Added test inference from exllamav2, added gguf-py from llama.cpp

gguf/__init__.py

@@ -0,0 +1,9 @@
+from .constants import *
+from .lazy import *
+from .gguf_reader import *
+from .gguf_writer import *
+from .quants import *
+from .tensor_mapping import *
+from .vocab import *
+from .utility import *
+from .metadata import *

gguf/constants.py

@@ -0,0 +1,1609 @@
+from __future__ import annotations
+
+from enum import Enum, IntEnum, auto
+from typing import Any
+
+#
+# constants
+#
+
+GGUF_MAGIC             = 0x46554747  # "GGUF"
+GGUF_VERSION           = 3
+GGUF_DEFAULT_ALIGNMENT = 32
+GGML_QUANT_VERSION     = 2  # GGML_QNT_VERSION from ggml.h
+
+#
+# metadata keys
+#
+
+
+class Keys:
+    class General:
+        TYPE                       = "general.type"
+        ARCHITECTURE               = "general.architecture"
+        QUANTIZATION_VERSION       = "general.quantization_version"
+        ALIGNMENT                  = "general.alignment"
+        FILE_TYPE                  = "general.file_type"
+
+        # Authorship Metadata
+        NAME                       = "general.name"
+        AUTHOR                     = "general.author"
+        VERSION                    = "general.version"
+        ORGANIZATION               = "general.organization"
+
+        FINETUNE                   = "general.finetune"
+        BASENAME                   = "general.basename"
+
+        DESCRIPTION                = "general.description"
+        QUANTIZED_BY               = "general.quantized_by"
+
+        SIZE_LABEL                 = "general.size_label"
+
+        # Licensing details
+        LICENSE                    = "general.license"
+        LICENSE_NAME               = "general.license.name"
+        LICENSE_LINK               = "general.license.link"
+
+        # Typically represents the converted GGUF repo (Unless native)
+        URL                        = "general.url" # Model Website/Paper
+        DOI                        = "general.doi"
+        UUID                       = "general.uuid"
+        REPO_URL                   = "general.repo_url" # Model Source Repository (git/svn/etc...)
+
+        # Model Source during conversion
+        SOURCE_URL                 = "general.source.url" # Model Website/Paper
+        SOURCE_DOI                 = "general.source.doi"
+        SOURCE_UUID                = "general.source.uuid"
+        SOURCE_REPO_URL            = "general.source.repo_url" # Model Source Repository (git/svn/etc...)
+
+        # Base Model Source. There can be more than one source if it's a merged
+        # model like with 'Mistral-7B-Merge-14-v0.1'. This will assist in
+        # tracing linage of models as it is finetuned or merged over time.
+        BASE_MODEL_COUNT           = "general.base_model.count"
+        BASE_MODEL_NAME            = "general.base_model.{id}.name"
+        BASE_MODEL_AUTHOR          = "general.base_model.{id}.author"
+        BASE_MODEL_VERSION         = "general.base_model.{id}.version"
+        BASE_MODEL_ORGANIZATION    = "general.base_model.{id}.organization"
+        BASE_MODEL_URL             = "general.base_model.{id}.url" # Model Website/Paper
+        BASE_MODEL_DOI             = "general.base_model.{id}.doi"
+        BASE_MODEL_UUID            = "general.base_model.{id}.uuid"
+        BASE_MODEL_REPO_URL        = "general.base_model.{id}.repo_url" # Model Source Repository (git/svn/etc...)
+
+        # Array based KV stores
+        TAGS                       = "general.tags"
+        LANGUAGES                  = "general.languages"
+        DATASETS                   = "general.datasets"
+
+    class LLM:
+        VOCAB_SIZE                        = "{arch}.vocab_size"
+        CONTEXT_LENGTH                    = "{arch}.context_length"
+        EMBEDDING_LENGTH                  = "{arch}.embedding_length"
+        BLOCK_COUNT                       = "{arch}.block_count"
+        LEADING_DENSE_BLOCK_COUNT         = "{arch}.leading_dense_block_count"
+        FEED_FORWARD_LENGTH               = "{arch}.feed_forward_length"
+        EXPERT_FEED_FORWARD_LENGTH        = "{arch}.expert_feed_forward_length"
+        EXPERT_SHARED_FEED_FORWARD_LENGTH = "{arch}.expert_shared_feed_forward_length"
+        USE_PARALLEL_RESIDUAL             = "{arch}.use_parallel_residual"
+        TENSOR_DATA_LAYOUT                = "{arch}.tensor_data_layout"
+        EXPERT_COUNT                      = "{arch}.expert_count"
+        EXPERT_USED_COUNT                 = "{arch}.expert_used_count"
+        EXPERT_SHARED_COUNT               = "{arch}.expert_shared_count"
+        EXPERT_WEIGHTS_SCALE              = "{arch}.expert_weights_scale"
+        POOLING_TYPE                      = "{arch}.pooling_type"
+        LOGIT_SCALE                       = "{arch}.logit_scale"
+        DECODER_START_TOKEN_ID            = "{arch}.decoder_start_token_id"
+        ATTN_LOGIT_SOFTCAPPING            = "{arch}.attn_logit_softcapping"
+        FINAL_LOGIT_SOFTCAPPING           = "{arch}.final_logit_softcapping"
+        SWIN_NORM                         = "{arch}.swin_norm"
+        RESCALE_EVERY_N_LAYERS            = "{arch}.rescale_every_n_layers"
+        TIME_MIX_EXTRA_DIM                = "{arch}.time_mix_extra_dim"
+        TIME_DECAY_EXTRA_DIM              = "{arch}.time_decay_extra_dim"
+        RESIDUAL_SCALE                    = "{arch}.residual_scale"
+        EMBEDDING_SCALE                   = "{arch}.embedding_scale"
+
+    class Attention:
+        HEAD_COUNT        = "{arch}.attention.head_count"
+        HEAD_COUNT_KV     = "{arch}.attention.head_count_kv"
+        MAX_ALIBI_BIAS    = "{arch}.attention.max_alibi_bias"
+        CLAMP_KQV         = "{arch}.attention.clamp_kqv"
+        KEY_LENGTH        = "{arch}.attention.key_length"
+        VALUE_LENGTH      = "{arch}.attention.value_length"
+        LAYERNORM_EPS     = "{arch}.attention.layer_norm_epsilon"
+        LAYERNORM_RMS_EPS = "{arch}.attention.layer_norm_rms_epsilon"
+        CAUSAL            = "{arch}.attention.causal"
+        Q_LORA_RANK       = "{arch}.attention.q_lora_rank"
+        KV_LORA_RANK      = "{arch}.attention.kv_lora_rank"
+        REL_BUCKETS_COUNT = "{arch}.attention.relative_buckets_count"
+        SLIDING_WINDOW    = "{arch}.attention.sliding_window"
+        SCALE             = "{arch}.attention.scale"
+
+    class Rope:
+        DIMENSION_COUNT         = "{arch}.rope.dimension_count"
+        FREQ_BASE               = "{arch}.rope.freq_base"
+        SCALING_TYPE            = "{arch}.rope.scaling.type"
+        SCALING_FACTOR          = "{arch}.rope.scaling.factor"
+        SCALING_ATTN_FACTOR     = "{arch}.rope.scaling.attn_factor"
+        SCALING_ORIG_CTX_LEN    = "{arch}.rope.scaling.original_context_length"
+        SCALING_FINETUNED       = "{arch}.rope.scaling.finetuned"
+        SCALING_YARN_LOG_MUL    = "{arch}.rope.scaling.yarn_log_multiplier"
+
+    class Split:
+        LLM_KV_SPLIT_NO            = "split.no"
+        LLM_KV_SPLIT_COUNT         = "split.count"
+        LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count"
+
+    class SSM:
+        CONV_KERNEL    = "{arch}.ssm.conv_kernel"
+        INNER_SIZE     = "{arch}.ssm.inner_size"
+        STATE_SIZE     = "{arch}.ssm.state_size"
+        TIME_STEP_RANK = "{arch}.ssm.time_step_rank"
+        DT_B_C_RMS     = "{arch}.ssm.dt_b_c_rms"
+
+    class WKV:
+        HEAD_SIZE = "{arch}.wkv.head_size"
+
+    class Tokenizer:
+        MODEL                = "tokenizer.ggml.model"
+        PRE                  = "tokenizer.ggml.pre"
+        LIST                 = "tokenizer.ggml.tokens"
+        TOKEN_TYPE           = "tokenizer.ggml.token_type"
+        TOKEN_TYPE_COUNT     = "tokenizer.ggml.token_type_count"  # for BERT-style token types
+        SCORES               = "tokenizer.ggml.scores"
+        MERGES               = "tokenizer.ggml.merges"
+        BOS_ID               = "tokenizer.ggml.bos_token_id"
+        EOS_ID               = "tokenizer.ggml.eos_token_id"
+        EOT_ID               = "tokenizer.ggml.eot_token_id"
+        EOM_ID               = "tokenizer.ggml.eom_token_id"
+        UNK_ID               = "tokenizer.ggml.unknown_token_id"
+        SEP_ID               = "tokenizer.ggml.seperator_token_id"
+        PAD_ID               = "tokenizer.ggml.padding_token_id"
+        CLS_ID               = "tokenizer.ggml.cls_token_id"
+        MASK_ID              = "tokenizer.ggml.mask_token_id"
+        ADD_BOS              = "tokenizer.ggml.add_bos_token"
+        ADD_EOS              = "tokenizer.ggml.add_eos_token"
+        ADD_PREFIX           = "tokenizer.ggml.add_space_prefix"
+        REMOVE_EXTRA_WS      = "tokenizer.ggml.remove_extra_whitespaces"
+        PRECOMPILED_CHARSMAP = "tokenizer.ggml.precompiled_charsmap"
+        HF_JSON              = "tokenizer.huggingface.json"
+        RWKV                 = "tokenizer.rwkv.world"
+        CHAT_TEMPLATE        = "tokenizer.chat_template"
+        CHAT_TEMPLATE_N      = "tokenizer.chat_template.{name}"
+        CHAT_TEMPLATES       = "tokenizer.chat_templates"
+        # FIM/Infill special tokens constants
+        FIM_PRE_ID           = "tokenizer.ggml.fim_pre_token_id"
+        FIM_SUF_ID           = "tokenizer.ggml.fim_suf_token_id"
+        FIM_MID_ID           = "tokenizer.ggml.fim_mid_token_id"
+        FIM_PAD_ID           = "tokenizer.ggml.fim_pad_token_id"
+        FIM_REP_ID           = "tokenizer.ggml.fim_rep_token_id"
+        FIM_SEP_ID           = "tokenizer.ggml.fim_sep_token_id"
+        # deprecated:
+        PREFIX_ID            = "tokenizer.ggml.prefix_token_id"
+        SUFFIX_ID            = "tokenizer.ggml.suffix_token_id"
+        MIDDLE_ID            = "tokenizer.ggml.middle_token_id"
+
+    class Adapter:
+        TYPE       = "adapter.type"
+        LORA_ALPHA = "adapter.lora.alpha"
+
+#
+# recommended mapping of model tensor names for storage in gguf
+#
+
+
+class GGUFType:
+    MODEL   = "model"
+    ADAPTER = "adapter"
+
+
+class MODEL_ARCH(IntEnum):
+    LLAMA        = auto()
+    FALCON       = auto()
+    BAICHUAN     = auto()
+    GROK         = auto()
+    GPT2         = auto()
+    GPTJ         = auto()
+    GPTNEOX      = auto()
+    MPT          = auto()
+    STARCODER    = auto()
+    REFACT       = auto()
+    BERT         = auto()
+    NOMIC_BERT   = auto()
+    JINA_BERT_V2 = auto()
+    BLOOM        = auto()
+    STABLELM     = auto()
+    QWEN         = auto()
+    QWEN2        = auto()
+    QWEN2MOE     = auto()
+    PHI2         = auto()
+    PHI3         = auto()
+    PLAMO        = auto()
+    CODESHELL    = auto()
+    ORION        = auto()
+    INTERNLM2    = auto()
+    MINICPM      = auto()
+    MINICPM3     = auto()
+    GEMMA        = auto()
+    GEMMA2       = auto()
+    STARCODER2   = auto()
+    RWKV6        = auto()
+    MAMBA        = auto()
+    XVERSE       = auto()
+    COMMAND_R    = auto()
+    DBRX         = auto()
+    OLMO         = auto()
+    OLMOE        = auto()
+    OPENELM      = auto()
+    ARCTIC       = auto()
+    DEEPSEEK2    = auto()
+    CHATGLM      = auto()
+    BITNET       = auto()
+    T5           = auto()
+    T5ENCODER    = auto()
+    JAIS         = auto()
+    NEMOTRON     = auto()
+    EXAONE       = auto()
+    GRANITE      = auto()
+    GRANITE_MOE  = auto()
+    CHAMELEON    = auto()
+
+
+class MODEL_TENSOR(IntEnum):
+    TOKEN_EMBD           = auto()
+    TOKEN_EMBD_NORM      = auto()
+    TOKEN_TYPES          = auto()
+    POS_EMBD             = auto()
+    OUTPUT               = auto()
+    OUTPUT_NORM          = auto()
+    ROPE_FREQS           = auto()
+    ROPE_FACTORS_LONG    = auto()
+    ROPE_FACTORS_SHORT   = auto()
+    ATTN_Q               = auto()
+    ATTN_K               = auto()
+    ATTN_V               = auto()
+    ATTN_QKV             = auto()
+    ATTN_OUT             = auto()
+    ATTN_NORM            = auto()
+    ATTN_NORM_2          = auto()
+    ATTN_OUT_NORM        = auto()
+    ATTN_POST_NORM       = auto()
+    ATTN_ROT_EMBD        = auto()
+    FFN_GATE_INP         = auto()
+    FFN_GATE_INP_SHEXP   = auto()
+    FFN_NORM             = auto()
+    FFN_PRE_NORM         = auto()
+    FFN_POST_NORM        = auto()
+    FFN_GATE             = auto()
+    FFN_DOWN             = auto()
+    FFN_UP               = auto()
+    FFN_ACT              = auto()
+    FFN_NORM_EXP         = auto()
+    FFN_GATE_EXP         = auto()
+    FFN_DOWN_EXP         = auto()
+    FFN_UP_EXP           = auto()
+    FFN_GATE_SHEXP       = auto()
+    FFN_DOWN_SHEXP       = auto()
+    FFN_UP_SHEXP         = auto()
+    ATTN_Q_NORM          = auto()
+    ATTN_K_NORM          = auto()
+    LAYER_OUT_NORM       = auto()
+    SSM_IN               = auto()
+    SSM_CONV1D           = auto()
+    SSM_X                = auto()
+    SSM_DT               = auto()
+    SSM_A                = auto()
+    SSM_D                = auto()
+    SSM_OUT              = auto()
+    TIME_MIX_W1          = auto()
+    TIME_MIX_W2          = auto()
+    TIME_MIX_LERP_X      = auto()
+    TIME_MIX_LERP_K      = auto()
+    TIME_MIX_LERP_V      = auto()
+    TIME_MIX_LERP_R      = auto()
+    TIME_MIX_LERP_G      = auto()
+    TIME_MIX_LERP_W      = auto()
+    TIME_MIX_FIRST       = auto()
+    TIME_MIX_DECAY       = auto()
+    TIME_MIX_DECAY_W1    = auto()
+    TIME_MIX_DECAY_W2    = auto()
+    TIME_MIX_KEY         = auto()
+    TIME_MIX_VALUE       = auto()
+    TIME_MIX_RECEPTANCE  = auto()
+    TIME_MIX_GATE        = auto()
+    TIME_MIX_LN          = auto()
+    TIME_MIX_OUTPUT      = auto()
+    CHANNEL_MIX_LERP_K   = auto()
+    CHANNEL_MIX_LERP_R   = auto()
+    CHANNEL_MIX_KEY      = auto()
+    CHANNEL_MIX_RECEPTANCE = auto()
+    CHANNEL_MIX_VALUE    = auto()
+    ATTN_Q_A             = auto()
+    ATTN_Q_B             = auto()
+    ATTN_KV_A_MQA        = auto()
+    ATTN_KV_B            = auto()
+    ATTN_Q_A_NORM        = auto()
+    ATTN_KV_A_NORM       = auto()
+    FFN_SUB_NORM         = auto()
+    ATTN_SUB_NORM        = auto()
+    DEC_ATTN_NORM        = auto()
+    DEC_ATTN_Q           = auto()
+    DEC_ATTN_K           = auto()
+    DEC_ATTN_V           = auto()
+    DEC_ATTN_OUT         = auto()
+    DEC_ATTN_REL_B       = auto()
+    DEC_CROSS_ATTN_NORM  = auto()
+    DEC_CROSS_ATTN_Q     = auto()
+    DEC_CROSS_ATTN_K     = auto()
+    DEC_CROSS_ATTN_V     = auto()
+    DEC_CROSS_ATTN_OUT   = auto()
+    DEC_CROSS_ATTN_REL_B = auto()
+    DEC_FFN_NORM         = auto()
+    DEC_FFN_GATE         = auto()
+    DEC_FFN_DOWN         = auto()
+    DEC_FFN_UP           = auto()
+    DEC_OUTPUT_NORM      = auto()
+    ENC_ATTN_NORM        = auto()
+    ENC_ATTN_Q           = auto()
+    ENC_ATTN_K           = auto()
+    ENC_ATTN_V           = auto()
+    ENC_ATTN_OUT         = auto()
+    ENC_ATTN_REL_B       = auto()
+    ENC_FFN_NORM         = auto()
+    ENC_FFN_GATE         = auto()
+    ENC_FFN_DOWN         = auto()
+    ENC_FFN_UP           = auto()
+    ENC_OUTPUT_NORM      = auto()
+    CLS                  = auto() # classifier
+    CLS_OUT              = auto() # classifier output projection
+
+
+MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
+    MODEL_ARCH.LLAMA:          "llama",
+    MODEL_ARCH.FALCON:         "falcon",
+    MODEL_ARCH.BAICHUAN:       "baichuan",
+    MODEL_ARCH.GROK:           "grok",
+    MODEL_ARCH.GPT2:           "gpt2",
+    MODEL_ARCH.GPTJ:           "gptj",
+    MODEL_ARCH.GPTNEOX:        "gptneox",
+    MODEL_ARCH.MPT:            "mpt",
+    MODEL_ARCH.STARCODER:      "starcoder",
+    MODEL_ARCH.REFACT:         "refact",
+    MODEL_ARCH.BERT:           "bert",
+    MODEL_ARCH.NOMIC_BERT:     "nomic-bert",
+    MODEL_ARCH.JINA_BERT_V2:   "jina-bert-v2",
+    MODEL_ARCH.BLOOM:          "bloom",
+    MODEL_ARCH.STABLELM:       "stablelm",
+    MODEL_ARCH.QWEN:           "qwen",
+    MODEL_ARCH.QWEN2:          "qwen2",
+    MODEL_ARCH.QWEN2MOE:       "qwen2moe",
+    MODEL_ARCH.PHI2:           "phi2",
+    MODEL_ARCH.PHI3:           "phi3",
+    MODEL_ARCH.PLAMO:          "plamo",
+    MODEL_ARCH.CODESHELL:      "codeshell",
+    MODEL_ARCH.ORION:          "orion",
+    MODEL_ARCH.INTERNLM2:      "internlm2",
+    MODEL_ARCH.MINICPM:        "minicpm",
+    MODEL_ARCH.MINICPM3:       "minicpm3",
+    MODEL_ARCH.GEMMA:          "gemma",
+    MODEL_ARCH.GEMMA2:         "gemma2",
+    MODEL_ARCH.STARCODER2:     "starcoder2",
+    MODEL_ARCH.RWKV6:          "rwkv6",
+    MODEL_ARCH.MAMBA:          "mamba",
+    MODEL_ARCH.XVERSE:         "xverse",
+    MODEL_ARCH.COMMAND_R:      "command-r",
+    MODEL_ARCH.DBRX:           "dbrx",
+    MODEL_ARCH.OLMO:           "olmo",
+    MODEL_ARCH.OLMOE:          "olmoe",
+    MODEL_ARCH.OPENELM:        "openelm",
+    MODEL_ARCH.ARCTIC:         "arctic",
+    MODEL_ARCH.DEEPSEEK2:      "deepseek2",
+    MODEL_ARCH.CHATGLM:        "chatglm",
+    MODEL_ARCH.BITNET:         "bitnet",
+    MODEL_ARCH.T5:             "t5",
+    MODEL_ARCH.T5ENCODER:      "t5encoder",
+    MODEL_ARCH.JAIS:           "jais",
+    MODEL_ARCH.NEMOTRON:       "nemotron",
+    MODEL_ARCH.EXAONE:         "exaone",
+    MODEL_ARCH.GRANITE:        "granite",
+    MODEL_ARCH.GRANITE_MOE:    "granitemoe",
+    MODEL_ARCH.CHAMELEON:      "chameleon",
+}
+
+TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
+    MODEL_TENSOR.TOKEN_EMBD:                "token_embd",
+    MODEL_TENSOR.TOKEN_EMBD_NORM:           "token_embd_norm",
+    MODEL_TENSOR.TOKEN_TYPES:               "token_types",
+    MODEL_TENSOR.POS_EMBD:                  "position_embd",
+    MODEL_TENSOR.OUTPUT_NORM:               "output_norm",
+    MODEL_TENSOR.OUTPUT:                    "output",
+    MODEL_TENSOR.ROPE_FREQS:                "rope_freqs",
+    MODEL_TENSOR.ROPE_FACTORS_LONG:         "rope_factors_long",
+    MODEL_TENSOR.ROPE_FACTORS_SHORT:        "rope_factors_short",
+    MODEL_TENSOR.ATTN_NORM:                 "blk.{bid}.attn_norm",
+    MODEL_TENSOR.ATTN_NORM_2:               "blk.{bid}.attn_norm_2",
+    MODEL_TENSOR.ATTN_QKV:                  "blk.{bid}.attn_qkv",
+    MODEL_TENSOR.ATTN_Q:                    "blk.{bid}.attn_q",
+    MODEL_TENSOR.ATTN_K:                    "blk.{bid}.attn_k",
+    MODEL_TENSOR.ATTN_V:                    "blk.{bid}.attn_v",
+    MODEL_TENSOR.ATTN_OUT:                  "blk.{bid}.attn_output",
+    MODEL_TENSOR.ATTN_ROT_EMBD:             "blk.{bid}.attn_rot_embd",
+    MODEL_TENSOR.ATTN_Q_NORM:               "blk.{bid}.attn_q_norm",
+    MODEL_TENSOR.ATTN_K_NORM:               "blk.{bid}.attn_k_norm",
+    MODEL_TENSOR.ATTN_OUT_NORM:             "blk.{bid}.attn_output_norm",
+    MODEL_TENSOR.ATTN_POST_NORM:            "blk.{bid}.post_attention_norm",
+    MODEL_TENSOR.FFN_GATE_INP:              "blk.{bid}.ffn_gate_inp",
+    MODEL_TENSOR.FFN_GATE_INP_SHEXP:        "blk.{bid}.ffn_gate_inp_shexp",
+    MODEL_TENSOR.FFN_NORM:                  "blk.{bid}.ffn_norm",
+    MODEL_TENSOR.FFN_PRE_NORM:              "blk.{bid}.ffn_norm",
+    MODEL_TENSOR.FFN_POST_NORM:             "blk.{bid}.post_ffw_norm",
+    MODEL_TENSOR.FFN_GATE:                  "blk.{bid}.ffn_gate",
+    MODEL_TENSOR.FFN_DOWN:                  "blk.{bid}.ffn_down",
+    MODEL_TENSOR.FFN_UP:                    "blk.{bid}.ffn_up",
+    MODEL_TENSOR.FFN_GATE_SHEXP:            "blk.{bid}.ffn_gate_shexp",
+    MODEL_TENSOR.FFN_DOWN_SHEXP:            "blk.{bid}.ffn_down_shexp",
+    MODEL_TENSOR.FFN_UP_SHEXP:              "blk.{bid}.ffn_up_shexp",
+    MODEL_TENSOR.FFN_ACT:                   "blk.{bid}.ffn",
+    MODEL_TENSOR.FFN_NORM_EXP:              "blk.{bid}.ffn_norm_exps",
+    MODEL_TENSOR.FFN_GATE_EXP:              "blk.{bid}.ffn_gate_exps",
+    MODEL_TENSOR.FFN_DOWN_EXP:              "blk.{bid}.ffn_down_exps",
+    MODEL_TENSOR.FFN_UP_EXP:                "blk.{bid}.ffn_up_exps",
+    MODEL_TENSOR.LAYER_OUT_NORM:            "blk.{bid}.layer_output_norm",
+    MODEL_TENSOR.SSM_IN:                    "blk.{bid}.ssm_in",
+    MODEL_TENSOR.SSM_CONV1D:                "blk.{bid}.ssm_conv1d",
+    MODEL_TENSOR.SSM_X:                     "blk.{bid}.ssm_x",
+    MODEL_TENSOR.SSM_DT:                    "blk.{bid}.ssm_dt",
+    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.TIME_MIX_W1:               "blk.{bid}.time_mix_w1",
+    MODEL_TENSOR.TIME_MIX_W2:               "blk.{bid}.time_mix_w2",
+    MODEL_TENSOR.TIME_MIX_LERP_X:           "blk.{bid}.time_mix_lerp_x",
+    MODEL_TENSOR.TIME_MIX_LERP_K:           "blk.{bid}.time_mix_lerp_k",
+    MODEL_TENSOR.TIME_MIX_LERP_V:           "blk.{bid}.time_mix_lerp_v",
+    MODEL_TENSOR.TIME_MIX_LERP_R:           "blk.{bid}.time_mix_lerp_r",
+    MODEL_TENSOR.TIME_MIX_LERP_G:           "blk.{bid}.time_mix_lerp_g",
+    MODEL_TENSOR.TIME_MIX_LERP_W:           "blk.{bid}.time_mix_lerp_w",
+    MODEL_TENSOR.TIME_MIX_FIRST:            "blk.{bid}.time_mix_first",
+    MODEL_TENSOR.TIME_MIX_DECAY:            "blk.{bid}.time_mix_decay",
+    MODEL_TENSOR.TIME_MIX_DECAY_W1:         "blk.{bid}.time_mix_decay_w1",
+    MODEL_TENSOR.TIME_MIX_DECAY_W2:         "blk.{bid}.time_mix_decay_w2",
+    MODEL_TENSOR.TIME_MIX_KEY:              "blk.{bid}.time_mix_key",
+    MODEL_TENSOR.TIME_MIX_VALUE:            "blk.{bid}.time_mix_value",
+    MODEL_TENSOR.TIME_MIX_RECEPTANCE:       "blk.{bid}.time_mix_receptance",
+    MODEL_TENSOR.TIME_MIX_GATE:             "blk.{bid}.time_mix_gate",
+    MODEL_TENSOR.TIME_MIX_LN:               "blk.{bid}.time_mix_ln",
+    MODEL_TENSOR.TIME_MIX_OUTPUT:           "blk.{bid}.time_mix_output",
+    MODEL_TENSOR.CHANNEL_MIX_LERP_K:        "blk.{bid}.channel_mix_lerp_k",
+    MODEL_TENSOR.CHANNEL_MIX_LERP_R:        "blk.{bid}.channel_mix_lerp_r",
+    MODEL_TENSOR.CHANNEL_MIX_KEY:           "blk.{bid}.channel_mix_key",
+    MODEL_TENSOR.CHANNEL_MIX_RECEPTANCE:    "blk.{bid}.channel_mix_receptance",
+    MODEL_TENSOR.CHANNEL_MIX_VALUE:         "blk.{bid}.channel_mix_value",
+    MODEL_TENSOR.ATTN_Q_A:                  "blk.{bid}.attn_q_a",
+    MODEL_TENSOR.ATTN_Q_B:                  "blk.{bid}.attn_q_b",
+    MODEL_TENSOR.ATTN_KV_A_MQA:             "blk.{bid}.attn_kv_a_mqa",
+    MODEL_TENSOR.ATTN_KV_B:                 "blk.{bid}.attn_kv_b",
+    MODEL_TENSOR.ATTN_Q_A_NORM:             "blk.{bid}.attn_q_a_norm",
+    MODEL_TENSOR.ATTN_KV_A_NORM:            "blk.{bid}.attn_kv_a_norm",
+    MODEL_TENSOR.ATTN_SUB_NORM:             "blk.{bid}.attn_sub_norm",
+    MODEL_TENSOR.FFN_SUB_NORM:              "blk.{bid}.ffn_sub_norm",
+    MODEL_TENSOR.DEC_ATTN_NORM:             "dec.blk.{bid}.attn_norm",
+    MODEL_TENSOR.DEC_ATTN_Q:                "dec.blk.{bid}.attn_q",
+    MODEL_TENSOR.DEC_ATTN_K:                "dec.blk.{bid}.attn_k",
+    MODEL_TENSOR.DEC_ATTN_V:                "dec.blk.{bid}.attn_v",
+    MODEL_TENSOR.DEC_ATTN_OUT:              "dec.blk.{bid}.attn_o",
+    MODEL_TENSOR.DEC_ATTN_REL_B:            "dec.blk.{bid}.attn_rel_b",
+    MODEL_TENSOR.DEC_CROSS_ATTN_NORM:       "dec.blk.{bid}.cross_attn_norm",
+    MODEL_TENSOR.DEC_CROSS_ATTN_Q:          "dec.blk.{bid}.cross_attn_q",
+    MODEL_TENSOR.DEC_CROSS_ATTN_K:          "dec.blk.{bid}.cross_attn_k",
+    MODEL_TENSOR.DEC_CROSS_ATTN_V:          "dec.blk.{bid}.cross_attn_v",
+    MODEL_TENSOR.DEC_CROSS_ATTN_OUT:        "dec.blk.{bid}.cross_attn_o",
+    MODEL_TENSOR.DEC_CROSS_ATTN_REL_B:      "dec.blk.{bid}.cross_attn_rel_b",
+    MODEL_TENSOR.DEC_FFN_NORM:              "dec.blk.{bid}.ffn_norm",
+    MODEL_TENSOR.DEC_FFN_GATE:              "dec.blk.{bid}.ffn_gate",
+    MODEL_TENSOR.DEC_FFN_DOWN:              "dec.blk.{bid}.ffn_down",
+    MODEL_TENSOR.DEC_FFN_UP:                "dec.blk.{bid}.ffn_up",
+    MODEL_TENSOR.DEC_OUTPUT_NORM:           "dec.output_norm",
+    MODEL_TENSOR.ENC_ATTN_NORM:             "enc.blk.{bid}.attn_norm",
+    MODEL_TENSOR.ENC_ATTN_Q:                "enc.blk.{bid}.attn_q",
+    MODEL_TENSOR.ENC_ATTN_K:                "enc.blk.{bid}.attn_k",
+    MODEL_TENSOR.ENC_ATTN_V:                "enc.blk.{bid}.attn_v",
+    MODEL_TENSOR.ENC_ATTN_OUT:              "enc.blk.{bid}.attn_o",
+    MODEL_TENSOR.ENC_ATTN_REL_B:            "enc.blk.{bid}.attn_rel_b",
+    MODEL_TENSOR.ENC_FFN_NORM:              "enc.blk.{bid}.ffn_norm",
+    MODEL_TENSOR.ENC_FFN_GATE:              "enc.blk.{bid}.ffn_gate",
+    MODEL_TENSOR.ENC_FFN_DOWN:              "enc.blk.{bid}.ffn_down",
+    MODEL_TENSOR.ENC_FFN_UP:                "enc.blk.{bid}.ffn_up",
+    MODEL_TENSOR.ENC_OUTPUT_NORM:           "enc.output_norm",
+    MODEL_TENSOR.CLS:                       "cls",
+    MODEL_TENSOR.CLS_OUT:                   "cls.output",
+}
+
+MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
+    MODEL_ARCH.LLAMA: [
+        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_ARCH.GROK: [
+        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.ATTN_OUT_NORM,
+        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.LAYER_OUT_NORM,
+    ],
+    MODEL_ARCH.GPTNEOX: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.FALCON: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_NORM_2,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.BAICHUAN: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.STARCODER: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.POS_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.BERT: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.TOKEN_EMBD_NORM,
+        MODEL_TENSOR.TOKEN_TYPES,
+        MODEL_TENSOR.POS_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_OUT_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.LAYER_OUT_NORM,
+        MODEL_TENSOR.CLS,
+        MODEL_TENSOR.CLS_OUT,
+    ],
+    MODEL_ARCH.NOMIC_BERT: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.TOKEN_EMBD_NORM,
+        MODEL_TENSOR.TOKEN_TYPES,
+        MODEL_TENSOR.POS_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_OUT_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.LAYER_OUT_NORM,
+    ],
+    MODEL_ARCH.JINA_BERT_V2: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.TOKEN_EMBD_NORM,
+        MODEL_TENSOR.TOKEN_TYPES,
+        MODEL_TENSOR.ATTN_NORM_2,
+        MODEL_TENSOR.ATTN_OUT_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.LAYER_OUT_NORM,
+        MODEL_TENSOR.CLS,
+    ],
+    MODEL_ARCH.MPT: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_ACT,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.POS_EMBD,
+    ],
+    MODEL_ARCH.GPTJ: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.REFACT: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.BLOOM: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.TOKEN_EMBD_NORM,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.STABLELM: [
+        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.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K_NORM,
+    ],
+    MODEL_ARCH.QWEN: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.QWEN2: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.QWEN2MOE: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_GATE_INP_SHEXP,
+        MODEL_TENSOR.FFN_GATE_SHEXP,
+        MODEL_TENSOR.FFN_DOWN_SHEXP,
+        MODEL_TENSOR.FFN_UP_SHEXP,
+    ],
+    MODEL_ARCH.PLAMO: [
+        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,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.GPT2: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.POS_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.PHI2: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.PHI3: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FACTORS_LONG,
+        MODEL_TENSOR.ROPE_FACTORS_SHORT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.CODESHELL: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.POS_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.ORION: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.INTERNLM2: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.MINICPM: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.OUTPUT_NORM,
+        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_ARCH.MINICPM3: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FACTORS_LONG,
+        MODEL_TENSOR.ROPE_FACTORS_SHORT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q_A,
+        MODEL_TENSOR.ATTN_Q_B,
+        MODEL_TENSOR.ATTN_KV_A_MQA,
+        MODEL_TENSOR.ATTN_KV_B,
+        MODEL_TENSOR.ATTN_Q_A_NORM,
+        MODEL_TENSOR.ATTN_KV_A_NORM,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.GEMMA: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_NORM,
+    ],
+    MODEL_ARCH.GEMMA2: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_POST_NORM,
+        MODEL_TENSOR.FFN_PRE_NORM,
+        MODEL_TENSOR.FFN_POST_NORM,
+    ],
+    MODEL_ARCH.STARCODER2: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.RWKV6: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.TOKEN_EMBD_NORM,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_NORM_2,
+        MODEL_TENSOR.TIME_MIX_W1,
+        MODEL_TENSOR.TIME_MIX_W2,
+        MODEL_TENSOR.TIME_MIX_LERP_X,
+        MODEL_TENSOR.TIME_MIX_LERP_K,
+        MODEL_TENSOR.TIME_MIX_LERP_V,
+        MODEL_TENSOR.TIME_MIX_LERP_R,
+        MODEL_TENSOR.TIME_MIX_LERP_G,
+        MODEL_TENSOR.TIME_MIX_LERP_W,
+        MODEL_TENSOR.TIME_MIX_FIRST,
+        MODEL_TENSOR.TIME_MIX_DECAY,
+        MODEL_TENSOR.TIME_MIX_DECAY_W1,
+        MODEL_TENSOR.TIME_MIX_DECAY_W2,
+        MODEL_TENSOR.TIME_MIX_KEY,
+        MODEL_TENSOR.TIME_MIX_VALUE,
+        MODEL_TENSOR.TIME_MIX_RECEPTANCE,
+        MODEL_TENSOR.TIME_MIX_GATE,
+        MODEL_TENSOR.TIME_MIX_LN,
+        MODEL_TENSOR.TIME_MIX_OUTPUT,
+        MODEL_TENSOR.CHANNEL_MIX_LERP_K,
+        MODEL_TENSOR.CHANNEL_MIX_LERP_R,
+        MODEL_TENSOR.CHANNEL_MIX_KEY,
+        MODEL_TENSOR.CHANNEL_MIX_RECEPTANCE,
+        MODEL_TENSOR.CHANNEL_MIX_VALUE,
+    ],
+    MODEL_ARCH.MAMBA: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.SSM_IN,
+        MODEL_TENSOR.SSM_CONV1D,
+        MODEL_TENSOR.SSM_X,
+        MODEL_TENSOR.SSM_DT,
+        MODEL_TENSOR.SSM_A,
+        MODEL_TENSOR.SSM_D,
+        MODEL_TENSOR.SSM_OUT,
+    ],
+    MODEL_ARCH.XVERSE: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.COMMAND_R: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_Q_NORM,
+    ],
+    MODEL_ARCH.DBRX: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_OUT_NORM,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+    ],
+    MODEL_ARCH.OLMO: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.OLMOE: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+    ],
+    MODEL_ARCH.OPENELM: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        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_NORM_EXP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+    ],
+    MODEL_ARCH.DEEPSEEK2: [
+        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_Q_A,
+        MODEL_TENSOR.ATTN_Q_B,
+        MODEL_TENSOR.ATTN_KV_A_MQA,
+        MODEL_TENSOR.ATTN_KV_B,
+        MODEL_TENSOR.ATTN_Q_A_NORM,
+        MODEL_TENSOR.ATTN_KV_A_NORM,
+        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_GATE_SHEXP,
+        MODEL_TENSOR.FFN_DOWN_SHEXP,
+        MODEL_TENSOR.FFN_UP_SHEXP,
+    ],
+    MODEL_ARCH.CHATGLM : [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.BITNET: [
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.ATTN_SUB_NORM,
+        MODEL_TENSOR.FFN_SUB_NORM,
+    ],
+    MODEL_ARCH.T5: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.DEC_ATTN_NORM,
+        MODEL_TENSOR.DEC_ATTN_Q,
+        MODEL_TENSOR.DEC_ATTN_K,
+        MODEL_TENSOR.DEC_ATTN_V,
+        MODEL_TENSOR.DEC_ATTN_OUT,
+        MODEL_TENSOR.DEC_ATTN_REL_B,
+        MODEL_TENSOR.DEC_CROSS_ATTN_NORM,
+        MODEL_TENSOR.DEC_CROSS_ATTN_Q,
+        MODEL_TENSOR.DEC_CROSS_ATTN_K,
+        MODEL_TENSOR.DEC_CROSS_ATTN_V,
+        MODEL_TENSOR.DEC_CROSS_ATTN_OUT,
+        MODEL_TENSOR.DEC_CROSS_ATTN_REL_B,
+        MODEL_TENSOR.DEC_FFN_NORM,
+        MODEL_TENSOR.DEC_FFN_GATE,
+        MODEL_TENSOR.DEC_FFN_DOWN,
+        MODEL_TENSOR.DEC_FFN_UP,
+        MODEL_TENSOR.DEC_OUTPUT_NORM,
+        MODEL_TENSOR.ENC_ATTN_NORM,
+        MODEL_TENSOR.ENC_ATTN_Q,
+        MODEL_TENSOR.ENC_ATTN_K,
+        MODEL_TENSOR.ENC_ATTN_V,
+        MODEL_TENSOR.ENC_ATTN_OUT,
+        MODEL_TENSOR.ENC_ATTN_REL_B,
+        MODEL_TENSOR.ENC_FFN_NORM,
+        MODEL_TENSOR.ENC_FFN_GATE,
+        MODEL_TENSOR.ENC_FFN_DOWN,
+        MODEL_TENSOR.ENC_FFN_UP,
+        MODEL_TENSOR.ENC_OUTPUT_NORM,
+    ],
+    MODEL_ARCH.T5ENCODER: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ENC_ATTN_NORM,
+        MODEL_TENSOR.ENC_ATTN_Q,
+        MODEL_TENSOR.ENC_ATTN_K,
+        MODEL_TENSOR.ENC_ATTN_V,
+        MODEL_TENSOR.ENC_ATTN_OUT,
+        MODEL_TENSOR.ENC_ATTN_REL_B,
+        MODEL_TENSOR.ENC_FFN_NORM,
+        MODEL_TENSOR.ENC_FFN_GATE,
+        MODEL_TENSOR.ENC_FFN_DOWN,
+        MODEL_TENSOR.ENC_FFN_UP,
+        MODEL_TENSOR.ENC_OUTPUT_NORM,
+    ],
+    MODEL_ARCH.JAIS: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.NEMOTRON: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.EXAONE: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.GRANITE: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.GRANITE_MOE: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+    ],
+    MODEL_ARCH.CHAMELEON: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    # TODO
+}
+
+# tensors that will not be serialized
+MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
+    MODEL_ARCH.LLAMA: [
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+    ],
+    MODEL_ARCH.BAICHUAN: [
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+    ],
+    MODEL_ARCH.QWEN: [
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+    ],
+    MODEL_ARCH.CODESHELL: [
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+    ],
+    MODEL_ARCH.ORION: [
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+    ],
+    MODEL_ARCH.STARCODER2: [
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+    ],
+    MODEL_ARCH.XVERSE: [
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+    ],
+    MODEL_ARCH.DEEPSEEK2: [
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+    ],
+    MODEL_ARCH.CHATGLM: [
+        MODEL_TENSOR.ROPE_FREQS,
+    ],
+    MODEL_ARCH.NEMOTRON: [
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+    ],
+}
+
+#
+# types
+#
+
+
+class TokenType(IntEnum):
+    NORMAL       = 1
+    UNKNOWN      = 2
+    CONTROL      = 3
+    USER_DEFINED = 4
+    UNUSED       = 5
+    BYTE         = 6
+
+
+class RopeScalingType(Enum):
+    NONE   = 'none'
+    LINEAR = 'linear'
+    YARN   = 'yarn'
+
+
+class PoolingType(IntEnum):
+    NONE = 0
+    MEAN = 1
+    CLS  = 2
+
+
+class GGMLQuantizationType(IntEnum):
+    F32     = 0
+    F16     = 1
+    Q4_0    = 2
+    Q4_1    = 3
+    Q5_0    = 6
+    Q5_1    = 7
+    Q8_0    = 8
+    Q8_1    = 9
+    Q2_K    = 10
+    Q3_K    = 11
+    Q4_K    = 12
+    Q5_K    = 13
+    Q6_K    = 14
+    Q8_K    = 15
+    IQ2_XXS = 16
+    IQ2_XS  = 17
+    IQ3_XXS = 18
+    IQ1_S   = 19
+    IQ4_NL  = 20
+    IQ3_S   = 21
+    IQ2_S   = 22
+    IQ4_XS  = 23
+    I8      = 24
+    I16     = 25
+    I32     = 26
+    I64     = 27
+    F64     = 28
+    IQ1_M   = 29
+    BF16    = 30
+    Q4_0_4_4 = 31
+    Q4_0_4_8 = 32
+    Q4_0_8_8 = 33
+    TQ1_0   = 34
+    TQ2_0   = 35
+
+
+# TODO: add GGMLFileType from ggml_ftype in ggml.h
+
+
+# from llama_ftype in llama.h
+# ALL VALUES SHOULD BE THE SAME HERE AS THEY ARE OVER THERE.
+class LlamaFileType(IntEnum):
+    ALL_F32              = 0
+    MOSTLY_F16           = 1   # except 1d tensors
+    MOSTLY_Q4_0          = 2   # except 1d tensors
+    MOSTLY_Q4_1          = 3   # except 1d tensors
+    # MOSTLY_Q4_1_SOME_F16 = 4   # tok_embeddings.weight and output.weight are F16
+    # MOSTLY_Q4_2        = 5   # support has been removed
+    # MOSTLY_Q4_3        = 6   # support has been removed
+    MOSTLY_Q8_0          = 7   # except 1d tensors
+    MOSTLY_Q5_0          = 8   # except 1d tensors
+    MOSTLY_Q5_1          = 9   # except 1d tensors
+    MOSTLY_Q2_K          = 10  # except 1d tensors
+    MOSTLY_Q3_K_S        = 11  # except 1d tensors
+    MOSTLY_Q3_K_M        = 12  # except 1d tensors
+    MOSTLY_Q3_K_L        = 13  # except 1d tensors
+    MOSTLY_Q4_K_S        = 14  # except 1d tensors
+    MOSTLY_Q4_K_M        = 15  # except 1d tensors
+    MOSTLY_Q5_K_S        = 16  # except 1d tensors
+    MOSTLY_Q5_K_M        = 17  # except 1d tensors
+    MOSTLY_Q6_K          = 18  # except 1d tensors
+    MOSTLY_IQ2_XXS       = 19  # except 1d tensors
+    MOSTLY_IQ2_XS        = 20  # except 1d tensors
+    MOSTLY_Q2_K_S        = 21  # except 1d tensors
+    MOSTLY_IQ3_XS        = 22  # except 1d tensors
+    MOSTLY_IQ3_XXS       = 23  # except 1d tensors
+    MOSTLY_IQ1_S         = 24  # except 1d tensors
+    MOSTLY_IQ4_NL        = 25  # except 1d tensors
+    MOSTLY_IQ3_S         = 26  # except 1d tensors
+    MOSTLY_IQ3_M         = 27  # except 1d tensors
+    MOSTLY_IQ2_S         = 28  # except 1d tensors
+    MOSTLY_IQ2_M         = 29  # except 1d tensors
+    MOSTLY_IQ4_XS        = 30  # except 1d tensors
+    MOSTLY_IQ1_M         = 31  # except 1d tensors
+    MOSTLY_BF16          = 32  # except 1d tensors
+    MOSTLY_Q4_0_4_4      = 33  # except 1d tensors
+    MOSTLY_Q4_0_4_8      = 34  # except 1d tensors
+    MOSTLY_Q4_0_8_8      = 35  # except 1d tensors
+    MOSTLY_TQ1_0         = 36  # except 1d tensors
+    MOSTLY_TQ2_0         = 37  # except 1d tensors
+
+    GUESSED              = 1024  # not specified in the model file
+
+
+class GGUFEndian(IntEnum):
+    LITTLE = 0
+    BIG = 1
+
+
+class GGUFValueType(IntEnum):
+    UINT8   = 0
+    INT8    = 1
+    UINT16  = 2
+    INT16   = 3
+    UINT32  = 4
+    INT32   = 5
+    FLOAT32 = 6
+    BOOL    = 7
+    STRING  = 8
+    ARRAY   = 9
+    UINT64  = 10
+    INT64   = 11
+    FLOAT64 = 12
+
+    @staticmethod
+    def get_type(val: Any) -> GGUFValueType:
+        if isinstance(val, (str, bytes, bytearray)):
+            return GGUFValueType.STRING
+        elif isinstance(val, list):
+            return GGUFValueType.ARRAY
+        elif isinstance(val, float):
+            return GGUFValueType.FLOAT32
+        elif isinstance(val, bool):
+            return GGUFValueType.BOOL
+        elif isinstance(val, int):
+            return GGUFValueType.INT32
+        # TODO: need help with 64-bit types in Python
+        else:
+            raise ValueError(f"Unknown type: {type(val)}")
+
+
+# Items here are (block size, type size)
+QK_K = 256
+GGML_QUANT_SIZES: dict[GGMLQuantizationType, tuple[int, int]] = {
+    GGMLQuantizationType.F32:     (1, 4),
+    GGMLQuantizationType.F16:     (1, 2),
+    GGMLQuantizationType.Q4_0:    (32, 2 + 16),
+    GGMLQuantizationType.Q4_1:    (32, 2 + 2 + 16),
+    GGMLQuantizationType.Q5_0:    (32, 2 + 4 + 16),
+    GGMLQuantizationType.Q5_1:    (32, 2 + 2 + 4 + 16),
+    GGMLQuantizationType.Q8_0:    (32, 2 + 32),
+    GGMLQuantizationType.Q8_1:    (32, 4 + 4 + 32),
+    GGMLQuantizationType.Q2_K:    (256, 2 + 2 + QK_K // 16 + QK_K // 4),
+    GGMLQuantizationType.Q3_K:    (256, 2 + QK_K // 4 + QK_K // 8 + 12),
+    GGMLQuantizationType.Q4_K:    (256, 2 + 2 + QK_K // 2 + 12),
+    GGMLQuantizationType.Q5_K:    (256, 2 + 2 + QK_K // 2 + QK_K // 8 + 12),
+    GGMLQuantizationType.Q6_K:    (256, 2 + QK_K // 2 + QK_K // 4 + QK_K // 16),
+    GGMLQuantizationType.Q8_K:    (256, 4 + QK_K + QK_K // 8),
+    GGMLQuantizationType.IQ2_XXS: (256, 2 + QK_K // 4),
+    GGMLQuantizationType.IQ2_XS:  (256, 2 + QK_K // 4 + QK_K // 32),
+    GGMLQuantizationType.IQ3_XXS: (256, 2 + QK_K // 4 + QK_K // 8),
+    GGMLQuantizationType.IQ1_S:   (256, 2 + QK_K // 8 + QK_K // 16),
+    GGMLQuantizationType.IQ4_NL:  (32, 2 + 16),
+    GGMLQuantizationType.IQ3_S:   (256, 2 + QK_K // 4 + QK_K // 8 + QK_K // 32 + 4),
+    GGMLQuantizationType.IQ2_S:   (256, 2 + QK_K // 4 + QK_K // 16),
+    GGMLQuantizationType.IQ4_XS:  (256, 2 + 2 + QK_K // 2 + QK_K // 64),
+    GGMLQuantizationType.I8:      (1, 1),
+    GGMLQuantizationType.I16:     (1, 2),
+    GGMLQuantizationType.I32:     (1, 4),
+    GGMLQuantizationType.I64:     (1, 8),
+    GGMLQuantizationType.F64:     (1, 8),
+    GGMLQuantizationType.IQ1_M:   (256, QK_K // 8 + QK_K // 16  + QK_K // 32),
+    GGMLQuantizationType.BF16:    (1, 2),
+    GGMLQuantizationType.Q4_0_4_4:(32, 2 + 16),
+    GGMLQuantizationType.Q4_0_4_8:(32, 2 + 16),
+    GGMLQuantizationType.Q4_0_8_8:(32, 2 + 16),
+    GGMLQuantizationType.TQ1_0:   (256, 2 + 4 * 13),
+    GGMLQuantizationType.TQ2_0:   (256, 2 + 64),
+}
+
+
+# Aliases for backward compatibility.
+
+# general
+KEY_GENERAL_ARCHITECTURE         = Keys.General.ARCHITECTURE
+KEY_GENERAL_QUANTIZATION_VERSION = Keys.General.QUANTIZATION_VERSION
+KEY_GENERAL_ALIGNMENT            = Keys.General.ALIGNMENT
+KEY_GENERAL_NAME                 = Keys.General.NAME
+KEY_GENERAL_AUTHOR               = Keys.General.AUTHOR
+KEY_GENERAL_URL                  = Keys.General.URL
+KEY_GENERAL_DESCRIPTION          = Keys.General.DESCRIPTION
+KEY_GENERAL_LICENSE              = Keys.General.LICENSE
+KEY_GENERAL_SOURCE_URL           = Keys.General.SOURCE_URL
+KEY_GENERAL_FILE_TYPE            = Keys.General.FILE_TYPE
+
+# LLM
+KEY_VOCAB_SIZE            = Keys.LLM.VOCAB_SIZE
+KEY_CONTEXT_LENGTH        = Keys.LLM.CONTEXT_LENGTH
+KEY_EMBEDDING_LENGTH      = Keys.LLM.EMBEDDING_LENGTH
+KEY_BLOCK_COUNT           = Keys.LLM.BLOCK_COUNT
+KEY_FEED_FORWARD_LENGTH   = Keys.LLM.FEED_FORWARD_LENGTH
+KEY_USE_PARALLEL_RESIDUAL = Keys.LLM.USE_PARALLEL_RESIDUAL
+KEY_TENSOR_DATA_LAYOUT    = Keys.LLM.TENSOR_DATA_LAYOUT
+
+# attention
+KEY_ATTENTION_HEAD_COUNT        = Keys.Attention.HEAD_COUNT
+KEY_ATTENTION_HEAD_COUNT_KV     = Keys.Attention.HEAD_COUNT_KV
+KEY_ATTENTION_MAX_ALIBI_BIAS    = Keys.Attention.MAX_ALIBI_BIAS
+KEY_ATTENTION_CLAMP_KQV         = Keys.Attention.CLAMP_KQV
+KEY_ATTENTION_LAYERNORM_EPS     = Keys.Attention.LAYERNORM_EPS
+KEY_ATTENTION_LAYERNORM_RMS_EPS = Keys.Attention.LAYERNORM_RMS_EPS
+
+# RoPE
+KEY_ROPE_DIMENSION_COUNT      = Keys.Rope.DIMENSION_COUNT
+KEY_ROPE_FREQ_BASE            = Keys.Rope.FREQ_BASE
+KEY_ROPE_SCALING_TYPE         = Keys.Rope.SCALING_TYPE
+KEY_ROPE_SCALING_FACTOR       = Keys.Rope.SCALING_FACTOR
+KEY_ROPE_SCALING_ORIG_CTX_LEN = Keys.Rope.SCALING_ORIG_CTX_LEN
+KEY_ROPE_SCALING_FINETUNED    = Keys.Rope.SCALING_FINETUNED
+
+# SSM
+KEY_SSM_CONV_KERNEL    = Keys.SSM.CONV_KERNEL
+KEY_SSM_INNER_SIZE     = Keys.SSM.INNER_SIZE
+KEY_SSM_STATE_SIZE     = Keys.SSM.STATE_SIZE
+KEY_SSM_TIME_STEP_RANK = Keys.SSM.TIME_STEP_RANK
+KEY_SSM_DT_B_C_RMS     = Keys.SSM.DT_B_C_RMS
+
+# tokenization
+KEY_TOKENIZER_MODEL      = Keys.Tokenizer.MODEL
+KEY_TOKENIZER_PRE        = Keys.Tokenizer.PRE
+KEY_TOKENIZER_LIST       = Keys.Tokenizer.LIST
+KEY_TOKENIZER_TOKEN_TYPE = Keys.Tokenizer.TOKEN_TYPE
+KEY_TOKENIZER_SCORES     = Keys.Tokenizer.SCORES
+KEY_TOKENIZER_MERGES     = Keys.Tokenizer.MERGES
+KEY_TOKENIZER_BOS_ID     = Keys.Tokenizer.BOS_ID
+KEY_TOKENIZER_EOS_ID     = Keys.Tokenizer.EOS_ID
+KEY_TOKENIZER_EOT_ID     = Keys.Tokenizer.EOT_ID
+KEY_TOKENIZER_EOM_ID     = Keys.Tokenizer.EOM_ID
+KEY_TOKENIZER_UNK_ID     = Keys.Tokenizer.UNK_ID
+KEY_TOKENIZER_SEP_ID     = Keys.Tokenizer.SEP_ID
+KEY_TOKENIZER_PAD_ID     = Keys.Tokenizer.PAD_ID
+KEY_TOKENIZER_CLS_ID     = Keys.Tokenizer.CLS_ID
+KEY_TOKENIZER_MASK_ID    = Keys.Tokenizer.MASK_ID
+KEY_TOKENIZER_HF_JSON    = Keys.Tokenizer.HF_JSON
+KEY_TOKENIZER_RWKV       = Keys.Tokenizer.RWKV
+
+KEY_TOKENIZER_FIM_PRE_ID = Keys.Tokenizer.FIM_PRE_ID
+KEY_TOKENIZER_FIM_SUF_ID = Keys.Tokenizer.FIM_SUF_ID
+KEY_TOKENIZER_FIM_MID_ID = Keys.Tokenizer.FIM_MID_ID
+KEY_TOKENIZER_FIM_PAD_ID = Keys.Tokenizer.FIM_PAD_ID
+KEY_TOKENIZER_FIM_REP_ID = Keys.Tokenizer.FIM_REP_ID
+KEY_TOKENIZER_FIM_SEP_ID = Keys.Tokenizer.FIM_SEP_ID
+
+# deprecated
+KEY_TOKENIZER_PREFIX_ID  = Keys.Tokenizer.PREFIX_ID
+KEY_TOKENIZER_SUFFIX_ID  = Keys.Tokenizer.SUFFIX_ID
+KEY_TOKENIZER_MIDDLE_ID  = Keys.Tokenizer.MIDDLE_ID

