EngineX-Mthreads/enginex-mthreads-llama.cpp
4
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

sync from b7516

Browse Source
This commit is contained in:
xiezhongtao
2026-01-16 11:16:14 +08:00
parent f4ae4cc7da
commit 6ee41dd9e3
380 changed files with 18435 additions and 38806 deletions
Download Patch File Download Diff File Expand all files Collapse all files

243
gguf-py/gguf/constants.py
View File

@@ -104,7 +104,6 @@ class Keys:
VOCAB_SIZE = "{arch}.vocab_size"
CONTEXT_LENGTH = "{arch}.context_length"
EMBEDDING_LENGTH = "{arch}.embedding_length"
EMBEDDING_LENGTH_OUT = "{arch}.embedding_length_out"
FEATURES_LENGTH = "{arch}.features_length"
BLOCK_COUNT = "{arch}.block_count"
LEADING_DENSE_BLOCK_COUNT = "{arch}.leading_dense_block_count"
@@ -276,13 +275,12 @@ class Keys:
DATASETS = "imatrix.datasets"
class Clip:
PROJECTOR_TYPE = "clip.projector_type"
HAS_VISION_ENCODER = "clip.has_vision_encoder"
HAS_AUDIO_ENCODER = "clip.has_audio_encoder"
HAS_LLAVA_PROJECTOR = "clip.has_llava_projector"
PROJECTOR_TYPE = "clip.projector_type"
HAS_VISION_ENCODER = "clip.has_vision_encoder"
HAS_AUDIO_ENCODER = "clip.has_audio_encoder"
HAS_LLAVA_PROJECTOR = "clip.has_llava_projector"
class ClipVision:
PROJECTOR_TYPE = "clip.vision.projector_type" # for mixed modality models
IMAGE_SIZE = "clip.vision.image_size"
PREPROC_IMAGE_SIZE = "clip.vision.preproc_image_size"
PATCH_SIZE = "clip.vision.patch_size"
@@ -296,9 +294,7 @@ class Keys:
USE_GELU = "clip.use_gelu"
USE_SILU = "clip.use_silu"
N_WA_PATTERN = "clip.vision.n_wa_pattern" # used by qwen2.5vl
WA_LAYER_INDEXES = "clip.vision.wa_layer_indexes" # used by youtuvl
IS_DEEPSTACK_LAYERS = "clip.vision.is_deepstack_layers"
WINDOW_SIZE = "clip.vision.window_size"
class Attention:
HEAD_COUNT = "clip.vision.attention.head_count"
@@ -308,7 +304,6 @@ class Keys:
SCALE_FACTOR = "clip.vision.projector.scale_factor"
class ClipAudio:
PROJECTOR_TYPE = "clip.audio.projector_type" # for mixed modality models
NUM_MEL_BINS = "clip.audio.num_mel_bins"
EMBEDDING_LENGTH = "clip.audio.embedding_length"
FEED_FORWARD_LENGTH = "clip.audio.feed_forward_length"
@@ -382,7 +377,6 @@ class MODEL_ARCH(IntEnum):
PHIMOE = auto()
PLAMO = auto()
PLAMO2 = auto()
PLAMO3 = auto()
CODESHELL = auto()
ORION = auto()
INTERNLM2 = auto()
@@ -424,7 +418,6 @@ class MODEL_ARCH(IntEnum):
NEMOTRON_H_MOE = auto()
EXAONE = auto()
EXAONE4 = auto()
EXAONE_MOE = auto()
GRANITE = auto()
GRANITE_MOE = auto()
GRANITE_HYBRID = auto()
@@ -456,9 +449,6 @@ class MODEL_ARCH(IntEnum):
RND1 = auto()
PANGU_EMBED = auto()
MISTRAL3 = auto()
MIMO2 = auto()
LLAMA_EMBED = auto()
MAINCODER = auto()
class VISION_PROJECTOR_TYPE(IntEnum):
@@ -468,7 +458,6 @@ class VISION_PROJECTOR_TYPE(IntEnum):
RESAMPLER = auto()
GLM_EDGE = auto()
MERGER = auto()
GEMMA3N = auto()
GEMMA3 = auto()
QWEN3VL = auto()
COGVLM = auto()
@@ -679,15 +668,6 @@ class MODEL_TENSOR(IntEnum):
V_MM_INP_NORM = auto()
V_MM_INP_PROJ = auto() # gemma3
V_MM_SOFT_EMB_NORM = auto() # gemma3
V_MM_EMBEDDING = auto() # gemma3n
V_MM_HARD_EMB_NORM = auto() # gemma3n
V_ENC_CONV_STEM = auto() # gemma3n
V_ENC_CONV_STEM_NORM = auto() # gemma3n
V_ENC_MSFA_EXP = auto() # gemma3n
V_ENC_MSFA_EXP_NORM = auto() # gemma3n
V_ENC_MSFA_PROJ = auto() # gemma3n
V_ENC_MSFA_PROJ_NORM = auto() # gemma3n
V_ENC_MSFA_NORM = auto() # gemma3n
V_RESMPL_POS_EMBD_K = auto() # minicpmv
V_RESMPL_ATTN_Q = auto() # minicpmv
V_RESMPL_ATTN_K = auto() # minicpmv
