Instructions to use nota-ai/Solar-Open-100B-NotaMoEQuant-NVFP4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use nota-ai/Solar-Open-100B-NotaMoEQuant-NVFP4 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="nota-ai/Solar-Open-100B-NotaMoEQuant-NVFP4", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("nota-ai/Solar-Open-100B-NotaMoEQuant-NVFP4", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("nota-ai/Solar-Open-100B-NotaMoEQuant-NVFP4", trust_remote_code=True, device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use nota-ai/Solar-Open-100B-NotaMoEQuant-NVFP4 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "nota-ai/Solar-Open-100B-NotaMoEQuant-NVFP4"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "nota-ai/Solar-Open-100B-NotaMoEQuant-NVFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/nota-ai/Solar-Open-100B-NotaMoEQuant-NVFP4

SGLang

How to use nota-ai/Solar-Open-100B-NotaMoEQuant-NVFP4 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "nota-ai/Solar-Open-100B-NotaMoEQuant-NVFP4" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "nota-ai/Solar-Open-100B-NotaMoEQuant-NVFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "nota-ai/Solar-Open-100B-NotaMoEQuant-NVFP4" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "nota-ai/Solar-Open-100B-NotaMoEQuant-NVFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use nota-ai/Solar-Open-100B-NotaMoEQuant-NVFP4 with Docker Model Runner:
```
docker model run hf.co/nota-ai/Solar-Open-100B-NotaMoEQuant-NVFP4
```