gguf/gguf.py

@@ -0,0 +1,15 @@
+# This file left for compatibility. If you want to use the GGUF API from Python
+# then don't import gguf/gguf.py directly. If you're looking for examples, see the
+# examples/ directory for gguf-py
+
+import importlib
+import sys
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).parent.parent))
+
+# Compatibility for people trying to import gguf/gguf.py directly instead of as a package.
+importlib.invalidate_caches()
+import gguf  # noqa: E402
+
+importlib.reload(gguf)

gguf/gguf_reader.py

@@ -0,0 +1,317 @@
+#
+# GGUF file reading/modification support. For API usage information,
+# please see the files scripts/ for some fairly simple examples.
+#
+from __future__ import annotations
+
+import logging
+import os
+from collections import OrderedDict
+from typing import Any, Literal, NamedTuple, TypeVar, Union
+
+import numpy as np
+import numpy.typing as npt
+
+from .quants import quant_shape_to_byte_shape
+
+if __name__ == "__main__":
+    import sys
+    from pathlib import Path
+
+    # Allow running file in package as a script.
+    sys.path.insert(0, str(Path(__file__).parent.parent))
+
+from gguf.constants import (
+    GGML_QUANT_SIZES,
+    GGUF_DEFAULT_ALIGNMENT,
+    GGUF_MAGIC,
+    GGUF_VERSION,
+    GGMLQuantizationType,
+    GGUFValueType,
+)
+
+logger = logging.getLogger(__name__)
+
+READER_SUPPORTED_VERSIONS = [2, GGUF_VERSION]
+
+
+class ReaderField(NamedTuple):
+    # Offset to start of this field.
+    offset: int
+
+    # Name of the field (not necessarily from file data).
+    name: str
+
+    # Data parts. Some types have multiple components, such as strings
+    # that consist of a length followed by the string data.
+    parts: list[npt.NDArray[Any]] = []
+
+    # Indexes into parts that we can call the actual data. For example
+    # an array of strings will be populated with indexes to the actual
+    # string data.
+    data: list[int] = [-1]
+
+    types: list[GGUFValueType] = []
+
+
+class ReaderTensor(NamedTuple):
+    name: str
+    tensor_type: GGMLQuantizationType
+    shape: npt.NDArray[np.uint32]
+    n_elements: int
+    n_bytes: int
+    data_offset: int
+    data: npt.NDArray[Any]
+    field: ReaderField
+
+
+class GGUFReader:
+    # I - same as host, S - swapped
+    byte_order: Literal['I', 'S'] = 'I'
+    alignment: int = GGUF_DEFAULT_ALIGNMENT
+    data_offset: int
+
+    # Note: Internal helper, API may change.
+    gguf_scalar_to_np: dict[GGUFValueType, type[np.generic]] = {
+        GGUFValueType.UINT8:   np.uint8,
+        GGUFValueType.INT8:    np.int8,
+        GGUFValueType.UINT16:  np.uint16,
+        GGUFValueType.INT16:   np.int16,
+        GGUFValueType.UINT32:  np.uint32,
+        GGUFValueType.INT32:   np.int32,
+        GGUFValueType.FLOAT32: np.float32,
+        GGUFValueType.UINT64:  np.uint64,
+        GGUFValueType.INT64:   np.int64,
+        GGUFValueType.FLOAT64: np.float64,
+        GGUFValueType.BOOL:    np.bool_,
+    }
+
+    def __init__(self, path: os.PathLike[str] | str, mode: Literal['r', 'r+', 'c'] = 'r'):
+        self.data = np.memmap(path, mode = mode)
+        offs = 0
+
+        # Check for GGUF magic
+        if self._get(offs, np.uint32, override_order = '<')[0] != GGUF_MAGIC:
+            raise ValueError('GGUF magic invalid')
+        offs += 4
+
+        # Check GGUF version
+        temp_version = self._get(offs, np.uint32)
+        if temp_version[0] & 65535 == 0:
+            # If we get 0 here that means it's (probably) a GGUF file created for
+            # the opposite byte order of the machine this script is running on.
+            self.byte_order = 'S'
+            temp_version = temp_version.newbyteorder(self.byte_order)
+        version = temp_version[0]
+        if version not in READER_SUPPORTED_VERSIONS:
+            raise ValueError(f'Sorry, file appears to be version {version} which we cannot handle')
+        self.fields: OrderedDict[str, ReaderField] = OrderedDict()
+        self.tensors: list[ReaderTensor] = []
+        offs += self._push_field(ReaderField(offs, 'GGUF.version', [temp_version], [0], [GGUFValueType.UINT32]))
+
+        # Check tensor count and kv count
+        temp_counts = self._get(offs, np.uint64, 2)
+        offs += self._push_field(ReaderField(offs, 'GGUF.tensor_count', [temp_counts[:1]], [0], [GGUFValueType.UINT64]))
+        offs += self._push_field(ReaderField(offs, 'GGUF.kv_count', [temp_counts[1:]], [0], [GGUFValueType.UINT64]))
+        tensor_count, kv_count = temp_counts
+        offs = self._build_fields(offs, kv_count)
+
+        # Build Tensor Info Fields
+        offs, tensors_fields = self._build_tensor_info(offs, tensor_count)
+        new_align = self.fields.get('general.alignment')
+        if new_align is not None:
+            if new_align.types != [GGUFValueType.UINT32]:
+                raise ValueError('Bad type for general.alignment field')
+            self.alignment = new_align.parts[-1][0]
+        padding = offs % self.alignment
+        if padding != 0:
+            offs += self.alignment - padding
+        self.data_offset = offs
+        self._build_tensors(offs, tensors_fields)
+
+    _DT = TypeVar('_DT', bound = npt.DTypeLike)
+
+    # Fetch a key/value metadata field by key.
+    def get_field(self, key: str) -> Union[ReaderField, None]:
+        return self.fields.get(key, None)
+
+    # Fetch a tensor from the list by index.
+    def get_tensor(self, idx: int) -> ReaderTensor:
+        return self.tensors[idx]
+
+    def _get(
+        self, offset: int, dtype: npt.DTypeLike, count: int = 1, override_order: None | Literal['I', 'S', '<'] = None,
+    ) -> npt.NDArray[Any]:
+        count = int(count)
+        itemsize = int(np.empty([], dtype = dtype).itemsize)
+        end_offs = offset + itemsize * count
+        return (
+            self.data[offset:end_offs]
+            .view(dtype = dtype)[:count]
+            .newbyteorder(override_order or self.byte_order)
+        )
+
+    def _push_field(self, field: ReaderField, skip_sum: bool = False) -> int:
+        if field.name in self.fields:
+            # TODO: add option to generate error on duplicate keys
+            # raise KeyError(f'Duplicate {field.name} already in list at offset {field.offset}')
+
+            logger.warning(f'Duplicate key {field.name} at offset {field.offset}')
+            self.fields[field.name + '_{}'.format(field.offset)] = field
+        else:
+            self.fields[field.name] = field
+        return 0 if skip_sum else sum(int(part.nbytes) for part in field.parts)
+
+    def _get_str(self, offset: int) -> tuple[npt.NDArray[np.uint64], npt.NDArray[np.uint8]]:
+        slen = self._get(offset, np.uint64)
+        return slen, self._get(offset + 8, np.uint8, slen[0])
+
+    def _get_field_parts(
+        self, orig_offs: int, raw_type: int,
+    ) -> tuple[int, list[npt.NDArray[Any]], list[int], list[GGUFValueType]]:
+        offs = orig_offs
+        types: list[GGUFValueType] = []
+        gtype = GGUFValueType(raw_type)
+        types.append(gtype)
+        # Handle strings.
+        if gtype == GGUFValueType.STRING:
+            sparts: list[npt.NDArray[Any]] = list(self._get_str(offs))
+            size = sum(int(part.nbytes) for part in sparts)
+            return size, sparts, [1], types
+        # Check if it's a simple scalar type.
+        nptype = self.gguf_scalar_to_np.get(gtype)
+        if nptype is not None:
+            val = self._get(offs, nptype)
+            return int(val.nbytes), [val], [0], types
+        # Handle arrays.
+        if gtype == GGUFValueType.ARRAY:
+            raw_itype = self._get(offs, np.uint32)
+            offs += int(raw_itype.nbytes)
+            alen = self._get(offs, np.uint64)
+            offs += int(alen.nbytes)
+            aparts: list[npt.NDArray[Any]] = [raw_itype, alen]
+            data_idxs: list[int] = []
+            for idx in range(alen[0]):
+                curr_size, curr_parts, curr_idxs, curr_types = self._get_field_parts(offs, raw_itype[0])
+                if idx == 0:
+                    types += curr_types
+                idxs_offs = len(aparts)
+                aparts += curr_parts
+                data_idxs += (idx + idxs_offs for idx in curr_idxs)
+                offs += curr_size
+            return offs - orig_offs, aparts, data_idxs, types
+        # We can't deal with this one.
+        raise ValueError('Unknown/unhandled field type {gtype}')
+
+    def _get_tensor_info_field(self, orig_offs: int) -> ReaderField:
+        offs = orig_offs
+
+        # Get Tensor Name
+        name_len, name_data = self._get_str(offs)
+        offs += int(name_len.nbytes + name_data.nbytes)
+
+        # Get Tensor Dimensions Count
+        n_dims = self._get(offs, np.uint32)
+        offs += int(n_dims.nbytes)
+
+        # Get Tensor Dimension Array
+        dims = self._get(offs, np.uint64, n_dims[0])
+        offs += int(dims.nbytes)
+
+        # Get Tensor Encoding Scheme Type
+        raw_dtype = self._get(offs, np.uint32)
+        offs += int(raw_dtype.nbytes)
+
+        # Get Tensor Offset
+        offset_tensor = self._get(offs, np.uint64)
+        offs += int(offset_tensor.nbytes)
+
+        return ReaderField(
+            orig_offs,
+            str(bytes(name_data), encoding = 'utf-8'),
+            [name_len, name_data, n_dims, dims, raw_dtype, offset_tensor],
+            [1, 3, 4, 5],
+        )
+
+    def _build_fields(self, offs: int, count: int) -> int:
+        for _ in range(count):
+            orig_offs = offs
+            kv_klen, kv_kdata = self._get_str(offs)
+            offs += int(kv_klen.nbytes + kv_kdata.nbytes)
+            raw_kv_type = self._get(offs, np.uint32)
+            offs += int(raw_kv_type.nbytes)
+            parts: list[npt.NDArray[Any]] = [kv_klen, kv_kdata, raw_kv_type]
+            idxs_offs = len(parts)
+            field_size, field_parts, field_idxs, field_types = self._get_field_parts(offs, raw_kv_type[0])
+            parts += field_parts
+            self._push_field(ReaderField(
+                orig_offs,
+                str(bytes(kv_kdata), encoding = 'utf-8'),
+                parts,
+                [idx + idxs_offs for idx in field_idxs],
+                field_types,
+            ), skip_sum = True)
+            offs += field_size
+        return offs
+
+    def _build_tensor_info(self, offs: int, count: int) -> tuple[int, list[ReaderField]]:
+        tensor_fields = []
+        for _ in range(count):
+            field = self._get_tensor_info_field(offs)
+            offs += sum(int(part.nbytes) for part in field.parts)
+            tensor_fields.append(field)
+        return offs, tensor_fields
+
+    def _build_tensors(self, start_offs: int, fields: list[ReaderField]) -> None:
+        tensors = []
+        tensor_names = set() # keep track of name to prevent duplicated tensors
+        for field in fields:
+            _name_len, name_data, _n_dims, dims, raw_dtype, offset_tensor = field.parts
+            # check if there's any tensor having same name already in the list
+            tensor_name = str(bytes(name_data), encoding = 'utf-8')
+            if tensor_name in tensor_names:
+                raise ValueError(f'Found duplicated tensor with name {tensor_name}')
+            tensor_names.add(tensor_name)
+            ggml_type = GGMLQuantizationType(raw_dtype[0])
+            n_elems = int(np.prod(dims))
+            np_dims = tuple(reversed(dims.tolist()))
+            block_size, type_size = GGML_QUANT_SIZES[ggml_type]
+            n_bytes = n_elems * type_size // block_size
+            data_offs = int(start_offs + offset_tensor[0])
+            item_type: npt.DTypeLike
+            if ggml_type == GGMLQuantizationType.F16:
+                item_count = n_elems
+                item_type = np.float16
+            elif ggml_type == GGMLQuantizationType.F32:
+                item_count = n_elems
+                item_type = np.float32
+            elif ggml_type == GGMLQuantizationType.F64:
+                item_count = n_elems
+                item_type = np.float64
+            elif ggml_type == GGMLQuantizationType.I8:
+                item_count = n_elems
+                item_type = np.int8
+            elif ggml_type == GGMLQuantizationType.I16:
+                item_count = n_elems
+                item_type = np.int16
+            elif ggml_type == GGMLQuantizationType.I32:
+                item_count = n_elems
+                item_type = np.int32
+            elif ggml_type == GGMLQuantizationType.I64:
+                item_count = n_elems
+                item_type = np.int64
+            else:
+                item_count = n_bytes
+                item_type = np.uint8
+                np_dims = quant_shape_to_byte_shape(np_dims, ggml_type)
+            tensors.append(ReaderTensor(
+                name = tensor_name,
+                tensor_type = ggml_type,
+                shape = dims,
+                n_elements = n_elems,
+                n_bytes = n_bytes,
+                data_offset = data_offs,
+                data = self._get(data_offs, item_type, item_count).reshape(np_dims),
+                field = field,
+            ))
+        self.tensors = tensors

