tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-26 19:55:04 +00:00
Code Issues Packages Projects Releases Wiki Activity

model : gemma3n text-only (#14400)

Browse Source 
* gemma3n

* add llm_graph_input_one
This commit is contained in:
Xuan-Son Nguyen
2025-06-26 19:34:02 +02:00
committed by GitHub
parent a01047b041
commit 8846aace49
13 changed files with 960 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

75
gguf-py/gguf/constants.py
View File

@ -118,6 +118,10 @@ class Keys:
EMBEDDING_SCALE = "{arch}.embedding_scale"
TOKEN_SHIFT_COUNT = "{arch}.token_shift_count"
INTERLEAVE_MOE_LAYER_STEP = "{arch}.interleave_moe_layer_step"
ACTIVATION_SPARSITY_SCALE = "{arch}.activation_sparsity_scale"
ALTUP_ACTIVE_IDX = "{arch}.altup.active_idx"
ALTUP_NUM_INPUTS = "{arch}.altup.num_inputs"
EMBD_LENGTH_PER_LAYER_INP = "{arch}.embedding_length_per_layer_input"
class Attention:
HEAD_COUNT = "{arch}.attention.head_count"
@ -142,6 +146,8 @@ class Keys:
SCALE = "{arch}.attention.scale"
KEY_LENGTH_MLA = "{arch}.attention.key_length_mla"
VALUE_LENGTH_MLA = "{arch}.attention.value_length_mla"
SHARED_KV_LAYERS = "{arch}.attention.shared_kv_layers"
SLIDING_WINDOW_PATTERN = "{arch}.attention.sliding_window_pattern"
class Rope:
DIMENSION_COUNT = "{arch}.rope.dimension_count"
@ -314,6 +320,7 @@ class MODEL_ARCH(IntEnum):
GEMMA = auto()
GEMMA2 = auto()
GEMMA3 = auto()
GEMMA3N = auto()
STARCODER2 = auto()
RWKV6 = auto()
RWKV6QWEN2 = auto()
@ -399,6 +406,22 @@ class MODEL_TENSOR(IntEnum):
ATTN_Q_NORM = auto()
ATTN_K_NORM = auto()
LAYER_OUT_NORM = auto()
PER_LAYER_TOKEN_EMBD = auto() # gemma3n
PER_LAYER_MODEL_PROJ = auto() # gemma3n
PER_LAYER_INP_GATE = auto() # gemma3n
PER_LAYER_PROJ = auto() # gemma3n
PER_LAYER_PROJ_NORM = auto() # gemma3n
PER_LAYER_POST_NORM = auto() # gemma3n
ALTUP_PROJ = auto() # gemma3n
ALTUP_UNEMBD_PROJ = auto() # gemma3n
ALTUP_CORRECT_COEF = auto() # gemma3n
ALTUP_CORRECT_SCALE = auto() # gemma3n
ALTUP_PREDICT_COEF = auto() # gemma3n
ALTUP_ROUTER = auto() # gemma3n
ALTUP_ROUTER_NORM = auto() # gemma3n
LAUREL_L = auto() # gemma3n
LAUREL_R = auto() # gemma3n
LAUREL_POST_NORM = auto() # gemma3n
SSM_IN = auto()
SSM_CONV1D = auto()
SSM_X = auto()
@ -597,6 +620,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.GEMMA: "gemma",
MODEL_ARCH.GEMMA2: "gemma2",
MODEL_ARCH.GEMMA3: "gemma3",
MODEL_ARCH.GEMMA3N: "gemma3n",
MODEL_ARCH.STARCODER2: "starcoder2",
MODEL_ARCH.RWKV6: "rwkv6",
MODEL_ARCH.RWKV6QWEN2: "rwkv6qwen2",
@ -682,6 +706,22 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
MODEL_TENSOR.FFN_UP_EXP: "blk.{bid}.ffn_up_exps",
MODEL_TENSOR.FFN_EXP_PROBS_B: "blk.{bid}.exp_probs_b",
MODEL_TENSOR.LAYER_OUT_NORM: "blk.{bid}.layer_output_norm",
MODEL_TENSOR.PER_LAYER_TOKEN_EMBD: "per_layer_token_embd", # gemma3n
MODEL_TENSOR.PER_LAYER_MODEL_PROJ: "per_layer_model_proj", # gemma3n
MODEL_TENSOR.PER_LAYER_PROJ_NORM: "per_layer_proj_norm", # gemma3n
MODEL_TENSOR.ALTUP_UNEMBD_PROJ: "altup_unembd_proj", # gemma3n
MODEL_TENSOR.ALTUP_PROJ: "altup_proj", # gemma3n
MODEL_TENSOR.PER_LAYER_INP_GATE: "blk.{bid}.inp_gate", # gemma3n
MODEL_TENSOR.PER_LAYER_PROJ: "blk.{bid}.proj", # gemma3n
MODEL_TENSOR.PER_LAYER_POST_NORM: "blk.{bid}.post_norm", # gemma3n
MODEL_TENSOR.ALTUP_CORRECT_COEF: "blk.{bid}.altup_correct_coef", # gemma3n
MODEL_TENSOR.ALTUP_CORRECT_SCALE: "blk.{bid}.altup_correct_scale", # gemma3n
MODEL_TENSOR.ALTUP_PREDICT_COEF: "blk.{bid}.altup_predict_coef", # gemma3n
MODEL_TENSOR.ALTUP_ROUTER: "blk.{bid}.altup_router", # gemma3n
MODEL_TENSOR.ALTUP_ROUTER_NORM: "blk.{bid}.altup_router_norm", # gemma3n
MODEL_TENSOR.LAUREL_L: "blk.{bid}.laurel_l", # gemma3n
MODEL_TENSOR.LAUREL_R: "blk.{bid}.laurel_r", # gemma3n
MODEL_TENSOR.LAUREL_POST_NORM: "blk.{bid}.laurel_post_norm", # gemma3n
MODEL_TENSOR.SSM_IN: "blk.{bid}.ssm_in",
MODEL_TENSOR.SSM_CONV1D: "blk.{bid}.ssm_conv1d",
MODEL_TENSOR.SSM_X: "blk.{bid}.ssm_x",
@ -1486,6 +1526,41 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_PRE_NORM,
MODEL_TENSOR.FFN_POST_NORM,
],
MODEL_ARCH.GEMMA3N: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.OUTPUT_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_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,
# altup / laurel
MODEL_TENSOR.PER_LAYER_TOKEN_EMBD,
MODEL_TENSOR.PER_LAYER_MODEL_PROJ,
MODEL_TENSOR.PER_LAYER_INP_GATE,
MODEL_TENSOR.PER_LAYER_PROJ,
MODEL_TENSOR.PER_LAYER_PROJ_NORM,
MODEL_TENSOR.PER_LAYER_POST_NORM,
MODEL_TENSOR.ALTUP_PROJ,
MODEL_TENSOR.ALTUP_UNEMBD_PROJ,
MODEL_TENSOR.ALTUP_CORRECT_COEF,
MODEL_TENSOR.ALTUP_CORRECT_SCALE,
MODEL_TENSOR.ALTUP_PREDICT_COEF,
MODEL_TENSOR.ALTUP_ROUTER,
MODEL_TENSOR.ALTUP_ROUTER_NORM,
MODEL_TENSOR.LAUREL_L,
MODEL_TENSOR.LAUREL_R,
MODEL_TENSOR.LAUREL_POST_NORM,
],
MODEL_ARCH.STARCODER2: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,