hancheolp commited on Mar 11

Commit

812ca4f

verified ·

1 Parent(s): 5969c57

Add patches

Browse files

Files changed (2) hide show

patches/registry.py +1304 -0
patches/solar_open.py +693 -0

patches/registry.py ADDED Viewed

	@@ -0,0 +1,1304 @@

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Whenever you add an architecture to this page, please also update
+`tests/models/registry.py` with example HuggingFace models for it.
+"""
+import importlib
+import json
+import os
+import pickle
+import subprocess
+import sys
+import tempfile
+from abc import ABC, abstractmethod
+from collections.abc import Callable, Set
+from dataclasses import asdict, dataclass, field
+from functools import lru_cache
+from pathlib import Path
+from typing import TYPE_CHECKING, Any, TypeVar
+import torch.nn as nn
+import transformers
+from vllm import envs
+from vllm.config import (
+    ModelConfig,
+    iter_architecture_defaults,
+    try_match_architecture_defaults,
+)
+from vllm.logger import init_logger
+from vllm.logging_utils import logtime
+from vllm.transformers_utils.dynamic_module import try_get_class_from_dynamic_module
+from vllm.utils.hashing import safe_hash
+if TYPE_CHECKING:
+    from vllm.config.model import AttnTypeStr
+    from vllm.config.pooler import SequencePoolingType, TokenPoolingType
+else:
+    AttnTypeStr = Any
+    SequencePoolingType = Any
+    TokenPoolingType = Any
+from .interfaces import (
+    has_inner_state,
+    has_noops,
+    is_attention_free,
+    is_hybrid,
+    requires_raw_input_tokens,
+    supports_cross_encoding,
+    supports_late_interaction,
+    supports_mamba_prefix_caching,
+    supports_multimodal,
+    supports_multimodal_encoder_tp_data,
+    supports_multimodal_raw_input_only,
+    supports_pp,
+    supports_transcription,
+)
+from .interfaces_base import (
+    get_attn_type,
+    get_default_seq_pooling_type,
+    get_default_tok_pooling_type,
+    is_pooling_model,
+    is_text_generation_model,
+)
+logger = init_logger(__name__)
+_TEXT_GENERATION_MODELS = {
+    # [Decoder-only]
+    "AfmoeForCausalLM": ("afmoe", "AfmoeForCausalLM"),
+    "ApertusForCausalLM": ("apertus", "ApertusForCausalLM"),
+    "AquilaModel": ("llama", "LlamaForCausalLM"),
+    "AquilaForCausalLM": ("llama", "LlamaForCausalLM"),  # AquilaChat2
+    "ArceeForCausalLM": ("arcee", "ArceeForCausalLM"),
+    "ArcticForCausalLM": ("arctic", "ArcticForCausalLM"),
+    "AXK1ForCausalLM": ("AXK1", "AXK1ForCausalLM"),
+    # baichuan-7b, upper case 'C' in the class name
+    "BaiChuanForCausalLM": ("baichuan", "BaiChuanForCausalLM"),
+    # baichuan-13b, lower case 'c' in the class name
+    "BaichuanForCausalLM": ("baichuan", "BaichuanForCausalLM"),
+    "BailingMoeForCausalLM": ("bailing_moe", "BailingMoeForCausalLM"),
+    "BailingMoeV2ForCausalLM": ("bailing_moe", "BailingMoeV2ForCausalLM"),
+    "BailingMoeV2_5ForCausalLM": ("bailing_moe_linear", "BailingMoeV25ForCausalLM"),
+    "BambaForCausalLM": ("bamba", "BambaForCausalLM"),
+    "BloomForCausalLM": ("bloom", "BloomForCausalLM"),
+    "ChatGLMModel": ("chatglm", "ChatGLMForCausalLM"),
+    "ChatGLMForConditionalGeneration": ("chatglm", "ChatGLMForCausalLM"),
+    "CohereForCausalLM": ("commandr", "CohereForCausalLM"),
+    "Cohere2ForCausalLM": ("commandr", "CohereForCausalLM"),
+    "CwmForCausalLM": ("llama", "LlamaForCausalLM"),
+    "DbrxForCausalLM": ("dbrx", "DbrxForCausalLM"),
+    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
+    "DeepseekForCausalLM": ("deepseek_v2", "DeepseekForCausalLM"),