gguf/gguf_writer.py

@@ -0,0 +1,903 @@
+from __future__ import annotations
+
+import logging
+import os
+import shutil
+import struct
+import tempfile
+from dataclasses import dataclass
+from enum import Enum, auto
+from math import prod
+from pathlib import Path
+from io import BufferedWriter
+from typing import IO, Any, Sequence, Mapping
+from string import ascii_letters, digits
+
+import numpy as np
+
+from .constants import (
+    GGUF_DEFAULT_ALIGNMENT,
+    GGUF_MAGIC,
+    GGUF_VERSION,
+    GGMLQuantizationType,
+    GGUFEndian,
+    GGUFValueType,
+    Keys,
+    RopeScalingType,
+    PoolingType,
+    TokenType,
+)
+
+from .quants import quant_shape_from_byte_shape
+
+logger = logging.getLogger(__name__)
+
+
+SHARD_NAME_FORMAT = "{:s}-{:05d}-of-{:05d}.gguf"
+
+
+@dataclass
+class TensorInfo:
+    shape: Sequence[int]
+    dtype: GGMLQuantizationType
+    nbytes: int
+    tensor: np.ndarray[Any, Any] | None = None
+
+
+@dataclass
+class GGUFValue:
+    value: Any
+    type: GGUFValueType
+
+
+class WriterState(Enum):
+    NO_FILE = auto()
+    EMPTY   = auto()
+    HEADER  = auto()
+    KV_DATA = auto()
+    TI_DATA = auto()
+    WEIGHTS = auto()
+
+
+class GGUFWriter:
+    fout: list[BufferedWriter] | None
+    path: Path | None
+    temp_file: tempfile.SpooledTemporaryFile[bytes] | None
+    tensors: list[dict[str, TensorInfo]]
+    kv_data: list[dict[str, GGUFValue]]
+    state: WriterState
+    _simple_value_packing = {
+        GGUFValueType.UINT8:   "B",
+        GGUFValueType.INT8:    "b",
+        GGUFValueType.UINT16:  "H",
+        GGUFValueType.INT16:   "h",
+        GGUFValueType.UINT32:  "I",
+        GGUFValueType.INT32:   "i",
+        GGUFValueType.FLOAT32: "f",
+        GGUFValueType.UINT64:  "Q",
+        GGUFValueType.INT64:   "q",
+        GGUFValueType.FLOAT64: "d",
+        GGUFValueType.BOOL:    "?",
+    }
+
+    def __init__(
+        self, path: os.PathLike[str] | str | None, arch: str, use_temp_file: bool = False, endianess: GGUFEndian = GGUFEndian.LITTLE,
+        split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, small_first_shard: bool = False
+    ):
+        self.fout = None
+        self.path = Path(path) if path else None
+        self.arch = arch
+        self.endianess = endianess
+        self.data_alignment = GGUF_DEFAULT_ALIGNMENT
+        self.use_temp_file = use_temp_file
+        self.temp_file = None
+        self.tensors = [{}]
+        self.kv_data = [{}]
+        self.split_max_tensors = split_max_tensors
+        self.split_max_size = split_max_size
+        self.dry_run = dry_run
+        self.small_first_shard = small_first_shard
+        logger.info("gguf: This GGUF file is for {0} Endian only".format(
+            "Big" if self.endianess == GGUFEndian.BIG else "Little",
+        ))
+        self.state = WriterState.NO_FILE
+
+        if self.small_first_shard:
+            self.tensors.append({})
+
+        self.add_architecture()
+
+    def get_total_parameter_count(self) -> tuple[int, int, int, int]:
+        total_params = 0
+        shared_params = 0
+        expert_params = 0
+
+        expert_sum = 0
+        n_expert_tensors = 0
+
+        last_lora_a: tuple[str, TensorInfo] | None = None
+
+        for tensors in self.tensors:
+            for name, info in tensors.items():
+
+                shape = info.shape
+
+                if name.endswith(".lora_a"):
+                    last_lora_a = (name, info)
+                    continue
+                elif name.endswith(".lora_b"):
+                    if last_lora_a is None or last_lora_a[0] != name[:-1] + "a":
+                        # Bail when the LoRA pair can't be found trivially
+                        logger.warning("can't measure LoRA size correctly, tensor order is unusual")
+                        return 0, 0, 0, 0
+                    else:
+                        shape = (*shape[:-1], last_lora_a[1].shape[-1])
+
+                size = prod(shape)
+
+                if "_exps." in name:
+                    expert_params += (size // shape[-3])
+                    expert_sum += shape[-3]
+                    n_expert_tensors += 1
+                else:
+                    shared_params += size
+
+                total_params += size
+
+        # Hopefully this should work even for variable-expert-count models
+        expert_count = (expert_sum // n_expert_tensors) if n_expert_tensors > 0 else 0
+
+        # Negate the total to signal it's likely not exact
+        if last_lora_a is not None:
+            total_params = -total_params
+
+        # NOTE: keep the output in the same order as accepted by 'size_label' in gguf-py/gguf/utility.py
+        return total_params, shared_params, expert_params, expert_count
+
+    def format_shard_names(self, path: Path) -> list[Path]:
+        if len(self.tensors) == 1:
+            return [path]
+        return [path.with_name(SHARD_NAME_FORMAT.format(path.stem, i + 1, len(self.tensors))) for i in range(len(self.tensors))]
+
+    def open_output_file(self, path: Path | None = None) -> None:
+        if self.state is WriterState.EMPTY and self.fout is not None and (path is None or path == self.path):
+            # allow calling this multiple times as long as the path is the same
+            return
+
+        if self.state is not WriterState.NO_FILE:
+            raise ValueError(f'Expected output file to be not yet opened, got {self.state}')
+
+        if path is not None:
+            self.path = path
+
+        if self.path is not None:
+            filenames = self.print_plan()
+            self.fout = [open(filename, "wb") for filename in filenames]
+            self.state = WriterState.EMPTY
+
+    def print_plan(self) -> list[Path]:
+        logger.info("Writing the following files:")
+        assert self.path is not None
+        filenames = self.format_shard_names(self.path)
+        assert len(filenames) == len(self.tensors)
+        for name, tensors in zip(filenames, self.tensors):
+            logger.info(f"{name}: n_tensors = {len(tensors)}, total_size = {GGUFWriter.format_n_bytes_to_str(sum(ti.nbytes for ti in tensors.values()))}")
+
+        if self.dry_run:
+            logger.info("Dry run, not writing files")
+            for name in filenames:
+                print(name)  # noqa: NP100
+            exit()
+
+        return filenames
+
+    def add_shard_kv_data(self) -> None:
+        if len(self.tensors) == 1:
+            return
+
+        total_tensors = sum(len(t) for t in self.tensors)
+        assert self.fout is not None
+        total_splits = len(self.fout)
+        self.kv_data.extend({} for _ in range(len(self.kv_data), total_splits))
+        for i, kv_data in enumerate(self.kv_data):
+            kv_data[Keys.Split.LLM_KV_SPLIT_NO] = GGUFValue(i, GGUFValueType.UINT16)
+            kv_data[Keys.Split.LLM_KV_SPLIT_COUNT] = GGUFValue(total_splits, GGUFValueType.UINT16)
+            kv_data[Keys.Split.LLM_KV_SPLIT_TENSORS_COUNT] = GGUFValue(total_tensors, GGUFValueType.INT32)
+
+    def write_header_to_file(self, path: Path | None = None) -> None:
+        if len(self.tensors) == 1 and (self.split_max_tensors != 0 or self.split_max_size != 0):
+            logger.warning("Model fails split requirements, not splitting")
+
+        self.open_output_file(path)
+
+        if self.state is not WriterState.EMPTY:
+            raise ValueError(f'Expected output file to be empty, got {self.state}')
+
+        assert self.fout is not None
+        assert len(self.fout) == len(self.tensors)
+        assert len(self.kv_data) == 1
+
+        self.add_shard_kv_data()
+
+        for fout, tensors, kv_data in zip(self.fout, self.tensors, self.kv_data):
+            fout.write(self._pack("<I", GGUF_MAGIC, skip_pack_prefix = True))
+            fout.write(self._pack("I", GGUF_VERSION))
+            fout.write(self._pack("Q", len(tensors)))
+            fout.write(self._pack("Q", len(kv_data)))
+            fout.flush()
+        self.state = WriterState.HEADER
+
+    def write_kv_data_to_file(self) -> None:
+        if self.state is not WriterState.HEADER:
+            raise ValueError(f'Expected output file to contain the header, got {self.state}')
+        assert self.fout is not None
+
+        for fout, kv_data in zip(self.fout, self.kv_data):
+            kv_bytes = bytearray()
+
+            for key, val in kv_data.items():
+                kv_bytes += self._pack_val(key, GGUFValueType.STRING, add_vtype=False)
+                kv_bytes += self._pack_val(val.value, val.type, add_vtype=True)
+
+            fout.write(kv_bytes)
+
+        self.flush()
+        self.state = WriterState.KV_DATA
+
+    def write_ti_data_to_file(self) -> None:
+        if self.state is not WriterState.KV_DATA:
+            raise ValueError(f'Expected output file to contain KV data, got {self.state}')
+        assert self.fout is not None
+
+        for fout, tensors in zip(self.fout, self.tensors):
+            ti_data = bytearray()
+            offset_tensor = 0
+
+            for name, ti in tensors.items():
+                ti_data += self._pack_val(name, GGUFValueType.STRING, add_vtype=False)
+                n_dims = len(ti.shape)
+                ti_data += self._pack("I", n_dims)
+                for j in range(n_dims):
+                    ti_data += self._pack("Q", ti.shape[n_dims - 1 - j])
+                ti_data += self._pack("I", ti.dtype)
+                ti_data += self._pack("Q", offset_tensor)
+                offset_tensor += GGUFWriter.ggml_pad(ti.nbytes, self.data_alignment)
+
+            fout.write(ti_data)
+            fout.flush()
+        self.state = WriterState.TI_DATA
+
+    def add_key_value(self, key: str, val: Any, vtype: GGUFValueType) -> None:
+        if any(key in kv_data for kv_data in self.kv_data):
+            raise ValueError(f'Duplicated key name {key!r}')
+
+        self.kv_data[0][key] = GGUFValue(value=val, type=vtype)
+
+    def add_uint8(self, key: str, val: int) -> None:
+        self.add_key_value(key,val, GGUFValueType.UINT8)
+
+    def add_int8(self, key: str, val: int) -> None:
+        self.add_key_value(key, val, GGUFValueType.INT8)
+
+    def add_uint16(self, key: str, val: int) -> None:
+        self.add_key_value(key, val, GGUFValueType.UINT16)
+
+    def add_int16(self, key: str, val: int) -> None:
+        self.add_key_value(key, val, GGUFValueType.INT16)
+
+    def add_uint32(self, key: str, val: int) -> None:
+        self.add_key_value(key, val, GGUFValueType.UINT32)
+
+    def add_int32(self, key: str, val: int) -> None:
+        self.add_key_value(key, val, GGUFValueType.INT32)
+
+    def add_float32(self, key: str, val: float) -> None:
+        self.add_key_value(key, val, GGUFValueType.FLOAT32)
+
+    def add_uint64(self, key: str, val: int) -> None:
+        self.add_key_value(key, val, GGUFValueType.UINT64)
+
+    def add_int64(self, key: str, val: int) -> None:
+        self.add_key_value(key, val, GGUFValueType.INT64)
+
+    def add_float64(self, key: str, val: float) -> None:
+        self.add_key_value(key, val, GGUFValueType.FLOAT64)
+
+    def add_bool(self, key: str, val: bool) -> None:
+        self.add_key_value(key, val, GGUFValueType.BOOL)
+
+    def add_string(self, key: str, val: str) -> None:
+        if not val:
+            return
+        self.add_key_value(key, val, GGUFValueType.STRING)
+
+    def add_array(self, key: str, val: Sequence[Any]) -> None:
+        if len(val) == 0:
+            return
+        self.add_key_value(key, val, GGUFValueType.ARRAY)
+
+    @staticmethod
+    def ggml_pad(x: int, n: int) -> int:
+        return ((x + n - 1) // n) * n
+
+    def add_tensor_info(
+        self, name: str, tensor_shape: Sequence[int], tensor_dtype: np.dtype,
+        tensor_nbytes: int, raw_dtype: GGMLQuantizationType | None = None,
+    ) -> None:
+        if self.state is not WriterState.NO_FILE:
+            raise ValueError(f'Expected output file to be not yet opened, got {self.state}')
+
+        if any(name in tensors for tensors in self.tensors):
+            raise ValueError(f'Duplicated tensor name {name!r}')
+
+        if raw_dtype is None:
+            if tensor_dtype == np.float16:
+                dtype = GGMLQuantizationType.F16
+            elif tensor_dtype == np.float32:
+                dtype = GGMLQuantizationType.F32
+            elif tensor_dtype == np.float64:
+                dtype = GGMLQuantizationType.F64
+            elif tensor_dtype == np.int8:
+                dtype = GGMLQuantizationType.I8
+            elif tensor_dtype == np.int16:
+                dtype = GGMLQuantizationType.I16
+            elif tensor_dtype == np.int32:
+                dtype = GGMLQuantizationType.I32
+            elif tensor_dtype == np.int64:
+                dtype = GGMLQuantizationType.I64
+            else:
+                raise ValueError("Only F16, F32, F64, I8, I16, I32, I64 tensors are supported for now")
+        else:
+            dtype = raw_dtype
+            if tensor_dtype == np.uint8:
+                tensor_shape = quant_shape_from_byte_shape(tensor_shape, raw_dtype)
+
+        # make sure there is at least one tensor before splitting
+        if len(self.tensors[-1]) > 0:
+            if (  # split when over tensor limit
+                self.split_max_tensors != 0
+                and len(self.tensors[-1]) >= self.split_max_tensors
+            ) or (   # split when over size limit
+                self.split_max_size != 0
+                and sum(ti.nbytes for ti in self.tensors[-1].values()) + tensor_nbytes > self.split_max_size
+            ):
+                self.tensors.append({})
+
+        self.tensors[-1][name] = TensorInfo(shape=tensor_shape, dtype=dtype, nbytes=tensor_nbytes)
+
+    def add_tensor(
+        self, name: str, tensor: np.ndarray[Any, Any], raw_shape: Sequence[int] | None = None,
+        raw_dtype: GGMLQuantizationType | None = None,
+    ) -> None:
+        if self.endianess == GGUFEndian.BIG:
+            tensor.byteswap(inplace=True)
+        if self.use_temp_file and self.temp_file is None:
+            fp = tempfile.SpooledTemporaryFile(mode="w+b", max_size=256 * 1024 * 1024)
+            fp.seek(0)
+            self.temp_file = fp
+
+        shape: Sequence[int] = raw_shape if raw_shape is not None else tensor.shape
+        self.add_tensor_info(name, shape, tensor.dtype, tensor.nbytes, raw_dtype=raw_dtype)
+
+        if self.temp_file is None:
+            self.tensors[-1][name].tensor = tensor
+            return
+
+        tensor.tofile(self.temp_file)
+        self.write_padding(self.temp_file, tensor.nbytes)
+
+    def write_padding(self, fp: IO[bytes], n: int, align: int | None = None) -> None:
+        pad = GGUFWriter.ggml_pad(n, align if align is not None else self.data_alignment) - n
+        if pad != 0:
+            fp.write(bytes([0] * pad))
+
+    def write_tensor_data(self, tensor: np.ndarray[Any, Any]) -> None:
+        if self.state is not WriterState.TI_DATA and self.state is not WriterState.WEIGHTS:
+            raise ValueError(f'Expected output file to contain tensor info or weights, got {self.state}')
+        assert self.fout is not None
+
+        if self.endianess == GGUFEndian.BIG:
+            tensor.byteswap(inplace=True)
+
+        file_id = -1
+        for i, tensors in enumerate(self.tensors):
+            if len(tensors) > 0:
+                file_id = i
+                break
+
+        fout = self.fout[file_id]
+
+        # pop the first tensor info
+        # TODO: cleaner way to get the first key
+        first_tensor_name = [name for name, _ in zip(self.tensors[file_id].keys(), range(1))][0]
+        ti = self.tensors[file_id].pop(first_tensor_name)
+        assert ti.nbytes == tensor.nbytes
+
+        self.write_padding(fout, fout.tell())
+        tensor.tofile(fout)
+        self.write_padding(fout, tensor.nbytes)
+
+        self.state = WriterState.WEIGHTS
+
+    def write_tensors_to_file(self, *, progress: bool = False) -> None:
+        self.write_ti_data_to_file()
+
+        assert self.fout is not None
+
+        for fout in self.fout:
+            self.write_padding(fout, fout.tell())
+
+        if self.temp_file is None:
+            shard_bar = None
+            bar = None
+
+            if progress:
+                from tqdm import tqdm
+
+                total_bytes = sum(ti.nbytes for t in self.tensors for ti in t.values())
+
+                if len(self.fout) > 1:
+                    shard_bar = tqdm(desc=f"Shard (0/{len(self.fout)})", total=None, unit="byte", unit_scale=True)
+                bar = tqdm(desc="Writing", total=total_bytes, unit="byte", unit_scale=True)
+
+            for i, (fout, tensors) in enumerate(zip(self.fout, self.tensors)):
+                if shard_bar is not None:
+                    shard_bar.set_description(f"Shard ({i + 1}/{len(self.fout)})")
+                    total = sum(ti.nbytes for ti in tensors.values())
+                    shard_bar.reset(total=(total if total > 0 else None))
+
+                # relying on the fact that Python dicts preserve insertion order (since 3.7)
+                for ti in tensors.values():
+                    assert ti.tensor is not None  # can only iterate once over the tensors
+                    assert ti.tensor.nbytes == ti.nbytes
+                    ti.tensor.tofile(fout)
+                    if shard_bar is not None:
+                        shard_bar.update(ti.nbytes)
+                    if bar is not None:
+                        bar.update(ti.nbytes)
+                    self.write_padding(fout, ti.nbytes)
+                    ti.tensor = None
+        else:
+            self.temp_file.seek(0)
+
+            shutil.copyfileobj(self.temp_file, self.fout[0 if not self.small_first_shard else 1])
+            self.flush()
+            self.temp_file.close()
+
+        self.state = WriterState.WEIGHTS
+
+    def flush(self) -> None:
+        assert self.fout is not None
+        for fout in self.fout:
+            fout.flush()
+
+    def close(self) -> None:
+        if self.fout is not None:
+            for fout in self.fout:
+                fout.close()
+            self.fout = None
+
+    def add_type(self, type_name: str) -> None:
+        self.add_string(Keys.General.TYPE, type_name)
+
+    def add_architecture(self) -> None:
+        self.add_string(Keys.General.ARCHITECTURE, self.arch)
+
+    def add_quantization_version(self, quantization_version: int) -> None:
+        self.add_uint32(Keys.General.QUANTIZATION_VERSION, quantization_version)
+
+    def add_custom_alignment(self, alignment: int) -> None:
+        self.data_alignment = alignment
+        self.add_uint32(Keys.General.ALIGNMENT, alignment)
+
+    def add_file_type(self, ftype: int) -> None:
+        self.add_uint32(Keys.General.FILE_TYPE, ftype)
+
+    def add_name(self, name: str) -> None:
+        self.add_string(Keys.General.NAME, name)
+
+    def add_author(self, author: str) -> None:
+        self.add_string(Keys.General.AUTHOR, author)
+
+    def add_version(self, version: str) -> None:
+        self.add_string(Keys.General.VERSION, version)
+
+    def add_organization(self, organization: str) -> None:
+        self.add_string(Keys.General.ORGANIZATION, organization)
+
+    def add_finetune(self, finetune: str) -> None:
+        self.add_string(Keys.General.FINETUNE, finetune)
+
+    def add_basename(self, basename: str) -> None:
+        self.add_string(Keys.General.BASENAME, basename)
+
+    def add_description(self, description: str) -> None:
+        self.add_string(Keys.General.DESCRIPTION, description)
+
+    def add_quantized_by(self, quantized: str) -> None:
+        self.add_string(Keys.General.QUANTIZED_BY, quantized)
+
+    def add_size_label(self, size_label: str) -> None:
+        self.add_string(Keys.General.SIZE_LABEL, size_label)
+
+    def add_license(self, license: str) -> None:
+        self.add_string(Keys.General.LICENSE, license)
+
+    def add_license_name(self, license: str) -> None:
+        self.add_string(Keys.General.LICENSE_NAME, license)
+
+    def add_license_link(self, license: str) -> None:
+        self.add_string(Keys.General.LICENSE_LINK, license)
+
+    def add_url(self, url: str) -> None:
+        self.add_string(Keys.General.URL, url)
+
+    def add_doi(self, doi: str) -> None:
+        self.add_string(Keys.General.DOI, doi)
+
+    def add_uuid(self, uuid: str) -> None:
+        self.add_string(Keys.General.UUID, uuid)
+
+    def add_repo_url(self, repo_url: str) -> None:
+        self.add_string(Keys.General.REPO_URL, repo_url)
+
+    def add_source_url(self, url: str) -> None:
+        self.add_string(Keys.General.SOURCE_URL, url)
+
+    def add_source_doi(self, doi: str) -> None:
+        self.add_string(Keys.General.SOURCE_DOI, doi)
+
+    def add_source_uuid(self, uuid: str) -> None:
+        self.add_string(Keys.General.SOURCE_UUID, uuid)
+
+    def add_source_repo_url(self, repo_url: str) -> None:
+        self.add_string(Keys.General.SOURCE_REPO_URL, repo_url)
+
+    def add_base_model_count(self, source_count: int) -> None:
+        self.add_uint32(Keys.General.BASE_MODEL_COUNT, source_count)
+
+    def add_base_model_name(self, source_id: int, name: str) -> None:
+        self.add_string(Keys.General.BASE_MODEL_NAME.format(id=source_id), name)
+
+    def add_base_model_author(self, source_id: int, author: str) -> None:
+        self.add_string(Keys.General.BASE_MODEL_AUTHOR.format(id=source_id), author)
+
+    def add_base_model_version(self, source_id: int, version: str) -> None:
+        self.add_string(Keys.General.BASE_MODEL_VERSION.format(id=source_id), version)
+
+    def add_base_model_organization(self, source_id: int, organization: str) -> None:
+        self.add_string(Keys.General.BASE_MODEL_ORGANIZATION.format(id=source_id), organization)
+
+    def add_base_model_url(self, source_id: int, url: str) -> None:
+        self.add_string(Keys.General.BASE_MODEL_URL.format(id=source_id), url)
+
+    def add_base_model_doi(self, source_id: int, doi: str) -> None:
+        self.add_string(Keys.General.BASE_MODEL_DOI.format(id=source_id), doi)
+
+    def add_base_model_uuid(self, source_id: int, uuid: str) -> None:
+        self.add_string(Keys.General.BASE_MODEL_UUID.format(id=source_id), uuid)
+
+    def add_base_model_repo_url(self, source_id: int, repo_url: str) -> None:
+        self.add_string(Keys.General.BASE_MODEL_REPO_URL.format(id=source_id), repo_url)
+
+    def add_tags(self, tags: Sequence[str]) -> None:
+        self.add_array(Keys.General.TAGS, tags)
+
+    def add_languages(self, languages: Sequence[str]) -> None:
+        self.add_array(Keys.General.LANGUAGES, languages)
+
+    def add_datasets(self, datasets: Sequence[str]) -> None:
+        self.add_array(Keys.General.DATASETS, datasets)
+
+    def add_tensor_data_layout(self, layout: str) -> None:
+        self.add_string(Keys.LLM.TENSOR_DATA_LAYOUT.format(arch=self.arch), layout)
+
+    def add_vocab_size(self, size: int) -> None:
+        self.add_uint32(Keys.LLM.VOCAB_SIZE.format(arch=self.arch), size)
+
+    def add_context_length(self, length: int) -> None:
+        self.add_uint32(Keys.LLM.CONTEXT_LENGTH.format(arch=self.arch), length)
+
+    def add_embedding_length(self, length: int) -> None:
+        self.add_uint32(Keys.LLM.EMBEDDING_LENGTH.format(arch=self.arch), length)
+
+    def add_block_count(self, length: int) -> None:
+        self.add_uint32(Keys.LLM.BLOCK_COUNT.format(arch=self.arch), length)
+
+    def add_leading_dense_block_count(self, length: int) -> None:
+        self.add_uint32(Keys.LLM.LEADING_DENSE_BLOCK_COUNT.format(arch=self.arch), length)
+
+    def add_feed_forward_length(self, length: int | Sequence[int]) -> None:
+        if isinstance(length, int):
+            self.add_uint32(Keys.LLM.FEED_FORWARD_LENGTH.format(arch=self.arch), length)
+        else:
+            self.add_array(Keys.LLM.FEED_FORWARD_LENGTH.format(arch=self.arch), length)
+
+    def add_expert_feed_forward_length(self, length: int) -> None:
+        self.add_uint32(Keys.LLM.EXPERT_FEED_FORWARD_LENGTH.format(arch=self.arch), length)
+
+    def add_expert_shared_feed_forward_length(self, length: int) -> None:
+        self.add_uint32(Keys.LLM.EXPERT_SHARED_FEED_FORWARD_LENGTH.format(arch=self.arch), length)
+
+    def add_parallel_residual(self, use: bool) -> None:
+        self.add_bool(Keys.LLM.USE_PARALLEL_RESIDUAL.format(arch=self.arch), use)
+
+    def add_decoder_start_token_id(self, id: int) -> None:
+        self.add_uint32(Keys.LLM.DECODER_START_TOKEN_ID.format(arch=self.arch), id)
+
+    def add_head_count(self, count: int | Sequence[int]) -> None:
+        if isinstance(count, int):
+            self.add_uint32(Keys.Attention.HEAD_COUNT.format(arch=self.arch), count)
+        else:
+            self.add_array(Keys.Attention.HEAD_COUNT.format(arch=self.arch), count)
+
+    def add_head_count_kv(self, count: int | Sequence[int]) -> None:
+        if isinstance(count, int):
+            self.add_uint32(Keys.Attention.HEAD_COUNT_KV.format(arch=self.arch), count)
+        else:
+            self.add_array(Keys.Attention.HEAD_COUNT_KV.format(arch=self.arch), count)
+
+    def add_key_length(self, length: int) -> None:
+        self.add_uint32(Keys.Attention.KEY_LENGTH.format(arch=self.arch), length)
+
+    def add_value_length(self, length: int) -> None:
+        self.add_uint32(Keys.Attention.VALUE_LENGTH.format(arch=self.arch), length)
+
+    def add_max_alibi_bias(self, bias: float) -> None:
+        self.add_float32(Keys.Attention.MAX_ALIBI_BIAS.format(arch=self.arch), bias)
+
+    def add_clamp_kqv(self, value: float) -> None:
+        self.add_float32(Keys.Attention.CLAMP_KQV.format(arch=self.arch), value)
+
+    def add_logit_scale(self, value: float) -> None:
+        self.add_float32(Keys.LLM.LOGIT_SCALE.format(arch=self.arch), value)
+
+    def add_attn_logit_softcapping(self, value: float) -> None:
+        self.add_float32(Keys.LLM.ATTN_LOGIT_SOFTCAPPING.format(arch=self.arch), value)
+
+    def add_final_logit_softcapping(self, value: float) -> None:
+        self.add_float32(Keys.LLM.FINAL_LOGIT_SOFTCAPPING.format(arch=self.arch), value)
+
+    def add_expert_count(self, count: int) -> None:
+        self.add_uint32(Keys.LLM.EXPERT_COUNT.format(arch=self.arch), count)
+
+    def add_expert_used_count(self, count: int) -> None:
+        self.add_uint32(Keys.LLM.EXPERT_USED_COUNT.format(arch=self.arch), count)
+
+    def add_expert_shared_count(self, count: int) -> None:
+        self.add_uint32(Keys.LLM.EXPERT_SHARED_COUNT.format(arch=self.arch), count)
+
+    def add_expert_weights_scale(self, value: float) -> None:
+        self.add_float32(Keys.LLM.EXPERT_WEIGHTS_SCALE.format(arch=self.arch), value)
+
+    def add_swin_norm(self, value: bool) -> None:
+        self.add_bool(Keys.LLM.SWIN_NORM.format(arch=self.arch), value)
+
+    def add_rescale_every_n_layers(self, count: int) -> None:
+        self.add_uint32(Keys.LLM.RESCALE_EVERY_N_LAYERS.format(arch=self.arch), count)
+
+    def add_time_mix_extra_dim(self, dim: int) -> None:
+        self.add_uint32(Keys.LLM.TIME_MIX_EXTRA_DIM.format(arch=self.arch), dim)
+
+    def add_time_decay_extra_dim(self, dim: int) -> None:
+        self.add_uint32(Keys.LLM.TIME_DECAY_EXTRA_DIM.format(arch=self.arch), dim)
+
+    def add_residual_scale(self, value: float) -> None:
+        self.add_float32(Keys.LLM.RESIDUAL_SCALE.format(arch=self.arch), value)
+
+    def add_embedding_scale(self, value: float) -> None:
+        self.add_float32(Keys.LLM.EMBEDDING_SCALE.format(arch=self.arch), value)
+
+    def add_wkv_head_size(self, size: int) -> None:
+        self.add_uint32(Keys.WKV.HEAD_SIZE.format(arch=self.arch), size)
+
+    def add_layer_norm_eps(self, value: float) -> None:
+        self.add_float32(Keys.Attention.LAYERNORM_EPS.format(arch=self.arch), value)
+
+    def add_layer_norm_rms_eps(self, value: float) -> None:
+        self.add_float32(Keys.Attention.LAYERNORM_RMS_EPS.format(arch=self.arch), value)
+
+    def add_causal_attention(self, value: bool) -> None:
+        self.add_bool(Keys.Attention.CAUSAL.format(arch=self.arch), value)
+
+    def add_q_lora_rank(self, length: int) -> None:
+        self.add_uint32(Keys.Attention.Q_LORA_RANK.format(arch=self.arch), length)
+
+    def add_kv_lora_rank(self, length: int) -> None:
+        self.add_uint32(Keys.Attention.KV_LORA_RANK.format(arch=self.arch), length)
+
+    def add_relative_attn_buckets_count(self, value: int) -> None:
+        self.add_uint32(Keys.Attention.REL_BUCKETS_COUNT.format(arch=self.arch), value)
+
+    def add_sliding_window(self, value: int) -> None:
+        self.add_uint32(Keys.Attention.SLIDING_WINDOW.format(arch=self.arch), value)
+
+    def add_attention_scale(self, value: float) -> None:
+        self.add_float32(Keys.Attention.SCALE.format(arch=self.arch), value)
+
+    def add_pooling_type(self, value: PoolingType) -> None:
+        self.add_uint32(Keys.LLM.POOLING_TYPE.format(arch=self.arch), value.value)
+
+    def add_rope_dimension_count(self, count: int) -> None:
+        self.add_uint32(Keys.Rope.DIMENSION_COUNT.format(arch=self.arch), count)
+
+    def add_rope_freq_base(self, value: float) -> None:
+        self.add_float32(Keys.Rope.FREQ_BASE.format(arch=self.arch), value)
+
+    def add_rope_scaling_type(self, value: RopeScalingType) -> None:
+        self.add_string(Keys.Rope.SCALING_TYPE.format(arch=self.arch), value.value)
+
+    def add_rope_scaling_factor(self, value: float) -> None:
+        self.add_float32(Keys.Rope.SCALING_FACTOR.format(arch=self.arch), value)
+
+    def add_rope_scaling_attn_factors(self, value: float) -> None:
+        self.add_float32(Keys.Rope.SCALING_ATTN_FACTOR.format(arch=self.arch), value)
+
+    def add_rope_scaling_orig_ctx_len(self, value: int) -> None:
+        self.add_uint32(Keys.Rope.SCALING_ORIG_CTX_LEN.format(arch=self.arch), value)
+
+    def add_rope_scaling_finetuned(self, value: bool) -> None:
+        self.add_bool(Keys.Rope.SCALING_FINETUNED.format(arch=self.arch), value)
+
+    def add_rope_scaling_yarn_log_mul(self, value: float) -> None:
+        self.add_float32(Keys.Rope.SCALING_YARN_LOG_MUL.format(arch=self.arch), value)
+
+    def add_ssm_conv_kernel(self, value: int) -> None:
+        self.add_uint32(Keys.SSM.CONV_KERNEL.format(arch=self.arch), value)
+
+    def add_ssm_inner_size(self, value: int) -> None:
+        self.add_uint32(Keys.SSM.INNER_SIZE.format(arch=self.arch), value)
+
+    def add_ssm_state_size(self, value: int) -> None:
+        self.add_uint32(Keys.SSM.STATE_SIZE.format(arch=self.arch), value)
+
+    def add_ssm_time_step_rank(self, value: int) -> None:
+        self.add_uint32(Keys.SSM.TIME_STEP_RANK.format(arch=self.arch), value)
+
+    def add_ssm_dt_b_c_rms(self, value: bool) -> None:
+        self.add_bool(Keys.SSM.DT_B_C_RMS.format(arch=self.arch), value)
+
+    def add_tokenizer_model(self, model: str) -> None:
+        self.add_string(Keys.Tokenizer.MODEL, model)
+
+    def add_tokenizer_pre(self, pre: str) -> None:
+        self.add_string(Keys.Tokenizer.PRE, pre)
+
+    def add_token_list(self, tokens: Sequence[str] | Sequence[bytes] | Sequence[bytearray]) -> None:
+        self.add_array(Keys.Tokenizer.LIST, tokens)
+
+    def add_token_merges(self, merges: Sequence[str] | Sequence[bytes] | Sequence[bytearray]) -> None:
+        self.add_array(Keys.Tokenizer.MERGES, merges)
+
+    def add_token_types(self, types: Sequence[TokenType] | Sequence[int]) -> None:
+        self.add_array(Keys.Tokenizer.TOKEN_TYPE, types)
+
+    def add_token_type_count(self, value: int) -> None:
+        self.add_uint32(Keys.Tokenizer.TOKEN_TYPE_COUNT, value)
+
+    def add_token_scores(self, scores: Sequence[float]) -> None:
+        self.add_array(Keys.Tokenizer.SCORES, scores)
+
+    def add_bos_token_id(self, id: int) -> None:
+        self.add_uint32(Keys.Tokenizer.BOS_ID, id)
+
+    def add_eos_token_id(self, id: int) -> None:
+        self.add_uint32(Keys.Tokenizer.EOS_ID, id)
+
+    def add_unk_token_id(self, id: int) -> None:
+        self.add_uint32(Keys.Tokenizer.UNK_ID, id)
+
+    def add_sep_token_id(self, id: int) -> None:
+        self.add_uint32(Keys.Tokenizer.SEP_ID, id)
+
+    def add_pad_token_id(self, id: int) -> None:
+        self.add_uint32(Keys.Tokenizer.PAD_ID, id)
+
+    def add_cls_token_id(self, id: int) -> None:
+        self.add_uint32(Keys.Tokenizer.CLS_ID, id)
+
+    def add_mask_token_id(self, id: int) -> None:
+        self.add_uint32(Keys.Tokenizer.MASK_ID, id)
+
+    def add_add_bos_token(self, value: bool) -> None:
+        self.add_bool(Keys.Tokenizer.ADD_BOS, value)
+
+    def add_add_eos_token(self, value: bool) -> None:
+        self.add_bool(Keys.Tokenizer.ADD_EOS, value)
+
+    def add_add_space_prefix(self, value: bool) -> None:
+        self.add_bool(Keys.Tokenizer.ADD_PREFIX, value)
+
+    def add_remove_extra_whitespaces(self, value: bool) -> None:
+        self.add_bool(Keys.Tokenizer.REMOVE_EXTRA_WS, value)
+
+    def add_precompiled_charsmap(self, charsmap: Sequence[bytes]) -> None:
+        self.add_array(Keys.Tokenizer.PRECOMPILED_CHARSMAP, charsmap)
+
+    def add_chat_template(self, value: str | Sequence[Mapping[str, str]]) -> None:
+        if not isinstance(value, str):
+            template_default = None
+            template_names = set()
+
+            for choice in value:
+                name = choice.get('name', '')
+                template = choice.get('template')
+
+                # Allowing non-alphanumerical characters in template name is probably not a good idea, so filter it
+                name = ''.join((c if c in ascii_letters + digits else '_' for c in name))
+
+                if name and template is not None:
+                    if name == 'default':
+                        template_default = template
+                    else:
+                        template_names.add(name)
+                        self.add_string(Keys.Tokenizer.CHAT_TEMPLATE_N.format(name=name), template)
+
+            if template_names:
+                self.add_array(Keys.Tokenizer.CHAT_TEMPLATES, list(template_names))
+
+            if template_default is None:
+                return
+
+            value = template_default
+
+        self.add_string(Keys.Tokenizer.CHAT_TEMPLATE, value)
+
+    def add_eot_token_id(self, id: int) -> None:
+        self.add_uint32(Keys.Tokenizer.EOT_ID, id)
+
+    def add_eom_token_id(self, id: int) -> None:
+        self.add_uint32(Keys.Tokenizer.EOM_ID, id)
+
+    def _pack(self, fmt: str, value: Any, skip_pack_prefix: bool = False) -> bytes:
+        pack_prefix = ''
+        if not skip_pack_prefix:
+            pack_prefix = '<' if self.endianess == GGUFEndian.LITTLE else '>'
+        return struct.pack(f'{pack_prefix}{fmt}', value)
+
+    def _pack_val(self, val: Any, vtype: GGUFValueType, add_vtype: bool) -> bytes:
+        kv_data = bytearray()
+
+        if add_vtype:
+            kv_data += self._pack("I", vtype)
+
+        pack_fmt = self._simple_value_packing.get(vtype)
+        if pack_fmt is not None:
+            kv_data += self._pack(pack_fmt, val, skip_pack_prefix = vtype == GGUFValueType.BOOL)
+        elif vtype == GGUFValueType.STRING:
+            encoded_val = val.encode("utf-8") if isinstance(val, str) else val
+            kv_data += self._pack("Q", len(encoded_val))
+            kv_data += encoded_val
+        elif vtype == GGUFValueType.ARRAY:
+
+            if not isinstance(val, Sequence):
+                raise ValueError("Invalid GGUF metadata array, expecting sequence")
+
+            if len(val) == 0:
+                raise ValueError("Invalid GGUF metadata array. Empty array")
+
+            if isinstance(val, bytes):
+                ltype = GGUFValueType.UINT8
+            else:
+                ltype = GGUFValueType.get_type(val[0])
+                if not all(GGUFValueType.get_type(i) is ltype for i in val[1:]):
+                    raise ValueError("All items in a GGUF array should be of the same type")
+            kv_data += self._pack("I", ltype)
+            kv_data += self._pack("Q", len(val))
+            for item in val:
+                kv_data += self._pack_val(item, ltype, add_vtype=False)
+        else:
+            raise ValueError("Invalid GGUF metadata value type or value")
+
+        return kv_data
+
+    @staticmethod
+    def format_n_bytes_to_str(num: int) -> str:
+        if num == 0:
+            return "negligible - metadata only"
+        fnum = float(num)
+        for unit in ("", "K", "M", "G"):
+            if abs(fnum) < 1000.0:
+                return f"{fnum:3.1f}{unit}"
+            fnum /= 1000.0
+        return f"{fnum:.1f}T - over 1TB, split recommended"