@@ -711,41 +691,30 @@ class MODEL_TENSOR(IntEnum):
V_TOK_BOI = auto() # cogvlm
V_TOK_EOI = auto() # cogvlm
# audio (mtmd)
A_ENC_EMBD_POS = auto()
A_ENC_EMBD_NORM = auto()
A_ENC_EMBD_TO_LOGITS = auto() # lfm2
A_ENC_CONV1D = auto()
A_ENC_CONV1D_NORM = auto() # gemma3n
A_PRE_NORM = auto()
A_POST_NORM = auto()
A_ENC_LAYER_PRE_NORM = auto() # gemma3n
A_ENC_ATTN_Q = auto()
A_ENC_ATTN_K = auto()
A_ENC_ATTN_V = auto()
A_ENC_PER_DIM_SCALE = auto() # gemma3n
A_ENC_INPUT_NORM = auto()
A_ENC_OUTPUT = auto()
A_ENC_OUTPUT_NORM = auto()
A_ENC_FFN_UP = auto()
A_ENC_FFN_NORM = auto()
A_ENC_FFN_POST_NORM = auto() # gemma3n
A_ENC_FFN_SCALE = auto() # gemma3n
A_ENC_FFN_GATE = auto()
A_ENC_FFN_DOWN = auto()
A_ENC_FFN_UP_1 = auto() # lfm2, gemma3n
A_ENC_FFN_NORM_1 = auto() # lfm2, gemma3n (pre-norm)
A_ENC_FFN_POST_NORM_1 = auto() # gemma3n
A_ENC_FFN_SCALE_1 = auto() # gemma3n
A_ENC_FFN_GATE_1 = auto() # lfm2, gemma3n
A_ENC_FFN_DOWN_1 = auto() # lfm2, gemma3n
A_MMPROJ = auto()
A_MMPROJ_FC = auto()
A_MM_NORM_PRE = auto()
A_MM_NORM_MID = auto()
A_MM_EMBEDDING = auto() # gemma3n
A_MM_HARD_EMB_NORM = auto() # gemma3n
A_MM_SOFT_EMB_NORM = auto() # gemma3n
A_MM_INP_PROJ = auto() # gemma3n
A_ENC_EMBD_POS = auto()
A_ENC_EMBD_NORM = auto()
A_ENC_EMBD_TO_LOGITS = auto()
A_ENC_CONV1D = auto()
A_PRE_NORM = auto()
A_POST_NORM = auto()
A_ENC_ATTN_Q = auto()
A_ENC_ATTN_K = auto()
A_ENC_ATTN_V = auto()
A_ENC_INPUT_NORM = auto()
A_ENC_OUTPUT = auto()
A_ENC_OUTPUT_NORM = auto()
A_ENC_FFN_UP = auto()
A_ENC_FFN_NORM = auto()
A_ENC_FFN_GATE = auto()
A_ENC_FFN_DOWN = auto()
A_ENC_FFN_UP_1 = auto()
A_ENC_FFN_NORM_1 = auto()
A_ENC_FFN_GATE_1 = auto()
A_ENC_FFN_DOWN_1 = auto()
A_MMPROJ = auto()
A_MMPROJ_FC = auto()
A_MM_NORM_PRE = auto()
A_MM_NORM_MID = auto()
# nextn/mtp
NEXTN_EH_PROJ = auto()
NEXTN_EMBED_TOKENS = auto()
@@ -802,7 +771,6 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.PHIMOE: "phimoe",
MODEL_ARCH.PLAMO: "plamo",
MODEL_ARCH.PLAMO2: "plamo2",
MODEL_ARCH.PLAMO3: "plamo3",
MODEL_ARCH.CODESHELL: "codeshell",
MODEL_ARCH.ORION: "orion",
MODEL_ARCH.INTERNLM2: "internlm2",
@@ -844,7 +812,6 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.NEMOTRON_H_MOE: "nemotron_h_moe",
MODEL_ARCH.EXAONE: "exaone",
MODEL_ARCH.EXAONE4: "exaone4",
MODEL_ARCH.EXAONE_MOE: "exaone-moe",
MODEL_ARCH.GRANITE: "granite",
MODEL_ARCH.GRANITE_MOE: "granitemoe",
MODEL_ARCH.GRANITE_HYBRID: "granitehybrid",
@@ -877,9 +844,6 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.RND1: "rnd1",
MODEL_ARCH.PANGU_EMBED: "pangu-embedded",
MODEL_ARCH.MISTRAL3: "mistral3",
MODEL_ARCH.MIMO2: "mimo2",
MODEL_ARCH.LLAMA_EMBED: "llama-embed",
MODEL_ARCH.MAINCODER: "maincoder",
}
VISION_PROJECTOR_TYPE_NAMES: dict[VISION_PROJECTOR_TYPE, str] = {
@@ -1096,16 +1060,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
MODEL_TENSOR.V_MM_POST_NORM: "mm.post_norm",
MODEL_TENSOR.V_MM_INP_PROJ: "mm.input_projection",
MODEL_TENSOR.V_MM_INP_NORM: "mm.input_norm",
MODEL_TENSOR.V_MM_SOFT_EMB_NORM: "mm.soft_emb_norm", # gemma3n
MODEL_TENSOR.V_MM_EMBEDDING: "mm.embedding", # gemma3n
MODEL_TENSOR.V_MM_HARD_EMB_NORM: "mm.hard_emb_norm", # gemma3n
MODEL_TENSOR.V_ENC_CONV_STEM: "v.conv_stem.conv", # gemma3n
MODEL_TENSOR.V_ENC_CONV_STEM_NORM: "v.conv_stem.bn", # gemma3n
MODEL_TENSOR.V_ENC_MSFA_EXP: "v.msfa.ffn.pw_exp.conv", # gemma3n
MODEL_TENSOR.V_ENC_MSFA_EXP_NORM: "v.msfa.ffn.pw_exp.bn", # gemma3n
MODEL_TENSOR.V_ENC_MSFA_PROJ: "v.msfa.ffn.pw_proj.conv", # gemma3n
