensor {name!r}")
-                sys.exit()
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            # if f32 desired, convert any float16 to float32
-            if self.ftype == 0 and data_dtype == np.float16:
-                data = data.astype(np.float32)
-
-            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-            if self.ftype == 1 and data_dtype == np.float16 and (n_dims == 1 or new_name.endswith("_norm.weight")):
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-                data = data.astype(np.float16)
-
-            print(f"{new_name}, n_dims = {n_dims}, shape = {data.shape}, {old_dtype} --> {data.dtype}")
-
-            self.gguf_writer.add_tensor(new_name, data)
-
-        if len(experts) > 0:
-            raise ValueError(f"Unprocessed experts: {experts.keys()}")
-=======
     _experts: list[dict[str, Tensor]] | None = None
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
@@ -2702,7 +1742,6 @@ class Qwen2MoeModel(Model):
             experts = [k for d in self._experts for k in d.keys()]
             if len(experts) > 0:
                 raise ValueError(f"Unprocessed experts: {experts}")
->>>>>>> uupstream/master
 
 
 @Model.register("GPT2LMHeadModel")
@@ -2735,42 +1774,11 @@ class GPT2Model(Model):
 
         tensors.append((new_name, data_torch))
 
-<<<<<<< HEAD
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                raise ValueError(f"Can not map tensor {name!r}")
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            # if f32 desired, convert any float16 to float32
-            if self.ftype == 0 and data_dtype == np.float16:
-                data = data.astype(np.float32)
-
-            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-                data = data.astype(np.float16)
-
-            logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-
-            self.gguf_writer.add_tensor(new_name, data)
-
-            # note: GPT2 output is tied to (same as) wte in original model
-            if new_name == "token_embd.weight":
-                logger.info(f"output.weight, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-                self.gguf_writer.add_tensor("output.weight", data)
-=======
         # note: GPT2 output is tied to (same as) wte in original model
         if new_name == self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD):
             tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch))
 
         return tensors
->>>>>>> uupstream/master
 
 
 @Model.register("PhiForCausalLM")
@@ -2810,12 +1818,8 @@ class Phi3MiniModel(Model):
         if not tokenizer_path.is_file():
             raise ValueError(f'Error: Missing {tokenizer_path}')
 
-<<<<<<< HEAD
-        tokenizer = SentencePieceProcessor(str(tokenizer_path))
-=======
         tokenizer = SentencePieceProcessor()
         tokenizer.LoadFromFile(str(tokenizer_path))
->>>>>>> uupstream/master
 
         vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
 
@@ -2825,20 +1829,6 @@ class Phi3MiniModel(Model):
 
         for token_id in range(tokenizer.vocab_size()):
 
-<<<<<<< HEAD
-            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):
-=======
             piece = tokenizer.IdToPiece(token_id)
             text = piece.encode("utf-8")
             score = tokenizer.GetScore(token_id)
@@ -2851,7 +1841,6 @@ class Phi3MiniModel(Model):
             elif tokenizer.IsUnused(token_id):
                 toktype = SentencePieceTokenTypes.UNUSED
             elif tokenizer.IsByte(token_id):
->>>>>>> uupstream/master
                 toktype = SentencePieceTokenTypes.BYTE
 
             tokens[token_id] = text
@@ -2873,8 +1862,6 @@ class Phi3MiniModel(Model):
                     scores[token_id] = -1000.0
                     toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
 
-<<<<<<< HEAD
-=======
         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:
@@ -2907,7 +1894,6 @@ class Phi3MiniModel(Model):
                     if foken_data.get("special"):
                         toktypes[token_id] = SentencePieceTokenTypes.CONTROL
 
->>>>>>> uupstream/master
         self.gguf_writer.add_tokenizer_model("llama")
         self.gguf_writer.add_tokenizer_pre("default")
         self.gguf_writer.add_token_list(tokens)
@@ -2920,25 +1906,6 @@ class Phi3MiniModel(Model):
     def set_gguf_parameters(self):
         block_count = self.find_hparam(["num_hidden_layers", "n_layer"])
 