+    "DeepseekV2ForCausalLM": ("deepseek_v2", "DeepseekV2ForCausalLM"),
+    "DeepseekV3ForCausalLM": ("deepseek_v2", "DeepseekV3ForCausalLM"),
+    "DeepseekV32ForCausalLM": ("deepseek_v2", "DeepseekV3ForCausalLM"),
+    "Dots1ForCausalLM": ("dots1", "Dots1ForCausalLM"),
+    "Ernie4_5ForCausalLM": ("ernie45", "Ernie4_5ForCausalLM"),
+    "Ernie4_5_MoeForCausalLM": ("ernie45_moe", "Ernie4_5_MoeForCausalLM"),
+    "ExaoneForCausalLM": ("exaone", "ExaoneForCausalLM"),
+    "Exaone4ForCausalLM": ("exaone4", "Exaone4ForCausalLM"),
+    "ExaoneMoEForCausalLM": ("exaone_moe", "ExaoneMoeForCausalLM"),
+    "Fairseq2LlamaForCausalLM": ("fairseq2_llama", "Fairseq2LlamaForCausalLM"),
+    "FalconForCausalLM": ("falcon", "FalconForCausalLM"),
+    "FalconMambaForCausalLM": ("mamba", "MambaForCausalLM"),
+    "FalconH1ForCausalLM": ("falcon_h1", "FalconH1ForCausalLM"),
+    "FlexOlmoForCausalLM": ("flex_olmo", "FlexOlmoForCausalLM"),
+    "GemmaForCausalLM": ("gemma", "GemmaForCausalLM"),
+    "Gemma2ForCausalLM": ("gemma2", "Gemma2ForCausalLM"),
+    "Gemma3ForCausalLM": ("gemma3", "Gemma3ForCausalLM"),
+    "Gemma3nForCausalLM": ("gemma3n", "Gemma3nForCausalLM"),
+    "Qwen3NextForCausalLM": ("qwen3_next", "Qwen3NextForCausalLM"),
+    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
+    "Glm4ForCausalLM": ("glm4", "Glm4ForCausalLM"),
+    "Glm4MoeForCausalLM": ("glm4_moe", "Glm4MoeForCausalLM"),
+    "Glm4MoeLiteForCausalLM": ("glm4_moe_lite", "Glm4MoeLiteForCausalLM"),
+    "GlmMoeDsaForCausalLM": ("deepseek_v2", "GlmMoeDsaForCausalLM"),
+    "GptOssForCausalLM": ("gpt_oss", "GptOssForCausalLM"),
+    "GPT2LMHeadModel": ("gpt2", "GPT2LMHeadModel"),
+    "GPTBigCodeForCausalLM": ("gpt_bigcode", "GPTBigCodeForCausalLM"),
+    "GPTJForCausalLM": ("gpt_j", "GPTJForCausalLM"),
+    "GPTNeoXForCausalLM": ("gpt_neox", "GPTNeoXForCausalLM"),
+    "GraniteForCausalLM": ("granite", "GraniteForCausalLM"),
+    "GraniteMoeForCausalLM": ("granitemoe", "GraniteMoeForCausalLM"),
+    "GraniteMoeHybridForCausalLM": ("granitemoehybrid", "GraniteMoeHybridForCausalLM"),  # noqa: E501
+    "GraniteMoeSharedForCausalLM": ("granitemoeshared", "GraniteMoeSharedForCausalLM"),  # noqa: E501
+    "GritLM": ("gritlm", "GritLM"),
+    "Grok1ModelForCausalLM": ("grok1", "GrokForCausalLM"),
+    "Grok1ForCausalLM": ("grok1", "GrokForCausalLM"),
+    "HunYuanMoEV1ForCausalLM": ("hunyuan_v1", "HunYuanMoEV1ForCausalLM"),
+    "HunYuanDenseV1ForCausalLM": ("hunyuan_v1", "HunYuanDenseV1ForCausalLM"),
+    "HCXVisionForCausalLM": ("hyperclovax_vision", "HCXVisionForCausalLM"),
+    "InternLMForCausalLM": ("llama", "LlamaForCausalLM"),
+    "InternLM2ForCausalLM": ("internlm2", "InternLM2ForCausalLM"),
+    "InternLM2VEForCausalLM": ("internlm2_ve", "InternLM2VEForCausalLM"),
+    "InternLM3ForCausalLM": ("llama", "LlamaForCausalLM"),
+    "IQuestCoderForCausalLM": ("llama", "LlamaForCausalLM"),
+    "IQuestLoopCoderForCausalLM": ("iquest_loopcoder", "IQuestLoopCoderForCausalLM"),
+    "JAISLMHeadModel": ("jais", "JAISLMHeadModel"),
+    "Jais2ForCausalLM": ("jais2", "Jais2ForCausalLM"),
+    "JambaForCausalLM": ("jamba", "JambaForCausalLM"),
+    "KimiLinearForCausalLM": ("kimi_linear", "KimiLinearForCausalLM"),  # noqa: E501
+    "Lfm2ForCausalLM": ("lfm2", "Lfm2ForCausalLM"),
+    "Lfm2MoeForCausalLM": ("lfm2_moe", "Lfm2MoeForCausalLM"),
+    "LlamaForCausalLM": ("llama", "LlamaForCausalLM"),
+    "Llama4ForCausalLM": ("llama4", "Llama4ForCausalLM"),
+    # For decapoda-research/llama-*