MODEL_TENSOR.V_ENC_MSFA_PROJ_NORM: "v.msfa.ffn.pw_proj.bn", # gemma3n
MODEL_TENSOR.V_ENC_MSFA_NORM: "v.msfa.norm", # gemma3n
MODEL_TENSOR.V_MM_SOFT_EMB_NORM: "mm.soft_emb_norm",
MODEL_TENSOR.V_RESMPL_POS_EMBD_K: "resampler.pos_embd_k",
MODEL_TENSOR.V_RESMPL_ATTN_Q: "resampler.attn.q",
MODEL_TENSOR.V_RESMPL_ATTN_K: "resampler.attn.k",
@@ -1134,26 +1089,19 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
MODEL_TENSOR.A_ENC_EMBD_NORM: "a.position_embd_norm",
MODEL_TENSOR.A_ENC_EMBD_TO_LOGITS: "a.embd_to_logits",
MODEL_TENSOR.A_ENC_CONV1D: "a.conv1d.{bid}",
MODEL_TENSOR.A_ENC_CONV1D_NORM: "a.conv1d.{bid}.norm",
MODEL_TENSOR.A_PRE_NORM: "a.pre_ln",
MODEL_TENSOR.A_POST_NORM: "a.post_ln",
MODEL_TENSOR.A_ENC_LAYER_PRE_NORM: "a.blk.{bid}.layer_pre_norm",
MODEL_TENSOR.A_ENC_ATTN_Q: "a.blk.{bid}.attn_q",
MODEL_TENSOR.A_ENC_ATTN_K: "a.blk.{bid}.attn_k",
MODEL_TENSOR.A_ENC_ATTN_V: "a.blk.{bid}.attn_v",
MODEL_TENSOR.A_ENC_PER_DIM_SCALE: "a.blk.{bid}.per_dim_scale",
MODEL_TENSOR.A_ENC_INPUT_NORM: "a.blk.{bid}.ln1",
MODEL_TENSOR.A_ENC_OUTPUT: "a.blk.{bid}.attn_out",
MODEL_TENSOR.A_ENC_OUTPUT_NORM: "a.blk.{bid}.ln2",
MODEL_TENSOR.A_ENC_FFN_NORM: "a.blk.{bid}.ffn_norm",
MODEL_TENSOR.A_ENC_FFN_POST_NORM: "a.blk.{bid}.ffn_post_norm",
MODEL_TENSOR.A_ENC_FFN_SCALE: "a.blk.{bid}.ffn_scale",
MODEL_TENSOR.A_ENC_FFN_UP: "a.blk.{bid}.ffn_up",
MODEL_TENSOR.A_ENC_FFN_GATE: "a.blk.{bid}.ffn_gate",
MODEL_TENSOR.A_ENC_FFN_DOWN: "a.blk.{bid}.ffn_down",
MODEL_TENSOR.A_ENC_FFN_NORM_1: "a.blk.{bid}.ffn_norm_1",
MODEL_TENSOR.A_ENC_FFN_POST_NORM_1: "a.blk.{bid}.ffn_post_norm_1",
MODEL_TENSOR.A_ENC_FFN_SCALE_1: "a.blk.{bid}.ffn_scale_1",
MODEL_TENSOR.A_ENC_FFN_UP_1: "a.blk.{bid}.ffn_up_1",
MODEL_TENSOR.A_ENC_FFN_GATE_1: "a.blk.{bid}.ffn_gate_1",
MODEL_TENSOR.A_ENC_FFN_DOWN_1: "a.blk.{bid}.ffn_down_1",
@@ -1161,10 +1109,6 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
MODEL_TENSOR.A_MMPROJ_FC: "mm.a.fc",
MODEL_TENSOR.A_MM_NORM_PRE: "mm.a.norm_pre",
MODEL_TENSOR.A_MM_NORM_MID: "mm.a.norm_mid",
MODEL_TENSOR.A_MM_INP_PROJ: "mm.a.input_projection", # gemma3n
MODEL_TENSOR.A_MM_SOFT_EMB_NORM: "mm.a.soft_emb_norm", # gemma3n
MODEL_TENSOR.A_MM_EMBEDDING: "mm.a.embedding", # gemma3n
MODEL_TENSOR.A_MM_HARD_EMB_NORM: "mm.a.hard_emb_norm", # gemma3n
# lfm2 audio
MODEL_TENSOR.A_ENC_NORM_CONV: "a.blk.{bid}.norm_conv",
MODEL_TENSOR.A_ENC_LINEAR_POS: "a.blk.{bid}.linear_pos",
@@ -1215,15 +1159,6 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.V_MM_INP_PROJ,
MODEL_TENSOR.V_MM_INP_NORM,
MODEL_TENSOR.V_MM_SOFT_EMB_NORM,
MODEL_TENSOR.V_MM_EMBEDDING,
MODEL_TENSOR.V_MM_HARD_EMB_NORM,
MODEL_TENSOR.V_ENC_CONV_STEM,
MODEL_TENSOR.V_ENC_CONV_STEM_NORM,
MODEL_TENSOR.V_ENC_MSFA_EXP,
MODEL_TENSOR.V_ENC_MSFA_EXP_NORM,
MODEL_TENSOR.V_ENC_MSFA_PROJ,
MODEL_TENSOR.V_ENC_MSFA_PROJ_NORM,
MODEL_TENSOR.V_ENC_MSFA_NORM,
MODEL_TENSOR.V_RESMPL_POS_EMBD_K,
MODEL_TENSOR.V_RESMPL_ATTN_Q,
MODEL_TENSOR.V_RESMPL_ATTN_K,
@@ -1251,26 +1186,19 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.A_ENC_EMBD_NORM,
MODEL_TENSOR.A_ENC_EMBD_TO_LOGITS,
MODEL_TENSOR.A_ENC_CONV1D,
MODEL_TENSOR.A_ENC_CONV1D_NORM,
MODEL_TENSOR.A_PRE_NORM,
MODEL_TENSOR.A_POST_NORM,
MODEL_TENSOR.A_ENC_LAYER_PRE_NORM,
MODEL_TENSOR.A_ENC_ATTN_Q,
MODEL_TENSOR.A_ENC_ATTN_K,
MODEL_TENSOR.A_ENC_ATTN_V,
MODEL_TENSOR.A_ENC_PER_DIM_SCALE,
MODEL_TENSOR.A_ENC_INPUT_NORM,
MODEL_TENSOR.A_ENC_OUTPUT,
MODEL_TENSOR.A_ENC_OUTPUT_NORM,
