[Model] Support `ArcticForCausalLM` architecture (Snowflake/snowflake-arctic-instruct) (#5078)

Co-authored-by: vincent-4 <vincentzhongy+githubvincent4@gmail.com>

docs/supported_models/generative_models.md

@@ -28,6 +28,7 @@ python3 -m sglang.launch_server \
 | **Command-R** (Cohere)              | `CohereForAI/c4ai-command-r-v01`                 | Cohere’s open conversational LLM (Command series) optimized for long context, retrieval-augmented generation, and tool use. |
 | **DBRX** (Databricks)              | `databricks/dbrx-instruct`                       | Databricks’ 132B-parameter MoE model (36B active) trained on 12T tokens; competes with GPT-3.5 quality as a fully open foundation model. |
 | **Grok** (xAI)                     | `xai-org/grok-1`                                | xAI’s grok-1 model known for vast size(314B parameters) and high quality; integrated in SGLang for high-performance inference. |
+| **Arctic** (Snowflake)               | `Snowflake/snowflake-arctic-instruct`            | Snowflake’s dense-MoE model (17B active, 480B total) with top-2 routing, built for enterprise-grade reasoning, code, and instruction tasks. |
 | **ChatGLM** (GLM-130B family)       | `THUDM/chatglm2-6b`                              | Zhipu AI’s bilingual chat model (6B) excelling at Chinese-English dialogue; fine-tuned for conversational quality and alignment. |
 | **InternLM 2** (7B, 20B)           | `internlm/internlm2-7b`                          | Next-gen InternLM (7B and 20B) from SenseTime, offering strong reasoning and ultra-long context support (up to 200K tokens). |
 | **ExaONE 3** (Korean-English)      | `LGAI-EXAONE/EXAONE-3.5-7.8B-Instruct`           | LG AI Research’s Korean-English model (7.8B) trained on 8T tokens; provides high-quality bilingual understanding and generation. |

python/sglang/srt/configs/__init__.py

+from sglang.srt.configs.arctic import ArcticConfig
 from sglang.srt.configs.chatglm import ChatGLMConfig
 from sglang.srt.configs.dbrx import DbrxConfig
 from sglang.srt.configs.deepseekvl2 import DeepseekVL2Config
@@ -5,6 +6,7 @@ from sglang.srt.configs.exaone import ExaoneConfig
 from sglang.srt.configs.janus_pro import MultiModalityConfig
 
 __all__ = [
+    "ArcticConfig",
     "ExaoneConfig",
     "ChatGLMConfig",
     "DbrxConfig",

python/sglang/srt/configs/arctic.py

+# SPDX-License-Identifier: Apache-2.0
+
+"""Arctic model configuration"""
+
+from typing import Any, Dict, Optional
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+ARCTIC_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "arctic": "https://huggingface.co/Snowflake/snowflake-arctic-instruct/tree/main/config.json",
+}
+
+
+class ArcticConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`ArcticModel`]. It is used to instantiate an
+    Arctic model according to the specified arguments, defining the model architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 32000):
+            Vocabulary size of the Arctic model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`ArcticModel`]
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 14336):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_key_value_heads (`int`, *optional*, defaults to 8):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 4096):
+            The maximum sequence length that this model might ever be used with.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-05):
+            The epsilon used by the rms normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models).
+        pad_token_id (`int`, *optional*):
+            The id of the padding token.
+        bos_token_id (`int`, *optional*, defaults to 1):
+            The id of the "beginning-of-sequence" token.
+        eos_token_id (`int`, *optional*, defaults to 2):
+            The id of the "end-of-sequence" token.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied.
+        rope_theta (`float`, *optional*, defaults to 1000000.0):
+            The base period of the RoPE embeddings.
+        sliding_window (`int`, *optional*):
+            Sliding window attention window size. If not specified, will default to `4096`.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        num_experts_per_tok (`int`, *optional*, defaults to 2):
+            The number of experts to root per-token, can be also interpreted as the `top-p` routing parameter
+        num_local_experts (`int`, *optional*, defaults to 8):
+            Number of experts per Sparse MLP layer.
+        moe_layer_frequency (`int`, *optional*, defaults to 2):
+            Frequency of MoE layers in the model.
+    """
+
+    model_type = "arctic"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=32000,
+        hidden_size=4096,
+        intermediate_size=14336,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=8,
+        hidden_act="silu",
+        max_position_embeddings=4096,
+        initializer_range=0.02,
+        rms_norm_eps=1e-5,
+        use_cache=True,
+        pad_token_id=None,
+        bos_token_id=1,
+        eos_token_id=2,
+        tie_word_embeddings=False,
+        rope_theta=1e6,
+        sliding_window=None,
+        attention_dropout=0.0,
+        num_experts_per_tok=1,
+        num_local_experts=8,
+        moe_layer_frequency=2,
+        **kwargs,
+    ):
+        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.sliding_window = sliding_window
+        self.num_key_value_heads = num_key_value_heads
+        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_dropout = attention_dropout
+        self.num_experts_per_tok = num_experts_per_tok
+        self.num_local_experts = num_local_experts
+        self.moe_layer_frequency = moe_layer_frequency
+
+        # For backward compatibility
+        self._attn_implementation = kwargs.pop("_attn_implementation", "eager")
+        self.use_residual = kwargs.pop("use_residual", True)
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

python/sglang/srt/hf_transformers_utils.py

@@ -31,6 +31,7 @@ from transformers import (
 from transformers.models.auto.modeling_auto import MODEL_FOR_CAUSAL_LM_MAPPING_NAMES
 
 from sglang.srt.configs import (
+    ArcticConfig,
     ChatGLMConfig,
     DbrxConfig,
     DeepseekVL2Config,
@@ -41,6 +42,7 @@ from sglang.srt.connector import create_remote_connector
 from sglang.srt.utils import is_remote_url
 
 _CONFIG_REGISTRY: Dict[str, Type[PretrainedConfig]] = {
+    ArcticConfig.model_type: ArcticConfig,
     ChatGLMConfig.model_type: ChatGLMConfig,
     DbrxConfig.model_type: DbrxConfig,
     ExaoneConfig.model_type: ExaoneConfig,