-<<<<<<< HEAD
-        rot_pct = 1.0
-        n_embd = self.find_hparam(["hidden_size", "n_embd"])
-        n_head = self.find_hparam(["num_attention_heads", "n_head"])
-        rms_eps = self.find_hparam(["rms_norm_eps"])
-
-        self.gguf_writer.add_name("Phi3")
-        self.gguf_writer.add_context_length(self.find_hparam(["n_positions", "max_position_embeddings"]))
-
-        self.gguf_writer.add_embedding_length(n_embd)
-        self.gguf_writer.add_feed_forward_length(8192)
-        self.gguf_writer.add_block_count(block_count)
-        self.gguf_writer.add_head_count(n_head)
-        self.gguf_writer.add_head_count_kv(n_head)
-        self.gguf_writer.add_layer_norm_rms_eps(rms_eps)
-        self.gguf_writer.add_rope_dimension_count(int(rot_pct * n_embd) // n_head)
-        self.gguf_writer.add_file_type(self.ftype)
-
-=======
         n_embd = self.find_hparam(["hidden_size", "n_embd"])
         n_head = self.find_hparam(["num_attention_heads", "n_head"])
         n_head_kv = self.find_hparam(["num_key_value_heads", "n_head_kv"])
@@ -2992,7 +1959,6 @@ class Phi3MiniModel(Model):
         self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_LONG]  + ".weight", np.array(long_factors, dtype=np.float32))
         self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT] + ".weight", np.array(short_factors, dtype=np.float32))
 
->>>>>>> uupstream/master
 
 @Model.register("PlamoForCausalLM")
 class PlamoModel(Model):
@@ -3034,45 +2000,6 @@ class PlamoModel(Model):
 
         new_name = self.map_tensor_name(name)
 
-<<<<<<< HEAD
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                raise ValueError(f"Can not map tensor {name!r}")
-
-            # shuffle for broadcasting of gqa in ggml_mul_mat
-            if new_name.endswith("attn_q.weight"):
-                data_torch = self.shuffle_attn_q_weight(data_torch)
-            elif new_name.endswith("attn_output.weight"):
-                data_torch = self.shuffle_attn_output_weight(data_torch)
-
-            old_dtype = data_torch.dtype
-
-            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
-                data_torch = data_torch.to(torch.float32)
-
-            data = data_torch.squeeze().numpy()
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            # if f32 desired, convert any float16 to float32
-            if self.ftype == 0 and data_dtype == np.float16:
-                data = data.astype(np.float32)
-
-            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-                data = data.astype(np.float16)
-
-            logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-
-            self.gguf_writer.add_tensor(new_name, data)
-=======
         # shuffle for broadcasting of gqa in ggml_mul_mat
         if new_name.endswith("attn_q.weight"):
             data_torch = self.shuffle_attn_q_weight(data_torch)
@@ -3080,7 +2007,6 @@ class PlamoModel(Model):
             data_torch = self.shuffle_attn_output_weight(data_torch)
 
         return [(new_name, data_torch)]
->>>>>>> uupstream/master
 
 
 @Model.register("CodeShellForCausalLM")
@@ -3113,41 +2039,11 @@ class CodeShellModel(Model):
         if new_name == self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD):
             assert self.tensor_names is not None
 
-<<<<<<< HEAD
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                raise ValueError(f"Can not map tensor {name!r}")
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            # if f32 desired, convert any float16 to float32
-            if self.ftype == 0 and data_dtype == np.float16:
-                data = data.astype(np.float32)
-
-            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-                data = data.astype(np.float16)
-
-            logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-
-            self.gguf_writer.add_tensor(new_name, data)
-
-            if not has_lm_head and name == "transformer.wte.weight":
-                self.gguf_writer.add_tensor("output.weight", data)
-                logger.info(name, f"=> output.weight, shape = {data.shape}, {old_dtype} --> {data.dtype}")
-=======
             if all(s not in self.tensor_names for s in ("lm_head.weight", "output.weight")):
                 # copy tok_embd.weight to output.weight
                 tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch))
 
         return tensors
->>>>>>> uupstream/master
 
 
 @Model.register("InternLM2ForCausalLM")
@@ -3188,13 +2084,8 @@ class InternLM2Model(Model):
             if text == b"\x00":
                 # (TODO): fixme
                 # Hack here and replace the \x00 characters.
-<<<<<<< HEAD
-                logger.debug(f"InternLM2 convert token '{text}' to '🐉'!")
-                text = "🐉"
-=======
                 logger.warning(f"InternLM2 convert token '{text}' to '🐉'!")
                 text = "🐉".encode("utf-8")
->>>>>>> uupstream/master
 
             toktype = SentencePieceTokenTypes.NORMAL
             if tokenizer.IsUnknown(token_id):
@@ -3270,51 +2161,10 @@ in chat mode so that the conversation can end normally.")
         self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"])
         self.gguf_writer.add_file_type(self.ftype)
 