MODEL_TENSOR.A_ENC_FFN_NORM,
MODEL_TENSOR.A_ENC_FFN_POST_NORM,
MODEL_TENSOR.A_ENC_FFN_SCALE,
MODEL_TENSOR.A_ENC_FFN_UP,
MODEL_TENSOR.A_ENC_FFN_GATE,
MODEL_TENSOR.A_ENC_FFN_DOWN,
MODEL_TENSOR.A_ENC_FFN_NORM_1,
MODEL_TENSOR.A_ENC_FFN_POST_NORM_1,
MODEL_TENSOR.A_ENC_FFN_SCALE_1,
MODEL_TENSOR.A_ENC_FFN_UP_1,
MODEL_TENSOR.A_ENC_FFN_GATE_1,
MODEL_TENSOR.A_ENC_FFN_DOWN_1,
@@ -1287,10 +1215,6 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.A_ENC_CONV_NORM,
MODEL_TENSOR.A_ENC_CONV_PW1,
MODEL_TENSOR.A_ENC_CONV_PW2,
MODEL_TENSOR.A_MM_INP_PROJ,
MODEL_TENSOR.A_MM_SOFT_EMB_NORM,
MODEL_TENSOR.A_MM_EMBEDDING,
MODEL_TENSOR.A_MM_HARD_EMB_NORM,
],
MODEL_ARCH.LLAMA: [
MODEL_TENSOR.TOKEN_EMBD,
@@ -1740,7 +1664,6 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.ATTN_POST_NORM,
MODEL_TENSOR.ATTN_GATE,
MODEL_TENSOR.ATTN_QKV,
MODEL_TENSOR.FFN_GATE_INP,
MODEL_TENSOR.FFN_GATE_INP_SHEXP,
MODEL_TENSOR.FFN_UP_SHEXP,
@@ -1836,21 +1759,6 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.SSM_B_NORM,
MODEL_TENSOR.SSM_C_NORM,
],
MODEL_ARCH.PLAMO3: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_QKV,
MODEL_TENSOR.ATTN_Q_NORM,
MODEL_TENSOR.ATTN_K_NORM,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.ATTN_POST_NORM,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.FFN_POST_NORM,
],
MODEL_ARCH.GPT2: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.POS_EMBD,
@@ -2756,38 +2664,6 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.FFN_POST_NORM,
],
MODEL_ARCH.EXAONE_MOE: [
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_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,
MODEL_TENSOR.FFN_GATE_INP,
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_TENSOR.FFN_EXP_PROBS_B,
# NextN/MTP tensors - preserved but unused
MODEL_TENSOR.NEXTN_EH_PROJ,
MODEL_TENSOR.NEXTN_EMBED_TOKENS,
MODEL_TENSOR.NEXTN_ENORM,
MODEL_TENSOR.NEXTN_HNORM,
MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD,
MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM,
],
MODEL_ARCH.GRANITE: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
@@ -3137,7 +3013,6 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.DENSE_2_OUT, # LFM2-ColBert-350M
],
MODEL_ARCH.LFM2MOE: [
MODEL_TENSOR.TOKEN_EMBD,
@@ -3321,62 +3196,6 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_DOWN_EXP,
MODEL_TENSOR.FFN_UP_EXP,
],
MODEL_ARCH.MIMO2: [
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_SINKS,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.FFN_GATE_INP,
MODEL_TENSOR.FFN_GATE_EXP,
MODEL_TENSOR.FFN_DOWN_EXP,
MODEL_TENSOR.FFN_UP_EXP,
MODEL_TENSOR.FFN_EXP_PROBS_B,
],
MODEL_ARCH.LLAMA_EMBED: [
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.MAINCODER: [
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
}
@@ -3594,8 +3413,6 @@ class GGUFValueType(IntEnum):
class VisionProjectorType:
GEMMA3 = "gemma3"
GEMMA3NV = "gemma3nv"
GEMMA3NA = "gemma3na"
IDEFICS3 = "idefics3"
PIXTRAL = "pixtral"
LLAMA4 = "llama4"
@@ -3614,9 +3431,7 @@ class VisionProjectorType:
COGVLM = "cogvlm"
JANUS_PRO = "janus_pro"
LFM2A = "lfm2a" # audio
MUSIC_FLAMINGO = "musicflamingo" # audio
GLM4V = "glm4v"
YOUTUVL = "youtuvl"
# Items here are (block size, type size)

38
gguf-py/gguf/gguf_writer.py
View File

@@ -681,9 +681,6 @@ class GGUFWriter:
def add_embedding_length(self, length: int) -> None:
self.add_uint32(Keys.LLM.EMBEDDING_LENGTH.format(arch=self.arch), length)
def add_embedding_length_out(self, length: int) -> None:
