[New Model] Laguna XS.2 implementation (#41129)

Signed-off-by: Joe Rowell <joerowell4@gmail.com>
Signed-off-by: Robert Shaw <robertgshaw2@gmail.com>
Co-authored-by: Robert Shaw <robertgshaw2@gmail.com>
This commit is contained in:
Joe Rowell
2026-04-28 14:23:00 -04:00
committed by GitHub
co-authored by Robert Shaw
parent 358a755e43
commit 0899f436aa
10 changed files with 1701 additions and 16 deletions
+28 -16
View File
@@ -127,6 +127,7 @@ _CONFIG_REGISTRY: dict[str, type[PretrainedConfig]] = LazyConfigDict(
qwen3_next="Qwen3NextConfig",
qwen3_5="Qwen3_5Config",
qwen3_5_moe="Qwen3_5MoeConfig",
laguna="LagunaConfig",
lfm2_moe="Lfm2MoeConfig",
tarsier2="Tarsier2Config",
)
@@ -409,22 +410,33 @@ def patch_rope_parameters(config: PretrainedConfig) -> None:
ompe = getattr(config, "original_max_position_embeddings", None)
if Version(version("transformers")) < Version("5.0.0"):
# Transformers v4 installed, legacy config fields may be present
if (rope_scaling := getattr(config, "rope_scaling", None)) is not None:
config.rope_parameters = rope_scaling
if (
rope_theta is not None
or partial_rotary_factor is not None
or ompe is not None
) and not getattr(config, "rope_parameters", None):
config.rope_parameters = {"rope_type": "default"}
# Patch legacy fields into rope_parameters
if rope_theta is not None:
config.rope_parameters["rope_theta"] = rope_theta
if partial_rotary_factor is not None:
config.rope_parameters["partial_rotary_factor"] = partial_rotary_factor
if ompe is not None:
config.rope_parameters["original_max_position_embeddings"] = ompe
# Transformers v4 installed, legacy config fields may be present.
existing_rp = getattr(config, "rope_parameters", None)
if isinstance(existing_rp, dict) and is_rope_parameters_nested(existing_rp):
# Interleaved-attention models (e.g. Laguna-XS.2) ship a nested
# {layer_type: {...}} rope_parameters that the model code indexes
# by layer_type. The per-layer-type sub-dicts already carry the
# correct rope_theta / partial_rotary_factor / ompe (the converter
# places top-level legacy fields inside full_attention), so don't
# merge top-level fields here — that would shadow the per-type
# values and break sliding-attention layers.
pass
else:
if (rope_scaling := getattr(config, "rope_scaling", None)) is not None:
config.rope_parameters = rope_scaling
if (
rope_theta is not None
or partial_rotary_factor is not None
or ompe is not None
) and not getattr(config, "rope_parameters", None):
config.rope_parameters = {"rope_type": "default"}
# Patch legacy fields into rope_parameters
if rope_theta is not None:
config.rope_parameters["rope_theta"] = rope_theta
if partial_rotary_factor is not None:
config.rope_parameters["partial_rotary_factor"] = partial_rotary_factor
if ompe is not None:
config.rope_parameters["original_max_position_embeddings"] = ompe
elif rope_theta is not None or getattr(config, "rope_parameters", None):