-<<<<<<< HEAD
-    def post_write_tensors(self, tensor_map, name, data_torch):
-        old_dtype = data_torch.dtype
-
-        # convert any unsupported data types to float32
-        if data_torch.dtype not in (torch.float16, torch.float32):
-            data_torch = data_torch.to(torch.float32)
-
-        data = data_torch.squeeze().numpy()
-
-        # map tensor names
-        new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-        if new_name is None:
-            raise ValueError(f"Can not map tensor {name!r}")
-
-        n_dims = len(data.shape)
-        data_dtype = data.dtype
-
-        # if f32 desired, convert any float16 to float32
-        if self.ftype == 0 and data_dtype == np.float16:
-            data = data.astype(np.float32)
-
-        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-        if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-            data = data.astype(np.float32)
-
-        # if f16 desired, convert any float32 2-dim weight tensors to float16
-        if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-            data = data.astype(np.float16)
-
-        logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-        self.gguf_writer.add_tensor(new_name, data)
-
-    def write_tensors(self):
-        from einops import rearrange
-
-        num_heads = self.hparams.get("num_attention_heads")
-        num_kv_heads = self.hparams.get("num_key_value_heads")
-        hidden_size = self.hparams.get("hidden_size")
-=======
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         num_heads = self.hparams["num_attention_heads"]
         num_kv_heads = self.hparams["num_key_value_heads"]
         hidden_size = self.hparams["hidden_size"]
->>>>>>> uupstream/master
         q_per_kv = num_heads // num_kv_heads
         head_dim = hidden_size // num_heads
         num_groups = num_heads // q_per_kv
@@ -3408,43 +2258,11 @@ class BertModel(Model):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused
 
-<<<<<<< HEAD
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                raise ValueError(f"Can not map tensor {name!r}")
-
-            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
-                data_torch = data_torch.to(torch.float32)
-
-            data = data_torch.squeeze().numpy()
-            n_dims = len(data.shape)
-            new_dtype: type[np.floating[Any]]
-
-            if (
-                self.ftype == 1 and name.endswith(".weight") and n_dims == 2
-                and name != "embeddings.token_type_embeddings.weight"  # not used with get_rows, must be F32
-            ):
-                # if f16 desired, convert any float32 2-dim weight tensors to float16
-                new_dtype = np.float16
-            else:
-                # if f32 desired, convert any float16 to float32
-                new_dtype = np.float32
-
-            logger.info(f"{new_name}, n_dims = {n_dims}, {data_torch.dtype} --> {new_dtype}")
-
-            if data.dtype != new_dtype:
-                data = data.astype(new_dtype)
-
-            self.gguf_writer.add_tensor(new_name, data)
-=======
         # we are only using BERT for embeddings so we don't need the pooling layer
         if name in ("embeddings.position_ids", "pooler.dense.weight", "pooler.dense.bias"):
             return [] # we don't need these
 
         return [(self.map_tensor_name(name), data_torch)]
->>>>>>> uupstream/master
 
 
 @Model.register("NomicBertModel")
@@ -3513,53 +2331,17 @@ class GemmaModel(Model):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused
 
-<<<<<<< HEAD
-        for name, data_torch in self.get_tensors():
-            # lm_head is not used in llama.cpp, while autoawq will include this tensor in model
-            # To prevent errors, skip loading lm_head.weight.
-            if name == "lm_head.weight":
-                logger.debug(f"Skipping get tensor {name!r} in safetensors so that convert can end normally.")
-                continue
-
-            old_dtype = data_torch.dtype
-=======
         # lm_head is not used in llama.cpp, while autoawq will include this tensor in model
         # To prevent errors, skip loading lm_head.weight.
         if name == "lm_head.weight":
             logger.debug(f"Skipping get tensor {name!r} in safetensors so that convert can end normally.")
             return []
->>>>>>> uupstream/master
 
         # ref: https://github.com/huggingface/transformers/blob/fc37f38915372c15992b540dfcbbe00a916d4fc6/src/transformers/models/gemma/modeling_gemma.py#L89
         if name.endswith("norm.weight"):
             data_torch = data_torch + 1
 