self.add_uint32(Keys.LLM.EMBEDDING_LENGTH_OUT.format(arch=self.arch), length)
def add_features_length(self, length: int) -> None:
self.add_uint32(Keys.LLM.FEATURES_LENGTH.format(arch=self.arch), length)
@@ -1086,9 +1083,6 @@ class GGUFWriter:
def add_clip_projector_type(self, value: str) -> None:
self.add_string(Keys.Clip.PROJECTOR_TYPE, value)
def add_clip_vision_projector_type(self, value: str) -> None:
self.add_string(Keys.ClipVision.PROJECTOR_TYPE, value)
def add_vision_projection_dim(self, value: int) -> None:
self.add_uint32(Keys.ClipVision.PROJECTION_DIM, value)
@@ -1135,45 +1129,13 @@ class GGUFWriter:
self.add_uint32(Keys.ClipVision.Projector.SCALE_FACTOR, value)
def add_vision_n_wa_pattern(self, value: int) -> None:
"""Add window attention pattern interval for vision models.
This defines the pattern interval for window attention vs full attention layers.
For example, if n_wa_pattern=4, then layers 3, 7, 11, ... use full attention,
while other layers use window attention.
Used by models like Qwen2.5-VL where full attention layers follow a regular pattern.
"""
self.add_uint32(Keys.ClipVision.N_WA_PATTERN, value)
def add_vision_wa_layer_indexes(self, layers: Sequence[int]) -> None:
"""Add explicit layer indexes that use full attention in vision models.
This specifies the exact layer indices (0-based) that should use full attention
instead of window attention. All other layers will use window attention.
Args:
layers: List of layer indices that use full attention (e.g., [3, 7, 11, 15])
Used by models like YoutuVL where full attention layers are explicitly specified
rather than following a regular pattern.
Difference from add_vision_n_wa_pattern:
- n_wa_pattern: Defines a regular interval pattern (every Nth layer uses full attention)
- wa_layer_indexes: Explicitly lists which layers use full attention (irregular pattern)
"""
self.add_array(Keys.ClipVision.WA_LAYER_INDEXES, layers)
def add_vision_is_deepstack_layers(self, layers: Sequence[bool]) -> None:
self.add_array(Keys.ClipVision.IS_DEEPSTACK_LAYERS, layers)
def add_vision_window_size(self, value: int) -> None:
self.add_uint32(Keys.ClipVision.WINDOW_SIZE, value)
# audio models
def add_clip_audio_projector_type(self, value: str) -> None:
self.add_string(Keys.ClipAudio.PROJECTOR_TYPE, value)
def add_audio_projection_dim(self, value: int) -> None:
self.add_uint32(Keys.ClipAudio.PROJECTION_DIM, value)

115
gguf-py/gguf/tensor_mapping.py
View File

@@ -123,40 +123,6 @@ class TensorNameMap:
MODEL_TENSOR.CONV1D: (
"backbone.embed", # roberta
),
MODEL_TENSOR.V_MM_EMBEDDING: (
"model.embed_vision.embedding", # gemma3n
),
MODEL_TENSOR.V_MM_HARD_EMB_NORM: (
"model.embed_vision.hard_embedding_norm", # gemma3n
),
MODEL_TENSOR.V_MM_INP_PROJ: (
"model.embed_vision.embedding_projection", # gemma3n
),
MODEL_TENSOR.V_MM_SOFT_EMB_NORM: (
"model.embed_vision.soft_embedding_norm", # gemma3n
),
MODEL_TENSOR.V_ENC_CONV_STEM: (
"model.vision_tower.timm_model.conv_stem.conv", # gemma3n
),
MODEL_TENSOR.V_ENC_CONV_STEM_NORM: (
"model.vision_tower.timm_model.conv_stem.bn", # gemma3n
),
MODEL_TENSOR.V_ENC_MSFA_EXP: (
"model.vision_tower.timm_model.msfa.ffn.pw_exp.conv", # gemma3n
),
MODEL_TENSOR.V_ENC_MSFA_EXP_NORM: (
"model.vision_tower.timm_model.msfa.ffn.pw_exp.bn", # gemma3n
),
MODEL_TENSOR.V_ENC_MSFA_PROJ: (
"model.vision_tower.timm_model.msfa.ffn.pw_proj.conv", # gemma3n
),
MODEL_TENSOR.V_ENC_MSFA_PROJ_NORM: (
