usermma/Quark-72M-Instruct-bucket / configuration_quark.py
usermma's picture
download
raw
1.15 kB
"""
Quark model configuration — compatibile con AutoConfig / AutoModel HuggingFace.
"""
from transformers import PretrainedConfig
class QuarkConfig(PretrainedConfig):
model_type = "quark"
def __init__(
self,
vocab_size = 65_536,
d_model = 512,
n_heads = 8,
n_kv_heads = 2,
n_layers = 14,
d_ff = 1344,
head_dim = 64,
max_seq_len = 2048,
rope_theta = 10_000.0,
rms_eps = 1e-5,
qkv_bias = True,
dropout = 0.0,
tie_word_embeddings = True,
**kwargs,
):
self.vocab_size = vocab_size
self.d_model = d_model
self.n_heads = n_heads
self.n_kv_heads = n_kv_heads
self.n_layers = n_layers
self.d_ff = d_ff
self.head_dim = head_dim
self.max_seq_len = max_seq_len
self.rope_theta = rope_theta
self.rms_eps = rms_eps
self.qkv_bias = qkv_bias
self.dropout = dropout
super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)

Xet Storage Details

Size:
1.15 kB
·
Xet hash:
9c5293018c76a26cdbefc2f4de2827d143f029f59e5c6abc50703055ffabf67a

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.