-<<<<<<< HEAD
-            # ref: https://github.com/huggingface/transformers/blob/fc37f38915372c15992b540dfcbbe00a916d4fc6/src/transformers/models/gemma/modeling_gemma.py#L89
-            if name.endswith("norm.weight"):
-                data_torch = data_torch + 1
-            data = data_torch.squeeze().numpy()
-
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                raise ValueError(f"Can not map tensor {name!r}")
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-                data = data.astype(np.float16)
-
-            logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-
-            self.gguf_writer.add_tensor(new_name, data)
-=======
         return [(self.map_tensor_name(name), data_torch)]
->>>>>>> uupstream/master
 
 
 @Model.register("Starcoder2ForCausalLM")
@@ -3591,17 +2373,10 @@ class MambaModel(Model):
             neox_reader = gguf.GGUFReader(tokenizer_path, "r")
 
             field = neox_reader.get_field(gguf.Keys.Tokenizer.MODEL)
-<<<<<<< HEAD
-            self.gguf_writer.add_tokenizer_model(bytes(field.parts[-1]))
-
-            field = neox_reader.get_field(gguf.Keys.Tokenizer.PRE)
-            self.gguf_writer.add_tokenizer_pre(bytes(field.parts[-1]))
-=======
             self.gguf_writer.add_tokenizer_model(bytes(field.parts[-1]).decode("utf-8") if field else "gpt2")
 
             field = neox_reader.get_field(gguf.Keys.Tokenizer.PRE)
             self.gguf_writer.add_tokenizer_pre(bytes(field.parts[-1]).decode("utf-8") if field else "mpt")
->>>>>>> uupstream/master
 
             field = neox_reader.get_field(gguf.Keys.Tokenizer.LIST)
             assert field
@@ -3616,17 +2391,10 @@ class MambaModel(Model):
             self.gguf_writer.add_token_merges([bytes(field.parts[i]) for i in field.data])
 
             field = neox_reader.get_field(gguf.Keys.Tokenizer.BOS_ID)
-<<<<<<< HEAD
-            self.gguf_writer.add_bos_token_id(field.parts[-1].tolist()[0])
-
-            field = neox_reader.get_field(gguf.Keys.Tokenizer.EOS_ID)
-            self.gguf_writer.add_eos_token_id(field.parts[-1].tolist()[0])
-=======
             self.gguf_writer.add_bos_token_id(field.parts[-1].tolist()[0] if field else 1)
 
             field = neox_reader.get_field(gguf.Keys.Tokenizer.EOS_ID)
             self.gguf_writer.add_eos_token_id(field.parts[-1].tolist()[0] if field else 0)
->>>>>>> uupstream/master
 
             field = neox_reader.get_field(gguf.Keys.Tokenizer.UNK_ID)
             self.gguf_writer.add_unk_token_id(field.parts[-1].tolist()[0] if field else 0)
@@ -3671,24 +2439,6 @@ class MambaModel(Model):
 
         new_name = self.map_tensor_name(name)
 
-<<<<<<< HEAD
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                raise ValueError(f"Can not map tensor {name!r}")
-
-            if name.endswith(".A_log"):
-                logger.debug("A_log --> A ==> " + new_name)
-                data_torch = -torch.exp(data_torch)
-
-            # assuming token_embd.weight is seen before output.weight
-            if tok_embd is not None and new_name == output_name:
-                if torch.equal(tok_embd, data_torch):
-                    logger.debug(f"{output_name} is equivalent to {tok_embd_name}, omitting")
-                    continue
-            if new_name == tok_embd_name:
-                tok_embd = data_torch
-=======
         if name.endswith(".A_log"):
             logger.debug("A_log --> A ==> " + new_name)
             data_torch = -torch.exp(data_torch)
@@ -3702,32 +2452,10 @@ class MambaModel(Model):
             self._tok_embd = data_torch
 
         return [(new_name, data_torch)]
->>>>>>> uupstream/master
 
     def extra_f32_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
         del n_dims  # unused
 