"model.vision_tower.timm_model.msfa.ffn.pw_proj.bn", # gemma3n
),
MODEL_TENSOR.V_ENC_MSFA_NORM: (
"model.vision_tower.timm_model.msfa.norm", # gemma3n
),
}
block_mappings_cfg: dict[MODEL_TENSOR, tuple[str, ...]] = {
@@ -354,7 +320,6 @@ class TensorNameMap:
MODEL_TENSOR.ATTN_SINKS: (
"model.layers.{bid}.self_attn.sinks", # openai-moe
"model.layers.{bid}.self_attn.attention_sink_bias", # mimov2
),
MODEL_TENSOR.ATTN_GATE: (
@@ -436,8 +401,7 @@ class TensorNameMap:
"model.layers.{bid}.mlp.expert_bias", # afmoe
"model.layers.{bid}.feed_forward.expert_bias", # lfm2moe
"model.layers.{bid}.block_sparse_moe.e_score_correction", # minimax-m2
"backbone.layers.{bid}.mixer.gate.e_score_correction", # nemotron-h-moe
"model.layers.{bid}.mlp.e_score_correction", # exaone-moe
"backbone.layers.{bid}.mixer.gate.e_score_correction" # nemotron-h-moe
),
# Feed-forward up
@@ -630,7 +594,6 @@ class TensorNameMap:
"encoder.layer.{bid}.attention.self.layer_norm_q", # jina-bert-v2
"transformer.layers.{bid}.attn.q_norm", # openelm
"model.layers.layers.{bid}.mixer.q", # plamo2
"model.layers.layers.{bid}.mixer.q_norm", # plamo3
"layers.{bid}.self_attn.q_norm", # qwen3-embedding
"model.layers.{bid}.attention.query_layernorm", # apertus
),
@@ -646,7 +609,6 @@ class TensorNameMap:
"encoder.layer.{bid}.attention.self.layer_norm_k", # jina-bert-v2
"transformer.layers.{bid}.attn.k_norm", # openelm
"model.layers.layers.{bid}.mixer.k", # plamo2
"model.layers.layers.{bid}.mixer.k_norm", # plamo3
"layers.{bid}.self_attn.k_norm", # qwen3-embedding
"model.layers.{bid}.attention.key_layernorm", # apertus
),
@@ -1256,7 +1218,6 @@ class TensorNameMap:
MODEL_TENSOR.V_MMPROJ: (
"multi_modal_projector.linear_{bid}",
"visual.merger.mlp.{bid}", # qwen2vl
"merger.mlp.{bid}",
),
MODEL_TENSOR.V_MMPROJ_FC: (
@@ -1294,7 +1255,6 @@ class TensorNameMap:
"visual.patch_embed.proj", # qwen2vl
"vision_tower.patch_embed.proj", # kimi-vl
"model.vision.patch_embedding.proj", # cogvlm
"siglip2.vision_model.embeddings.patch_embedding",
),
MODEL_TENSOR.V_ENC_EMBD_NORM: (
@@ -1328,7 +1288,6 @@ class TensorNameMap:
"vision_encoder.transformer.layers.{bid}.attention.wq", # pixtral
"visual.blocks.{bid}.attn.q", # qwen2vl, generated
"vision_tower.encoder.blocks.{bid}.wq", # kimi-vl, generated
"siglip2.vision_model.encoder.layers.{bid}.self_attn.q_proj", # youtuvl
),
MODEL_TENSOR.V_ENC_ATTN_Q_NORM: (
@@ -1346,7 +1305,6 @@ class TensorNameMap:
"vision_encoder.transformer.layers.{bid}.attention.wk", # pixtral
"visual.blocks.{bid}.attn.k", # qwen2vl, generated
"vision_tower.encoder.blocks.{bid}.wk", # kimi-vl, generated
"siglip2.vision_model.encoder.layers.{bid}.self_attn.k_proj",
),
MODEL_TENSOR.V_ENC_ATTN_K_NORM: (
@@ -1364,7 +1322,6 @@ class TensorNameMap:
"vision_encoder.transformer.layers.{bid}.attention.wv", # pixtral
"visual.blocks.{bid}.attn.v", # qwen2vl, generated
"vision_tower.encoder.blocks.{bid}.wv", # kimi-vl, generated
"siglip2.vision_model.encoder.layers.{bid}.self_attn.v_proj",
),
MODEL_TENSOR.V_ENC_INPUT_NORM: (
@@ -1379,7 +1336,6 @@ class TensorNameMap:
"visual.blocks.{bid}.norm1", # qwen2vl
"vision_tower.encoder.blocks.{bid}.norm0", # kimi-vl (norm0/norm1)
"model.vision.transformer.layers.{bid}.input_layernorm", # cogvlm
"siglip2.vision_model.encoder.layers.{bid}.layer_norm1",
),
MODEL_TENSOR.V_ENC_ATTN_O: (
@@ -1395,7 +1351,6 @@ class TensorNameMap:
"visual.blocks.{bid}.attn.proj", # qwen2vl