gguf/lazy.py

@@ -0,0 +1,213 @@
+from __future__ import annotations
+from abc import ABC, ABCMeta, abstractmethod
+
+import logging
+from typing import Any, Callable
+
+import numpy as np
+from numpy.typing import DTypeLike
+
+
+logger = logging.getLogger(__name__)
+
+
+class LazyMeta(ABCMeta):
+
+    def __new__(cls, name: str, bases: tuple[type, ...], namespace: dict[str, Any], **kwargs):
+        def __getattr__(self, name: str) -> Any:
+            meta_attr = getattr(self._meta, name)
+            if callable(meta_attr):
+                return type(self)._wrap_fn(
+                    (lambda s, *args, **kwargs: getattr(s, name)(*args, **kwargs)),
+                    use_self=self,
+                )
+            elif isinstance(meta_attr, self._tensor_type):
+                # e.g. self.T with torch.Tensor should still be wrapped
+                return type(self)._wrap_fn(lambda s: getattr(s, name))(self)
+            else:
+                # no need to wrap non-tensor properties,
+                # and they likely don't depend on the actual contents of the tensor
+                return meta_attr
+
+        namespace["__getattr__"] = __getattr__
+
+        # need to make a builder for the wrapped wrapper to copy the name,
+        # or else it fails with very cryptic error messages,
+        # because somehow the same string would end up in every closures
+        def mk_wrap(op_name: str, *, meta_noop: bool = False):
+            # need to wrap the wrapper to get self
+            def wrapped_special_op(self, *args, **kwargs):
+                return type(self)._wrap_fn(
+                    getattr(type(self)._tensor_type, op_name),
+                    meta_noop=meta_noop,
+                )(self, *args, **kwargs)
+            return wrapped_special_op
+
+        # special methods bypass __getattr__, so they need to be added manually
+        # ref: https://docs.python.org/3/reference/datamodel.html#special-lookup
+        # NOTE: doing this from a metaclass is very convenient
+        # TODO: make this even more comprehensive
+        for binary_op in (
+            "lt", "le", "eq", "ne", "ge", "gt", "not"
+            "abs", "add", "and", "floordiv", "invert", "lshift", "mod", "mul", "matmul",
+            "neg", "or", "pos", "pow", "rshift", "sub", "truediv", "xor",
+            "iadd", "iand", "ifloordiv", "ilshift", "imod", "imul", "ior", "irshift", "isub", "ixor",
+            "radd", "rand", "rfloordiv", "rmul", "ror", "rpow", "rsub", "rtruediv", "rxor",
+        ):
+            attr_name = f"__{binary_op}__"
+            # the result of these operators usually has the same shape and dtype as the input,
+            # so evaluation on the meta tensor can be skipped.
+            namespace[attr_name] = mk_wrap(attr_name, meta_noop=True)
+
+        for special_op in (
+            "getitem", "setitem", "len",
+        ):
+            attr_name = f"__{special_op}__"
+            namespace[attr_name] = mk_wrap(attr_name, meta_noop=False)
+
+        return super().__new__(cls, name, bases, namespace, **kwargs)
+
+
+# Tree of lazy tensors
+class LazyBase(ABC, metaclass=LazyMeta):
+    _tensor_type: type
+    _meta: Any
+    _data: Any | None
+    _args: tuple
+    _kwargs: dict[str, Any]
+    _func: Callable[[Any], Any] | None
+
+    def __init__(self, *, meta: Any, data: Any | None = None, args: tuple = (), kwargs: dict[str, Any] | None = None, func: Callable[[Any], Any] | None = None):
+        super().__init__()
+        self._meta = meta
+        self._data = data
+        self._args = args
+        self._kwargs = kwargs if kwargs is not None else {}
+        self._func = func
+        assert self._func is not None or self._data is not None
+
+    def __init_subclass__(cls) -> None:
+        if "_tensor_type" not in cls.__dict__:
+            raise TypeError(f"property '_tensor_type' must be defined for {cls!r}")
+        return super().__init_subclass__()
+
+    @staticmethod
+    def _recurse_apply(o: Any, fn: Callable[[Any], Any]) -> Any:
+        # TODO: dict and set
+        if isinstance(o, (list, tuple)):
+            L = []
+            for item in o:
+                L.append(LazyBase._recurse_apply(item, fn))
+            if isinstance(o, tuple):
+                L = tuple(L)
+            return L
+        elif isinstance(o, LazyBase):
+            return fn(o)
+        else:
+            return o
+
+    @classmethod
+    def _wrap_fn(cls, fn: Callable, *, use_self: LazyBase | None = None, meta_noop: bool | DTypeLike | tuple[DTypeLike, Callable[[tuple[int, ...]], tuple[int, ...]]] = False) -> Callable[[Any], Any]:
+        def wrapped_fn(*args, **kwargs):
+            if kwargs is None:
+                kwargs = {}
+            args = ((use_self,) if use_self is not None else ()) + args
+
+            meta_args = LazyBase._recurse_apply(args, lambda t: t._meta)
+            # TODO: maybe handle tensors in kwargs too
+
+            if isinstance(meta_noop, bool) and not meta_noop:
+                try:
+                    res = fn(*meta_args, **kwargs)
+                except NotImplementedError:
+                    # running some operations on PyTorch's Meta tensors can cause this exception
+                    res = None
+            else:
+                # some operators don't need to actually run on the meta tensors
+                assert len(args) > 0
+                res = args[0]
+                assert isinstance(res, cls)
+                res = res._meta
+                # allow operations to override the dtype and shape
+                if meta_noop is not True:
+                    if isinstance(meta_noop, tuple):
+                        dtype, shape = meta_noop
+                        assert callable(shape)
+                        res = cls.meta_with_dtype_and_shape(dtype, shape(res.shape))
+                    else:
+                        res = cls.meta_with_dtype_and_shape(meta_noop, res.shape)
+
+            if isinstance(res, cls._tensor_type):
+                return cls(meta=cls.eager_to_meta(res), args=args, kwargs=kwargs, func=fn)
+            else:
+                del res  # not needed
+                # non-tensor return likely relies on the contents of the args
+                # (e.g. the result of torch.equal)
+                eager_args = cls.to_eager(args)
+                return fn(*eager_args, **kwargs)
+        return wrapped_fn
+
+    @classmethod
+    def to_eager(cls, t: Any) -> Any:
+        def simple_to_eager(_t: LazyBase) -> Any:
+            if _t._data is not None:
+                return _t._data
+
+            # NOTE: there's a recursion limit in Python (usually 1000)
+
+            assert _t._func is not None
+            _t._args = cls._recurse_apply(_t._args, simple_to_eager)
+            _t._data = _t._func(*_t._args, **_t._kwargs)
+            # sanity check
+            assert _t._data is not None
+            assert _t._data.dtype == _t._meta.dtype
+            assert _t._data.shape == _t._meta.shape
+
+            return _t._data
+
+        # recurse into lists and/or tuples, keeping their structure
+        return cls._recurse_apply(t, simple_to_eager)
+
+    @classmethod
+    def eager_to_meta(cls, t: Any) -> Any:
+        return cls.meta_with_dtype_and_shape(t.dtype, t.shape)
+
+    # must be overridden, meta tensor init is backend-specific
+    @classmethod
+    @abstractmethod
+    def meta_with_dtype_and_shape(cls, dtype: Any, shape: Any) -> Any: pass
+
+    @classmethod
+    def from_eager(cls, t: Any) -> Any:
+        if type(t) is cls:
+            # already lazy
+            return t
+        elif isinstance(t, cls._tensor_type):
+            return cls(meta=cls.eager_to_meta(t), data=t)
+        else:
+            return TypeError(f"{type(t)!r} is not compatible with {cls._tensor_type!r}")
+
+
+class LazyNumpyTensor(LazyBase):
+    _tensor_type = np.ndarray
+
+    shape: tuple[int, ...]  # Makes the type checker happy in quants.py
+
+    @classmethod
+    def meta_with_dtype_and_shape(cls, dtype: DTypeLike, shape: tuple[int, ...]) -> np.ndarray[Any, Any]:
+        # The initial idea was to use np.nan as the fill value,
+        # but non-float types like np.int16 can't use that.
+        # So zero it is.
+        cheat = np.zeros(1, dtype)
+        return np.lib.stride_tricks.as_strided(cheat, shape, (0 for _ in shape))
+
+    def astype(self, dtype, *args, **kwargs):
+        meta = type(self).meta_with_dtype_and_shape(dtype, self._meta.shape)
+        full_args = (self, dtype,) + args
+        return type(self)(meta=meta, args=full_args, kwargs=kwargs, func=(lambda a, *args, **kwargs: a.astype(*args, **kwargs)))
+
+    def tofile(self, *args, **kwargs):
+        eager = LazyNumpyTensor.to_eager(self)
+        return eager.tofile(*args, **kwargs)
+
+    # TODO: __array_function__

gguf/metadata.py

@@ -0,0 +1,510 @@
+from __future__ import annotations
+
+import re
+import json
+import yaml
+import logging
+from pathlib import Path
+from typing import Any, Literal, Optional
+from dataclasses import dataclass
+
+from .constants import Keys
+
+import gguf
+
+logger = logging.getLogger("metadata")
+
+
+@dataclass
+class Metadata:
+    # Authorship Metadata to be written to GGUF KV Store
+    name: Optional[str] = None
+    author: Optional[str] = None
+    version: Optional[str] = None
+    organization: Optional[str] = None
+    finetune: Optional[str] = None
+    basename: Optional[str] = None
+    description: Optional[str] = None
+    quantized_by: Optional[str] = None
+    size_label: Optional[str] = None
+    url: Optional[str] = None
+    doi: Optional[str] = None
+    uuid: Optional[str] = None
+    repo_url: Optional[str] = None
+    source_url: Optional[str] = None
+    source_doi: Optional[str] = None
+    source_uuid: Optional[str] = None
+    source_repo_url: Optional[str] = None
+    license: Optional[str] = None
+    license_name: Optional[str] = None
+    license_link: Optional[str] = None
+    base_models: Optional[list[dict]] = None
+    tags: Optional[list[str]] = None
+    languages: Optional[list[str]] = None
+    datasets: Optional[list[str]] = None
+
+    @staticmethod
+    def load(metadata_override_path: Optional[Path] = None, model_path: Optional[Path] = None, model_name: Optional[str] = None, total_params: int = 0) -> Metadata:
+        # This grabs as many contextual authorship metadata as possible from the model repository
+        # making any conversion as required to match the gguf kv store metadata format
+        # as well as giving users the ability to override any authorship metadata that may be incorrect
+
+        # Create a new Metadata instance
+        metadata = Metadata()
+
+        model_card = Metadata.load_model_card(model_path)
+        hf_params = Metadata.load_hf_parameters(model_path)
+        # TODO: load adapter_config.json when possible, it usually contains the base model of the LoRA adapter
+
+        # heuristics
+        metadata = Metadata.apply_metadata_heuristic(metadata, model_card, hf_params, model_path, total_params)
+
+        # Metadata Override File Provided
+        # This is based on LLM_KV_NAMES mapping in llama.cpp
+        metadata_override = Metadata.load_metadata_override(metadata_override_path)
+
+        metadata.name            = metadata_override.get(Keys.General.NAME,            metadata.name)
+        metadata.author          = metadata_override.get(Keys.General.AUTHOR,          metadata.author)
+        metadata.version         = metadata_override.get(Keys.General.VERSION,         metadata.version)
+        metadata.organization    = metadata_override.get(Keys.General.ORGANIZATION,    metadata.organization)
+
+        metadata.finetune        = metadata_override.get(Keys.General.FINETUNE,        metadata.finetune)
+        metadata.basename        = metadata_override.get(Keys.General.BASENAME,        metadata.basename)
+
+        metadata.description     = metadata_override.get(Keys.General.DESCRIPTION,     metadata.description)
+        metadata.quantized_by    = metadata_override.get(Keys.General.QUANTIZED_BY,    metadata.quantized_by)
+
+        metadata.size_label      = metadata_override.get(Keys.General.SIZE_LABEL,      metadata.size_label)
+        metadata.license_name    = metadata_override.get(Keys.General.LICENSE_NAME,    metadata.license_name)
+        metadata.license_link    = metadata_override.get(Keys.General.LICENSE_LINK,    metadata.license_link)
+
+        metadata.url             = metadata_override.get(Keys.General.URL,             metadata.url)
+        metadata.doi             = metadata_override.get(Keys.General.DOI,             metadata.doi)
+        metadata.uuid            = metadata_override.get(Keys.General.UUID,            metadata.uuid)
+        metadata.repo_url        = metadata_override.get(Keys.General.REPO_URL,        metadata.repo_url)
+
+        metadata.source_url      = metadata_override.get(Keys.General.SOURCE_URL,      metadata.source_url)
+        metadata.source_doi      = metadata_override.get(Keys.General.SOURCE_DOI,      metadata.source_doi)
+        metadata.source_uuid     = metadata_override.get(Keys.General.SOURCE_UUID,     metadata.source_uuid)
+        metadata.source_repo_url = metadata_override.get(Keys.General.SOURCE_REPO_URL, metadata.source_repo_url)
+
+        # Base Models is received here as an array of models
+        metadata.base_models     = metadata_override.get("general.base_models",        metadata.base_models)
+
+        metadata.tags            = metadata_override.get(Keys.General.TAGS,            metadata.tags)
+        metadata.languages       = metadata_override.get(Keys.General.LANGUAGES,       metadata.languages)
+        metadata.datasets        = metadata_override.get(Keys.General.DATASETS,        metadata.datasets)
+
+        # Direct Metadata Override (via direct cli argument)
+        if model_name is not None:
+            metadata.name = model_name
+
+        return metadata
+
+    @staticmethod
+    def load_metadata_override(metadata_override_path: Optional[Path] = None) -> dict[str, Any]:
+        if metadata_override_path is None or not metadata_override_path.is_file():
+            return {}
+
+        with open(metadata_override_path, "r", encoding="utf-8") as f:
+            return json.load(f)
+
+    @staticmethod
+    def load_model_card(model_path: Optional[Path] = None) -> dict[str, Any]:
+        if model_path is None or not model_path.is_dir():
+            return {}
+
+        model_card_path = model_path / "README.md"
+
+        if not model_card_path.is_file():
+            return {}
+
+        # The model card metadata is assumed to always be in YAML
+        # ref: https://github.com/huggingface/transformers/blob/a5c642fe7a1f25d3bdcd76991443ba6ff7ee34b2/src/transformers/modelcard.py#L468-L473
+        with open(model_card_path, "r", encoding="utf-8") as f:
+            if f.readline() == "---\n":
+                raw = f.read().partition("---\n")[0]
+                data = yaml.safe_load(raw)
+                if isinstance(data, dict):
+                    return data
+                else:
+                    logger.error(f"while reading YAML model card frontmatter, data is {type(data)} instead of dict")
+                    return {}
+            else:
+                return {}
+
+    @staticmethod
+    def load_hf_parameters(model_path: Optional[Path] = None) -> dict[str, Any]:
+        if model_path is None or not model_path.is_dir():
+            return {}
+
+        config_path = model_path / "config.json"
+
+        if not config_path.is_file():
+            return {}
+
+        with open(config_path, "r", encoding="utf-8") as f:
+            return json.load(f)
+
+    @staticmethod
+    def id_to_title(string):
+        # Convert capitalization into title form unless acronym or version number
+        return ' '.join([w.title() if w.islower() and not re.match(r'^(v\d+(?:\.\d+)*|\d.*)$', w) else w for w in string.strip().replace('-', ' ').split()])
+
+    @staticmethod
+    def get_model_id_components(model_id: Optional[str] = None, total_params: int = 0) -> tuple[str | None, str | None, str | None, str | None, str | None, str | None]:
+        # Huggingface often store model id as '<org>/<model name>'
+        # so let's parse it and apply some heuristics if possible for model name components
+
+        if model_id is None:
+            # model ID missing
+            return None, None, None, None, None, None
+
+        if ' ' in model_id:
+            # model ID is actually a normal human sentence
+            # which means its most likely a normal model name only
+            # not part of the hugging face naming standard, but whatever
+            return model_id, None, None, None, None, None
+
+        if '/' in model_id:
+            # model ID (huggingface style)
+            org_component, model_full_name_component = model_id.split('/', 1)
+        else:
+            # model ID but missing org components
+            org_component, model_full_name_component = None, model_id
+
+        # Check if we erroneously matched against './' or '../' etc...
+        if org_component is not None and len(org_component) > 0 and org_component[0] == '.':
+            org_component = None
+
+        name_parts: list[str] = model_full_name_component.split('-')
+
+        # Remove empty parts
+        for i in reversed(range(len(name_parts))):
+            if len(name_parts[i]) == 0:
+                del name_parts[i]
+
+        name_types: list[
+            set[Literal["basename", "size_label", "finetune", "version", "type"]]
+        ] = [set() for _ in name_parts]
+
+        # Annotate the name
+        for i, part in enumerate(name_parts):
+            # Version
+            if re.fullmatch(r'(v|iter)?\d+([.]\d+)*', part, re.IGNORECASE):
+                name_types[i].add("version")
+            # Quant type (should not be there for base models, but still annotated)
+            elif re.fullmatch(r'i?q\d(_\w)*|b?fp?(16|32)', part, re.IGNORECASE):
+                name_types[i].add("type")
+                name_parts[i] = part.upper()
+            # Model size
+            elif i > 0 and re.fullmatch(r'(([A]|\d+[x])?\d+([._]\d+)?[KMBT][\d]?|small|mini|medium|large|x?xl)', part, re.IGNORECASE):
+                part = part.replace("_", ".")
+                # Handle weird bloom-7b1 notation
+                if part[-1].isdecimal():
+                    part = part[:-2] + "." + part[-1] + part[-2]
+                # Normalize the size suffixes
+                if len(part) > 1 and part[-2].isdecimal():
+                    if part[-1] in "kmbt":
+                        part = part[:-1] + part[-1].upper()
+                if total_params != 0:
+                    try:
+                        label_params = float(part[:-1]) * pow(1000, " KMBT".find(part[-1]))
+                        # Only use it as a size label if it's close or bigger than the model size
+                        # Note that LoRA adapters don't necessarily include all layers,
+                        # so this is why bigger label sizes are accepted.
+                        # Do not use the size label when it's smaller than 1/8 of the model size
+                        if (total_params < 0 and label_params < abs(total_params) // 8) or (
+                            # Check both directions when the current model isn't a LoRA adapter
+                            total_params > 0 and abs(label_params - total_params) > 7 * total_params // 8
+                        ):
+                            # Likely a context length
+                            name_types[i].add("finetune")
+                            # Lowercase the size when it's a context length
+                            part = part[:-1] + part[-1].lower()
+                    except ValueError:
+                        # Failed to convert the size label to float, use it anyway
+                        pass
+                if len(name_types[i]) == 0:
+                    name_types[i].add("size_label")
+                name_parts[i] = part
+            # Some easy to recognize finetune names
+            elif i > 0 and re.fullmatch(r'chat|instruct|vision|lora', part, re.IGNORECASE):
+                if total_params < 0 and part.lower() == "lora":
+                    # ignore redundant "lora" in the finetune part when the output is a lora adapter
+                    name_types[i].add("type")
+                else:
+                    name_types[i].add("finetune")
+
+        # Ignore word-based size labels when there is at least a number-based one present
+        # TODO: should word-based size labels always be removed instead?
+        if any(c.isdecimal() for n, t in zip(name_parts, name_types) if "size_label" in t for c in n):
+            for n, t in zip(name_parts, name_types):
+                if "size_label" in t:
+                    if all(c.isalpha() for c in n):
+                        t.remove("size_label")
+
+        at_start = True
+        # Find the basename through the annotated name
+        for part, t in zip(name_parts, name_types):
+            if at_start and ((len(t) == 0 and part[0].isalpha()) or "version" in t):
+                t.add("basename")
+            else:
+                if at_start:
+                    at_start = False
+                if len(t) == 0:
+                    t.add("finetune")
+
+        # Remove the basename annotation from trailing version
+        for part, t in zip(reversed(name_parts), reversed(name_types)):
+            if "basename" in t and len(t) > 1:
+                t.remove("basename")
+            else:
+                break
+
+        basename = "-".join(n for n, t in zip(name_parts, name_types) if "basename" in t) or None
+        # Deduplicate size labels using order-preserving 'dict' ('set' seems to sort the keys)
+        size_label = "-".join(dict.fromkeys(s for s, t in zip(name_parts, name_types) if "size_label" in t).keys()) or None
+        finetune = "-".join(f for f, t in zip(name_parts, name_types) if "finetune" in t) or None
+        # TODO: should the basename version always be excluded?
+        # NOTE: multiple finetune versions are joined together
+        version = "-".join(v for v, t, in zip(name_parts, name_types) if "version" in t and "basename" not in t) or None
+
+        if size_label is None and finetune is None and version is None:
+            # Too ambiguous, output nothing
+            basename = None
+
+        return model_full_name_component, org_component, basename, finetune, version, size_label
+
+    @staticmethod
+    def apply_metadata_heuristic(metadata: Metadata, model_card: Optional[dict] = None, hf_params: Optional[dict] = None, model_path: Optional[Path] = None, total_params: int = 0) -> Metadata:
+        # Reference Model Card Metadata: https://github.com/huggingface/hub-docs/blob/main/modelcard.md?plain=1
+
+        # Model Card Heuristics
+        ########################
+        if model_card is not None:
+
+            def use_model_card_metadata(metadata_key: str, model_card_key: str):
+                if model_card_key in model_card and getattr(metadata, metadata_key, None) is None:
+                    setattr(metadata, metadata_key, model_card.get(model_card_key))
+
+            def use_array_model_card_metadata(metadata_key: str, model_card_key: str):
+                # Note: Will append rather than replace if already exist
+                tags_value = model_card.get(model_card_key, None)
+                if tags_value is None:
+                    return
+
+                current_value = getattr(metadata, metadata_key, None)
+                if current_value is None:
+                    current_value = []
+
+                if isinstance(tags_value, str):
+                    current_value.append(tags_value)
+                elif isinstance(tags_value, list):
+                    current_value.extend(tags_value)
+
+                setattr(metadata, metadata_key, current_value)
+
+            # LLAMA.cpp's direct internal convention
+            # (Definitely not part of hugging face formal/informal standard)
+            #########################################
+            use_model_card_metadata("name", "name")
+            use_model_card_metadata("author", "author")
+            use_model_card_metadata("version", "version")
+            use_model_card_metadata("organization", "organization")
+            use_model_card_metadata("description", "description")
+            use_model_card_metadata("finetune", "finetune")
+            use_model_card_metadata("basename", "basename")
+            use_model_card_metadata("size_label", "size_label")
+            use_model_card_metadata("source_url", "url")
+            use_model_card_metadata("source_doi", "doi")
+            use_model_card_metadata("source_uuid", "uuid")
+            use_model_card_metadata("source_repo_url", "repo_url")
+
+            # LLAMA.cpp's huggingface style convention
+            # (Definitely not part of hugging face formal/informal standard... but with model_ appended to match their style)
+            ###########################################
+            use_model_card_metadata("name", "model_name")
+            use_model_card_metadata("author", "model_author")
+            use_model_card_metadata("version", "model_version")
+            use_model_card_metadata("organization", "model_organization")
+            use_model_card_metadata("description", "model_description")
+            use_model_card_metadata("finetune", "model_finetune")
+            use_model_card_metadata("basename", "model_basename")
+            use_model_card_metadata("size_label", "model_size_label")
+            use_model_card_metadata("source_url", "model_url")
+            use_model_card_metadata("source_doi", "model_doi")
+            use_model_card_metadata("source_uuid", "model_uuid")
+            use_model_card_metadata("source_repo_url", "model_repo_url")
+
+            # Hugging Face Direct Convention
+            #################################
+
+            # Not part of huggingface model card standard but notice some model creator using it
+            # such as TheBloke in 'TheBloke/Mistral-7B-Instruct-v0.2-GGUF'
+            use_model_card_metadata("name", "model_name")
+            use_model_card_metadata("author", "model_creator")
+            use_model_card_metadata("basename", "model_type")
+
+            if "base_model" in model_card:
+                # This represents the parent models that this is based on
+                # Example: stabilityai/stable-diffusion-xl-base-1.0. Can also be a list (for merges)
+                # Example of merges: https://huggingface.co/EmbeddedLLM/Mistral-7B-Merge-14-v0.1/blob/main/README.md
+                metadata_base_models = []
+                base_model_value = model_card.get("base_model", None)
+
+                if base_model_value is not None:
+                    if isinstance(base_model_value, str):
+                        metadata_base_models.append(base_model_value)
+                    elif isinstance(base_model_value, list):
+                        metadata_base_models.extend(base_model_value)
+
+                if metadata.base_models is None:
+                    metadata.base_models = []
+
+                for model_id in metadata_base_models:
+                    # NOTE: model size of base model is assumed to be similar to the size of the current model
+                    model_full_name_component, org_component, basename, finetune, version, size_label = Metadata.get_model_id_components(model_id, total_params)
+                    base_model = {}
+                    if model_full_name_component is not None:
+                        base_model["name"] = Metadata.id_to_title(model_full_name_component)
+                    if org_component is not None:
+                        base_model["organization"] = Metadata.id_to_title(org_component)
+                    if version is not None:
+                        base_model["version"] = version
+                    if org_component is not None and model_full_name_component is not None:
+                        base_model["repo_url"] = f"https://huggingface.co/{org_component}/{model_full_name_component}"
+                    metadata.base_models.append(base_model)
+
+            use_model_card_metadata("license", "license")
+            use_model_card_metadata("license_name", "license_name")
+            use_model_card_metadata("license_link", "license_link")
+
+            use_array_model_card_metadata("tags", "tags")
+            use_array_model_card_metadata("tags", "pipeline_tag")
+
+            use_array_model_card_metadata("languages", "languages")
+            use_array_model_card_metadata("languages", "language")
+
+            use_array_model_card_metadata("datasets", "datasets")
+            use_array_model_card_metadata("datasets", "dataset")
+
+        # Hugging Face Parameter Heuristics
+        ####################################
+
+        if hf_params is not None:
+
+            hf_name_or_path = hf_params.get("_name_or_path")
+            if hf_name_or_path is not None and hf_name_or_path.count('/') <= 1:
+                # Use _name_or_path only if its actually a model name and not some computer path
+                # e.g. 'meta-llama/Llama-2-7b-hf'
+                model_id = hf_name_or_path
+                model_full_name_component, org_component, basename, finetune, version, size_label = Metadata.get_model_id_components(model_id, total_params)
+                if metadata.name is None and model_full_name_component is not None:
+                    metadata.name = Metadata.id_to_title(model_full_name_component)
+                if metadata.organization is None and org_component is not None:
+                    metadata.organization = Metadata.id_to_title(org_component)
+                if metadata.basename is None and basename is not None:
+                    metadata.basename = basename
+                if metadata.finetune is None and finetune is not None:
+                    metadata.finetune = finetune
+                if metadata.version is None and version is not None:
+                    metadata.version = version
+                if metadata.size_label is None and size_label is not None:
+                    metadata.size_label = size_label
+
+        # Directory Folder Name Fallback Heuristics
+        ############################################
+        if model_path is not None:
+            model_id = model_path.name
+            model_full_name_component, org_component, basename, finetune, version, size_label = Metadata.get_model_id_components(model_id, total_params)
+            if metadata.name is None and model_full_name_component is not None:
+                metadata.name = Metadata.id_to_title(model_full_name_component)
+            if metadata.organization is None and org_component is not None:
+                metadata.organization = Metadata.id_to_title(org_component)
+            if metadata.basename is None and basename is not None:
+                metadata.basename = basename
+            if metadata.finetune is None and finetune is not None:
+                metadata.finetune = finetune
+            if metadata.version is None and version is not None:
+                metadata.version = version
+            if metadata.size_label is None and size_label is not None:
+                metadata.size_label = size_label
+
+        return metadata
+
+    def set_gguf_meta_model(self, gguf_writer: gguf.GGUFWriter):
+        assert self.name is not None
+        gguf_writer.add_name(self.name)
+
+        if self.author is not None:
+            gguf_writer.add_author(self.author)
+        if self.version is not None:
+            gguf_writer.add_version(self.version)
+        if self.organization is not None:
+            gguf_writer.add_organization(self.organization)
+
+        if self.finetune is not None:
+            gguf_writer.add_finetune(self.finetune)
+        if self.basename is not None:
+            gguf_writer.add_basename(self.basename)
+
+        if self.description is not None:
+            gguf_writer.add_description(self.description)
+        if self.quantized_by is not None:
+            gguf_writer.add_quantized_by(self.quantized_by)
+
+        if self.size_label is not None:
+            gguf_writer.add_size_label(self.size_label)
+
+        if self.license is not None:
+            gguf_writer.add_license(self.license)
+        if self.license_name is not None:
+            gguf_writer.add_license_name(self.license_name)
+        if self.license_link is not None:
+            gguf_writer.add_license_link(self.license_link)
+
+        if self.url is not None:
+            gguf_writer.add_url(self.url)
+        if self.doi is not None:
+            gguf_writer.add_doi(self.doi)
+        if self.uuid is not None:
+            gguf_writer.add_uuid(self.uuid)
+        if self.repo_url is not None:
+            gguf_writer.add_repo_url(self.repo_url)
+
+        if self.source_url is not None:
+            gguf_writer.add_source_url(self.source_url)
+        if self.source_doi is not None:
+            gguf_writer.add_source_doi(self.source_doi)
+        if self.source_uuid is not None:
+            gguf_writer.add_source_uuid(self.source_uuid)
+        if self.source_repo_url is not None:
+            gguf_writer.add_source_repo_url(self.source_repo_url)
+
+        if self.base_models is not None:
+            gguf_writer.add_base_model_count(len(self.base_models))
+            for key, base_model_entry in enumerate(self.base_models):
+                if "name" in base_model_entry:
+                    gguf_writer.add_base_model_name(key, base_model_entry["name"])
+                if "author" in base_model_entry:
+                    gguf_writer.add_base_model_author(key, base_model_entry["author"])
+                if "version" in base_model_entry:
+                    gguf_writer.add_base_model_version(key, base_model_entry["version"])
+                if "organization" in base_model_entry:
+                    gguf_writer.add_base_model_organization(key, base_model_entry["organization"])
+                if "url" in base_model_entry:
+                    gguf_writer.add_base_model_url(key, base_model_entry["url"])
+                if "doi" in base_model_entry:
+                    gguf_writer.add_base_model_doi(key, base_model_entry["doi"])
+                if "uuid" in base_model_entry:
+                    gguf_writer.add_base_model_uuid(key, base_model_entry["uuid"])
+                if "repo_url" in base_model_entry:
+                    gguf_writer.add_base_model_repo_url(key, base_model_entry["repo_url"])
+
+        if self.tags is not None:
+            gguf_writer.add_tags(self.tags)
+        if self.languages is not None:
+            gguf_writer.add_languages(self.languages)
+        if self.datasets is not None:
+            gguf_writer.add_datasets(self.datasets)