-<<<<<<< HEAD
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            # if f32 desired, convert any float16 to float32
-            if self.ftype == 0 and data_dtype == np.float16:
-                data = data.astype(np.float32)
-
-            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert big float32 2-dim weight tensors to float16
-            new_weight_name = new_name[:-len(".weight")] if new_name.endswith(".weight") else ""
-            if self.ftype == 1 and data_dtype == np.float32 and new_weight_name.endswith((".ssm_in", ".ssm_out", "token_embd", "output")) and n_dims == 2:
-                data = data.astype(np.float16)
-
-            logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-
-            self.gguf_writer.add_tensor(new_name, data)
-=======
         return bid is not None and new_name in (
             self.format_tensor_name(n, bid, ".weight" if name.endswith(".weight") else "") for n in [
                 gguf.MODEL_TENSOR.SSM_CONV1D,
@@ -3737,7 +2465,6 @@ class MambaModel(Model):
                 gguf.MODEL_TENSOR.SSM_D,
             ]
         )
->>>>>>> uupstream/master
 
 
 @Model.register("CohereForCausalLM")
@@ -3772,52 +2499,6 @@ class OlmoModel(Model):
 
     # Same as super class, but permuting q_proj, k_proj
     # Copied from: LlamaModel
-<<<<<<< HEAD
-    def write_tensors(self):
-        block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
-        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
-        n_head = self.hparams.get("num_attention_heads")
-        n_kv_head = self.hparams.get("num_key_value_heads")
-        for name, data_torch in self.get_tensors():
-            old_dtype = data_torch.dtype
-
-            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
-                data_torch = data_torch.to(torch.float32)
-
-            data = data_torch.numpy()
-
-            if name.endswith("q_proj.weight"):
-                data = permute(data, n_head, n_head)
-            if name.endswith("k_proj.weight"):
-                data = permute(data, n_head, n_kv_head)
-
-            data = data.squeeze()
-
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                raise ValueError(f"Can not map tensor {name!r}")
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            # if f32 desired, convert any float16 to float32
-            if self.ftype == 0 and data_dtype == np.float16:
-                data = data.astype(np.float32)
-
-            # 1d tensors need to be converted to float32
-            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and n_dims == 2:
-                data = data.astype(np.float16)
-
-            logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-
-            self.gguf_writer.add_tensor(new_name, data)
-=======
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused
 
@@ -4018,7 +2699,6 @@ class ArcticModel(Model):
             experts = [k for d in self._experts for k in d.keys()]
             if len(experts) > 0:
                 raise ValueError(f"Unprocessed experts: {experts}")
->>>>>>> uupstream/master
 
 
 ###### CONVERSION LOGIC ######
@@ -4090,11 +2770,6 @@ def parse_args() -> argparse.Namespace:
         "model", type=Path,
         help="directory containing model file",
     )
-<<<<<<< HEAD
-    parser.add_argument("--use-temp-file", action="store_true", help="use the tempfile library while processing (helpful when running out of memory, process killed)")
-    parser.add_argument("--model-name", type=str, default=None, help="name of the model")
-    parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
-=======
     parser.add_argument(
         "--use-temp-file", action="store_true",
         help="use the tempfile library while processing (helpful when running out of memory, process killed)",
@@ -4111,7 +2786,6 @@ def parse_args() -> argparse.Namespace:
         "--verbose", action="store_true",
         help="increase output verbosity",
     )
->>>>>>> uupstream/master
 
     return parser.parse_args()
 
@@ -4160,11 +2834,7 @@ def main() -> None:
 
     with torch.inference_mode():
         model_class = Model.from_model_architecture(hparams["architectures"][0])
-<<<<<<< HEAD
-        model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian, args.use_temp_file)
-=======
         model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian, args.use_temp_file, args.no_lazy)
->>>>>>> uupstream/master
 
         logger.info("Set model parameters")
         model_instance.set_gguf_parameters()
@@ -4175,15 +2845,6 @@ def main() -> None:
         model_instance.gguf_writer.add_quantization_version(gguf.GGML_QUANT_VERSION)
 
         if args.vocab_only:
-<<<<<<< HEAD
-            logger.info(f"Exporting model vocab to '{fname_out}'")
-            model_instance.write_vocab()
-        else:
-            logger.info(f"Exporting model to '{fname_out}'")
-            model_instance.write()
-
-        logger.info(f"Model successfully exported to '{fname_out}'")
-=======
             logger.info(f"Exporting model vocab to '{model_instance.fname_out}'")
             model_instance.write_vocab()
         else:
@@ -4191,7 +2852,6 @@ def main() -> None:
             model_instance.write()
 
         logger.info(f"Model successfully exported to '{model_instance.fname_out}'")
->>>>>>> uupstream/master
 
 
 if __name__ == '__main__':