"vision_tower.encoder.blocks.{bid}.wo", # kimi-vl
"model.vision.transformer.layers.{bid}.attention.dense", # cogvlm
"siglip2.vision_model.encoder.layers.{bid}.self_attn.out_proj", # youtuvl
),
MODEL_TENSOR.V_ENC_POST_ATTN_NORM: (
@@ -1410,7 +1365,6 @@ class TensorNameMap:
"visual.blocks.{bid}.norm2", # qwen2vl
"vision_tower.encoder.blocks.{bid}.norm1", # kimi-vl (norm0/norm1)
"model.vision.transformer.layers.{bid}.post_attention_layernorm", # cogvlm
"siglip2.vision_model.encoder.layers.{bid}.layer_norm2",
),
MODEL_TENSOR.V_ENC_FFN_UP: (
@@ -1426,7 +1380,6 @@ class TensorNameMap:
"visual.blocks.{bid}.mlp.linear_fc1", # qwen3vl
"vision_tower.encoder.blocks.{bid}.mlp.fc0", # kimi-vl (fc0/fc1)
"model.vision.transformer.layers.{bid}.mlp.fc1", # cogvlm
"siglip2.vision_model.encoder.layers.{bid}.mlp.fc1",
),
MODEL_TENSOR.V_ENC_FFN_GATE: (
@@ -1448,7 +1401,6 @@ class TensorNameMap:
"visual.blocks.{bid}.mlp.linear_fc2", # qwen3vl
"vision_tower.encoder.blocks.{bid}.mlp.fc1", # kimi-vl (fc0/fc1)
"model.vision.transformer.layers.{bid}.mlp.fc2", # cogvlm
"siglip2.vision_model.encoder.layers.{bid}.mlp.fc2",
),
MODEL_TENSOR.V_LAYER_SCALE_1: (
@@ -1475,7 +1427,6 @@ class TensorNameMap:
"visual.merger.ln_q", # qwen2vl
"vision_tower.encoder.final_layernorm", # kimi-vl
"visual.post_layernorm", # glm4v
"siglip2.vision_model.post_layernorm",
),
MODEL_TENSOR.V_MM_POST_NORM: (
@@ -1492,7 +1443,6 @@ class TensorNameMap:
"multi_modal_projector.pre_norm",
"pre_mm_projector_norm",
"model.vision.linear_proj.norm1", # cogvlm
"merger.ln_q",
),
MODEL_TENSOR.V_MM_SOFT_EMB_NORM: (
@@ -1610,11 +1560,6 @@ class TensorNameMap:
MODEL_TENSOR.A_ENC_CONV1D: (
"audio_tower.conv{bid}", # ultravox
"conformer.pre_encode.conv.{bid}", # lfm2
"model.audio_tower.subsample_conv_projection.conv_{bid}.conv", # gemma3n
),
MODEL_TENSOR.A_ENC_CONV1D_NORM: (
"model.audio_tower.subsample_conv_projection.conv_{bid}.norm", # gemma3n
),
MODEL_TENSOR.A_PRE_NORM: (),
@@ -1627,64 +1572,40 @@ class TensorNameMap:
MODEL_TENSOR.A_ENC_ATTN_Q: (
"audio_tower.layers.{bid}.self_attn.q_proj", # ultravox
"conformer.layers.{bid}.self_attn.linear_q", # lfm2
"conformer.layers.{bid}.attention.attn.q_proj", # gemma3n
),
MODEL_TENSOR.A_ENC_ATTN_K: (
"audio_tower.layers.{bid}.self_attn.k_proj", # ultravox
"conformer.layers.{bid}.self_attn.linear_k", # lfm2
"conformer.layers.{bid}.attention.attn.k_proj", # gemma3n
),
MODEL_TENSOR.A_ENC_ATTN_V: (
"audio_tower.layers.{bid}.self_attn.v_proj", # ultravox
"conformer.layers.{bid}.self_attn.linear_v", # lfm2
"conformer.layers.{bid}.attention.attn.v_proj", # gemma3n
),
MODEL_TENSOR.A_ENC_PER_DIM_SCALE: (
"conformer.layers.{bid}.attention.attn.per_dim_scale", # gemma3n
),
MODEL_TENSOR.A_ENC_LAYER_PRE_NORM: (
"conformer.layers.{bid}.norm", # gemma3n
),
MODEL_TENSOR.A_ENC_INPUT_NORM: (
"audio_tower.layers.{bid}.self_attn_layer_norm", # ultravox
"conformer.layers.{bid}.norm_self_att", # lfm2
"conformer.layers.{bid}.attention.pre_attn_norm", # gemma3n
),
MODEL_TENSOR.A_ENC_OUTPUT: (
"audio_tower.layers.{bid}.self_attn.out_proj", # ultravox
"conformer.layers.{bid}.self_attn.linear_out", # lfm2
"conformer.layers.{bid}.attention.post", # gemma3n
),
MODEL_TENSOR.A_ENC_OUTPUT_NORM: (
"audio_tower.layers.{bid}.final_layer_norm", # ultravox
"conformer.layers.{bid}.norm_out", # lfm2
"conformer.layers.{bid}.attention.post_norm", # gemma3n
),
MODEL_TENSOR.A_ENC_FFN_NORM: (
"conformer.layers.{bid}.norm_feed_forward1", # lfm2