gguf/quants.py

@@ -0,0 +1,1269 @@
+from __future__ import annotations
+from abc import ABC, abstractmethod
+from typing import Any, Callable, Sequence
+from math import log2, ceil
+
+from numpy.typing import DTypeLike
+
+from .constants import GGML_QUANT_SIZES, GGMLQuantizationType, QK_K
+from .lazy import LazyNumpyTensor
+
+import numpy as np
+
+
+def quant_shape_to_byte_shape(shape: Sequence[int], quant_type: GGMLQuantizationType) -> tuple[int, ...]:
+    block_size, type_size = GGML_QUANT_SIZES[quant_type]
+    if shape[-1] % block_size != 0:
+        raise ValueError(f"Quantized tensor row size ({shape[-1]}) is not a multiple of {quant_type.name} block size ({block_size})")
+    return (*shape[:-1], shape[-1] // block_size * type_size)
+
+
+def quant_shape_from_byte_shape(shape: Sequence[int], quant_type: GGMLQuantizationType) -> tuple[int, ...]:
+    block_size, type_size = GGML_QUANT_SIZES[quant_type]
+    if shape[-1] % type_size != 0:
+        raise ValueError(f"Quantized tensor bytes per row ({shape[-1]}) is not a multiple of {quant_type.name} type size ({type_size})")
+    return (*shape[:-1], shape[-1] // type_size * block_size)
+
+
+# This is faster than np.vectorize and np.apply_along_axis because it works on more than one row at a time
+def _apply_over_grouped_rows(func: Callable[[np.ndarray], np.ndarray], arr: np.ndarray, otype: DTypeLike, oshape: tuple[int, ...]) -> np.ndarray:
+    rows = arr.reshape((-1, arr.shape[-1]))
+    osize = 1
+    for dim in oshape:
+        osize *= dim
+    out = np.empty(shape=osize, dtype=otype)
+    # compute over groups of 16 rows (arbitrary, but seems good for performance)
+    n_groups = (rows.shape[0] // 16) or 1
+    np.concatenate([func(group).ravel() for group in np.array_split(rows, n_groups)], axis=0, out=out)
+    return out.reshape(oshape)
+
+
+# round away from zero
+# ref: https://stackoverflow.com/a/59143326/22827863
+def np_roundf(n: np.ndarray) -> np.ndarray:
+    a = abs(n)
+    floored = np.floor(a)
+    b = floored + np.floor(2 * (a - floored))
+    return np.sign(n) * b
+
+
+class QuantError(Exception): ...
+
+
+_type_traits: dict[GGMLQuantizationType, type[__Quant]] = {}
+
+
+def quantize(data: np.ndarray, qtype: GGMLQuantizationType) -> np.ndarray:
+    if qtype == GGMLQuantizationType.F32:
+        return data.astype(np.float32, copy=False)
+    elif qtype == GGMLQuantizationType.F16:
+        return data.astype(np.float16, copy=False)
+    elif (q := _type_traits.get(qtype)) is not None:
+        return q.quantize(data)
+    else:
+        raise NotImplementedError(f"Quantization for {qtype.name} is not yet implemented")
+
+
+def dequantize(data: np.ndarray, qtype: GGMLQuantizationType) -> np.ndarray:
+    if qtype == GGMLQuantizationType.F32:
+        return data.view(np.float32)
+    elif qtype == GGMLQuantizationType.F16:
+        return data.view(np.float16).astype(np.float32)
+    elif (q := _type_traits.get(qtype)) is not None:
+        return q.dequantize(data)
+    else:
+        raise NotImplementedError(f"Dequantization for {qtype.name} is not yet implemented")
+
+
+class __Quant(ABC):
+    qtype: GGMLQuantizationType
+    block_size: int
+    type_size: int
+
+    grid: np.ndarray[Any, np.dtype[np.float32]] | None = None
+    grid_shape: tuple[int, int] = (0, 0)
+    grid_map: tuple[int | float, ...] = ()
+    grid_hex: bytes | None = None
+
+    def __init__(self):
+        return TypeError("Quant conversion classes can't have instances")
+
+    def __init_subclass__(cls, qtype: GGMLQuantizationType) -> None:
+        cls.qtype = qtype
+        cls.block_size, cls.type_size = GGML_QUANT_SIZES[qtype]
+        cls.__quantize_lazy = LazyNumpyTensor._wrap_fn(
+            cls.__quantize_array,
+            meta_noop=(np.uint8, cls.__shape_to_bytes)
+        )
+        cls.__dequantize_lazy = LazyNumpyTensor._wrap_fn(
+            cls.__dequantize_array,
+            meta_noop=(np.float32, cls.__shape_from_bytes)
+        )
+        assert qtype not in _type_traits
+        _type_traits[qtype] = cls
+
+    @classmethod
+    def init_grid(cls):
+        if cls.grid is not None or cls.grid_hex is None:
+            return
+
+        bits_per_elem = ceil(log2(len(cls.grid_map)))
+        assert bits_per_elem != 0, cls.qtype.name
+        elems_per_byte = 8 // bits_per_elem
+
+        grid = np.frombuffer(cls.grid_hex, dtype=np.uint8)
+        # decode hexadecimal chars from grid
+        grid = grid.reshape((-1, 2))
+        grid = (np.where(grid > 0x40, grid + 9, grid) & 0x0F) << np.array([4, 0], dtype=np.uint8).reshape((1, 2))
+        grid = grid[..., 0] | grid[..., 1]
+        # unpack the grid values
+        grid = grid.reshape((-1, 1)) >> np.array([i for i in range(0, 8, 8 // elems_per_byte)], dtype=np.uint8).reshape((1, elems_per_byte))
+        grid = (grid & ((1 << bits_per_elem) - 1)).reshape((-1, 1))
+        grid_map = np.array(cls.grid_map, dtype=np.float32).reshape((1, -1))
+        grid = np.take_along_axis(grid_map, grid, axis=-1)
+        cls.grid = grid.reshape((1, 1, *cls.grid_shape))
+
+    @classmethod
+    @abstractmethod
+    def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        raise NotImplementedError
+
+    @classmethod
+    @abstractmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        raise NotImplementedError
+
+    @classmethod
+    def quantize_rows(cls, rows: np.ndarray) -> np.ndarray:
+        rows = rows.astype(np.float32, copy=False)
+        shape = rows.shape
+        n_blocks = rows.size // cls.block_size
+        blocks = rows.reshape((n_blocks, cls.block_size))
+        blocks = cls.quantize_blocks(blocks)
+        assert blocks.dtype == np.uint8
+        assert blocks.shape[-1] == cls.type_size
+        return blocks.reshape(cls.__shape_to_bytes(shape))
+
+    @classmethod
+    def dequantize_rows(cls, rows: np.ndarray) -> np.ndarray:
+        rows = rows.view(np.uint8)
+        shape = rows.shape
+        n_blocks = rows.size // cls.type_size
+        blocks = rows.reshape((n_blocks, cls.type_size))
+        blocks = cls.dequantize_blocks(blocks)
+        assert blocks.dtype == np.float32
+        assert blocks.shape[-1] == cls.block_size
+        return blocks.reshape(cls.__shape_from_bytes(shape))
+
+    @classmethod
+    def __shape_to_bytes(cls, shape: Sequence[int]):
+        return quant_shape_to_byte_shape(shape, cls.qtype)
+
+    @classmethod
+    def __shape_from_bytes(cls, shape: Sequence[int]):
+        return quant_shape_from_byte_shape(shape, cls.qtype)
+
+    @classmethod
+    def __quantize_array(cls, array: np.ndarray) -> np.ndarray:
+        return _apply_over_grouped_rows(cls.quantize_rows, arr=array, otype=np.uint8, oshape=cls.__shape_to_bytes(array.shape))
+
+    @classmethod
+    def __dequantize_array(cls, array: np.ndarray) -> np.ndarray:
+        cls.init_grid()
+        return _apply_over_grouped_rows(cls.dequantize_rows, arr=array, otype=np.float32, oshape=cls.__shape_from_bytes(array.shape))
+
+    @classmethod
+    def __quantize_lazy(cls, lazy_tensor: LazyNumpyTensor, /) -> Any:
+        pass
+
+    @classmethod
+    def __dequantize_lazy(cls, lazy_tensor: LazyNumpyTensor, /) -> Any:
+        pass
+
+    @classmethod
+    def can_quantize(cls, tensor: np.ndarray | LazyNumpyTensor) -> bool:
+        return tensor.shape[-1] % cls.block_size == 0
+
+    @classmethod
+    def quantize(cls, tensor: np.ndarray | LazyNumpyTensor) -> np.ndarray:
+        if not cls.can_quantize(tensor):
+            raise QuantError(f"Can't quantize tensor with shape {tensor.shape} to {cls.qtype.name}")
+        if isinstance(tensor, LazyNumpyTensor):
+            return cls.__quantize_lazy(tensor)
+        else:
+            return cls.__quantize_array(tensor)
+
+    @classmethod
+    def dequantize(cls, tensor: np.ndarray | LazyNumpyTensor) -> np.ndarray:
+        if isinstance(tensor, LazyNumpyTensor):
+            return cls.__dequantize_lazy(tensor)
+        else:
+            return cls.__dequantize_array(tensor)
+
+
+class BF16(__Quant, qtype=GGMLQuantizationType.BF16):
+    @classmethod
+    # same as ggml_compute_fp32_to_bf16 in ggml-impl.h
+    def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n = blocks.view(np.uint32)
+        # force nan to quiet
+        n = np.where((n & 0x7fffffff) > 0x7f800000, (n & np.uint32(0xffff0000)) | np.uint32(64 << 16), n)
+        # round to nearest even
+        n = (np.uint64(n) + (0x7fff + ((n >> 16) & 1))) >> 16
+        return n.astype(np.uint16).view(np.uint8)
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        return (blocks.view(np.int16).astype(np.int32) << 16).view(np.float32)
+
+
+class Q4_0(__Quant, qtype=GGMLQuantizationType.Q4_0):
+    @classmethod
+    def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        imax = abs(blocks).argmax(axis=-1, keepdims=True)
+        max = np.take_along_axis(blocks, imax, axis=-1)
+
+        d = max / -8
+        with np.errstate(divide="ignore"):
+            id = np.where(d == 0, 0, 1 / d)
+        # FIXME: Q4_0's reference rounding is cursed and depends on FMA
+        qs = np.trunc((np.float64(blocks) * np.float64(id)) + np.float64(8.5), dtype=np.float32).astype(np.uint8).clip(0, 15)
+
+        qs = qs.reshape((n_blocks, 2, cls.block_size // 2))
+        qs = qs[..., 0, :] | (qs[..., 1, :] << np.uint8(4))
+
+        d = d.astype(np.float16).view(np.uint8)
+
+        return np.concatenate([d, qs], axis=-1)
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d, qs = np.hsplit(blocks, [2])
+
+        d = d.view(np.float16).astype(np.float32)
+
+        qs = qs.reshape((n_blocks, -1, 1, cls.block_size // 2)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1))
+        qs = (qs & np.uint8(0x0F)).reshape((n_blocks, -1)).astype(np.int8) - np.int8(8)
+
+        return (d * qs.astype(np.float32))
+
+
+class Q4_1(__Quant, qtype=GGMLQuantizationType.Q4_1):
+    @classmethod
+    def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        max = blocks.max(axis=-1, keepdims=True)
+        min = blocks.min(axis=-1, keepdims=True)
+
+        d = (max - min) / 15
+        with np.errstate(divide="ignore"):
+            id = np.where(d == 0, 0, 1 / d)
+        qs = np.trunc((blocks - min) * id + np.float32(0.5), dtype=np.float32).astype(np.uint8).clip(0, 15)
+
+        qs = qs.reshape((n_blocks, 2, cls.block_size // 2))
+        qs = qs[..., 0, :] | (qs[..., 1, :] << np.uint8(4))
+
+        d = d.astype(np.float16).view(np.uint8)
+        m = min.astype(np.float16).view(np.uint8)
+
+        return np.concatenate([d, m, qs], axis=-1)
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d, rest = np.hsplit(blocks, [2])
+        m, qs = np.hsplit(rest, [2])
+
+        d = d.view(np.float16).astype(np.float32)
+        m = m.view(np.float16).astype(np.float32)
+
+        qs = qs.reshape((n_blocks, -1, 1, cls.block_size // 2)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1))
+        qs = (qs & np.uint8(0x0F)).reshape((n_blocks, -1)).astype(np.float32)
+
+        return (d * qs) + m
+
+
+class Q5_0(__Quant, qtype=GGMLQuantizationType.Q5_0):
+    @classmethod
+    def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        imax = abs(blocks).argmax(axis=-1, keepdims=True)
+        max = np.take_along_axis(blocks, imax, axis=-1)
+
+        d = max / -16
+        with np.errstate(divide="ignore"):
+            id = np.where(d == 0, 0, 1 / d)
+        # FIXME: Q5_0's reference rounding is cursed and depends on FMA
+        q = np.trunc((np.float64(blocks) * np.float64(id)) + np.float64(16.5), dtype=np.float32).astype(np.uint8).clip(0, 31)
+
+        qs = q.reshape((n_blocks, 2, cls.block_size // 2))
+        qs = (qs[..., 0, :] & np.uint8(0x0F)) | (qs[..., 1, :] << np.uint8(4))
+
+        qh = np.packbits(q.reshape((n_blocks, 1, 32)) >> np.uint8(4), axis=-1, bitorder="little").reshape(n_blocks, 4)
+
+        d = d.astype(np.float16).view(np.uint8)
+
+        return np.concatenate([d, qh, qs], axis=-1)
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d, rest = np.hsplit(blocks, [2])
+        qh, qs = np.hsplit(rest, [4])
+
+        d = d.view(np.float16).astype(np.float32)
+        qh = qh.view(np.uint32)
+
+        qh = qh.reshape((n_blocks, 1)) >> np.array([i for i in range(32)], dtype=np.uint32).reshape((1, 32))
+        ql = qs.reshape((n_blocks, -1, 1, cls.block_size // 2)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1))
+        qh = (qh & np.uint32(0x01)).astype(np.uint8)
+        ql = (ql & np.uint8(0x0F)).reshape((n_blocks, -1))
+
+        qs = (ql | (qh << np.uint8(4))).astype(np.int8) - np.int8(16)
+
+        return (d * qs.astype(np.float32))
+
+
+class Q5_1(__Quant, qtype=GGMLQuantizationType.Q5_1):
+    @classmethod
+    def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        max = blocks.max(axis=-1, keepdims=True)
+        min = blocks.min(axis=-1, keepdims=True)
+
+        d = (max - min) / 31
+        with np.errstate(divide="ignore"):
+            id = np.where(d == 0, 0, 1 / d)
+        q = np.trunc((blocks - min) * id + np.float32(0.5), dtype=np.float32).astype(np.uint8).clip(0, 31)
+
+        qs = q.reshape((n_blocks, 2, cls.block_size // 2))
+        qs = (qs[..., 0, :] & np.uint8(0x0F)) | (qs[..., 1, :] << np.uint8(4))
+
+        qh = np.packbits(q.reshape((n_blocks, 1, 32)) >> np.uint8(4), axis=-1, bitorder="little").reshape(n_blocks, 4)
+
+        d = d.astype(np.float16).view(np.uint8)
+        m = min.astype(np.float16).view(np.uint8)
+
+        return np.concatenate([d, m, qh, qs], axis=-1)
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d, rest = np.hsplit(blocks, [2])
+        m, rest = np.hsplit(rest, [2])
+        qh, qs = np.hsplit(rest, [4])
+
+        d = d.view(np.float16).astype(np.float32)
+        m = m.view(np.float16).astype(np.float32)
+        qh = qh.view(np.uint32)
+
+        qh = qh.reshape((n_blocks, 1)) >> np.array([i for i in range(32)], dtype=np.uint32).reshape((1, 32))
+        ql = qs.reshape((n_blocks, -1, 1, cls.block_size // 2)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1))
+        qh = (qh & np.uint32(0x01)).astype(np.uint8)
+        ql = (ql & np.uint8(0x0F)).reshape((n_blocks, -1))
+
+        qs = (ql | (qh << np.uint8(4))).astype(np.float32)
+
+        return (d * qs) + m
+
+
+class Q8_0(__Quant, qtype=GGMLQuantizationType.Q8_0):
+    @classmethod
+    # Implementation of Q8_0 with bit-exact same results as reference implementation in ggml-quants.c
+    def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+
+        d = abs(blocks).max(axis=1, keepdims=True) / 127
+        with np.errstate(divide="ignore"):
+            id = np.where(d == 0, 0, 1 / d)
+        qs = np_roundf(blocks * id)
+
+        # (n_blocks, 2)
+        d = d.astype(np.float16).view(np.uint8)
+        # (n_blocks, block_size)
+        qs = qs.astype(np.int8).view(np.uint8)
+
+        return np.concatenate([d, qs], axis=1)
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        d, x = np.split(blocks, [2], axis=1)
+        d = d.view(np.float16).astype(np.float32)
+        x = x.view(np.int8).astype(np.float32)
+
+        return (x * d)
+
+
+class Q2_K(__Quant, qtype=GGMLQuantizationType.Q2_K):
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        scales, rest = np.hsplit(blocks, [QK_K // 16])
+        qs, rest = np.hsplit(rest, [QK_K // 4])
+        d, dmin = np.hsplit(rest, [2])
+
+        d = d.view(np.float16).astype(np.float32)
+        dmin = dmin.view(np.float16).astype(np.float32)
+
+        # (n_blocks, 16, 1)
+        dl = (d * (scales & 0xF).astype(np.float32)).reshape((n_blocks, QK_K // 16, 1))
+        ml = (dmin * (scales >> 4).astype(np.float32)).reshape((n_blocks, QK_K // 16, 1))
+
+        shift = np.array([0, 2, 4, 6], dtype=np.uint8).reshape((1, 1, 4, 1))
+
+        qs = (qs.reshape((n_blocks, -1, 1, 32)) >> shift) & np.uint8(3)
+
+        qs = qs.reshape((n_blocks, QK_K // 16, 16)).astype(np.float32)
+
+        qs = dl * qs - ml
+
+        return qs.reshape((n_blocks, -1))
+
+
+class Q3_K(__Quant, qtype=GGMLQuantizationType.Q3_K):
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        hmask, rest = np.hsplit(blocks, [QK_K // 8])
+        qs, rest = np.hsplit(rest, [QK_K // 4])
+        scales, d = np.hsplit(rest, [12])
+
+        d = d.view(np.float16).astype(np.float32)
+
+        # The scales are packed at 6-bit each in this pattern:
+        #  0: IIIIAAAA
+        #  1: JJJJBBBB
+        #  2: KKKKCCCC
+        #  3: LLLLDDDD
+        #  4: MMMMEEEE
+        #  5: NNNNFFFF
+        #  6: OOOOGGGG
+        #  7: PPPPHHHH
+        #  8: MMIIEEAA
+        #  9: NNJJFFBB
+        # 10: OOKKGGCC
+        # 11: PPLLHHDD
+        lscales, hscales = np.hsplit(scales, [8])
+        lscales = lscales.reshape((n_blocks, 1, 8)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 2, 1))
+        lscales = lscales.reshape((n_blocks, 16))
+        hscales = hscales.reshape((n_blocks, 1, 4)) >> np.array([0, 2, 4, 6], dtype=np.uint8).reshape((1, 4, 1))
+        hscales = hscales.reshape((n_blocks, 16))
+        scales = (lscales & np.uint8(0x0F)) | ((hscales & np.uint8(0x03)) << np.uint8(4))
+        scales = (scales.astype(np.int8) - np.int8(32)).astype(np.float32)
+
+        dl = (d * scales).reshape((n_blocks, 16, 1))
+
+        ql = qs.reshape((n_blocks, -1, 1, 32)) >> np.array([0, 2, 4, 6], dtype=np.uint8).reshape((1, 1, 4, 1))
+        qh = hmask.reshape(n_blocks, -1, 1, 32) >> np.array([i for i in range(8)], dtype=np.uint8).reshape((1, 1, 8, 1))
+        ql = ql.reshape((n_blocks, 16, QK_K // 16)) & np.uint8(3)
+        qh = (qh.reshape((n_blocks, 16, QK_K // 16)) & np.uint8(1))
+        qh = qh ^ np.uint8(1)  # strangely, the offset is zero when the bitmask is 1
+        q = (ql.astype(np.int8) - (qh << np.uint8(2)).astype(np.int8)).astype(np.float32)
+
+        return (dl * q).reshape((n_blocks, QK_K))
+
+
+class Q4_K(__Quant, qtype=GGMLQuantizationType.Q4_K):
+    K_SCALE_SIZE = 12
+
+    @staticmethod
+    def get_scale_min(scales: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
+        n_blocks = scales.shape[0]
+        scales = scales.view(np.uint8)
+        ### Unpacking the following: ###
+        #  0 EEAAAAAA
+        #  1 FFBBBBBB
+        #  2 GGCCCCCC
+        #  3 HHDDDDDD
+        #  4 eeaaaaaa
+        #  5 ffbbbbbb
+        #  6 ggcccccc
+        #  7 hhdddddd
+        #  8 eeeeEEEE
+        #  9 ffffFFFF
+        # 10 ggggGGGG
+        # 11 hhhhHHHH
+        scales = scales.reshape((n_blocks, 3, 4))
+        d, m, m_d = np.split(scales, 3, axis=-2)
+
+        sc = np.concatenate([d & 0x3F, (m_d & 0x0F) | ((d >> 2) & 0x30)], axis=-1)
+        min = np.concatenate([m & 0x3F, (m_d >> 4) | ((m >> 2) & 0x30)], axis=-1)
+
+        return (sc.reshape((n_blocks, 8)), min.reshape((n_blocks, 8)))
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d, rest = np.hsplit(blocks, [2])
+        dmin, rest = np.hsplit(rest, [2])
+        scales, qs = np.hsplit(rest, [cls.K_SCALE_SIZE])
+
+        d = d.view(np.float16).astype(np.float32)
+        dmin = dmin.view(np.float16).astype(np.float32)
+
+        sc, m = Q4_K.get_scale_min(scales)
+
+        d = (d * sc.astype(np.float32)).reshape((n_blocks, -1, 1))
+        dm = (dmin * m.astype(np.float32)).reshape((n_blocks, -1, 1))
+
+        qs = qs.reshape((n_blocks, -1, 1, 32)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1))
+        qs = (qs & np.uint8(0x0F)).reshape((n_blocks, -1, 32)).astype(np.float32)
+
+        return (d * qs - dm).reshape((n_blocks, QK_K))
+
+
+class Q5_K(__Quant, qtype=GGMLQuantizationType.Q5_K):
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d, rest = np.hsplit(blocks, [2])
+        dmin, rest = np.hsplit(rest, [2])
+        scales, rest = np.hsplit(rest, [Q4_K.K_SCALE_SIZE])
+        qh, qs = np.hsplit(rest, [QK_K // 8])
+
+        d = d.view(np.float16).astype(np.float32)
+        dmin = dmin.view(np.float16).astype(np.float32)
+
+        sc, m = Q4_K.get_scale_min(scales)
+
+        d = (d * sc.astype(np.float32)).reshape((n_blocks, -1, 1))
+        dm = (dmin * m.astype(np.float32)).reshape((n_blocks, -1, 1))
+
+        ql = qs.reshape((n_blocks, -1, 1, 32)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1))
+        qh = qh.reshape((n_blocks, -1, 1, 32)) >> np.array([i for i in range(8)], dtype=np.uint8).reshape((1, 1, 8, 1))
+        ql = (ql & np.uint8(0x0F)).reshape((n_blocks, -1, 32))
+        qh = (qh & np.uint8(0x01)).reshape((n_blocks, -1, 32))
+        q = (ql | (qh << np.uint8(4))).astype(np.float32)
+
+        return (d * q - dm).reshape((n_blocks, QK_K))
+
+
+class Q6_K(__Quant, qtype=GGMLQuantizationType.Q6_K):
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        ql, rest = np.hsplit(blocks, [QK_K // 2])
+        qh, rest = np.hsplit(rest, [QK_K // 4])
+        scales, d = np.hsplit(rest, [QK_K // 16])
+
+        scales = scales.view(np.int8).astype(np.float32)
+        d = d.view(np.float16).astype(np.float32)
+        d = (d * scales).reshape((n_blocks, QK_K // 16, 1))
+
+        ql = ql.reshape((n_blocks, -1, 1, 64)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1))
+        ql = (ql & np.uint8(0x0F)).reshape((n_blocks, -1, 32))
+        qh = qh.reshape((n_blocks, -1, 1, 32)) >> np.array([0, 2, 4, 6], dtype=np.uint8).reshape((1, 1, 4, 1))
+        qh = (qh & np.uint8(0x03)).reshape((n_blocks, -1, 32))
+        q = (ql | (qh << np.uint8(4))).astype(np.int8) - np.int8(32)
+        q = q.reshape((n_blocks, QK_K // 16, -1)).astype(np.float32)
+
+        return (d * q).reshape((n_blocks, QK_K))
+
+
+class TQ1_0(__Quant, qtype=GGMLQuantizationType.TQ1_0):
+    @classmethod
+    def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d = abs(blocks).max(axis=-1, keepdims=True)
+        with np.errstate(divide="ignore"):
+            id = np.where(d == 0, 0, 1 / d)
+        qs = np_roundf(blocks * id)
+        qs = (qs.astype(np.int8) + np.int8(1)).astype(np.uint8)
+
+        qs0, qs1, qh = qs[..., :(32 * 5)], qs[..., (32 * 5):(48 * 5)], qs[..., (48 * 5):]
+        qs0 = qs0.reshape((n_blocks, -1, 5, 32)) * np.array([81, 27, 9, 3, 1], dtype=np.uint8).reshape((1, 1, 5, 1))
+        qs0 = np.sum(qs0, axis=-2).reshape((n_blocks, -1))
+        qs1 = qs1.reshape((n_blocks, -1, 5, 16)) * np.array([81, 27, 9, 3, 1], dtype=np.uint8).reshape((1, 1, 5, 1))
+        qs1 = np.sum(qs1, axis=-2).reshape((n_blocks, -1))
+        qh = qh.reshape((n_blocks, -1, 4, 4)) * np.array([81, 27, 9, 3], dtype=np.uint8).reshape((1, 1, 4, 1))
+        qh = np.sum(qh, axis=-2).reshape((n_blocks, -1))
+        qs = np.concatenate([qs0, qs1, qh], axis=-1)
+        qs = (qs.astype(np.uint16) * 256 + (243 - 1)) // 243
+
+        qs = qs.astype(np.uint8)
+        d = d.astype(np.float16).view(np.uint8)
+
+        return np.concatenate([qs, d], axis=-1)
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        qs, rest = np.hsplit(blocks, [(QK_K - 4 * QK_K // 64) // 5])
+        qh, d = np.hsplit(rest, [QK_K // 64])
+
+        d = d.view(np.float16).astype(np.float32)
+
+        qs0, qs1 = qs[..., :32], qs[..., 32:]
+        qs0 = qs0.reshape((n_blocks, -1, 1, 32)) * np.array([1, 3, 9, 27, 81], dtype=np.uint8).reshape((1, 1, 5, 1))
+        qs0 = qs0.reshape((n_blocks, -1))
+        qs1 = qs1.reshape((n_blocks, -1, 1, 16)) * np.array([1, 3, 9, 27, 81], dtype=np.uint8).reshape((1, 1, 5, 1))
+        qs1 = qs1.reshape((n_blocks, -1))
+        qh = qh.reshape((n_blocks, -1, 1, 4)) * np.array([1, 3, 9, 27], dtype=np.uint8).reshape((1, 1, 4, 1))
+        qh = qh.reshape((n_blocks, -1))
+        qs = np.concatenate([qs0, qs1, qh], axis=-1)
+        qs = ((qs.astype(np.uint16) * 3) >> 8).astype(np.int8) - np.int8(1)
+
+        return (d * qs.astype(np.float32))
+
+
+class TQ2_0(__Quant, qtype=GGMLQuantizationType.TQ2_0):
+    @classmethod
+    def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d = abs(blocks).max(axis=-1, keepdims=True)
+        with np.errstate(divide="ignore"):
+            id = np.where(d == 0, 0, 1 / d)
+        qs = np_roundf(blocks * id)
+        qs = (qs.astype(np.int8) + np.int8(1)).astype(np.uint8)
+
+        qs = qs.reshape((n_blocks, -1, 4, 32)) << np.array([0, 2, 4, 6], dtype=np.uint8).reshape((1, 1, 4, 1))
+        qs = qs[..., 0, :] | qs[..., 1, :] | qs[..., 2, :] | qs[..., 3, :]
+        qs = qs.reshape((n_blocks, -1))
+
+        d = d.astype(np.float16).view(np.uint8)
+
+        return np.concatenate([qs, d], axis=-1)
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        qs, d = np.hsplit(blocks, [QK_K // 4])
+
+        d = d.view(np.float16).astype(np.float32)
+
+        qs = qs.reshape((n_blocks, -1, 1, 32)) >> np.array([0, 2, 4, 6], dtype=np.uint8).reshape((1, 1, 4, 1))
+        qs = (qs & 0x03).reshape((n_blocks, -1)).astype(np.int8) - np.int8(1)
+
+        return (d * qs.astype(np.float32))
+
+
+class IQ2_XXS(__Quant, qtype=GGMLQuantizationType.IQ2_XXS):
+    ksigns: bytes = (
+        b"\x00\x81\x82\x03\x84\x05\x06\x87\x88\x09\x0a\x8b\x0c\x8d\x8e\x0f"
+        b"\x90\x11\x12\x93\x14\x95\x96\x17\x18\x99\x9a\x1b\x9c\x1d\x1e\x9f"
+        b"\xa0\x21\x22\xa3\x24\xa5\xa6\x27\x28\xa9\xaa\x2b\xac\x2d\x2e\xaf"
+        b"\x30\xb1\xb2\x33\xb4\x35\x36\xb7\xb8\x39\x3a\xbb\x3c\xbd\xbe\x3f"
+        b"\xc0\x41\x42\xc3\x44\xc5\xc6\x47\x48\xc9\xca\x4b\xcc\x4d\x4e\xcf"
+        b"\x50\xd1\xd2\x53\xd4\x55\x56\xd7\xd8\x59\x5a\xdb\x5c\xdd\xde\x5f"
+        b"\x60\xe1\xe2\x63\xe4\x65\x66\xe7\xe8\x69\x6a\xeb\x6c\xed\xee\x6f"
+        b"\xf0\x71\x72\xf3\x74\xf5\xf6\x77\x78\xf9\xfa\x7b\xfc\x7d\x7e\xff"
+    )
+
+    # iq2xxs_grid, but with each byte of the original packed in 2 bits,
+    # by mapping 0x08 to 0, 0x19 to 1, and 0x2b to 2.
+    grid_shape = (256, 8)
+    grid_map = (0x08, 0x19, 0x2b)
+    grid_hex = (
+        b"00000200050008000a00110014002000220028002a0041004400500058006100"
+        b"6400800082008a00a20001010401100115014001840198010002020222028202"
+        b"010404041004210424044004420448046004810484049004a404000502050805"
+        b"200546056905800591050906100640068406a406000805080808140828084108"
+        b"440850085208880804094009020a140a01100410101021104010601084109010"
+        b"951000110811201150115a118011241245120014081420142514491480141815"
+        b"6215001616160118041810184018811800190519a019511a002002200a204420"
+        b"6120802082202921482100220222012404241024402456240025412564259026"
+        b"082820289428442a014004401040184021402440404048405640604081408440"
+        b"9040004120416141804185410142104248425642684200440844204480449944"
+        b"124524450046014804481048404845480049584961498249454a904a00500850"
+        b"1150195020508050885004514251a4519152905492540a550156545600581158"
+        b"195864584059085a046010604060686000615561186260620064056410651265"
+        b"84654268008002800a8041808280048118814081118201840484108415844084"
+        b"608400854685948509864086608602880489118a0490109024904090a1901691"
+        b"8091459200942294449451958198209902a050a085a009a100a218a450a804a9"
+    )
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d, qs = np.hsplit(blocks, [2])
+
+        d = d.view(np.float16).astype(np.float32)
+
+        qs = qs.view(np.uint32).reshape(n_blocks, -1, 2)
+
+        db = d * (np.float32(0.5) + (qs[..., 1] >> 28).astype(np.float32)) * np.float32(0.25)
+        db = db.reshape((n_blocks, -1, 1, 1))
+
+        # get the sign indices and unpack the bits
+        signs = qs[..., 1].reshape((n_blocks, -1, 1)) >> np.array([0, 7, 14, 21], dtype=np.uint32).reshape((1, 1, 4))
+        ksigns = np.frombuffer(cls.ksigns, dtype=np.uint8).reshape((1, 1, 1, 128))
+        signs = (signs & np.uint32(0x7F)).reshape((n_blocks, -1, 4, 1))
+        signs = np.take_along_axis(ksigns, signs, axis=-1)
+        signs = signs.reshape((n_blocks, -1, 4, 1)) >> np.array([i for i in range(8)], dtype=np.uint8).reshape((1, 1, 1, 8))
+        signs = signs & np.uint8(0x01)
+        signs = np.where(signs == 0, np.float32(1), np.float32(-1))
+        signs = signs.reshape((n_blocks, -1, 4, 8))
+
+        assert cls.grid is not None
+        grid = np.take_along_axis(cls.grid, qs[..., 0].copy().view(np.uint8).reshape((n_blocks, -1, 1, 1)), axis=-2)
+        grid = grid.reshape((n_blocks, -1, 4, 8))
+
+        return (db * grid * signs).reshape((n_blocks, -1))
+
+
+class IQ2_XS(__Quant, qtype=GGMLQuantizationType.IQ2_XS):
+    # iq2xs_grid, but with each byte of the original packed in 2 bits,
+    # by mapping 0x08 to 0, 0x19 to 1, and 0x2b to 2.
+    grid_shape = (512, 8)
+    grid_map = (0x08, 0x19, 0x2b)
+    grid_hex = (
+        b"00000200050008000a0011001400160019002000220025002800410044004600"
+        b"49005000520055005800610064008000820085008800910094009900a0000101"
+        b"04010601090110011201150118011a0121012401400142014501480151015401"
+        b"6001680181018401900100020202050208021102140220024102440250025502"
+        b"80028a0201040404060409041004120415041804210424044004420445044804"
+        b"5104540456046004810484049004000502050505080511051405200541054405"
+        b"500561058005010604061006260640064206840600080208050808080a081108"
+        b"14082008250841084408500858088008a008aa08010904091009400981098909"
+        b"000a200a280a960aa00a01100410061009101010121015101810211024104010"
+        b"4210451048105110541060106a10811084109010001102110511081111111411"
+        b"2011411144115011801194119611011204120612101240126012001402140514"
+        b"0814111414142014411444144914501464148014011504151015401500161416"
+        b"49160118041810181218401854188618001905196619511aa91a002002200520"
+        b"08200a201120142020204120442050208020a020012104211021402148216521"
+        b"002222228022a82201240424102429244024002541255225992501261a26a626"
+        b"002808280a28202855288828a22868299029082a202a822a882a8a2a01400440"
+        b"0640094010401240154018402140244040404240454048404a40514054406040"
+        b"6540814084409040004102410541084111411441204141414441504180418541"
+        b"a241014204421042124229424042004402440544084411441444194420444144"