+    "LLaMAForCausalLM": ("llama", "LlamaForCausalLM"),
+    "LongcatFlashForCausalLM": ("longcat_flash", "LongcatFlashForCausalLM"),
+    "MambaForCausalLM": ("mamba", "MambaForCausalLM"),
+    "Mamba2ForCausalLM": ("mamba2", "Mamba2ForCausalLM"),
+    "MiniCPMForCausalLM": ("minicpm", "MiniCPMForCausalLM"),
+    "MiniCPM3ForCausalLM": ("minicpm3", "MiniCPM3ForCausalLM"),
+    "MiniMaxForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
+    "MiniMaxText01ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
+    "MiniMaxM1ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
+    "MiniMaxM2ForCausalLM": ("minimax_m2", "MiniMaxM2ForCausalLM"),
+    "MistralForCausalLM": ("mistral", "MistralForCausalLM"),
+    "MistralLarge3ForCausalLM": ("mistral_large_3", "MistralLarge3ForCausalLM"),
+    "MixtralForCausalLM": ("mixtral", "MixtralForCausalLM"),
+    # transformers's mpt class has lower case
+    "MptForCausalLM": ("mpt", "MPTForCausalLM"),
+    "MPTForCausalLM": ("mpt", "MPTForCausalLM"),
+    "MiMoForCausalLM": ("mimo", "MiMoForCausalLM"),
+    "MiMoV2FlashForCausalLM": ("mimo_v2_flash", "MiMoV2FlashForCausalLM"),
+    "NemotronForCausalLM": ("nemotron", "NemotronForCausalLM"),
+    "NemotronHForCausalLM": ("nemotron_h", "NemotronHForCausalLM"),
+    "NemotronHPuzzleForCausalLM": ("nemotron_h", "NemotronHForCausalLM"),
+    "OlmoForCausalLM": ("olmo", "OlmoForCausalLM"),
+    "Olmo2ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
+    "Olmo3ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
+    "OlmoeForCausalLM": ("olmoe", "OlmoeForCausalLM"),
+    "OPTForCausalLM": ("opt", "OPTForCausalLM"),
+    "OrionForCausalLM": ("orion", "OrionForCausalLM"),
+    "OuroForCausalLM": ("ouro", "OuroForCausalLM"),
+    "PanguEmbeddedForCausalLM": ("openpangu", "PanguEmbeddedForCausalLM"),
+    "PanguProMoEV2ForCausalLM": ("openpangu", "PanguProMoEV2ForCausalLM"),
+    "PanguUltraMoEForCausalLM": ("openpangu", "PanguUltraMoEForCausalLM"),
+    "PersimmonForCausalLM": ("persimmon", "PersimmonForCausalLM"),
+    "PhiForCausalLM": ("phi", "PhiForCausalLM"),
+    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
+    "PhiMoEForCausalLM": ("phimoe", "PhiMoEForCausalLM"),
+    "Plamo2ForCausalLM": ("plamo2", "Plamo2ForCausalLM"),
+    "Plamo3ForCausalLM": ("plamo3", "Plamo3ForCausalLM"),
+    "QWenLMHeadModel": ("qwen", "QWenLMHeadModel"),
+    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
+    "Qwen2MoeForCausalLM": ("qwen2_moe", "Qwen2MoeForCausalLM"),
+    "Qwen3ForCausalLM": ("qwen3", "Qwen3ForCausalLM"),
+    "Qwen3MoeForCausalLM": ("qwen3_moe", "Qwen3MoeForCausalLM"),
+    "RWForCausalLM": ("falcon", "FalconForCausalLM"),
+    "SeedOssForCausalLM": ("seed_oss", "SeedOssForCausalLM"),
+    "Step1ForCausalLM": ("step1", "Step1ForCausalLM"),
+    "Step3TextForCausalLM": ("step3_text", "Step3TextForCausalLM"),
+    "Step3p5ForCausalLM": ("step3p5", "Step3p5ForCausalLM"),
+    "StableLMEpochForCausalLM": ("stablelm", "StablelmForCausalLM"),
+    "StableLmForCausalLM": ("stablelm", "StablelmForCausalLM"),
+    "Starcoder2ForCausalLM": ("starcoder2", "Starcoder2ForCausalLM"),
+    "SolarForCausalLM": ("solar", "SolarForCausalLM"),
+    "SolarOpenForCausalLM": ("solar_open", "SolarOpenForCausalLM"),
+    "TeleChatForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
+    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
+    "TeleFLMForCausalLM": ("teleflm", "TeleFLMForCausalLM"),
+    "XverseForCausalLM": ("llama", "LlamaForCausalLM"),
+    "Zamba2ForCausalLM": ("zamba2", "Zamba2ForCausalLM"),