"conformer.layers.{bid}.ffw_layer_start.pre_layer_norm", # gemma3n
),
MODEL_TENSOR.A_ENC_FFN_POST_NORM: (
"conformer.layers.{bid}.ffw_layer_start.post_layer_norm", # gemma3n
),
MODEL_TENSOR.A_ENC_FFN_SCALE: (
"conformer.layers.{bid}.ffw_layer_start.post_layer_scale", # gemma3n
),
MODEL_TENSOR.A_ENC_FFN_UP: (
"audio_tower.layers.{bid}.fc1", # ultravox
"conformer.layers.{bid}.feed_forward1.linear1", # lfm2
"conformer.layers.{bid}.ffw_layer_start.ffw_layer_1", # gemma3n
),
MODEL_TENSOR.A_ENC_FFN_GATE: (),
@@ -1692,35 +1613,22 @@ class TensorNameMap:
MODEL_TENSOR.A_ENC_FFN_DOWN: (
"audio_tower.layers.{bid}.fc2", # ultravox
"conformer.layers.{bid}.feed_forward1.linear2", # lfm2
"conformer.layers.{bid}.ffw_layer_start.ffw_layer_2", # gemma3n
),
MODEL_TENSOR.A_ENC_FFN_UP_1: (
"conformer.layers.{bid}.feed_forward2.linear1", # lfm2
"conformer.layers.{bid}.ffw_layer_end.ffw_layer_1", # gemma3n
),
MODEL_TENSOR.A_ENC_FFN_DOWN_1: (
"conformer.layers.{bid}.feed_forward2.linear2", # lfm2
"conformer.layers.{bid}.ffw_layer_end.ffw_layer_2", # gemma3n
),
MODEL_TENSOR.A_ENC_FFN_NORM_1: (
"conformer.layers.{bid}.norm_feed_forward2", # lfm2
"conformer.layers.{bid}.ffw_layer_end.pre_layer_norm", # gemma3n
),
MODEL_TENSOR.A_ENC_FFN_POST_NORM_1: (
"conformer.layers.{bid}.ffw_layer_end.post_layer_norm", # gemma3n
),
MODEL_TENSOR.A_ENC_FFN_SCALE_1: (
"conformer.layers.{bid}.ffw_layer_end.post_layer_scale", # gemma3n
),
MODEL_TENSOR.A_ENC_LINEAR_POS: (
"conformer.layers.{bid}.self_attn.linear_pos", # lfm2
"conformer.layers.{bid}.attention.attn.relative_position_embedding.pos_proj", # gemma3n
),
MODEL_TENSOR.A_ENC_POS_BIAS_U: (
@@ -1733,7 +1641,6 @@ class TensorNameMap:
MODEL_TENSOR.A_ENC_OUT: (
"conformer.pre_encode.out", # lfm2
"model.audio_tower.subsample_conv_projection.input_proj_linear", # gemma3n
),
# note: some tensors below has "audio." pseudo-prefix, to prevent conflicts with vision tensors
@@ -1759,43 +1666,25 @@ class TensorNameMap:
MODEL_TENSOR.A_ENC_CONV_DW: (
"conformer.layers.{bid}.conv.depthwise_conv", # lfm2
"conformer.layers.{bid}.lconv1d.depthwise_conv1d", # gemma3n
),
MODEL_TENSOR.A_ENC_CONV_NORM: (
"conformer.layers.{bid}.conv.batch_norm", # lfm2
"conformer.layers.{bid}.lconv1d.pre_layer_norm", # gemma3n
),
MODEL_TENSOR.A_ENC_CONV_PW1: (
"conformer.layers.{bid}.conv.pointwise_conv1", # lfm2
"conformer.layers.{bid}.lconv1d.linear_start", # gemma3n
),
MODEL_TENSOR.A_ENC_CONV_PW2: (
"conformer.layers.{bid}.conv.pointwise_conv2", # lfm2
"conformer.layers.{bid}.lconv1d.linear_end", # gemma3n
),
MODEL_TENSOR.A_ENC_NORM_CONV: (
"conformer.layers.{bid}.norm_conv", # lfm2
"conformer.layers.{bid}.lconv1d.conv_norm", # gemma3n
),
MODEL_TENSOR.A_MM_EMBEDDING: (
"model.embed_audio.embedding", # gemma3n
),
MODEL_TENSOR.A_MM_HARD_EMB_NORM: (
"model.embed_audio.hard_embedding_norm", # gemma3n
),
MODEL_TENSOR.A_MM_INP_PROJ: (
"model.embed_audio.embedding_projection", # gemma3n
),
MODEL_TENSOR.A_MM_SOFT_EMB_NORM: (
"model.embed_audio.soft_embedding_norm", # gemma3n
),
# NextN/MTP tensors
# NextN/MTP tensors for GLM4_MOE
MODEL_TENSOR.NEXTN_EH_PROJ: (
"model.layers.{bid}.eh_proj",
),

1
gguf-py/pyproject.toml
View File

@@ -22,7 +22,6 @@ python = ">=3.8"
numpy = ">=1.17"
tqdm = ">=4.27"
pyyaml = ">=5.1"
requests = ">=2.25"
sentencepiece = { version = ">=0.1.98,<=0.2.0", optional = true }
PySide6 = { version = "^6.9", python = ">=3.9,<3.14", optional = true }