vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

llama : initial Mamba-2 support

Browse source
This commit is contained in:
Francis Couture-Harpin 2024-08-01 10:43:42 -04:00
parent a1631e53f6
commit 1f0fea70fb
7 changed files with 490 additions and 82 deletions
Download patch file Download diff file Expand all files Collapse all files

19
gguf-py/gguf/constants.py
View file

@ -130,6 +130,7 @@ class Keys:
INNER_SIZE = "{arch}.ssm.inner_size"
STATE_SIZE = "{arch}.ssm.state_size"
TIME_STEP_RANK = "{arch}.ssm.time_step_rank"
GROUP_COUNT = "{arch}.ssm.group_count"
DT_B_C_RMS = "{arch}.ssm.dt_b_c_rms"
class Tokenizer:
@ -208,6 +209,7 @@ class MODEL_ARCH(IntEnum):
GEMMA2 = auto()
STARCODER2 = auto()
MAMBA = auto()
MAMBA2 = auto()
XVERSE = auto()
COMMAND_R = auto()
DBRX = auto()
@ -269,6 +271,7 @@ class MODEL_TENSOR(IntEnum):
SSM_DT = auto()
SSM_A = auto()
SSM_D = auto()
SSM_NORM = auto()
SSM_OUT = auto()
ATTN_Q_A = auto()
ATTN_Q_B = auto()
@ -338,6 +341,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.GEMMA2: "gemma2",
MODEL_ARCH.STARCODER2: "starcoder2",
MODEL_ARCH.MAMBA: "mamba",
MODEL_ARCH.MAMBA2: "mamba2",
MODEL_ARCH.XVERSE: "xverse",
MODEL_ARCH.COMMAND_R: "command-r",
MODEL_ARCH.DBRX: "dbrx",
@ -399,6 +403,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
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_NORM: "blk.{bid}.ssm_norm",
MODEL_TENSOR.SSM_OUT: "blk.{bid}.ssm_out",
MODEL_TENSOR.ATTN_Q_A: "blk.{bid}.attn_q_a",
MODEL_TENSOR.ATTN_Q_B: "blk.{bid}.attn_q_b",
@ -869,6 +874,19 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.SSM_D,
MODEL_TENSOR.SSM_OUT,
],
MODEL_ARCH.MAMBA2: [
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_DT,
MODEL_TENSOR.SSM_A,
MODEL_TENSOR.SSM_D,
MODEL_TENSOR.SSM_NORM,
MODEL_TENSOR.SSM_OUT,
],
MODEL_ARCH.XVERSE: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
@ -1373,6 +1391,7 @@ 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_GROUP_COUNT = Keys.SSM.GROUP_COUNT
KEY_SSM_DT_B_C_RMS = Keys.SSM.DT_B_C_RMS
# tokenization

3
gguf-py/gguf/gguf_writer.py
View file

@ -730,6 +730,9 @@ class GGUFWriter:
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_group_count(self, value: int) -> None:
self.add_uint32(Keys.SSM.GROUP_COUNT.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)

6
gguf-py/gguf/tensor_mapping.py
View file

@ -396,7 +396,7 @@ class TensorNameMap:
"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
"encoder.layer.{bid}.layer_norm_2", # jina-v2-code
),
MODEL_TENSOR.SSM_IN: (
@ -429,6 +429,10 @@ class TensorNameMap:
"backbone.layers.{bid}.mixer.D",
),
MODEL_TENSOR.SSM_NORM: (
"backbone.layers.{bid}.mixer.norm", # mamba2
),
MODEL_TENSOR.SSM_OUT: (
"model.layers.{bid}.out_proj",
"backbone.layers.{bid}.mixer.out_proj",