+        b"4444504480449444014504451045244540459a4500460a464446504601480448"
+        b"1048404845485448624800491149444950496949044a00500250055008501150"
+        b"145020502850415044505050805001510451105115514051425100524452aa52"
+        b"0154045410542154405460548154a154005508558055885521566856a1560058"
+        b"14584158505899581a5940594259855a0160046010604060546062608660a960"
+        b"006124624a62926200641664106540654565a46501686a682569066a546a626a"
+        b"00800280058008801180148020802a8041804480508080808280a880aa800181"
+        b"0481068110814081518159810082208280828282a082a8820184048410841284"
+        b"158440846084898400854485a58518866a860088088825885a8880888288a888"
+        b"0689228a808a888a968aa88a0190049010904090569084900091229164915692"
+        b"89920094059444945094589429959095929541965198a6984999159a609a00a0"
+        b"02a008a00aa020a02aa0a0a051a159a1a6a100a202a208a22aa280a2a0a240a4"
+        b"95a465a698a60aa820a822a828a8a0a8a8a804a984a986a928aa2aaa91aaaaaa"
+    )
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d, rest = np.hsplit(blocks, [2])
+        qs, scales = np.hsplit(rest, [2 * QK_K // 8])
+
+        d = d.view(np.float16).astype(np.float32)
+        qs = qs.view(np.uint16)
+
+        scales = scales.reshape((n_blocks, -1, 1)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2))
+        scales = (scales & 0x0F).reshape((n_blocks, -1))
+        db = d * (np.float32(0.5) + scales) * np.float32(0.25)
+        db = db.reshape((n_blocks, -1, 1, 1))
+
+        # get the sign indices and unpack the bits
+        signs = np.frombuffer(IQ2_XXS.ksigns, dtype=np.uint8).reshape(1, 1, 128)
+        signs = np.take_along_axis(signs, (qs >> 9).reshape((n_blocks, -1, 1)), axis=-1)
+        signs = signs.reshape((n_blocks, -1, 1)) >> np.array([i for i in range(8)], dtype=np.uint8).reshape((1, 1, 8))
+        signs = signs & np.uint8(0x01)
+        signs = np.where(signs == 0, np.float32(1), np.float32(-1))
+        signs = signs.reshape((n_blocks, -1, 2, 8))
+
+        assert cls.grid is not None
+        grid = np.take_along_axis(cls.grid, (qs & np.uint16(511)).reshape((n_blocks, -1, 1, 1)), axis=-2)
+        grid = grid.reshape((n_blocks, -1, 2, 8))
+
+        return (db * grid * signs).reshape((n_blocks, -1))
+
+
+class IQ2_S(__Quant, qtype=GGMLQuantizationType.IQ2_S):
+    # iq2s_grid, but with each byte of the original packed in 2 bits,
+    # by mapping 0x08 to 0, 0x19 to 1, and 0x2b to 2.
+    grid_shape = (1024, 8)
+    grid_map = (0x08, 0x19, 0x2b)
+    grid_hex = (
+        b"00000200050008000a0011001400160019002000220025002800410044004600"
+        b"490050005200550058006100640066006900800082008500880091009400a000"
+        b"a500aa0001010401060109011001120115011801210124014001420145014801"
+        b"510154015601590160016501680181018401900192019501a101a40100020202"
+        b"050208021102140220022a02410244024602490250025502800285028a029402"
+        b"a202010404040604090410041204150418042104240426042904400442044504"
+        b"48044a0451045404560459046004620465048104840486048904900495049804"
+        b"a104a40400050205050508050a05110514051605190520052505280541054405"
+        b"46054905500552055505580561056405800582058505880591059405a0050106"
+        b"0406060609061006150640064506480651065406600681068406900600080208"
+        b"050808081108140816081908200825082a084108440846084908500852085508"
+        b"580861086408800885089408aa08010904091009120915091809210940094509"
+        b"480951095409600981099009000a110a140a220a280a2a0a500a990a01100410"
+        b"0610091010101210151018102110241026104010421045104810511054105610"
+        b"59106010621065106810811084108610901095109810a110a410001102110511"
+        b"08110a1111111411161119112011221125112811411144114611491150115211"
+        b"5511581161116411801182118511881191119411011204120912101215122112"
+        b"2412401245125112541281128412901200140214051408141114141416141914"
+        b"2014251428144114441446144914501452145514581461146414801482148514"
+        b"881491149414a014011504150615091510151215151518152115241540154215"
+        b"4515481551155415601581158415901500160516081611161416201641164416"
+        b"50168016aa160118041806180918101815181818211840184218451848185118"
+        b"541860188118841800190219051908191119141920194119441950196919a219"
+        b"041a101a401a561a00200220052008201120142016201920202025202a204120"
+        b"4420502052205520642080208a209420aa200121042110211221152121214021"
+        b"4221452151215421602181218421902100220a22222228222a22442250228822"
+        b"8a22a82201240424062409241024152418242124242440244224452448245124"
+        b"5424602481248424902400250525082511251425202541254425502566258025"
+        b"0126042610264026592600280528112814284128442850288a28aa2801290429"
+        b"102995290a2a222a642a882a8a2a014004400640094010401240154018401a40"
+        b"21402440264040404240454048404a4051405440564059406040624065408140"
+        b"8440904095409840a140a4400041024105410841114114411641194120412241"
+        b"2541414144414641494150415241554158416141644180418241854188419141"
+        b"9441a04101420442104212421542184224424042454248425142544260428142"
+        b"844200440244054408440a441144144416441944204422442544284441444444"
+        b"46444944504452445544584461446444804482448544884491449444a0440145"
+        b"0445064509451045124515451845214524454045424545454845514554456045"
+        b"6a4581458445904500460246054608461146144620464146444650468046a546"
+        b"0148044809481048124815481848214824484048424845484848514854486048"
+        b"84489048004902490549084911491449204941494449504980499649014a044a"
+        b"104a404a00500250055008501150145016501950205022502550285041504450"
+        b"4650495050505250555058506150645080508250855088509150945001510451"
+        b"0651095110511251155118512151245140514251455148515151545160518151"
+        b"8451905100520552085211521452205241524452505269528052015404540654"
+        b"0954105412541554185421542454405442544554485451545454605481548454"
+        b"9054005502550555085511551455205541554455505580550156045610562656"
+        b"405600580258055808581158145820584158445850585a588058015904591059"
+        b"4059005a195a855aa85a01600460066010601260156018602160246040604560"
+        b"4860516054606060846090600061026105610861116114612061416144615061"
+        b"806199610462106240625662a162006405640864116414642064416444645064"
+        b"806401650465106540654a656865926500669466016804681068656898680069"
+        b"2a69426aa16a0080028005800880118014801980208025804180448050805280"
+        b"5580588061808080858091809480018104810981108112811581188121812481"
+        b"408142814581488151815481818184819081a981008205820a82118214824182"
+        b"4482508201840484068409841084128415841884218440844284458448845184"
+        b"5484608481848484908400850285058508851185148520854185448550858085"
+        b"8a85018604861086298640860088058811881488418844885088a28801890489"
+        b"40896589228a588a5a8a828aa28a019004900990109012901590189024904090"
+        b"4290459048905190549060908190849090900091059111911491419144915091"
+        b"5a910192049210924092a6920094029405940894119414942094419444945094"
+        b"8094969401950495109540959895a19500964696649601980498109826984098"
+        b"a998009949995299909a00a005a00aa014a022a02aa041a044a050a0a2a0aaa0"
+        b"40a165a102a20aa222a228a22aa282a288a28aa2a8a201a404a410a440a489a4"
+        b"a4a400a519a551a60aa828a8a2a854a986a908aa0aaa20aa22aa28aa88aaaaaa"
+    )
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d, rest = np.hsplit(blocks, [2])
+        qs, rest = np.hsplit(rest, [QK_K // 8])
+        signs, rest = np.hsplit(rest, [QK_K // 8])
+        qh, scales = np.hsplit(rest, [QK_K // 32])
+
+        d = d.view(np.float16).astype(np.float32)
+
+        scales = scales.reshape((n_blocks, -1, 1)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2))
+        scales = (scales & 0x0F).reshape((n_blocks, -1))
+        db = d * (np.float32(0.5) + scales) * np.float32(0.25)
+        db = db.reshape((n_blocks, -1, 1, 1))
+
+        # unpack the sign bits
+        signs = signs.reshape((n_blocks, -1, 1)) >> np.array([i for i in range(8)], dtype=np.uint8).reshape((1, 1, 8))
+        signs = signs & np.uint8(0x01)
+        signs = np.where(signs == 0, np.float32(1), np.float32(-1))
+        signs = signs.reshape((n_blocks, -1, 2, 8))
+
+        qh = qh.reshape((n_blocks, -1, 1)) >> np.array([0, 2, 4, 6], dtype=np.uint8).reshape((1, 1, 4))
+        qs = qs.astype(np.uint16) | ((qh & 0x03).astype(np.uint16) << 8).reshape((n_blocks, -1))
+
+        assert cls.grid is not None
+        grid = np.take_along_axis(cls.grid, qs.reshape((n_blocks, -1, 1, 1)), axis=-2)
+        grid = grid.reshape((n_blocks, -1, 2, 8))
+
+        return (db * grid * signs).reshape((n_blocks, -1))
+
+
+class IQ3_XXS(__Quant, qtype=GGMLQuantizationType.IQ3_XXS):
+    grid_shape = (256, 4)
+    grid_map = (0x04, 0x0c, 0x14, 0x1c, 0x24, 0x2c, 0x34, 0x3e)
+    grid_hex = (
+        b"0000020004001100130017002000220031004200730075000101030110011201"
+        b"2101250130013201410154017001000202020402110220022202310233023702"
+        b"5102570275020103070310031203250370031304370444045704730475040105"
+        b"0705320552053506640610071407160743076107011003101010121021102310"
+        b"3010321034104710501000110211111120112211011203121012121221123012"
+        b"7212001302132013311346136613011405145014201524154615711505162217"
+        b"4017002002201120132020202220262031204220012103210521102112212121"
+        b"3021632167217021002202221122172220222222372240225522012310231423"
+        b"7023742335245324032527254125742501270327162745270130103012302130"
+        b"2330503065307230003102312031313144314631013203321032253252327232"
+        b"1133333330344734723400350635223555351436363663363337603704401740"
+        b"3540374053405740744120423742404260426642074345430444514464442545"
+        b"4345704505471047124730471250415070500051065126515551145232527252"
+        b"0253535310542354275472540255315550562457425724604460466064602161"
+        b"6161176264623063366344640565526533660367216703700570077010703270"
+        b"5270267140711272457252720073157333736073217441740075027524753076"
+    )
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d, rest = np.hsplit(blocks, [2])
+        qs, scales = np.hsplit(rest, [QK_K // 4])
+
+        d = d.view(np.float16).astype(np.float32)
+        scales = scales.view(np.uint32)
+
+        db = d * (np.float32(0.5) + (scales >> 28).astype(np.float32)) * np.float32(0.5)
+        db = db.reshape((n_blocks, -1, 1, 1))
+
+        # get the sign indices and unpack the bits
+        signs = scales.reshape((n_blocks, -1, 1)) >> np.array([0, 7, 14, 21], dtype=np.uint32).reshape((1, 1, 4))
+        ksigns = np.frombuffer(IQ2_XXS.ksigns, dtype=np.uint8).reshape((1, 1, 1, 128))
+        signs = (signs & np.uint32(0x7F)).reshape((n_blocks, -1, 4, 1))
+        signs = np.take_along_axis(ksigns, signs, axis=-1)
+        signs = signs.reshape((n_blocks, -1, 4, 1)) >> np.array([i for i in range(8)], dtype=np.uint8).reshape((1, 1, 1, 8))
+        signs = signs & np.uint8(0x01)
+        signs = np.where(signs == 0, np.float32(1), np.float32(-1))
+        signs = signs.reshape((n_blocks, -1, 4, 8))
+
+        assert cls.grid is not None
+        grid = np.take_along_axis(cls.grid, qs.reshape((n_blocks, -1, 1, 1)), axis=-2)
+        grid = grid.reshape((n_blocks, -1, 4, 8))
+
+        return (db * grid * signs).reshape((n_blocks, -1))
+
+
+class IQ3_S(__Quant, qtype=GGMLQuantizationType.IQ3_S):
+    grid_shape = (512, 4)
+    grid_map = (0x01, 0x03, 0x05, 0x07, 0x09, 0x0b, 0x0d, 0x0f)
+    grid_hex = (
+        b"0000010002000500070010001100120014001600200021002500330040004200"
+        b"4500470051005300600062007100740077000001010102010401100111011501"
+        b"2001230127013101350144016101650172010002010205020702100213021602"
+        b"2102250230023402420245024702510253027002730203031103150320032203"
+        b"3103330336034403500352036703710375030004130417042104240432044004"
+        b"4304510470040205040520052205260533054105450547056605730506061106"
+        b"1306310652067106000702070407200722072607330750075407001001100210"
+        b"0410101011101310151017102010221031103410361054105610611072100011"
+        b"0111031106111011141121113011331141115011521170117611001212121512"
+        b"1712201224123212401243125512601272120113041307131013131321132713"
+        b"3013341341136213701303140514121414143114331442144614501454140115"
+        b"1015131521153015321551152016241627164416461601170317101712172117"
+        b"3517411762177017002001200320052007201020122014201620212023202720"
+        b"3020322041204320452050205220672070207320752000210221102113211721"
+        b"2221252131213421422151210122042207222122232230223722412253225722"
+        b"7122742200230223052311232223242331233323422350236623012407242024"
+        b"2324322435244124722475240425112522253725402553257025002602260726"
+        b"2126552661260527112726273027432750270230113013301530173022303130"
+        b"3330353042304430473051306330713001310331053114312131233140316031"
+        b"7231763100321232203232323432503201331033143321332333273330334133"
+        b"4333473355337333033411341634223431345234603464340135103512352535"
+        b"3235443556357335163641360137033720372237353700400440124020402440"
+        b"2740324041405040704002410741114113412241304135414341514155410142"
+        b"0342104215422142334240425742624270420443114313432043224331433543"
+        b"0044024424443744404471440545074521456245134634466046104715473047"
+        b"4347514702501050145022504050445047505250665074500151035105511251"
+        b"2151325172510052115223523052365253520253075310532753445351536553"
+        b"7353015404542054325446541255265551555355425602570457225711601360"
+        b"1560316033606060006120612761646112623462426255626262706200631463"
+        b"2163406325644364626400650365346560650566406611671367007004700770"
+        b"2070227036704070547062700271117124714371457101720472107216722172"
+        b"3072517202733273357353730174057413742074507422754275027631760077"
+    )
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d, rest = np.hsplit(blocks, [2])
+        qs, rest = np.hsplit(rest, [QK_K // 4])
+        qh, rest = np.hsplit(rest, [QK_K // 32])
+        signs, scales = np.hsplit(rest, [QK_K // 8])
+
+        d = d.view(np.float16).astype(np.float32)
+
+        scales = scales.reshape((n_blocks, -1, 1)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2))
+        scales = (scales & 0x0F).reshape((n_blocks, -1))
+        db = d * (1 + 2 * scales)
+        db = db.reshape((n_blocks, -1, 1, 1))
+
+        # unpack the sign bits
+        signs = signs.reshape((n_blocks, -1, 1)) >> np.array([i for i in range(8)], dtype=np.uint8).reshape((1, 1, 8))
+        signs = signs & np.uint8(0x01)
+        signs = np.where(signs == 0, np.float32(1), np.float32(-1))
+        signs = signs.reshape((n_blocks, -1, 4, 8))
+
+        qh = qh.reshape((n_blocks, -1, 1)) >> np.array([i for i in range(8)], dtype=np.uint8)
+        qh = (qh & 0x01).astype(np.uint16).reshape((n_blocks, -1))
+        qs = qs.astype(np.uint16) | (qh << 8)
+
+        assert cls.grid is not None
+        grid = np.take_along_axis(cls.grid, qs.reshape((n_blocks, -1, 1, 1)), axis=-2)
+        grid = grid.reshape((n_blocks, -1, 4, 8))
+
+        return (db * grid * signs).reshape((n_blocks, -1))
+
+
+class IQ1_S(__Quant, qtype=GGMLQuantizationType.IQ1_S):
+    # iq1s_grid, with each byte packed into 2 bits
+    # -1, 0, 1 <=> 0, 1, 2
+    grid_shape = (2048, 8)
+    grid_map = (-1, 0, 1)
+    grid_hex = (
+        b"00000200050008000a00110015002000220028002a0045005100540056006500"
+        b"8000820088008a009500a000a200a800aa000401050111011401160119011a01"
+        b"2501410146014901520155015a0161016401660168018501910194019601a501"
+        b"0002020208020a0215022002220228022a024502510259026402690280028202"
+        b"88028a02910295029902a002a202a802aa021104140416042504410449045504"
+        b"5a046404650491049904a5040105040505050605150518051a05290540054505"
+        b"4a0550055105540555055605590560056205650568056a058105910595059805"
+        b"9a05a105a405a505a605a9051406190641064406500652065506580660066106"
+        b"6606690685069106940699060008020808080a0815082008220828082a084508"
+        b"5108560865088008820888088a089508a008a208a808aa080509110914091909"
+        b"2409250941095009510955096109640969099109940996099909a509000a020a"
+        b"080a0a0a150a200a220a280a2a0a450a510a590a610a650a800a820a850a880a"
+        b"8a0a950aa00aa20aa80aaa0a1010111014101910241025104110441050105510"
+        b"58106110641065106910911094109610a110a510011104110611091110111211"
+        b"1511181121112411291145114a11501151115211541155115611591160116511"
+        b"841192119511a111a41111121412161225124012461249125212551258125a12"
+        b"641266128512911294129612a512011406140914141415141814191421142614"
+        b"41144514461448144a1451145414551456145914621465146814841489149014"
+        b"94149514981499149a14a114a414a514a914021505150a151115141515151615"
+        b"191520152215251528152a154115441545154615511552155415551556155915"
+        b"5a1561156415651566156915801582158415851588158a159015911594159515"
+        b"961599159a15a015a215a51501160416051606161516161618161a1621162616"
+        b"401642164416451648164a165116551656165816591661166416651668166916"
+        b"6a1686168a1692169516a416a916111816182518411844184618491850185518"
+        b"58185a1860186118641866186918851891189418a5181019121915191a192119"
+        b"25194219441945194819511954195519561959195a19601965196a1989199119"
+        b"921995199819a119a619a919091a161a241a261a441a461a491a501a521a551a"
+        b"581a611a661a691a851a911a961a9a1a0020022008200a201520202022202520"
+        b"28202a20452051205920612065208020822088208a209520a020a220a520a820"
+        b"aa2005211121142119212521422144214921552158215a216121642165216621"
+        b"8521902196219921a521012208220a22112215222022222228222a2245225122"
+        b"562259226522812288228a2291229522a022a222a822aa220524142416241924"
+        b"252444244524462449245224552458245a2466248524912494249924a124a524"
+        b"0925152521252925402545254825512554255525592562256525682589259025"
+        b"9425952598259a25a125a425a625a92505261026122619262526412649265526"
+        b"6026612669268426862690269a260028022808280a2815282028222828282a28"
+        b"45285128542865288028822888288a28a028a228a828aa280929112914291929"
+        b"2529462949295229552961296429662969298529902996299929a429a529002a"
+        b"022a082a0a2a202a222a282a2a2a452a512a562a592a652a802a822a882a8a2a"
+        b"952aa02aa22aa82aaa2a054011401640254049405240554058405a4061406440"
+        b"664094409940a140a6400041014104410641094112411541164118411a412141"
+        b"26412941454148414a41514154415541564159415a41654168416a4181418441"
+        b"8641904192419541a041a141a241054211421442164225424142524255425a42"
+        b"6442694289429442a5420144154419442944454448444a445144544455445644"
+        b"61446244654468446a44814486448944904492449544a044a144a94401450245"
+        b"05450a4511451445154516451945204525452a45414544454545464549455045"
+        b"5145544555455645584559456145644565456645694582458445854588459145"
+        b"94459545964599459a45a545a845aa450146054609461446154618461a462146"
+        b"2446294640464246454648465046514652465546564659466246654668468146"
+        b"85468a4694469546a146a446a6460548114815481a4825484248494850485548"
+        b"5848614864486648694885489148944896489948a5480149054906490a491049"
+        b"144915491849214924492649404945494a495149524954495549564959496049"
+        b"6249654966496a49864989499249954996499849a149a449a649a949164a444a"
+        b"464a494a554a584a5a4a644a694a944aa54a0150045005500650095012501550"
+        b"1a50215024502950405045504850515054505550565059506550685086508950"
+        b"95509850a050a150a650a9500551085109510a51115114511551165118511951"
+        b"20512551265128512a5141514451455146514951505151515251545155515651"
+        b"585159515a51615164516551665169518251855191519451955196519951a051"
+        b"a551aa5101520652125215521a5221522452425245524a525152545255525652"
+        b"595262526552855290529252955299529a52a452045405541154145415541654"
+        b"185419542154255428542a54415444544554465449544a545054515454545554"
+        b"5654585459545a54615462546454655466546954805488548a54915494549554"
+        b"96549954a154a454a554aa540155025504550555065509551055115512551455"
+        b"1555165519551a55215524552555265529554055415542554455455546554855"
+        b"4955505551555255545555555655585559555a55605561556455655566556855"
+        b"69556a5581558455855589558a559055915594559555965598559955a155a455"
+        b"a555a655a9550056015602560456065608560956115614561556185619562056"
+        b"2156225624562556265628562956415645564656485649564a56505651565256"
+        b"545655565656585659565a566156645665566956825685568656885689568a56"
+        b"915695569a56a256a556a656a856a95604580558065809581058155818582158"
+        b"2a58455848584a58515854585558565858585958605862586458655882588958"
+        b"9058925895589858a158a9580159025905590a59115914591559165919592559"
+        b"41594459455946594959505951595259545955595659585959595a5961596459"
+        b"655966596959815985598959915994599559965998599959a559045a085a155a"
+        b"1a5a205a255a265a295a455a485a495a515a555a565a585a595a625a655a685a"
+        b"6a5a815a8a5a925a955a965a985a9a5aa15a0560146016601960256044605060"
+        b"5560566058605a60616064606660696081609660a56001610461066109611261"
+        b"15612161226126612961456149615161556156615961656166616a6184618a61"
+        b"92619561a161a661a96111621662196240624162466255625662586260628562"
+        b"91629662a56211641264156416641a6421642664296440644264456448644a64"
+        b"516454645564566459645a646064626465648464856489649064926494649564"
+        b"966498649a64a164a464a964056508650a651165156516651965446545654665"
+        b"496550655165546555655665596561656465656566656965866589658a659165"
+        b"9565966599659a65a265a565a665a86502660966156620662666286629664066"
+        b"456648664a66516654665566566658665a666066656668668066826685668a66"
+        b"9466966698669966a066a466a666aa661668196825684168526855685a686168"
+        b"6968856891689868a66801690469106915692169246926692969406941694569"
+        b"4669486951695469556956695969606965696a69826984698a699569a169a469"
+        b"a569a969116a166a186a416a446a496a506a556a586a5a6a646a656a696a866a"
+        b"946a986a9a6aa66a0080028008800a802080228028802a804580508051805480"
+        b"5680598065808080828088808a809580a080a280a880aa800581118114811681"
+        b"1981258141814481498150815281558156815881598164816681698185818981"
+        b"948196819981a5810082028208820a8215822082228228822a82518254825982"
+        b"65828082828288828a829582a082a282a882aa82148419844184448451845584"
+        b"5a846184648469849484998401850985128515851a8526852985408541854585"
+        b"4885518554855585568559855a856585668568856a8581858485868589859085"
+        b"928595859885a68511861686198625864186448649864a865086558659865a86"
+        b"618666866a86858691869a86a4860088028808880a8815882088228828882a88"
+        b"41884588518854885988658869888088828888888a889588a088a288a888aa88"
+        b"05890689118914891689258941894489468949895089528955895a8961896489"
+        b"858996899989a589008a028a088a0a8a158a208a228a288a2a8a458a518a548a"
+        b"568a808a828a888a8a8a958aa08aa28aa88aaa8a059011901690189019902590"
+        b"419046904990559058905a9069906a9085909190949096909990a59001910491"
+        b"069109911091159118911a912191249126912991409145915091519154915591"
+        b"569159916291659184918691929195919891a191a491a691a991059211921492"
+        b"19922592449246924992509252925592589266926992859294929692a9920194"
+        b"04940694109415941894269440944a9451945494559456945894599460946194"
+        b"62946594849486949294949495949894a194a9940095059508950a9510951195"
+        b"14951595169519952195259529952a9541954495459546954995509551955295"
+        b"549555955695589559955a956195649565956695699581958595889591959295"
+        b"94959595969599959a95a095a295a595a895aa95019604961096159619962096"
+        b"2696299645964896499651965296559656965996659668968296849689968a96"
+        b"929694969596a496a696a9960598169819982598419846985098529855985698"
+        b"5a98649865988598919896989998a59804990699099910991299159918991a99"
+        b"209921992499269940994299459948994a995199549955995699599962996599"
+        b"66996a99819984999099929995999a99a199a699059a159a259a449a469a499a"
+        b"509a559a589a619a859a919a949a959a969a00a002a008a00aa015a020a022a0"
+        b"28a02aa045a051a054a056a059a080a082a088a08aa095a0a0a0a2a0a8a0aaa0"
+        b"05a109a111a114a116a119a11aa146a149a151a155a158a15aa161a164a185a1"
+        b"90a192a196a199a102a208a20aa210a219a222a228a22aa245a251a256a259a2"
+        b"65a280a282a288a28aa295a2a0a2a2a2a8a2aaa219a425a441a444a450a454a4"
+        b"55a458a45aa461a465a466a468a469a485a406a509a510a512a515a518a526a5"
+        b"29a542a545a551a554a555a556a559a565a56aa581a584a585a586a589a592a5"
+        b"95a598a505a611a616a61aa621a625a644a646a64aa652a655a656a658a660a6"
+        b"62a686a690a695a696a699a6a1a6a4a6a6a600a802a808a80aa820a822a828a8"
+        b"2aa851a854a856a859a880a882a888a88aa895a8a0a8a2a8a8a8aaa805a914a9"
+        b"19a921a925a941a950a955a95aa961a966a969a990a996a900aa02aa08aa0aaa"
+        b"20aa22aa28aa2aaa51aa54aa56aa80aa82aa88aa8aaa95aaa0aaa2aaa8aaaaaa"
+    )
+
+    delta = np.float32(0.125)
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d, rest = np.hsplit(blocks, [2])
+        qs, qh = np.hsplit(rest, [QK_K // 8])
+
+        d = d.view(np.float16).astype(np.float32)
+        qh = qh.view(np.uint16)
+
+        dl = d * (2 * ((qh >> 12) & 7) + 1)
+        dl = dl.reshape((n_blocks, -1, 1, 1))
+        delta = np.where((qh & np.uint16(0x8000)) == 0, cls.delta, -cls.delta)
+        delta = delta.reshape((n_blocks, -1, 1, 1))
+
+        qh = qh.reshape((n_blocks, -1, 1)) >> np.array([0, 3, 6, 9], dtype=np.uint16).reshape((1, 1, 4))
+        qs = qs.astype(np.uint16) | ((qh & 7) << 8).reshape((n_blocks, -1))
+
+        assert cls.grid is not None
+        grid = np.take_along_axis(cls.grid, qs.reshape((n_blocks, -1, 1, 1)), axis=-2)
+        grid = grid.reshape((n_blocks, -1, 4, 8))
+
+        return (dl * (grid + delta)).reshape((n_blocks, -1))
+
+
+class IQ1_M(__Quant, qtype=GGMLQuantizationType.IQ1_M):
+    grid_shape = IQ1_S.grid_shape
+    grid_map = IQ1_S.grid_map
+    grid_hex = IQ1_S.grid_hex
+
+    delta = IQ1_S.delta
+
+    # Okay *this* type is weird. It's the only one which stores the f16 scales in multiple parts.
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        qs, rest = np.hsplit(blocks, [QK_K // 8])
+        qh, scales = np.hsplit(rest, [QK_K // 16])
+
+        # The f16 scale is packed across multiple bytes
+        scales = scales.view(np.uint16)
+        d = (scales.reshape((n_blocks, 4)) & np.uint16(0xF000)) >> np.array([12, 8, 4, 0], dtype=np.uint16).reshape((1, 4))
+        d = d[..., 0] | d[..., 1] | d[..., 2] | d[..., 3]
+        d = d.view(np.float16).astype(np.float32).reshape((n_blocks, 1))
+
+        scales = scales.reshape(n_blocks, -1, 1) >> np.array([0, 3, 6, 9], dtype=np.uint16).reshape((1, 1, 4))
+        scales = (scales & 0x07).reshape((n_blocks, -1))
+        dl = d * (2 * scales + 1)
+        dl = dl.reshape((n_blocks, -1, 2, 1, 1))
+
+        qh = qh.reshape((n_blocks, -1, 1)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2))
+        qs = qs.astype(np.uint16) | ((qh & 0x07).astype(np.uint16) << 8).reshape((n_blocks, -1))
+
+        delta = np.where(qh & 0x08 == 0, cls.delta, -cls.delta)
+        delta = delta.reshape((n_blocks, -1, 2, 2, 1))
+
+        assert cls.grid is not None
+        grid = np.take_along_axis(cls.grid, qs.reshape((n_blocks, -1, 1, 1)), axis=-2)
+        grid = grid.reshape((n_blocks, -1, 2, 2, 8))
+
+        return (dl * (grid + delta)).reshape((n_blocks, -1))
+
+
+class IQ4_NL(__Quant, qtype=GGMLQuantizationType.IQ4_NL):
+    kvalues = (-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113)
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d, qs = np.hsplit(blocks, [2])
+
+        d = d.view(np.float16).astype(np.float32)
+
+        qs = qs.reshape((n_blocks, -1, 1, cls.block_size // 2)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1))
+
+        qs = (qs & np.uint8(0x0F)).reshape((n_blocks, -1, 1))
+
+        kvalues = np.array(cls.kvalues, dtype=np.int8).reshape(1, 1, 16)
+        qs = np.take_along_axis(kvalues, qs, axis=-1).astype(np.float32).reshape((n_blocks, -1))
+
+        return (d * qs)
+
+
+class IQ4_XS(__Quant, qtype=GGMLQuantizationType.IQ4_XS):
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d, rest = np.hsplit(blocks, [2])
+        scales_h, rest = np.hsplit(rest, [2])
+        scales_l, qs = np.hsplit(rest, [QK_K // 64])
+
+        d = d.view(np.float16).astype(np.float32)
+        scales_h = scales_h.view(np.uint16)
+
+        scales_l = scales_l.reshape((n_blocks, -1, 1)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2))
+        scales_h = scales_h.reshape((n_blocks, 1, -1)) >> np.array([2 * i for i in range(QK_K // 32)], dtype=np.uint16).reshape((1, -1, 1))
+        scales_l = scales_l.reshape((n_blocks, -1)) & np.uint8(0x0F)
+        scales_h = scales_h.reshape((n_blocks, -1)).astype(np.uint8) & np.uint8(0x03)
+
+        scales = (scales_l | (scales_h << np.uint8(4))).astype(np.int8) - np.int8(32)
+        dl = (d * scales.astype(np.float32)).reshape((n_blocks, -1, 1))
+
+        qs = qs.reshape((n_blocks, -1, 1, 16)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1))
+        qs = qs.reshape((n_blocks, -1, 32, 1)) & np.uint8(0x0F)
+
+        kvalues = np.array(IQ4_NL.kvalues, dtype=np.int8).reshape((1, 1, 1, -1))
+        qs = np.take_along_axis(kvalues, qs, axis=-1).astype(np.float32).reshape((n_blocks, -1, 32))
+
+        return (dl * qs).reshape((n_blocks, -1))