# Transformers v5 installed
# Patch these fields in case they used non-standard names
@@ -45,6 +45,7 @@ _CLASS_TO_MODULE: dict[str, str] = {
# `FalconConfig` class from the official HuggingFace transformers library.
"RWConfig": "vllm.transformers_utils.configs.falcon",
"JAISConfig": "vllm.transformers_utils.configs.jais",
"LagunaConfig": "vllm.transformers_utils.configs.laguna",
"Lfm2MoeConfig": "vllm.transformers_utils.configs.lfm2_moe",
"MedusaConfig": "vllm.transformers_utils.configs.medusa",
"MiDashengLMConfig": "vllm.transformers_utils.configs.midashenglm",
@@ -105,6 +106,7 @@ __all__ = [
"IsaacConfig",
"RWConfig",
"JAISConfig",
"LagunaConfig",
"Lfm2MoeConfig",
"MedusaConfig",
"MiDashengLMConfig",
+120
View File
@@ -0,0 +1,120 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from transformers.configuration_utils import PretrainedConfig
class LagunaConfig(PretrainedConfig):
model_type = "laguna"
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.g_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: int = 100352,
hidden_size: int = 2048,
intermediate_size: int = 8192,
num_hidden_layers: int = 40,
num_attention_heads: int = 48,
num_key_value_heads: int = 8,
head_dim: int = 128,
qkv_bias: bool = False,
attention_bias: bool = False,
gating: bool | str = True,
hidden_act: str = "silu",
max_position_embeddings: int = 131072,
initializer_range: float = 0.02,
rms_norm_eps: float = 1e-6,
use_cache: bool = True,
tie_word_embeddings: bool = False,
rope_theta: float = 500000.0,
rope_scaling: dict | None = None,
rope_parameters: dict | None = None,
partial_rotary_factor: float = 1.0,
attention_dropout: float = 0.0,
sliding_window: int | None = None,
layer_types: list[str] | None = None,
swa_attention_sink_enabled: bool = False,
swa_rope_parameters: dict | None = None,
num_attention_heads_per_layer: list[int] | None = None,
num_experts: int = 256,
num_experts_per_tok: int = 8,
moe_intermediate_size: int = 512,
shared_expert_intermediate_size: int = 512,
norm_topk_prob: bool = True,
decoder_sparse_step: int = 1,
mlp_only_layers: list[int] | None = None,
router_aux_loss_coef: float = 0.001,
output_router_logits: bool = False,
moe_routed_scaling_factor: float = 1.0,
moe_apply_router_weight_on_input: bool = False,
**kwargs,
):
if mlp_only_layers is None:
mlp_only_layers = [0]
# Accept either v4-style (rope_theta + rope_scaling) or v5-style
# (rope_parameters). Translate v5 → v4 so downstream code has one path.
if rope_parameters is not None:
rp = dict(rope_parameters)
rope_theta = float(rp.pop("rope_theta", rope_theta))
rt = rp.pop("rope_type", None)
if rt is not None and rt != "default":
rope_scaling = {"rope_type": rt, **rp}
elif rp and rope_scaling is None:
rope_scaling = {"rope_type": "default", **rp}
self.vocab_size = vocab_size
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.num_key_value_heads = num_key_value_heads
self.head_dim = head_dim
self.qkv_bias = qkv_bias
self.attention_bias = attention_bias
self.gating = gating
self.hidden_act = hidden_act
self.max_position_embeddings = max_position_embeddings
self.initializer_range = initializer_range
self.rms_norm_eps = rms_norm_eps
self.use_cache = use_cache
self.rope_theta = rope_theta
self.rope_scaling = rope_scaling
self.partial_rotary_factor = partial_rotary_factor
self.attention_dropout = attention_dropout
self.sliding_window = sliding_window
self.layer_types = layer_types
self.swa_attention_sink_enabled = swa_attention_sink_enabled
self.swa_rope_parameters = swa_rope_parameters
self.num_attention_heads_per_layer = num_attention_heads_per_layer
self.num_experts = num_experts
self.num_experts_per_tok = num_experts_per_tok
self.moe_intermediate_size = moe_intermediate_size
self.shared_expert_intermediate_size = shared_expert_intermediate_size
self.norm_topk_prob = norm_topk_prob
self.decoder_sparse_step = decoder_sparse_step
self.mlp_only_layers = mlp_only_layers
self.router_aux_loss_coef = router_aux_loss_coef
self.output_router_logits = output_router_logits
self.moe_routed_scaling_factor = moe_routed_scaling_factor
self.moe_apply_router_weight_on_input = moe_apply_router_weight_on_input
super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)
__all__ = ["LagunaConfig"]