161 lines
6.4 KiB
Python
161 lines
6.4 KiB
Python
# coding=utf-8
|
|
# Configuration class for Kanana-2 PD-series (Qwen3 architecture with
|
|
# sliding/full alternating attention and per-attention-type RoPE).
|
|
#
|
|
# Difference vs Qwen3:
|
|
# * `rope_parameters` is a dict keyed by attention type (`full_attention` /
|
|
# `sliding_attention`). Each entry is a self-contained RoPE config
|
|
# understood by `transformers.modeling_rope_utils.ROPE_INIT_FUNCTIONS`.
|
|
# This lets us apply YaRN to global-attention layers while keeping
|
|
# unscaled RoPE for sliding-attention layers.
|
|
# * Top-level `rope_scaling` is unused on this config; the modeling code
|
|
# builds per-attention-type sub-configs at construction time and sets
|
|
# `rope_scaling` on each sub-config so HF's standard rope init functions
|
|
# (which read `config.rope_scaling`) work unchanged.
|
|
|
|
from transformers.configuration_utils import PretrainedConfig, layer_type_validation
|
|
from transformers.modeling_rope_utils import rope_config_validation
|
|
from transformers.utils import logging
|
|
|
|
|
|
logger = logging.get_logger(__name__)
|
|
|
|
|
|
class Kanana2TinyConfig(PretrainedConfig):
|
|
"""Configuration for the Kanana-2 PD-series (Qwen3 + per-type RoPE)."""
|
|
|
|
model_type = "kanana2_tiny"
|
|
keys_to_ignore_at_inference = ["past_key_values"]
|
|
|
|
base_model_tp_plan = {
|
|
"layers.*.self_attn.q_proj": "colwise",
|
|
"layers.*.self_attn.k_proj": "colwise",
|
|
"layers.*.self_attn.v_proj": "colwise",
|
|
"layers.*.self_attn.o_proj": "rowwise",
|
|
"layers.*.mlp.gate_proj": "colwise",
|
|
"layers.*.mlp.up_proj": "colwise",
|
|
"layers.*.mlp.down_proj": "rowwise",
|
|
}
|
|
base_model_pp_plan = {
|
|
"embed_tokens": (["input_ids"], ["inputs_embeds"]),
|
|
"layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
|
|
"norm": (["hidden_states"], ["hidden_states"]),
|
|
}
|
|
|
|
def __init__(
|
|
self,
|
|
vocab_size=128256,
|
|
hidden_size=1024,
|
|
intermediate_size=4608,
|
|
num_hidden_layers=32,
|
|
num_attention_heads=32,
|
|
num_key_value_heads=8,
|
|
head_dim=128,
|
|
hidden_act="silu",
|
|
max_position_embeddings=35000,
|
|
initializer_range=0.02,
|
|
rms_norm_eps=1e-6,
|
|
use_cache=True,
|
|
tie_word_embeddings=True,
|
|
rope_theta=10000.0,
|
|
rope_parameters=None,
|
|
rope_scaling=None,
|
|
attention_bias=False,
|
|
use_sliding_window=True,
|
|
sliding_window=1024,
|
|
max_window_layers=32,
|
|
layer_types=None,
|
|
attention_dropout=0.0,
|
|
**kwargs,
|
|
):
|
|
# Standard Qwen3-ish fields
|
|
self.vocab_size = vocab_size
|
|
self.max_position_embeddings = max_position_embeddings
|
|
self.hidden_size = hidden_size
|
|
self.intermediate_size = intermediate_size
|
|
self.num_hidden_layers = num_hidden_layers
|
|
self.num_attention_heads = num_attention_heads
|
|
self.use_sliding_window = use_sliding_window
|
|
self.sliding_window = sliding_window if self.use_sliding_window else None
|
|
self.max_window_layers = max_window_layers
|
|
|
|
if num_key_value_heads is None:
|
|
num_key_value_heads = num_attention_heads
|
|
self.num_key_value_heads = num_key_value_heads
|
|
self.head_dim = head_dim
|
|
self.hidden_act = hidden_act
|
|
self.initializer_range = initializer_range
|
|
self.rms_norm_eps = rms_norm_eps
|
|
self.use_cache = use_cache
|
|
self.rope_theta = rope_theta
|
|
self.attention_bias = attention_bias
|
|
self.attention_dropout = attention_dropout
|
|
# Kept for HF helpers that probe the attribute. The per-attention RoPE
|
|
# config lives in `rope_parameters`; modeling code constructs sub-configs
|
|
# whose `rope_scaling` is the per-type dict at init time.
|
|
self.rope_scaling = rope_scaling
|
|
|
|
# Per-attention-type RoPE.
|
|
# Expected shape (defaults match the kanana-2-pd-series checkpoints):
|
|
# {
|
|
# "full_attention": {"rope_type": "yarn", "rope_theta": 10000,
|
|
# "factor": 40.0, "original_max_position_embeddings": 4096},
|
|
# "sliding_attention": {"rope_type": "default", "rope_theta": 10000.0},
|
|
# }
|
|
if rope_parameters is None:
|
|
rope_parameters = {
|
|
"full_attention": {
|
|
"rope_type": "default",
|
|
"rope_theta": rope_theta,
|
|
},
|
|
"sliding_attention": {
|
|
"rope_type": "default",
|
|
"rope_theta": rope_theta,
|
|
},
|
|
}
|
|
self.rope_parameters = rope_parameters
|
|
|
|
for attn_type, params in self.rope_parameters.items():
|
|
if not isinstance(params, dict) or "rope_type" not in params:
|
|
raise ValueError(
|
|
f"rope_parameters[{attn_type!r}] must be a dict with a 'rope_type' key, got {params!r}"
|
|
)
|
|
|
|
# Set layer_types BEFORE per-type rope validation: the layer types must
|
|
# exist for the validators that gate on layer_types.
|
|
self.layer_types = layer_types
|
|
if self.layer_types is None:
|
|
self.layer_types = [
|
|
"sliding_attention"
|
|
if self.sliding_window is not None and i >= self.max_window_layers
|
|
else "full_attention"
|
|
for i in range(self.num_hidden_layers)
|
|
]
|
|
layer_type_validation(self.layer_types, self.num_hidden_layers)
|
|
|
|
# Per-attention-type rope validation. The 4.57.1 validators read off
|
|
# `config.rope_scaling` (flat dict) and `config.rope_theta` (top-level),
|
|
# so for each per-type sub-dict we present it in that shape, run the
|
|
# validator, then restore. `rope_theta` is filtered out of the temporary
|
|
# `rope_scaling` because in 4.57.1's schema it lives at the top level.
|
|
for attn_type, params in self.rope_parameters.items():
|
|
if attn_type not in set(self.layer_types):
|
|
continue
|
|
saved_rope_scaling = self.rope_scaling
|
|
saved_rope_theta = self.rope_theta
|
|
try:
|
|
self.rope_scaling = {k: v for k, v in params.items() if k != "rope_theta"}
|
|
self.rope_theta = params.get("rope_theta", saved_rope_theta)
|
|
rope_config_validation(self)
|
|
finally:
|
|
self.rope_scaling = saved_rope_scaling
|
|
self.rope_theta = saved_rope_theta
|
|
|
|
super().__init__(
|
|
tie_word_embeddings=tie_word_embeddings,
|
|
**kwargs,
|
|
)
|
|
|
|
|
|
__all__ = ["Kanana2TinyConfig"]
|