gguf/tensor_mapping.py

@@ -0,0 +1,769 @@
+from __future__ import annotations
+
+from typing import Sequence
+
+from .constants import MODEL_ARCH, MODEL_TENSOR, MODEL_TENSORS, TENSOR_NAMES
+
+
+class TensorNameMap:
+    mappings_cfg: dict[MODEL_TENSOR, tuple[str, ...]] = {
+        # Token embeddings
+        MODEL_TENSOR.TOKEN_EMBD: (
+            "gpt_neox.embed_in",                         # gptneox
+            "transformer.wte",                           # gpt2 gpt-j mpt refact qwen dbrx jais exaone
+            "transformer.word_embeddings",               # falcon
+            "word_embeddings",                           # bloom
+            "model.embed_tokens",                        # llama-hf nemotron olmoe
+            "tok_embeddings",                            # llama-pth
+            "embeddings.word_embeddings",                # bert nomic-bert
+            "language_model.embedding.word_embeddings",  # persimmon
+            "wte",                                       # gpt2
+            "transformer.embd.wte",                      # phi2
+            "model.tok_embeddings",                      # internlm2
+            "model.embedding",                           # mamba-qbert
+            "backbone.embedding",                        # mamba
+            "backbone.embeddings",                       # mamba-hf
+            "transformer.in_out_embed",                  # Grok
+            "embedding.word_embeddings",                 # chatglm
+            "transformer.token_embeddings",              # openelm
+            "shared",                                    # t5
+            "rwkv.embeddings",                           # rwkv
+        ),
+
+        # Token type embeddings
+        MODEL_TENSOR.TOKEN_TYPES: (
+            "embeddings.token_type_embeddings",  # bert nomic-bert
+        ),
+
+        # Normalization of token embeddings
+        MODEL_TENSOR.TOKEN_EMBD_NORM: (
+            "word_embeddings_layernorm",  # bloom
+            "embeddings.LayerNorm",       # bert
+            "emb_ln",                     # nomic-bert
+            "transformer.norm",           # openelm
+            "rwkv.blocks.0.pre_ln",       # rwkv
+        ),
+
+        # Position embeddings
+        MODEL_TENSOR.POS_EMBD: (
+            "transformer.wpe",                 # gpt2
+            "embeddings.position_embeddings",  # bert
+            "wpe",                             # gpt2
+        ),
+
+        # Output
+        MODEL_TENSOR.OUTPUT: (
+            "embed_out",                 # gptneox
+            "lm_head",                   # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx jais nemotron exaone olmoe
+            "output",                    # llama-pth bloom internlm2
+            "word_embeddings_for_head",  # persimmon
+            "lm_head.linear",            # phi2
+            "output_layer",              # chatglm
+            "head",                      # rwkv
+        ),
+
+        # Output norm
+        MODEL_TENSOR.OUTPUT_NORM: (
+            "gpt_neox.final_layer_norm",               # gptneox
+            "transformer.ln_f",                        # gpt2 gpt-j falcon jais exaone
+            "model.norm",                              # llama-hf baichuan internlm2 olmoe
+            "norm",                                    # llama-pth
+            "transformer.norm_f",                      # mpt dbrx
+            "ln_f",                                    # refact bloom qwen gpt2
+            "language_model.encoder.final_layernorm",  # persimmon
+            "model.final_layernorm",                   # persimmon
+            "lm_head.ln",                              # phi2
+            "model.norm_f",                            # mamba-qbert
+            "backbone.norm_f",                         # mamba
+            "transformer.rms_norm",                    # Grok
+            "encoder.final_layernorm",                 # chatglm
+            "transformer.norm",                        # openelm
+            "model.norm",                              # nemotron
+            "rwkv.ln_out",                             # rwkv
+        ),
+
+        # Rope frequencies
+        MODEL_TENSOR.ROPE_FREQS: (
+            "rope.freqs",  # llama-pth
+            "rotary_pos_emb.inv_freq",  # chatglm
+        ),
+
+        MODEL_TENSOR.ROPE_FACTORS_LONG: (),
+        MODEL_TENSOR.ROPE_FACTORS_SHORT: (),
+    }
+
+    block_mappings_cfg: dict[MODEL_TENSOR, tuple[str, ...]] = {
+        # Attention norm
+        MODEL_TENSOR.ATTN_NORM: (
+            "gpt_neox.layers.{bid}.input_layernorm",                # gptneox
+            "transformer.h.{bid}.ln_1",                             # gpt2 gpt-j refact qwen jais exaone
+            "transformer.blocks.{bid}.norm_1",                      # mpt
+            "transformer.h.{bid}.input_layernorm",                  # falcon7b
+            "h.{bid}.input_layernorm",                              # bloom
+            "transformer.h.{bid}.ln_mlp",                           # falcon40b
+            "model.layers.{bid}.input_layernorm",                   # llama-hf nemotron olmoe
+            "layers.{bid}.attention_norm",                          # llama-pth
+            "language_model.encoder.layers.{bid}.input_layernorm",  # persimmon
+            "model.layers.{bid}.ln1",                               # yi
+            "h.{bid}.ln_1",                                         # gpt2
+            "transformer.h.{bid}.ln",                               # phi2
+            "model.layers.layers.{bid}.norm",                       # plamo
+            "model.layers.{bid}.attention_norm",                    # internlm2
+            "model.layers.{bid}.norm",                              # mamba-qbert
+            "backbone.layers.{bid}.norm",                           # mamba
+            "transformer.decoder_layer.{bid}.rms_norm",             # Grok
+            "transformer.blocks.{bid}.norm_attn_norm.norm_1",       # dbrx
+            "encoder.layers.{bid}.input_layernorm",                 # chatglm
+            "transformer.layers.{bid}.attn_norm",                   # openelm
+            "rwkv.blocks.{bid}.ln1",                                # rwkv
+        ),
+
+        # Attention norm 2
+        MODEL_TENSOR.ATTN_NORM_2: (
+            "transformer.h.{bid}.ln_attn",                  # falcon40b
+            "encoder.layer.{bid}.layer_norm_1",             # jina-v2-code
+            "rwkv.blocks.{bid}.ln2",                        # rwkv
+        ),
+
+        # Attention query-key-value
+        MODEL_TENSOR.ATTN_QKV: (
+            "gpt_neox.layers.{bid}.attention.query_key_value",                     # gptneox
+            "transformer.h.{bid}.attn.c_attn",                                     # gpt2 qwen jais
+            "transformer.blocks.{bid}.attn.Wqkv",                                  # mpt
+            "transformer.blocks.{bid}.norm_attn_norm.attn.Wqkv",                   # dbrx
+            "transformer.h.{bid}.self_attention.query_key_value",                  # falcon
+            "h.{bid}.self_attention.query_key_value",                              # bloom
+            "language_model.encoder.layers.{bid}.self_attention.query_key_value",  # persimmon
+            "model.layers.{bid}.self_attn.query_key_value",                        # persimmon
+            "h.{bid}.attn.c_attn",                                                 # gpt2
+            "transformer.h.{bid}.mixer.Wqkv",                                      # phi2
+            "encoder.layers.{bid}.attn.Wqkv",                                      # nomic-bert
+            "model.layers.{bid}.self_attn.qkv_proj",                               # phi3
+            "encoder.layers.{bid}.self_attention.query_key_value",                 # chatglm
+            "transformer.layers.{bid}.attn.qkv_proj",                              # openelm
+        ),
+
+        # Attention query
+        MODEL_TENSOR.ATTN_Q: (
+            "model.layers.{bid}.self_attn.q_proj",                       # llama-hf nemotron olmoe
+            "layers.{bid}.attention.wq",                                 # llama-pth
+            "encoder.layer.{bid}.attention.self.query",                  # bert
+            "transformer.h.{bid}.attn.q_proj",                           # gpt-j
+            "model.layers.layers.{bid}.self_attn.q_proj",                # plamo
+            "model.layers.{bid}.attention.wq",                           # internlm2
+            "transformer.decoder_layer.{bid}.multi_head_attention.query",# Grok
+            "transformer.h.{bid}.attn.attention.q_proj",                 # exaone
+        ),
+
+        # Attention key
+        MODEL_TENSOR.ATTN_K: (
+            "model.layers.{bid}.self_attn.k_proj",                     # llama-hf nemotron olmoe
+            "layers.{bid}.attention.wk",                               # llama-pth
+            "encoder.layer.{bid}.attention.self.key",                  # bert
+            "transformer.h.{bid}.attn.k_proj",                         # gpt-j
+            "transformer.h.{bid}.attn.k",                              # refact
+            "model.layers.layers.{bid}.self_attn.k_proj",              # plamo
+            "model.layers.{bid}.attention.wk",                         # internlm2
+            "transformer.decoder_layer.{bid}.multi_head_attention.key",# Grok
+            "transformer.h.{bid}.attn.attention.k_proj",               # exaone
+        ),
+
+        # Attention value
+        MODEL_TENSOR.ATTN_V: (
+            "model.layers.{bid}.self_attn.v_proj",                       # llama-hf nemotron olmoe
+            "layers.{bid}.attention.wv",                                 # llama-pth
+            "encoder.layer.{bid}.attention.self.value",                  # bert
+            "transformer.h.{bid}.attn.v_proj",                           # gpt-j
+            "transformer.h.{bid}.attn.v",                                # refact
+            "model.layers.layers.{bid}.self_attn.v_proj",                # plamo
+            "model.layers.{bid}.attention.wv",                           # internlm2
+            "transformer.decoder_layer.{bid}.multi_head_attention.value",# Grok
+            "transformer.h.{bid}.attn.attention.v_proj",                 # exaone
+        ),
+
+        # Attention output
+        MODEL_TENSOR.ATTN_OUT: (
+            "gpt_neox.layers.{bid}.attention.dense",                        # gptneox
+            "transformer.h.{bid}.attn.c_proj",                              # gpt2 refact qwen jais
+            "transformer.blocks.{bid}.attn.out_proj",                       # mpt
+            "transformer.h.{bid}.self_attention.dense",                     # falcon
+            "h.{bid}.self_attention.dense",                                 # bloom
+            "model.layers.{bid}.self_attn.o_proj",                          # llama-hf nemotron olmoe
+            "layers.{bid}.attention.wo",                                    # llama-pth
+            "encoder.layer.{bid}.attention.output.dense",                   # bert
+            "transformer.h.{bid}.attn.out_proj",                            # gpt-j
+            "language_model.encoder.layers.{bid}.self_attention.dense",     # persimmon
+            "model.layers.{bid}.self_attn.dense",                           # persimmon
+            "h.{bid}.attn.c_proj",                                          # gpt2
+            "transformer.h.{bid}.mixer.out_proj",                           # phi2
+            "model.layers.layers.{bid}.self_attn.o_proj",                   # plamo
+            "model.layers.{bid}.attention.wo",                              # internlm2
+            "encoder.layers.{bid}.attn.out_proj",                           # nomic-bert
+            "transformer.decoder_layer.{bid}.multi_head_attention.linear",  # Grok
+            "transformer.blocks.{bid}.norm_attn_norm.attn.out_proj",        # dbrx
+            "encoder.layers.{bid}.self_attention.dense",                    # chatglm
+            "transformer.layers.{bid}.attn.out_proj",                       # openelm
+            "transformer.h.{bid}.attn.attention.out_proj",                  # exaone
+        ),
+
+        # Attention output norm
+        MODEL_TENSOR.ATTN_OUT_NORM: (
+            "encoder.layer.{bid}.attention.output.LayerNorm",  # bert
+            "encoder.layers.{bid}.norm1",                      # nomic-bert
+            "transformer.decoder_layer.{bid}.rms_norm_1",      # Grok
+            "transformer.blocks.{bid}.norm_attn_norm.norm_2",  # dbrx
+        ),
+
+        MODEL_TENSOR.ATTN_POST_NORM: (
+            "model.layers.{bid}.post_attention_layernorm",     # gemma2
+        ),
+
+        # Rotary embeddings
+        MODEL_TENSOR.ATTN_ROT_EMBD: (
+            "model.layers.{bid}.self_attn.rotary_emb.inv_freq",        # llama-hf
+            "layers.{bid}.attention.inner_attention.rope.freqs",       # llama-pth
+            "model.layers.layers.{bid}.self_attn.rotary_emb.inv_freq", # plamo
+            "transformer.h.{bid}.attn.rotary_emb.inv_freq",            # codeshell
+        ),
+
+        # Feed-forward norm
+        MODEL_TENSOR.FFN_NORM: (
+            "gpt_neox.layers.{bid}.post_attention_layernorm",                # gptneox
+            "transformer.h.{bid}.ln_2",                                      # gpt2 refact qwen jais exaone
+            "h.{bid}.post_attention_layernorm",                              # bloom
+            "transformer.blocks.{bid}.norm_2",                               # mpt
+            "model.layers.{bid}.post_attention_layernorm",                   # llama-hf nemotron olmoe
+            "layers.{bid}.ffn_norm",                                         # llama-pth
+            "language_model.encoder.layers.{bid}.post_attention_layernorm",  # persimmon
+            "model.layers.{bid}.ln2",                                        # yi
+            "h.{bid}.ln_2",                                                  # gpt2
+            "model.layers.{bid}.ffn_norm",                                   # internlm2
+            "transformer.decoder_layer.{bid}.rms_norm_2",                    # Grok
+            "encoder.layers.{bid}.post_attention_layernorm",                 # chatglm
+            "transformer.layers.{bid}.ffn_norm",                             # openelm
+        ),
+
+        # Post feed-forward norm
+        MODEL_TENSOR.FFN_PRE_NORM: (
+            "model.layers.{bid}.pre_feedforward_layernorm", # gemma2
+        ),
+
+        # Post feed-forward norm
+        MODEL_TENSOR.FFN_POST_NORM: (
+            "model.layers.{bid}.post_feedforward_layernorm", # gemma2
+        ),
+
+        MODEL_TENSOR.FFN_GATE_INP: (
+            "layers.{bid}.feed_forward.gate",                   # mixtral
+            "model.layers.{bid}.block_sparse_moe.gate",         # mixtral
+            "model.layers.{bid}.mlp.gate",                      # qwen2moe olmoe
+            "transformer.decoder_layer.{bid}.router",           # Grok
+            "transformer.blocks.{bid}.ffn.router.layer",        # dbrx
+            "model.layers.{bid}.block_sparse_moe.router.layer", # granitemoe
+        ),
+
+        MODEL_TENSOR.FFN_GATE_INP_SHEXP: (
+            "model.layers.{bid}.mlp.shared_expert_gate", # qwen2moe
+        ),
+
+        # Feed-forward up
+        MODEL_TENSOR.FFN_UP: (
+            "gpt_neox.layers.{bid}.mlp.dense_h_to_4h",                # gptneox
+            "transformer.h.{bid}.mlp.c_fc",                           # gpt2 jais
+            "transformer.blocks.{bid}.ffn.up_proj",                   # mpt
+            "transformer.h.{bid}.mlp.dense_h_to_4h",                  # falcon
+            "h.{bid}.mlp.dense_h_to_4h",                              # bloom
+            "model.layers.{bid}.mlp.up_proj",                         # llama-hf refact nemotron
+            "layers.{bid}.feed_forward.w3",                           # llama-pth
+            "encoder.layer.{bid}.intermediate.dense",                 # bert
+            "transformer.h.{bid}.mlp.fc_in",                          # gpt-j
+            "transformer.h.{bid}.mlp.linear_3",                       # refact
+            "language_model.encoder.layers.{bid}.mlp.dense_h_to_4h",  # persimmon
+            "model.layers.{bid}.mlp.dense_h_to_4h",                   # persimmon
+            "transformer.h.{bid}.mlp.w1",                             # qwen
+            "h.{bid}.mlp.c_fc",                                       # gpt2
+            "transformer.h.{bid}.mlp.fc1",                            # phi2
+            "model.layers.{bid}.mlp.fc1",                             # phi2
+            "model.layers.{bid}.mlp.gate_up_proj",                    # phi3
+            "model.layers.layers.{bid}.mlp.up_proj",                  # plamo
+            "model.layers.{bid}.feed_forward.w3",                     # internlm2
+            "encoder.layers.{bid}.mlp.fc11",                          # nomic-bert
+            "model.layers.{bid}.mlp.c_fc",                            # starcoder2
+            "encoder.layer.{bid}.mlp.gated_layers_v",                 # jina-bert-v2
+            "model.layers.{bid}.residual_mlp.w3",                     # arctic
+            "encoder.layers.{bid}.mlp.dense_h_to_4h",                 # chatglm
+            "transformer.h.{bid}.mlp.c_fc_1",                         # exaone
+        ),
+
+        MODEL_TENSOR.FFN_UP_EXP: (
+            "layers.{bid}.feed_forward.experts.w3",          # mixtral (merged)
+            "transformer.decoder_layer.{bid}.moe.linear_v",  # Grok (merged)
+            "transformer.blocks.{bid}.ffn.experts.mlp.v1",   # dbrx
+            "model.layers.{bid}.mlp.experts.up_proj",        # qwen2moe olmoe (merged)
+        ),
+
+        MODEL_TENSOR.FFN_UP_SHEXP: (
+            "model.layers.{bid}.mlp.shared_expert.up_proj",  # qwen2moe
+            "model.layers.{bid}.mlp.shared_experts.up_proj", # deepseek2
+        ),
+
+        # AWQ-activation gate
+        MODEL_TENSOR.FFN_ACT: (
+            "transformer.blocks.{bid}.ffn.act",  # mpt
+        ),
+
+        # Feed-forward gate
+        MODEL_TENSOR.FFN_GATE: (
+            "model.layers.{bid}.mlp.gate_proj",           # llama-hf refact
+            "layers.{bid}.feed_forward.w1",               # llama-pth
+            "transformer.h.{bid}.mlp.w2",                 # qwen
+            "transformer.h.{bid}.mlp.c_fc2",              # jais
+            "model.layers.layers.{bid}.mlp.gate_proj",    # plamo
+            "model.layers.{bid}.feed_forward.w1",         # internlm2
+            "encoder.layers.{bid}.mlp.fc12",              # nomic-bert
+            "encoder.layer.{bid}.mlp.gated_layers_w",     # jina-bert-v2
+            "transformer.h.{bid}.mlp.linear_1",           # refact
+            "model.layers.{bid}.residual_mlp.w1",         # arctic
+            "transformer.h.{bid}.mlp.c_fc_0",             # exaone
+        ),
+
+        MODEL_TENSOR.FFN_GATE_EXP: (
+            "layers.{bid}.feed_forward.experts.w1",         # mixtral (merged)
+            "transformer.decoder_layer.{bid}.moe.linear",   # Grok (merged)
+            "transformer.blocks.{bid}.ffn.experts.mlp.w1",  # dbrx
+            "model.layers.{bid}.mlp.experts.gate_proj",     # qwen2moe olmoe (merged)
+        ),
+
+        MODEL_TENSOR.FFN_GATE_SHEXP: (
+            "model.layers.{bid}.mlp.shared_expert.gate_proj",  # qwen2moe
+            "model.layers.{bid}.mlp.shared_experts.gate_proj", # deepseek2
+        ),
+
+        # Feed-forward down
+        MODEL_TENSOR.FFN_DOWN: (
+            "gpt_neox.layers.{bid}.mlp.dense_4h_to_h",                # gptneox
+            "transformer.h.{bid}.mlp.c_proj",                         # gpt2 refact qwen jais
+            "transformer.blocks.{bid}.ffn.down_proj",                 # mpt
+            "transformer.h.{bid}.mlp.dense_4h_to_h",                  # falcon
+            "h.{bid}.mlp.dense_4h_to_h",                              # bloom
+            "model.layers.{bid}.mlp.down_proj",                       # llama-hf nemotron
+            "layers.{bid}.feed_forward.w2",                           # llama-pth
+            "encoder.layer.{bid}.output.dense",                       # bert
+            "transformer.h.{bid}.mlp.fc_out",                         # gpt-j
+            "language_model.encoder.layers.{bid}.mlp.dense_4h_to_h",  # persimmon
+            "model.layers.{bid}.mlp.dense_4h_to_h",                   # persimmon
+            "h.{bid}.mlp.c_proj",                                     # gpt2
+            "transformer.h.{bid}.mlp.fc2",                            # phi2
+            "model.layers.{bid}.mlp.fc2",                             # phi2
+            "model.layers.layers.{bid}.mlp.down_proj",                # plamo
+            "model.layers.{bid}.feed_forward.w2",                     # internlm2
+            "encoder.layers.{bid}.mlp.fc2",                           # nomic-bert
+            "model.layers.{bid}.mlp.c_proj",                          # starcoder2
+            "encoder.layer.{bid}.mlp.wo",                             # jina-bert-v2
+            "transformer.layers.{bid}.ffn.proj_2",                    # openelm
+            "model.layers.{bid}.residual_mlp.w2",                     # arctic
+            "encoder.layer.{bid}.mlp.down_layer",                     # jina-bert-v2
+            "encoder.layers.{bid}.mlp.dense_4h_to_h",                 # chatglm
+            "model.layers.h.{bid}.mlp.c_proj",                        # exaone
+        ),
+
+        MODEL_TENSOR.FFN_DOWN_EXP: (
+            "layers.{bid}.feed_forward.experts.w2",              # mixtral (merged)
+            "transformer.decoder_layer.{bid}.moe.linear_1",      # Grok (merged)
+            "transformer.blocks.{bid}.ffn.experts.mlp.w2",       # dbrx
+            "model.layers.{bid}.mlp.experts.down_proj",          # qwen2moe olmoe (merged)
+            "model.layers.{bid}.block_sparse_moe.output_linear", # granitemoe
+        ),
+
+        MODEL_TENSOR.FFN_DOWN_SHEXP: (
+            "model.layers.{bid}.mlp.shared_expert.down_proj",  # qwen2moe
+            "model.layers.{bid}.mlp.shared_experts.down_proj", # deepseek2
+        ),
+
+        MODEL_TENSOR.ATTN_Q_NORM: (
+            "language_model.encoder.layers.{bid}.self_attention.q_layernorm",
+            "model.layers.{bid}.self_attn.q_layernorm",                       # persimmon
+            "model.layers.{bid}.self_attn.q_norm",                            # cohere olmoe chameleon
+            "transformer.blocks.{bid}.attn.q_ln",                             # sea-lion
+            "encoder.layer.{bid}.attention.self.layer_norm_q",                # jina-bert-v2
+            "transformer.layers.{bid}.attn.q_norm",                           # openelm
+        ),
+
+        MODEL_TENSOR.ATTN_K_NORM: (
+            "language_model.encoder.layers.{bid}.self_attention.k_layernorm",
+            "model.layers.{bid}.self_attn.k_layernorm",                       # persimmon
+            "model.layers.{bid}.self_attn.k_norm",                            # cohere olmoe chameleon
+            "transformer.blocks.{bid}.attn.k_ln",                             # sea-lion
+            "encoder.layer.{bid}.attention.self.layer_norm_k",                # jina-bert-v2
+            "transformer.layers.{bid}.attn.k_norm",                           # openelm
+        ),
+
+        MODEL_TENSOR.ROPE_FREQS: (
+            "language_model.encoder.layers.{bid}.self_attention.rotary_emb.inv_freq",  # persimmon
+        ),
+
+        MODEL_TENSOR.LAYER_OUT_NORM: (
+            "encoder.layer.{bid}.output.LayerNorm",         # bert
+            "encoder.layers.{bid}.norm2",                   # nomic-bert
+            "transformer.decoder_layer.{bid}.rms_norm_3",   # Grok
+            "encoder.layer.{bid}.mlp.layernorm",            # jina-bert-v2
+            "encoder.layer.{bid}.layer_norm_2"              # jina-v2-code
+        ),
+
+        MODEL_TENSOR.SSM_IN: (
+            "model.layers.{bid}.in_proj",
+            "backbone.layers.{bid}.mixer.in_proj",
+        ),
+
+        MODEL_TENSOR.SSM_CONV1D: (
+            "model.layers.{bid}.conv1d",
+            "backbone.layers.{bid}.mixer.conv1d",
+        ),
+
+        MODEL_TENSOR.SSM_X: (
+            "model.layers.{bid}.x_proj",
+            "backbone.layers.{bid}.mixer.x_proj",
+        ),
+
+        MODEL_TENSOR.SSM_DT: (
+            "model.layers.{bid}.dt_proj",
+            "backbone.layers.{bid}.mixer.dt_proj",
+        ),
+
+        MODEL_TENSOR.SSM_A: (
+            "model.layers.{bid}.A_log",
+            "backbone.layers.{bid}.mixer.A_log",
+        ),
+
+        MODEL_TENSOR.SSM_D: (
+            "model.layers.{bid}.D",
+            "backbone.layers.{bid}.mixer.D",
+        ),
+
+        MODEL_TENSOR.SSM_OUT: (
+            "model.layers.{bid}.out_proj",
+            "backbone.layers.{bid}.mixer.out_proj",
+        ),
+
+        MODEL_TENSOR.TIME_MIX_W1: (
+            "rwkv.blocks.{bid}.attention.time_maa_w1",  # rwkv v6
+        ),
+
+        MODEL_TENSOR.TIME_MIX_W2: (
+            "rwkv.blocks.{bid}.attention.time_maa_w2",  # rwkv v6
+        ),
+
+        MODEL_TENSOR.TIME_MIX_LERP_X: (
+            "rwkv.blocks.{bid}.attention.time_maa_x",   # rwkv v6
+        ),
+
+        MODEL_TENSOR.TIME_MIX_LERP_K: (
+            "rwkv.blocks.{bid}.attention.time_maa_k",   # rwkv v6
+        ),
+
+        MODEL_TENSOR.TIME_MIX_LERP_V: (
+            "rwkv.blocks.{bid}.attention.time_maa_v",   # rwkv v6
+        ),
+
+        MODEL_TENSOR.TIME_MIX_LERP_R: (
+            "rwkv.blocks.{bid}.attention.time_maa_r",   # rwkv v6
+        ),
+
+        MODEL_TENSOR.TIME_MIX_LERP_G: (
+            "rwkv.blocks.{bid}.attention.time_maa_g",   # rwkv v6
+        ),
+
+        MODEL_TENSOR.TIME_MIX_LERP_W: (
+            "rwkv.blocks.{bid}.attention.time_maa_w",   # rwkv v6
+        ),
+
+        MODEL_TENSOR.TIME_MIX_FIRST: (
+            "rwkv.blocks.{bid}.attention.time_faaaa",   # rwkv v6
+        ),
+
+        MODEL_TENSOR.TIME_MIX_DECAY: (
+            "rwkv.blocks.{bid}.attention.time_decay",   # rwkv v6
+        ),
+
+        MODEL_TENSOR.TIME_MIX_DECAY_W1: (
+            "rwkv.blocks.{bid}.attention.time_decay_w1",  # rwkv v6
+        ),
+
+        MODEL_TENSOR.TIME_MIX_DECAY_W2: (
+            "rwkv.blocks.{bid}.attention.time_decay_w2",  # rwkv v6
+        ),
+
+        MODEL_TENSOR.TIME_MIX_KEY: (
+            "rwkv.blocks.{bid}.attention.key", # rwkv
+        ),
+
+        MODEL_TENSOR.TIME_MIX_VALUE: (
+            "rwkv.blocks.{bid}.attention.value", # rwkv
+        ),
+
+        MODEL_TENSOR.TIME_MIX_RECEPTANCE: (
+            "rwkv.blocks.{bid}.attention.receptance", # rwkv
+        ),
+
+        MODEL_TENSOR.TIME_MIX_GATE: (
+            "rwkv.blocks.{bid}.attention.gate", # rwkv
+        ),
+
+        MODEL_TENSOR.TIME_MIX_LN: (
+            "rwkv.blocks.{bid}.attention.ln_x", # rwkv
+        ),
+
+        MODEL_TENSOR.TIME_MIX_OUTPUT: (
+            "rwkv.blocks.{bid}.attention.output", # rwkv
+        ),
+
+        MODEL_TENSOR.CHANNEL_MIX_LERP_K: (
+            "rwkv.blocks.{bid}.feed_forward.time_maa_k", # rwkv v6
+        ),
+
+        MODEL_TENSOR.CHANNEL_MIX_LERP_R: (
+            "rwkv.blocks.{bid}.feed_forward.time_maa_r", # rwkv v6
+        ),
+
+        MODEL_TENSOR.CHANNEL_MIX_KEY: (
+            "rwkv.blocks.{bid}.feed_forward.key", # rwkv
+        ),
+
+        MODEL_TENSOR.CHANNEL_MIX_RECEPTANCE: (
+            "rwkv.blocks.{bid}.feed_forward.receptance", # rwkv
+        ),
+
+        MODEL_TENSOR.CHANNEL_MIX_VALUE: (
+            "rwkv.blocks.{bid}.feed_forward.value", # rwkv
+        ),
+
+        MODEL_TENSOR.ATTN_Q_A: (
+            "model.layers.{bid}.self_attn.q_a_proj", # deepseek2
+        ),
+
+        MODEL_TENSOR.ATTN_Q_B: (
+            "model.layers.{bid}.self_attn.q_b_proj", # deepseek2
+        ),
+
+        MODEL_TENSOR.ATTN_KV_A_MQA: (
+            "model.layers.{bid}.self_attn.kv_a_proj_with_mqa", # deepseek2
+        ),
+
+        MODEL_TENSOR.ATTN_KV_B: (
+            "model.layers.{bid}.self_attn.kv_b_proj", # deepseek2
+        ),
+
+        MODEL_TENSOR.ATTN_Q_A_NORM: (
+            "model.layers.{bid}.self_attn.q_a_layernorm", # deepseek2
+        ),
+
+        MODEL_TENSOR.ATTN_KV_A_NORM: (
+            "model.layers.{bid}.self_attn.kv_a_layernorm", # deepseek2
+        ),
+
+        MODEL_TENSOR.ATTN_SUB_NORM: (
+            "model.layers.{bid}.self_attn.inner_attn_ln",  # bitnet
+        ),
+
+        MODEL_TENSOR.FFN_SUB_NORM: (
+            "model.layers.{bid}.mlp.ffn_layernorm",  # bitnet
+        ),
+
+        MODEL_TENSOR.DEC_ATTN_NORM: (
+            "decoder.block.{bid}.layer.0.layer_norm", # t5
+        ),
+
+        MODEL_TENSOR.DEC_ATTN_Q: (
+            "decoder.block.{bid}.layer.0.SelfAttention.q", # t5
+        ),
+
+        MODEL_TENSOR.DEC_ATTN_K: (
+            "decoder.block.{bid}.layer.0.SelfAttention.k", # t5
+        ),
+
+        MODEL_TENSOR.DEC_ATTN_V: (
+            "decoder.block.{bid}.layer.0.SelfAttention.v", # t5
+        ),
+
+        MODEL_TENSOR.DEC_ATTN_OUT: (
+            "decoder.block.{bid}.layer.0.SelfAttention.o", # t5
+        ),
+
+        MODEL_TENSOR.DEC_ATTN_REL_B: (
+            "decoder.block.{bid}.layer.0.SelfAttention.relative_attention_bias", # t5
+        ),
+
+        MODEL_TENSOR.DEC_CROSS_ATTN_NORM: (
+            "decoder.block.{bid}.layer.1.layer_norm", # t5
+        ),
+
+        MODEL_TENSOR.DEC_CROSS_ATTN_Q: (
+            "decoder.block.{bid}.layer.1.EncDecAttention.q", # t5
+        ),
+
+        MODEL_TENSOR.DEC_CROSS_ATTN_K: (
+            "decoder.block.{bid}.layer.1.EncDecAttention.k", # t5
+        ),
+
+        MODEL_TENSOR.DEC_CROSS_ATTN_V: (
+            "decoder.block.{bid}.layer.1.EncDecAttention.v", # t5
+        ),
+
+        MODEL_TENSOR.DEC_CROSS_ATTN_OUT: (
+            "decoder.block.{bid}.layer.1.EncDecAttention.o", # t5
+        ),
+
+        MODEL_TENSOR.DEC_CROSS_ATTN_REL_B: (
+            "decoder.block.{bid}.layer.1.EncDecAttention.relative_attention_bias", # t5
+        ),
+
+        MODEL_TENSOR.DEC_FFN_NORM: (
+            "decoder.block.{bid}.layer.2.layer_norm", # t5
+        ),
+
+        MODEL_TENSOR.DEC_FFN_GATE: (
+            "decoder.block.{bid}.layer.2.DenseReluDense.wi_0", # flan-t5
+        ),
+
+        MODEL_TENSOR.DEC_FFN_UP: (
+            "decoder.block.{bid}.layer.2.DenseReluDense.wi",   # t5
+            "decoder.block.{bid}.layer.2.DenseReluDense.wi_1", # flan-t5
+        ),
+
+        MODEL_TENSOR.DEC_FFN_DOWN: (
+            "decoder.block.{bid}.layer.2.DenseReluDense.wo", # t5
+        ),
+
+        MODEL_TENSOR.DEC_OUTPUT_NORM: (
+            "decoder.final_layer_norm", # t5
+        ),
+
+        MODEL_TENSOR.ENC_ATTN_NORM: (
+            "encoder.block.{bid}.layer.0.layer_norm", # t5
+        ),
+
+        MODEL_TENSOR.ENC_ATTN_Q: (
+            "encoder.block.{bid}.layer.0.SelfAttention.q", # t5
+        ),
+
+        MODEL_TENSOR.ENC_ATTN_K: (
+            "encoder.block.{bid}.layer.0.SelfAttention.k", # t5
+        ),
+
+        MODEL_TENSOR.ENC_ATTN_V: (
+            "encoder.block.{bid}.layer.0.SelfAttention.v", # t5
+        ),
+
+        MODEL_TENSOR.ENC_ATTN_OUT: (
+            "encoder.block.{bid}.layer.0.SelfAttention.o", # t5
+        ),
+
+        MODEL_TENSOR.ENC_ATTN_REL_B: (
+            "encoder.block.{bid}.layer.0.SelfAttention.relative_attention_bias", # t5
+        ),
+
+        MODEL_TENSOR.ENC_FFN_NORM: (
+            "encoder.block.{bid}.layer.1.layer_norm", # t5
+        ),
+
+        MODEL_TENSOR.ENC_FFN_GATE: (
+            "encoder.block.{bid}.layer.1.DenseReluDense.wi_0", # flan-t5
+        ),
+
+        MODEL_TENSOR.ENC_FFN_UP: (
+            "encoder.block.{bid}.layer.1.DenseReluDense.wi",   # t5
+            "encoder.block.{bid}.layer.1.DenseReluDense.wi_1", # flan-t5
+        ),
+
+        MODEL_TENSOR.ENC_FFN_DOWN: (
+            "encoder.block.{bid}.layer.1.DenseReluDense.wo", # t5
+        ),
+
+        MODEL_TENSOR.ENC_OUTPUT_NORM: (
+            "encoder.final_layer_norm", # t5
+        ),
+
+        MODEL_TENSOR.CLS: (
+            "classifier",       # jina
+            "classifier.dense", # roberta
+        ),
+
+        MODEL_TENSOR.CLS_OUT: (
+            "classifier.out_proj", # roberta
+        ),
+    }
+
+    # architecture-specific block mappings
+    arch_block_mappings_cfg: dict[MODEL_ARCH, dict[MODEL_TENSOR, tuple[str, ...]]] = {
+        MODEL_ARCH.ARCTIC: {
+            MODEL_TENSOR.FFN_NORM: (
+                "model.layers.{bid}.residual_layernorm",
+            ),
+            MODEL_TENSOR.FFN_NORM_EXP: (
+                "model.layers.{bid}.post_attention_layernorm",
+            ),
+        },
+    }
+
+    mapping: dict[str, tuple[MODEL_TENSOR, str]]
+
+    def __init__(self, arch: MODEL_ARCH, n_blocks: int):
+        self.mapping = {}
+        for tensor, keys in self.mappings_cfg.items():
+            if tensor not in MODEL_TENSORS[arch]:
+                continue
+            tensor_name = TENSOR_NAMES[tensor]
+            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:
+            self.block_mappings_cfg.update(self.arch_block_mappings_cfg[arch])
+        for bid in range(n_blocks):
+            for tensor, keys in self.block_mappings_cfg.items():
+                if tensor not in MODEL_TENSORS[arch]:
+                    continue
+
+                tensor_name = TENSOR_NAMES[tensor].format(bid = bid)
+                self.mapping[tensor_name] = (tensor, tensor_name)
+                for key in keys:
+                    key = key.format(bid = bid)
+                    self.mapping[key] = (tensor, tensor_name)
+
+    def get_type_and_name(self, key: str, try_suffixes: Sequence[str] = ()) -> tuple[MODEL_TENSOR, str] | None:
+        result = self.mapping.get(key)
+        if result is not None:
+            return result
+        for suffix in try_suffixes:
+            if key.endswith(suffix):
+                result = self.mapping.get(key[:-len(suffix)])
+                if result is not None:
+                    return result[0], result[1] + suffix
+        return None
+
+    def get_name(self, key: str, try_suffixes: Sequence[str] = ()) -> str | None:
+        result = self.get_type_and_name(key, try_suffixes = try_suffixes)
+        if result is None:
+            return None
+        return result[1]
+
+    def get_type(self, key: str, try_suffixes: Sequence[str] = ()) -> MODEL_TENSOR | None:
+        result = self.get_type_and_name(key, try_suffixes = try_suffixes)
+        if result is None:
+            return None
+        return result[0]
+
+    def __getitem__(self, key: str) -> str:
+        try:
+            return self.mapping[key][1]
+        except KeyError:
+            raise KeyError(key)
+
+    def __contains__(self, key: str) -> bool:
+        return key in self.mapping
+
+    def __repr__(self) -> str:
+        return repr(self.mapping)
+
+
+def get_tensor_name_map(arch: MODEL_ARCH, n_blocks: int) -> TensorNameMap:
+    return TensorNameMap(arch, n_blocks)

gguf/utility.py

@@ -0,0 +1,69 @@
+from __future__ import annotations
+
+from typing import Literal
+
+
+def fill_templated_filename(filename: str, output_type: str | None) -> str:
+    # Given a file name fill in any type templates e.g. 'some-model-name.{ftype}.gguf'
+    ftype_lowercase: str = output_type.lower() if output_type is not None else ""
+    ftype_uppercase: str = output_type.upper() if output_type is not None else ""
+    return filename.format(ftype_lowercase,
+                           outtype=ftype_lowercase, ftype=ftype_lowercase,
+                           OUTTYPE=ftype_uppercase, FTYPE=ftype_uppercase)
+
+
+def model_weight_count_rounded_notation(model_params_count: int, min_digits: int = 2) -> str:
+    if model_params_count > 1e12 :
+        # Trillions Of Parameters
+        scaled_model_params = model_params_count * 1e-12
+        scale_suffix = "T"
+    elif model_params_count > 1e9 :
+        # Billions Of Parameters
+        scaled_model_params = model_params_count * 1e-9
+        scale_suffix = "B"
+    elif model_params_count > 1e6 :
+        # Millions Of Parameters
+        scaled_model_params = model_params_count * 1e-6
+        scale_suffix = "M"
+    else:
+        # Thousands Of Parameters
+        scaled_model_params = model_params_count * 1e-3
+        scale_suffix = "K"
+
+    fix = max(min_digits - len(str(round(scaled_model_params)).lstrip('0')), 0)
+
+    return f"{scaled_model_params:.{fix}f}{scale_suffix}"
+
+
+def size_label(total_params: int, shared_params: int, expert_params: int, expert_count: int) -> str:
+
+    if expert_count > 0:
+        pretty_size = model_weight_count_rounded_notation(abs(shared_params) + abs(expert_params), min_digits=2)
+        size_class = f"{expert_count}x{pretty_size}"
+    else:
+        size_class = model_weight_count_rounded_notation(abs(total_params), min_digits=2)
+
+    return size_class
+
+
+def naming_convention(model_name: str | None, base_name: str | None, finetune_string: str | None, version_string: str | None, size_label: str | None, output_type: str | None, model_type: Literal['vocab', 'LoRA'] | None = None) -> str:
+    # Reference: https://github.com/ggerganov/ggml/blob/master/docs/gguf.md#gguf-naming-convention
+
+    if base_name is not None:
+        name = base_name.strip().replace(' ', '-').replace('/', '-')
+    elif model_name is not None:
+        name = model_name.strip().replace(' ', '-').replace('/', '-')
+    else:
+        name = "ggml-model"
+
+    parameters = f"-{size_label}" if size_label is not None else ""
+
+    finetune = f"-{finetune_string.strip().replace(' ', '-')}" if finetune_string is not None else ""
+
+    version = f"-{version_string.strip().replace(' ', '-')}" if version_string is not None else ""
+
+    encoding = f"-{output_type.strip().replace(' ', '-').upper()}" if output_type is not None else ""
+
+    kind = f"-{model_type.strip().replace(' ', '-')}" if model_type is not None else ""
+
+    return f"{name}{parameters}{finetune}{version}{encoding}{kind}"