18
gguf-py/gguf/gguf_writer.py
View File

@ -672,6 +672,18 @@ class GGUFWriter:
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_embedding_length_per_layer_input(self, value: int) -> None:
self.add_uint32(Keys.LLM.EMBD_LENGTH_PER_LAYER_INP.format(arch=self.arch), value)
def add_altup_active_idx(self, val: int) -> None:
self.add_uint32(Keys.LLM.ALTUP_ACTIVE_IDX.format(arch=self.arch), val)
def add_altup_num_inputs(self, val: int) -> None:
self.add_uint32(Keys.LLM.ALTUP_NUM_INPUTS.format(arch=self.arch), val)
def add_activation_sparsity_scale(self, values: Sequence[float]) -> None:
self.add_array(Keys.LLM.ACTIVATION_SPARSITY_SCALE.format(arch=self.arch), values)
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)
@ -702,6 +714,12 @@ class GGUFWriter:
def add_clamp_kqv(self, value: float) -> None:
self.add_float32(Keys.Attention.CLAMP_KQV.format(arch=self.arch), value)
def add_shared_kv_layers(self, value: float) -> None:
self.add_float32(Keys.Attention.SHARED_KV_LAYERS.format(arch=self.arch), value)
def add_sliding_window_pattern(self, value: Sequence[bool]) -> None:
self.add_array(Keys.Attention.SLIDING_WINDOW_PATTERN.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)

64
gguf-py/gguf/tensor_mapping.py
View File

@ -480,6 +480,70 @@ class TensorNameMap:
"encoder.layer.{bid}.layer_norm_2" # jina-v2-code
),
MODEL_TENSOR.PER_LAYER_TOKEN_EMBD: (
"model.embed_tokens_per_layer", # gemma3n
),
MODEL_TENSOR.PER_LAYER_MODEL_PROJ: (
"model.per_layer_model_projection", # gemma3n
),
MODEL_TENSOR.PER_LAYER_PROJ_NORM: (
"model.per_layer_projection_norm", # gemma3n
),
MODEL_TENSOR.ALTUP_PROJ: (
"model.altup_projections", # gemma3n
),
MODEL_TENSOR.ALTUP_UNEMBD_PROJ: (
"model.altup_unembed_projections", # gemma3n
),
MODEL_TENSOR.PER_LAYER_INP_GATE: (
"model.layers.{bid}.per_layer_input_gate", # gemma3n
),
MODEL_TENSOR.PER_LAYER_PROJ: (
"model.layers.{bid}.per_layer_projection", # gemma3n
),
MODEL_TENSOR.PER_LAYER_POST_NORM: (
"model.layers.{bid}.post_per_layer_input_norm", # gemma3n
),
MODEL_TENSOR.ALTUP_CORRECT_COEF: (
"model.layers.{bid}.altup.correction_coefs", # gemma3n
),
MODEL_TENSOR.ALTUP_CORRECT_SCALE: (
"model.layers.{bid}.altup.correct_output_scale", # gemma3n
),
MODEL_TENSOR.ALTUP_PREDICT_COEF: (
"model.layers.{bid}.altup.prediction_coefs", # gemma3n
),
MODEL_TENSOR.ALTUP_ROUTER: (
"model.layers.{bid}.altup.modality_router", # gemma3n
),
MODEL_TENSOR.ALTUP_ROUTER_NORM: (
"model.layers.{bid}.altup.router_norm", # gemma3n
),
MODEL_TENSOR.LAUREL_L: (
"model.layers.{bid}.laurel.linear_left", # gemma3n
),
MODEL_TENSOR.LAUREL_R: (
"model.layers.{bid}.laurel.linear_right", # gemma3n
),
MODEL_TENSOR.LAUREL_POST_NORM: (
"model.layers.{bid}.laurel.post_laurel_norm", # gemma3n
),
MODEL_TENSOR.SSM_IN: (
"model.layers.{bid}.in_proj",
"backbone.layers.{bid}.mixer.in_proj",