gguf/vocab.py

@@ -0,0 +1,487 @@
+from __future__ import annotations
+
+import re
+import logging
+import json
+import os
+from pathlib import Path
+from typing import Any, Callable, Sequence, Mapping, Iterable, Protocol, ClassVar, runtime_checkable
+
+from sentencepiece import SentencePieceProcessor
+
+import gguf
+
+from .gguf_writer import GGUFWriter
+
+logger = logging.getLogger(__name__)
+
+
+class SpecialVocab:
+    merges: list[str]
+    add_special_token: dict[str, bool]
+    special_token_ids: dict[str, int]
+    chat_template: str | Sequence[Mapping[str, str]] | None
+
+    def __init__(
+        self, path: str | os.PathLike[str], load_merges: bool = False,
+        special_token_types: Iterable[str] | None = None,
+        n_vocab: int | None = None,
+    ):
+        self.special_token_ids = {}
+        self.add_special_token = {}
+        self.n_vocab = n_vocab
+        self.load_merges = load_merges
+        self.merges = []
+        self.chat_template = None
+        if special_token_types is not None:
+            self.special_token_types = special_token_types
+        else:
+            self.special_token_types = ('bos', 'eos', 'unk', 'sep', 'pad', 'cls', 'mask')
+        self._load(Path(path))
+
+    def __repr__(self) -> str:
+        return '<SpecialVocab with {} merges, special tokens {}, add special tokens {}>'.format(
+            len(self.merges), self.special_token_ids or "unset", self.add_special_token or "unset",
+        )
+
+    def add_to_gguf(self, gw: GGUFWriter, quiet: bool = False) -> None:
+        if self.merges:
+            if not quiet:
+                logger.info(f'Adding {len(self.merges)} merge(s).')
+            gw.add_token_merges(self.merges)
+        elif self.load_merges:
+            logger.warning('Adding merges requested but no merges found, output may be non-functional.')
+        for typ, tokid in self.special_token_ids.items():
+            id_handler: Callable[[int], None] | None = getattr(gw, f'add_{typ}_token_id', None)
+            if id_handler is None:
+                logger.warning(f'No handler for special token type {typ} with id {tokid} - skipping')
+                continue
+            if not quiet:
+                logger.info(f'Setting special token type {typ} to {tokid}')
+            id_handler(tokid)
+        for typ, value in self.add_special_token.items():
+            add_handler: Callable[[bool], None] | None = getattr(gw, f'add_add_{typ}_token', None)
+            if add_handler is None:
+                logger.warning(f'No handler for add_{typ}_token with value {value} - skipping')
+                continue
+            if not quiet:
+                logger.info(f'Setting add_{typ}_token to {value}')
+            add_handler(value)
+        if self.chat_template is not None:
+            if not quiet:
+                logger.info(f'Setting chat_template to {self.chat_template}')
+            gw.add_chat_template(self.chat_template)
+
+    def _load(self, path: Path) -> None:
+        self._try_load_from_tokenizer_json(path)
+        self._try_load_from_config_json(path)
+        if self.load_merges and not self.merges:
+            self._try_load_merges_txt(path)
+
+    def _try_load_merges_txt(self, path: Path) -> bool:
+        merges_file = path / 'merges.txt'
+        if not merges_file.is_file():
+            return False
+        with open(merges_file, 'r', encoding = 'utf-8') as fp:
+            first_line = next(fp, '').strip()
+            if not first_line.startswith('#'):
+                fp.seek(0)
+                line_num = 0
+            else:
+                line_num = 1
+            merges = []
+            for line in fp:
+                line_num += 1
+                line = line.strip()
+                if not line:
+                    continue
+                parts = line.split(None, 3)
+                if len(parts) != 2:
+                    logger.warning(f'{merges_file.name}: Line {line_num}: Entry malformed, ignoring')
+                    continue
+                merges.append(f'{parts[0]} {parts[1]}')
+        self.merges = merges
+        return True
+
+    def _set_special_token(self, typ: str, tid: Any) -> None:
+        if not isinstance(tid, int):
+            return
+        if tid < 0:
+            raise ValueError(f'invalid value for special token type {typ}: {tid}')
+        if self.n_vocab is None or tid < self.n_vocab:
+            if typ in self.special_token_ids:
+                return
+            self.special_token_ids[typ] = tid
+            return
+        logger.warning(f'Special token type {typ}, id {tid} out of range, must be under {self.n_vocab} - skipping')
+
+    def _try_load_from_tokenizer_json(self, path: Path) -> bool:
+        tokenizer_file = path / 'tokenizer.json'
+        if tokenizer_file.is_file():
+            with open(tokenizer_file, encoding = 'utf-8') as f:
+                tokenizer = json.load(f)
+            if self.load_merges:
+                merges = tokenizer.get('model', {}).get('merges')
+                if isinstance(merges, list) and merges:
+                    if isinstance(merges[0], str):
+                        self.merges = merges
+                    elif isinstance(merges[0], list) and len(merges[0]) == 2 and isinstance(merges[0][0], str):
+                        # New format since transformers 4.45 to support spaces in merges
+                        # ref: https://github.com/ggerganov/llama.cpp/issues/9692
+                        # TODO: internally store as the new format instead of converting to old
+                        if any(' ' in s for pair in merges for s in pair):
+                            logger.warning(f'Spaces in merges detected, encoding as {chr(ord(" ") + 256)!r}')
+                        self.merges = [
+                            ' '.join(
+                                [
+                                    # ensure the spaces are properly encoded
+                                    ''.join(
+                                        chr(ord(c) + 256) if c == ' ' else c
+                                        for c in part
+                                    )
+                                    for part in pair
+                                ]
+                            )
+                            for pair in merges
+                        ]
+                    else:
+                        raise ValueError("Unknown tokenizer merges format")
+            added_tokens = tokenizer.get('added_tokens', {})
+        else:
+            added_tokens = {}
+        tokenizer_config_file = path / 'tokenizer_config.json'
+        if not tokenizer_config_file.is_file():
+            return True
+        with open(tokenizer_config_file, encoding = 'utf-8') as f:
+            tokenizer_config = json.load(f)
+        chat_template = tokenizer_config.get('chat_template')
+        if chat_template is None or isinstance(chat_template, (str, list)):
+            self.chat_template = chat_template
+        else:
+            logger.warning(f'Bad type for chat_template field in {tokenizer_config_file!r} - ignoring')
+        for typ in self.special_token_types:
+            add_entry = tokenizer_config.get(f'add_{typ}_token')
+            if isinstance(add_entry, bool):
+                self.add_special_token[typ] = add_entry
+            entry = tokenizer_config.get(f'{typ}_token')
+            if isinstance(entry, str):
+                tc_content = entry
+            elif isinstance(entry, dict):
+                entry_content = entry.get('content')
+                if not isinstance(entry_content, str):
+                    continue
+                tc_content = entry_content
+            else:
+                continue
+            # We only need the first match here.
+            maybe_token_id = next(
+                (atok.get('id') for atok in added_tokens if atok.get('content') == tc_content),
+                None,
+            )
+            self._set_special_token(typ, maybe_token_id)
+        return True
+
+    def _try_load_from_config_json(self, path: Path) -> bool:
+        config_file = path / 'config.json'
+        if not config_file.is_file():
+            return False
+        with open(config_file, encoding = 'utf-8') as f:
+            config = json.load(f)
+        for typ in self.special_token_types:
+            self._set_special_token(typ, config.get(f'{typ}_token_id'))
+        return True
+
+
+@runtime_checkable
+class BaseVocab(Protocol):
+    tokenizer_model: ClassVar[str]
+    name: ClassVar[str]
+
+
+@runtime_checkable
+class Vocab(BaseVocab, Protocol):
+    vocab_size: int
+    added_tokens_dict: dict[str, int]
+    added_tokens_list: list[str]
+    fname_tokenizer: Path
+
+    def __init__(self, base_path: Path): ...
+    def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: ...
+
+
+class NoVocab(BaseVocab):
+    tokenizer_model = "no_vocab"
+    name = "no_vocab"
+
+    def __repr__(self) -> str:
+        return "<NoVocab for a model without integrated vocabulary>"
+
+
+class BpeVocab(Vocab):
+    tokenizer_model = "gpt2"
+    name = "bpe"
+
+    def __init__(self, base_path: Path):
+        added_tokens: dict[str, int] = {}
+
+        if (fname_tokenizer := base_path / 'vocab.json').exists():
+            # "slow" tokenizer
+            with open(fname_tokenizer, encoding="utf-8") as f:
+                self.vocab = json.load(f)
+
+            try:
+                # FIXME: Verify that added tokens here _cannot_ overlap with the main vocab.
+                with open(base_path / 'added_tokens.json', encoding="utf-8") as f:
+                    added_tokens = json.load(f)
+            except FileNotFoundError:
+                pass
+        else:
+            # "fast" tokenizer
+            fname_tokenizer = base_path / 'tokenizer.json'
+
+            # if this fails, FileNotFoundError propagates to caller
+            with open(fname_tokenizer, encoding="utf-8") as f:
+                tokenizer_json = json.load(f)
+
+            tokenizer_model: dict[str, Any] = tokenizer_json['model']
+            if (
+                tokenizer_model['type'] != 'BPE' or tokenizer_model.get('byte_fallback', False)
+                or tokenizer_json['decoder']['type'] != 'ByteLevel'
+            ):
+                raise FileNotFoundError('Cannot find GPT-2 BPE tokenizer')
+
+            self.vocab = tokenizer_model["vocab"]
+
+            if (added := tokenizer_json.get('added_tokens')) is not None:
+                # Added tokens here can be duplicates of the main vocabulary.
+                added_tokens = {item['content']: item['id']
+                                for item in added
+                                if item['content'] not in self.vocab}
+
+        vocab_size   = len(self.vocab)
+        expected_ids = list(range(vocab_size, vocab_size + len(added_tokens)))
+        actual_ids   = sorted(added_tokens.values())
+        if expected_ids != actual_ids:
+            expected_end_id = vocab_size + len(actual_ids) - 1
+            raise ValueError(f"Expected the {len(actual_ids)} added token ID(s) to be sequential in the range "
+                             f"{vocab_size} - {expected_end_id}; got {actual_ids}")
+
+        items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1])
+        self.added_tokens_dict    = added_tokens
+        self.added_tokens_list    = [text for (text, idx) in items]
+        self.vocab_size_base      = vocab_size
+        self.vocab_size           = self.vocab_size_base + len(self.added_tokens_list)
+        self.fname_tokenizer      = fname_tokenizer
+
+    def bpe_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
+        reverse_vocab = {id: encoded_tok for encoded_tok, id in self.vocab.items()}
+
+        for i, _ in enumerate(self.vocab):
+            yield reverse_vocab[i], 0.0, gguf.TokenType.NORMAL
+
+    def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
+        for text in self.added_tokens_list:
+            score = -1000.0
+            yield text.encode("utf-8"), score, gguf.TokenType.CONTROL
+
+    def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
+        yield from self.bpe_tokens()
+        yield from self.added_tokens()
+
+    def __repr__(self) -> str:
+        return f"<BpeVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
+
+
+class SentencePieceVocab(Vocab):
+    tokenizer_model = "llama"
+    name = "spm"
+
+    def __init__(self, base_path: Path):
+        added_tokens: dict[str, int] = {}
+        if (fname_tokenizer := base_path / 'tokenizer.model').exists():
+            # normal location
+            try:
+                with open(base_path / 'added_tokens.json', encoding="utf-8") as f:
+                    added_tokens = json.load(f)
+            except FileNotFoundError:
+                pass
+        elif not (fname_tokenizer := base_path.parent / 'tokenizer.model').exists():
+            # not found in alternate location either
+            raise FileNotFoundError('Cannot find tokenizer.model')
+
+        self.sentencepiece_tokenizer = SentencePieceProcessor()
+        self.sentencepiece_tokenizer.LoadFromFile(str(fname_tokenizer))
+        vocab_size = self.sentencepiece_tokenizer.vocab_size()
+
+        new_tokens       = {id: piece for piece, id in added_tokens.items() if id >= vocab_size}
+        expected_new_ids = list(range(vocab_size, vocab_size + len(new_tokens)))
+        actual_new_ids   = sorted(new_tokens.keys())
+
+        if expected_new_ids != actual_new_ids:
+            raise ValueError(f"Expected new token IDs {expected_new_ids} to be sequential; got {actual_new_ids}")
+
+        # Token pieces that were added to the base vocabulary.
+        self.added_tokens_dict  = added_tokens
+        self.added_tokens_list  = [new_tokens[id] for id in actual_new_ids]
+        self.vocab_size_base    = vocab_size
+        self.vocab_size         = self.vocab_size_base + len(self.added_tokens_list)
+        self.fname_tokenizer    = fname_tokenizer
+
+    def sentencepiece_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
+        tokenizer = self.sentencepiece_tokenizer
+        for i in range(tokenizer.vocab_size()):
+            piece = tokenizer.IdToPiece(i)
+            text         = piece.encode("utf-8")
+            score: float = tokenizer.GetScore(i)
+
+            toktype = gguf.TokenType.NORMAL
+            if tokenizer.IsUnknown(i):
+                toktype = gguf.TokenType.UNKNOWN
+            if tokenizer.IsControl(i):
+                toktype = gguf.TokenType.CONTROL
+
+            # NOTE: I think added_tokens are user defined.
+            # ref: https://github.com/google/sentencepiece/blob/master/src/sentencepiece_model.proto
+            # if tokenizer.is_user_defined(i): toktype = gguf.TokenType.USER_DEFINED
+
+            if tokenizer.IsUnused(i):
+                toktype = gguf.TokenType.UNUSED
+            if tokenizer.IsByte(i):
+                toktype = gguf.TokenType.BYTE
+
+            yield text, score, toktype
+
+    def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
+        for text in self.added_tokens_list:
+            score = -1000.0
+            yield text.encode("utf-8"), score, gguf.TokenType.USER_DEFINED
+
+    def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
+        yield from self.sentencepiece_tokens()
+        yield from self.added_tokens()
+
+    def __repr__(self) -> str:
+        return f"<SentencePieceVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
+
+
+class LlamaHfVocab(Vocab):
+    tokenizer_model = "llama"
+    name = "hfft"
+
+    def __init__(self, base_path: Path):
+        fname_tokenizer = base_path / 'tokenizer.json'
+        # if this fails, FileNotFoundError propagates to caller
+        with open(fname_tokenizer, encoding='utf-8') as f:
+            tokenizer_json = json.load(f)
+
+        # pre-check so we know if we need transformers
+        tokenizer_model: dict[str, Any] = tokenizer_json['model']
+        is_llama3 = (
+            tokenizer_model['type'] == 'BPE' and tokenizer_model.get('ignore_merges', False)
+            and not tokenizer_model.get('byte_fallback', True)
+        )
+        if is_llama3:
+            raise TypeError('Llama 3 must be converted with BpeVocab')
+
+        if not is_llama3 and (
+            tokenizer_model['type'] != 'BPE' or not tokenizer_model.get('byte_fallback', False)
+            or tokenizer_json['decoder']['type'] != 'Sequence'
+        ):
+            raise FileNotFoundError('Cannot find Llama BPE tokenizer')
+
+        try:
+            from transformers import AutoTokenizer
+        except ImportError as e:
+            raise ImportError(
+                "To use LlamaHfVocab, please install the `transformers` package. "
+                "You can install it with `pip install transformers`."
+            ) from e
+
+        # Allow the tokenizer to default to slow or fast versions.
+        # Explicitly set tokenizer to use local paths.
+        self.tokenizer = AutoTokenizer.from_pretrained(
+            base_path,
+            cache_dir=base_path,
+            local_files_only=True,
+        )
+        assert self.tokenizer.is_fast  # assume tokenizer.json is used
+
+        # Initialize lists and dictionaries for added tokens
+        self.added_tokens_list = []
+        self.added_tokens_dict = dict()
+        self.added_tokens_ids  = set()
+
+        # Process added tokens
+        for tok, tokidx in sorted(
+            self.tokenizer.get_added_vocab().items(), key=lambda x: x[1]
+        ):
+            # Only consider added tokens that are not in the base vocabulary
+            if tokidx >= self.tokenizer.vocab_size:
+                self.added_tokens_list.append(tok)
+                self.added_tokens_dict[tok] = tokidx
+                self.added_tokens_ids.add(tokidx)
+
+        # Store special tokens and their IDs
+        self.specials = {
+            tok: self.tokenizer.get_vocab()[tok]
+            for tok in self.tokenizer.all_special_tokens
+        }
+        self.special_ids = set(self.tokenizer.all_special_ids)
+
+        # Set vocabulary sizes
+        self.vocab_size_base = self.tokenizer.vocab_size
+        self.vocab_size      = self.vocab_size_base + len(self.added_tokens_list)
+
+        self.fname_tokenizer = fname_tokenizer
+
+    def hf_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
+        reverse_vocab = {
+            id: encoded_tok for encoded_tok, id in self.tokenizer.get_vocab().items()
+        }
+
+        for token_id in range(self.vocab_size_base):
+            # Skip processing added tokens here
+            if token_id in self.added_tokens_ids:
+                continue
+
+            # Convert token text to bytes
+            token_text = reverse_vocab[token_id].encode("utf-8")
+
+            # Yield token text, score, and type
+            yield token_text, self.get_token_score(token_id), self.get_token_type(
+                token_id, token_text, self.special_ids  # Reuse already stored special IDs
+            )
+
+    def get_token_type(self, token_id: int, token_text: bytes, special_ids: set[int]) -> gguf.TokenType:
+        # Special case for byte tokens
+        if re.fullmatch(br"<0x[0-9A-Fa-f]{2}>", token_text):
+            return gguf.TokenType.BYTE
+
+        # Determine token type based on whether it's a special token
+        return gguf.TokenType.CONTROL if token_id in special_ids else gguf.TokenType.NORMAL
+
+    def get_token_score(self, token_id: int) -> float:
+        # Placeholder for actual logic to determine the token's score
+        # This needs to be implemented based on specific requirements
+        return -1000.0  # Default score
+
+    def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
+        for text in self.added_tokens_list:
+            if text in self.specials:
+                toktype = self.get_token_type(self.specials[text], b'', self.special_ids)
+                score = self.get_token_score(self.specials[text])
+            else:
+                toktype = gguf.TokenType.USER_DEFINED
+                score = -1000.0
+
+            yield text.encode("utf-8"), score, toktype
+
+    def has_newline_token(self):
+        return "<0x0A>" in self.tokenizer.vocab or "\n" in self.tokenizer.vocab
+
+    def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
+        yield from self.hf_tokens()
+        yield from self.added_tokens()
+
+    def __repr__(self) -> str:
+        return f"<LlamaHfVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"

test_inference.py

@@ -0,0 +1,617 @@
+
+from exllamav2 import(
+    ExLlamaV2,
+    ExLlamaV2Config,
+    ExLlamaV2Cache,
+    ExLlamaV2Cache_8bit,
+    ExLlamaV2Cache_Q4,
+    ExLlamaV2Cache_Q6,
+    ExLlamaV2Cache_Q8,
+    ExLlamaV2Cache_TP,
+    ExLlamaV2Tokenizer,
+    model_init,
+)
+
+from exllamav2.generator import (
+    ExLlamaV2BaseGenerator,
+    ExLlamaV2Sampler
+)
+
+from exllamav2.attn import ExLlamaV2Attention
+from exllamav2.mlp import ExLlamaV2MLP
+from exllamav2.moe_mlp import ExLlamaV2MoEMLP
+from exllamav2.parallel_decoder import ExLlamaV2ParallelDecoder
+
+import argparse, os, math, time
+import torch
+import torch.nn.functional as F
+from exllamav2.conversion.tokenize import get_tokens
+from exllamav2.conversion.quantize import list_live_tensors
+import gc
+
+# from exllamav2.mlp import set_catch
+
+import sys
+import json
+
+torch.cuda._lazy_init()
+torch.set_printoptions(precision = 5, sci_mode = False, linewidth = 150)
+
+# torch.backends.cuda.matmul.allow_tf32 = True
+# torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction = True
+# torch.set_float32_matmul_precision("medium")
+
+# (!!!) NOTE: These go on top of the engine arguments that can be found in `model_init.py` (!!!)
+parser = argparse.ArgumentParser(description = "Test inference on ExLlamaV2 model")
+parser.add_argument("-ed", "--eval_dataset", type = str, help = "Perplexity evaluation dataset (.parquet file)")
+parser.add_argument("-er", "--eval_rows", type = int, default = 128, help = "Number of rows to apply from dataset")
+parser.add_argument("-el", "--eval_length", type = int, default = 2048, help = "Max no. tokens per sample")
+parser.add_argument("-et", "--eval_token", action = "store_true", help = "Evaluate perplexity on token-by-token inference using cache")
+parser.add_argument("-e8", "--eval_token_8bit", action = "store_true", help = "Evaluate perplexity on token-by-token inference using 8-bit (FP8) cache")
+parser.add_argument("-eq4", "--eval_token_q4", action = "store_true", help = "Evaluate perplexity on token-by-token inference using Q4 cache")
+parser.add_argument("-eq6", "--eval_token_q6", action = "store_true", help = "Evaluate perplexity on token-by-token inference using Q6 cache")
+parser.add_argument("-eq8", "--eval_token_q8", action = "store_true", help = "Evaluate perplexity on token-by-token inference using Q8 cache")
+parser.add_argument("-ecl", "--eval_context_lens", action = "store_true", help = "Evaluate perplexity at range of context lengths")
+# parser.add_argument("-eb", "--eval_bos", action = "store_true", help = "Add BOS token to every row in perplexity test (required by Gemma and maybe other models.)")
+parser.add_argument("-p", "--prompt", type = str, help = "Generate from prompt (basic sampling settings)")
+parser.add_argument("-pnb", "--prompt_no_bos", action = "store_true", help = "Don't add BOS token to prompt")
+parser.add_argument("-t", "--tokens", type = int, default = 128, help = "Max no. tokens")
+parser.add_argument("-ps", "--prompt_speed", action = "store_true", help = "Test prompt processing (batch) speed over context length")
+parser.add_argument("-s", "--speed", action = "store_true", help = "Test raw generation speed over context length")
+parser.add_argument("-mix", "--mix_layers", type = str, help = "Load replacement layers from secondary model. Example: --mix_layers 1,6-7:/mnt/models/other_model")
+parser.add_argument("-nwu", "--no_warmup", action = "store_true", help = "Skip warmup before testing model")
+parser.add_argument("-sl", "--stream_layers", action = "store_true", help = "Load model layer by layer (perplexity evaluation only)")
+parser.add_argument("-sp", "--standard_perplexity", choices = ["wiki2"], help = "Run standard (HF) perplexity test, stride 512 (experimental)")
+parser.add_argument("-rr", "--rank_reduce", type = str, help = "Rank-reduction for MLP layers of model, in reverse order (for experimentation)")
+parser.add_argument("-mol", "--max_output_len", type = int, help = "Set max output chunk size (incompatible with ppl tests)")
+
+# Initialize model and tokenizer
+
+model_init.add_args(parser)
+args = parser.parse_args()
+
+# Check conflicting settings
+
+if args.stream_layers:
+    if args.eval_token or args.eval_token_8bit or args.eval_token_q4 or args.eval_token_q6 or args.eval_token_q8:
+        print(" ## Can't test token ppl while streaming layers")
+        sys.exit()
+    if args.prompt:
+        print(" ## Can't generate while streaming layers")
+        sys.exit()
+    if args.speed or args.prompt_speed:
+        print(" ## Can't test speed while streaming layers")
+        sys.exit()
+    if args.gpu_split:
+        print(" ## Can only use one GPU when streaming layers")
+        sys.exit()
+    if args.eval_context_lens and args.stream_layers:
+        print(" ## eval_context_lens not compatible with stream_layers")
+        sys.exit()
+    if args.eval_dataset:
+        if args.length and args.eval_length != args.length:
+            print(" !! Overriding model context length to match eval row length")
+        args.length = args.eval_length
+
+# Init
+
+model_init.check_args(args)
+model_init.print_options(args)
+model, tokenizer = model_init.init(
+    args,
+    allow_auto_split = True,
+    skip_load = args.stream_layers,
+    benchmark = True,
+    max_output_len = args.max_output_len,
+    progress = True
+)
+cache = None
+
+# Auto split
+
+if not model.loaded and not args.stream_layers:
+
+    if args.mix_layers:
+        print(" !! Warning, auto split does not account for VRAM requirement of replacement layers")
+
+    print(" -- Loading model...")
+    cache = ExLlamaV2Cache(model, lazy = True)
+    t = time.time()
+    model.load_autosplit(cache, progress = True)
+    t = time.time() - t
+    print(f" -- Loaded model in {t:.4f} seconds")
+
+if args.stream_layers:
+
+    stream_batch_size = 2
+    model.config.max_batch_size = stream_batch_size
+    model.load(lazy = True)
+
+# Rank reduction
+
+if args.rank_reduce:
+
+    if args.stream_layers:
+        print(" ## --rank_reduce can not be combined with --stream_layers")
+        sys.exit()
+
+    rr = args.rank_reduce.split(",")
+    idx = len(model.modules) - 1
+    for r in rr:
+        k = float(r)
+
+        while True:
+            idx -= 1
+            module = model.modules[idx]
+            if isinstance(module, ExLlamaV2ParallelDecoder): break
+            if isinstance(module, ExLlamaV2MLP): break
+            if isinstance(module, ExLlamaV2MoEMLP): break
+            if idx < 0:
+                print(" ## Not enough layers")
+                sys.exit()
+
+        print(f" -- Reducing {module.key} ({module.name}) to {k * 100:.2f}%")
+        module.rank_reduce(k)
+
+# Replacement
+
+if args.mix_layers:
+    intervals_, extra_dir = args.mix_layers.split(":")
+
+    print(f" -- Loading replacement layers from: {extra_dir}")
+
+    extra_config = ExLlamaV2Config()
+    extra_config.model_dir = extra_dir
+    extra_config.prepare()
+    intervals = intervals_.split(",")
+    for interval in intervals:
+        ab = interval.split("-")
+        a, b = int(ab[0]), int(ab[-1])
+        for idx in range(a, b + 1):
+            print(f" --   Layer {idx}...")
+            layerkey = "model.layers." + str(idx) + "."
+            remove = [k for k in model.config.tensor_file_map.keys() if k.startswith(layerkey)]
+            replace = [k for k in extra_config.tensor_file_map.keys() if k.startswith(layerkey)]
+            # reload = [k for k in model.modules_dict.keys() if k.startswith(layerkey)]
+            for k in remove: del model.config.tensor_file_map[k]
+            for k in replace: model.config.tensor_file_map[k] = extra_config.tensor_file_map[k]
+            # for k in reload:
+            #     model.modules_dict[k].unload()
+            #     model.modules_dict[k].load()
+            if not args.stream_layers:
+                model.modules[idx * 2 + 1].reload()
+                model.modules[idx * 2 + 2].reload()
+
+# Test generation
+
+if args.prompt:
+
+    with torch.inference_mode():
+
+        if cache is None:
+            cache = ExLlamaV2Cache(model) if not model.tp_context else ExLlamaV2Cache_TP(model)
+
+        ids = tokenizer.encode(args.prompt)
+        tokens_prompt = ids.shape[-1]
+
+        print(f" -- Warmup...")
+
+        generator = ExLlamaV2BaseGenerator(model, cache, tokenizer)
+        if not args.no_warmup: generator.warmup()
+
+        print(f" -- Generating...")
+        print()
+
+        settings = ExLlamaV2Sampler.Settings()
+        settings.temperature = 1.0
+        settings.top_k = 0
+        settings.top_p = 0.8
+        settings.token_repetition_penalty = 1.02
+        settings.disallow_tokens(tokenizer, [tokenizer.eos_token_id])
+
+        time_begin = time.time()
+
+        output = generator.generate_simple(args.prompt, settings, args.tokens, token_healing = True, add_bos = not args.prompt_no_bos)
+
+        torch.cuda.synchronize()
+        time_prompt = time.time()
+
+        time_end = time.time()
+
+    print(output)
+    print()
+
+    total_gen = time_end - time_begin
+    print(f" -- Response generated in {total_gen:.2f} seconds, {args.tokens} tokens, {args.tokens / total_gen:.2f} tokens/second (includes prompt eval.)")
+
+
+# Test perplexity
+
+if args.eval_dataset or args.standard_perplexity:
+
+    with torch.inference_mode():
+
+        print(f" -- Running perplexity test")
+
+        if args.standard_perplexity:
+
+            eval_length = args.eval_length
+            if args.eval_dataset:
+                print(f" !! Note, overriding specified --eval_dataset with {args.standard_perplexity}")
+
+            from datasets import load_dataset
+
+            if args.standard_perplexity == "wiki2":
+                ds = "wikitext"
+                part = "wikitext-2-raw-v1"
+                split = "test"
+            # if args.standard_perplexity == "c4":
+            #     ds = "allenai/c4"
+            #     part = "allenai--c4"
+            #     split = "train"
+
+            print(f" -- Loading dataset {ds}, {part}, {split}...")
+            test = load_dataset(ds, part, split = split)
+
+            print(f" -- Tokenizing samples...")
+            text = "\n\n".join(test["text"])
+            eval_tokens = tokenizer.encode(text)
+
+            stride = 512
+            seqs = []
+            eval_len = []
+            a = 0
+            while True:
+                b = a + model.config.max_seq_len
+                if b > eval_tokens.shape[-1]: break
+                seqs.append(eval_tokens[:, a:b])
+                eval_len.append(b if a == 0 else stride)
+                a += stride
+
+            eval_tokens = torch.cat(seqs, dim = 0)
+
+        else:
+
+            eval_dataset = args.eval_dataset
+            eval_rows = args.eval_rows
+            eval_length = args.eval_length
+
+            print(f" -- Dataset: {eval_dataset}")
+            print(f" -- Tokenizing eval data, {eval_rows} rows x {eval_length} tokens...")
+
+            eval_tokens = get_tokens(eval_rows, eval_length, eval_dataset, tokenizer)
+            eval_len = [eval_tokens.shape[1]] * eval_tokens.shape[0]
+
+            # if args.eval_bos:
+            if model.config.arch.requires_bos:
+                boss = torch.full((eval_tokens.shape[0], 1), tokenizer.bos_token_id, dtype = torch.long)
+                eval_tokens = torch.cat((boss, eval_tokens[:, :-1]), dim = 1)
+
+        if args.eval_context_lens:
+            logprob_sum = []
+            logprob_count = []
+        else:
+            logprob_sum = 0.0
+            logprob_count = 0
+
+        def ppl(input_ids__, logits__, lengths__, bins = False):
+
+            logits_device = model.modules[-1].device() if not model.tp_context else \
+                            torch.device(model.tp_context.device)
+
+            if bins:
+                num_bins = (max(lengths__) + 255) // 256
+                logprob_sum_ = [0.0] * num_bins
+                logprob_count_ = [0] * num_bins
+            else:
+                logprob_sum_ = 0.0
+                logprob_count_ = 0
+
+            assert logits__.shape[0] == input_ids__.shape[0]
+            ll = logits__.shape[1]
+
+            for bi in range(logits__.shape[0]):
+                cl = max(ll - lengths__[bi], 0)
+                logits_ = logits__[bi:bi+1, cl:, :]
+                input_ids_ = input_ids__[bi:bi+1, cl:]
+
+                if bins:
+                    chunksize = 256
+                else:
+                    chunksize = logits_.shape[1] * 4000 // logits_.shape[2] + 1
+                b_ = 0
+                while b_ < logits_.shape[1]:
+                    a_ = b_
+                    b_ = min(b_ + chunksize, logits_.shape[1])
+
+                    logits_f = logits_[:, a_:b_, :].to(logits_device).float() + 1e-10
+                    target_ids = input_ids_[:, a_ + 1:b_ + 1].to(logits_f.device)
+
+                    log_probs = F.log_softmax(logits_f, dim=-1)
+                    token_log_probs = log_probs.gather(-1, target_ids.unsqueeze(-1)).squeeze(-1)
+                    if bins:
+                        # for cbin in range(a_ // 256 + 1):
+                        cbin = a_ // 256
+                        logprob_sum_[cbin] += token_log_probs.sum().item()
+                        logprob_count_[cbin] += target_ids.numel()
+                    else:
+                        logprob_sum_ += token_log_probs.sum().item()
+                        logprob_count_ += target_ids.numel()
+
+            return logprob_sum_, logprob_count_
+
+        if args.stream_layers:
+
+            print(f" -- Inference (streamed)", end = "")
+            sys.stdout.flush()
+
+            batch_size, seq_len = eval_tokens.shape
+            attn_params = ExLlamaV2Attention.Params(stream_batch_size, seq_len, 0, None, None)
+            # attn_mask = model.build_attn_mask(stream_batch_size, seq_len, 0, None, "cuda:0")
+
+            for idx, module in enumerate(model.modules):
+                module.set_device_idx(-1 if idx == 0 else 0)
+
+            model.modules[0].load()
+            hidden_state = model.modules[0].forward(eval_tokens)
+            model.modules[0].unload()
+
+            for idx, module in enumerate(model.modules):
+                if idx == 0: continue
+
+                print(".", end = "")
+                sys.stdout.flush()
+                module.load()
+
+                b = 0
+                while b < eval_tokens.shape[0]:
+                    a = b
+                    b = min(b + stream_batch_size, eval_tokens.shape[0])
+                    x = hidden_state[a:b, :, :].to("cuda:0")
+                    x = module.forward(x, cache = None, attn_params = attn_params, past_len = 0, loras = None)
+
+                    if idx < len(model.modules) - 1:
+                        hidden_state[a:b, :, :] = x.to("cpu")
+
+                    else:
+                        input_ids = eval_tokens[a:b, :]
+                        logits = x[:, :-1, :]
+
+                        # if model.config.logit_scale != 1:
+                        #     logits.mul_(model.config.logit_scale)
+
+                        logprob_sum__, logprob_count__ = ppl(input_ids, logits, eval_len[a:b])
+                        logprob_sum += logprob_sum__
+                        logprob_count += logprob_count__
+
+                module.unload()
+
+            print()
+
+        else:
+
+            print(f" -- Inference", end = "")
+            sys.stdout.flush()
+
+            if cache is None:
+                if eval_length > model.config.max_input_len:
+                    cache = ExLlamaV2Cache(model, max_seq_len = eval_length) if not model.tp_context else ExLlamaV2Cache_TP(model, max_seq_len = eval_length)
+                else:
+                    cache = None
+
+            for i in range(eval_tokens.shape[0]):
+
+                if i % 10 == 0: print(".", end = "")
+                sys.stdout.flush()
+
+                input_ids = eval_tokens[i:i+1, :]
+
+                input_ids = input_ids[:, :]
+                if cache is not None: cache.current_seq_len = 0
+                logits = model.forward(input_ids, cache, cpu_logits = input_ids.numel() > 2048)
+                logits = logits[:, :-1, :]
+
+                logprob_sum__, logprob_count__ = ppl(input_ids, logits, eval_len[i:i+1], args.eval_context_lens)
+                if args.eval_context_lens:
+                    while len(logprob_sum) < len(logprob_sum__):
+                        logprob_sum.append(0.0)
+                        logprob_count.append(0)
+                    for j in range(len(logprob_sum__)):
+                        logprob_sum[j] += logprob_sum__[j]
+                        logprob_count[j] += logprob_count__[j]
+                else:
+                    logprob_sum += logprob_sum__
+                    logprob_count += logprob_count__
+
+        if not args.eval_context_lens:
+            print()
+            mean_log_prob = logprob_sum / logprob_count
+            perplexity = math.exp(-mean_log_prob)
+            print(f" -- Evaluation perplexity: {perplexity:.4f}")
+        else:
+            print()
+            for j in range(len(logprob_sum__)):
+                mean_log_prob = logprob_sum[j] / logprob_count[j]
+                perplexity = math.exp(-mean_log_prob)
+                dl = min((j + 1) * 256, eval_length)
+                print(f" -- Evaluation perplexity: {dl} {perplexity:.4f}")
+
+        def test_ppl_token():
+            global logprob_sum, logprob_count, i, input_ids
+            global logits, target_ids, log_probs, token_log_probs
+            global mean_log_prob, perplexity
+
+            # set_catch("model.layers.3")
+
+            logprob_sum = 0
+            logprob_count = 0
+
+            for i in range(eval_tokens.shape[0]):
+
+                cache.current_seq_len = 0
+
+                for j in range(eval_tokens.shape[1] - 1):
+                    if j % 256 == 0: print(".", end = "")
+                    sys.stdout.flush()
+
+                    input_ids = eval_tokens[i:i + 1, j:j + 1]
+                    logits = model.forward(input_ids, cache)
+                    logits = logits.float() + 1e-10
+
+                    log_probs = F.log_softmax(logits, dim = -1)
+                    logprob_sum += log_probs[0, 0, eval_tokens[i, j+1]]
+                    logprob_count += 1
+
+                    # mean_log_prob = logprob_sum / logprob_count
+                    # perplexity = math.exp(-mean_log_prob)
+                    # print(f" -- Token {j}: {perplexity:.4f}")
+
+            print()
+
+            mean_log_prob = logprob_sum / logprob_count
+            perplexity = math.exp(-mean_log_prob)
+            print(f" -- Evaluation perplexity: {perplexity:.4f}")
+
+        if args.eval_token:
+            if args.standard_perplexity:
+                print(f" !! Note, can't evalutate token perplexity on standard test")
+            else:
+                print(f" -- Inference (token)", end = "")
+                sys.stdout.flush()
+                cache = ExLlamaV2Cache(model, max_seq_len = eval_length) if not model.tp_context else \
+                        ExLlamaV2Cache_TP(model, max_seq_len = eval_length)
+                test_ppl_token()
+
+        if args.eval_token_8bit:
+            if args.standard_perplexity:
+                print(f" !! Note, can't evalutate token perplexity on standard test")
+            else:
+                print(f" -- Inference (token, 8-bit cache)", end = "")
+                sys.stdout.flush()
+                cache = ExLlamaV2Cache_8bit(model, max_seq_len = eval_length) if not model.tp_context else \
+                        ExLlamaV2Cache_TP(model, max_seq_len = eval_length, base = ExLlamaV2Cache_8bit)
+                test_ppl_token()
+
+        if args.eval_token_q4:
+            if args.standard_perplexity:
+                print(f" !! Note, can't evalutate token perplexity on standard test")
+            else:
+                print(f" -- Inference (token, Q4 cache)", end = "")
+                sys.stdout.flush()
+                cache = ExLlamaV2Cache_Q4(model, max_seq_len = eval_length) if not model.tp_context else \
+                        ExLlamaV2Cache_TP(model, max_seq_len = eval_length, base = ExLlamaV2Cache_Q4)
+                # cache.calibrate(tokenizer)
+                test_ppl_token()
+
+        if args.eval_token_q6:
+            if args.standard_perplexity:
+                print(f" !! Note, can't evalutate token perplexity on standard test")
+            else:
+                print(f" -- Inference (token, Q6 cache)", end = "")
+                sys.stdout.flush()
+                cache = ExLlamaV2Cache_Q6(model, max_seq_len = eval_length) if not model.tp_context else \
+                        ExLlamaV2Cache_TP(model, max_seq_len = eval_length, base = ExLlamaV2Cache_Q6)
+                # cache.calibrate(tokenizer)
+                test_ppl_token()
+
+        if args.eval_token_q8:
+            if args.standard_perplexity:
+                print(f" !! Note, can't evalutate token perplexity on standard test")
+            else:
+                print(f" -- Inference (token, Q8 cache)", end = "")
+                sys.stdout.flush()
+                cache = ExLlamaV2Cache_Q8(model, max_seq_len = eval_length) if not model.tp_context else \
+                        ExLlamaV2Cache_TP(model, max_seq_len = eval_length, base = ExLlamaV2Cache_Q8)
+                # cache.calibrate(tokenizer)
+                test_ppl_token()
+
+
+# Test prompt speed
+
+if args.prompt_speed:
+
+    with torch.inference_mode():
+
+        if cache is None:
+            cache = ExLlamaV2Cache(model) if not model.tp_context else ExLlamaV2Cache_TP(model)
+
+        ids = torch.randint(0, model.config.vocab_size - 1, (1, model.config.max_seq_len))
+
+        print(f" -- Warmup...")
+
+        if not args.no_warmup:
+            model.forward(ids[:, -1:])
+
+        print(f" -- Measuring prompt speed...")
+
+        torch.cuda.synchronize()
+
+        current_len = 128
+        step = 128
+        prompt_iters = 3
+        while True:
+
+            total_time = 0
+            for i in range(prompt_iters):
+
+                torch.cuda.synchronize()
+                time_begin = time.time()
+
+                cache.current_seq_len = 0
+                model.forward(ids[:, :current_len], cache, preprocess_only = True)
+
+                torch.cuda.synchronize()
+                time_end = time.time()
+                total_time += time_end - time_begin
+
+            tps = current_len / (total_time / prompt_iters)
+
+            print(f" ** Length {current_len:>5} tokens: {tps:>11.4f} t/s")
+
+            if current_len >= 1024: step = 1024
+            if current_len >= 4096: step = 4096
+            if current_len >= 16384: step = 8192
+
+            current_len_ = current_len
+            current_len = min(current_len + step, model.config.max_seq_len)
+            if current_len == current_len_: break
+
+
+# Test token speed
+
+if args.speed:
+
+    with torch.inference_mode():
+
+        if cache is None:
+            cache = ExLlamaV2Cache(model) if not model.tp_context else ExLlamaV2Cache_TP(model)
+        cache.current_seq_len = 0
+
+        print(f" -- Measuring token speed...")
+        ids = tokenizer.encode("X")
+        model.forward(ids[:, :])
+
+        current_idx = ids.shape[-1]
+        next_stop = 128
+
+        while True:
+
+            time_begin = time.time()
+
+            tokens = next_stop - current_idx
+            for i in range(tokens):
+
+                logits = model.forward(ids[:, -1:], cache)
+                sample = torch.argmax(logits[0, -1]).cpu().unsqueeze(0).unsqueeze(0)
+                sample.clamp_(0, tokenizer.get_vocab_size() - 1)
+                ids = torch.cat((ids, sample), dim=-1)
+
+            time_end = time.time()
+            tps = tokens / (time_end - time_begin)
+
+            print(f" ** Position {current_idx:>5} + {tokens:>3} tokens: {tps:>9.4f} t/s")
+
+            current_idx = next_stop
+            next_stop = min(next_stop + 128, model.config.max_seq_len)
+            if next_stop == current_idx: break
+