+}
+_EMBEDDING_MODELS = {
+    # [Text-only]
+    "BertModel": ("bert", "BertEmbeddingModel"),
+    "BertSpladeSparseEmbeddingModel": ("bert", "BertSpladeSparseEmbeddingModel"),
+    "HF_ColBERT": ("colbert", "ColBERTModel"),
+    "ColBERTModernBertModel": ("colbert", "ColBERTModernBertModel"),
+    "ColBERTJinaRobertaModel": ("colbert", "ColBERTJinaRobertaModel"),
+    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
+    "Gemma2Model": ("gemma2", "Gemma2ForCausalLM"),
+    "Gemma3TextModel": ("gemma3", "Gemma3Model"),
+    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
+    "GPT2ForSequenceClassification": ("gpt2", "GPT2ForSequenceClassification"),
+    "GritLM": ("gritlm", "GritLM"),
+    "GteModel": ("bert_with_rope", "SnowflakeGteNewModel"),
+    "GteNewModel": ("bert_with_rope", "GteNewModel"),
+    "InternLM2ForRewardModel": ("internlm2", "InternLM2ForRewardModel"),
+    "JambaForSequenceClassification": ("jamba", "JambaForSequenceClassification"),  # noqa: E501
+    "LlamaBidirectionalModel": ("llama", "LlamaBidirectionalModel"),
+    "LlamaModel": ("llama", "LlamaForCausalLM"),
+    **{
+        # Multiple models share the same architecture, so we include them all
+        k: (mod, arch)
+        for k, (mod, arch) in _TEXT_GENERATION_MODELS.items()
+        if arch == "LlamaForCausalLM"
+    },
+    "MistralModel": ("llama", "LlamaForCausalLM"),
+    "ModernBertModel": ("modernbert", "ModernBertModel"),
+    "NomicBertModel": ("bert_with_rope", "NomicBertModel"),
+    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
+    "Qwen2Model": ("qwen2", "Qwen2ForCausalLM"),
+    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
+    "Qwen2ForRewardModel": ("qwen2_rm", "Qwen2ForRewardModel"),
+    "Qwen2ForProcessRewardModel": ("qwen2_rm", "Qwen2ForProcessRewardModel"),
+    "RobertaForMaskedLM": ("roberta", "RobertaEmbeddingModel"),
+    "RobertaModel": ("roberta", "RobertaEmbeddingModel"),
+    "TeleChatForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
+    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
+    "VoyageQwen3BidirectionalEmbedModel": (
+        "voyage",
+        "VoyageQwen3BidirectionalEmbedModel",
+    ),
+    "XLMRobertaModel": ("roberta", "RobertaEmbeddingModel"),
+    "BgeM3EmbeddingModel": ("roberta", "BgeM3EmbeddingModel"),
+    # [Multimodal]
+    "CLIPModel": ("clip", "CLIPEmbeddingModel"),
+    "ColModernVBertForRetrieval": ("colmodernvbert", "ColModernVBertForRetrieval"),
+    "LlavaNextForConditionalGeneration": (
+        "llava_next",
+        "LlavaNextForConditionalGeneration",
+    ),
+    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
+    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
+    "ColQwen3": ("colqwen3", "ColQwen3Model"),
+    "OpsColQwen3Model": ("colqwen3", "ColQwen3Model"),
+    "Qwen3VLNemotronEmbedModel": ("colqwen3", "ColQwen3Model"),
+    "SiglipModel": ("siglip", "SiglipEmbeddingModel"),
+    "LlamaNemotronVLModel": (
+        "nemotron_vl",
+        "LlamaNemotronVLForEmbedding",
+    ),
+    # Technically Terratorch models work on images, both in
+    # input and output. I am adding it here because it piggy-backs on embedding
+    # models for the time being.
+    "PrithviGeoSpatialMAE": ("terratorch", "Terratorch"),
+    "Terratorch": ("terratorch", "Terratorch"),
+}
+_CROSS_ENCODER_MODELS = {
+    "BertForSequenceClassification": ("bert", "BertForSequenceClassification"),
+    "BertForTokenClassification": ("bert", "BertForTokenClassification"),
+    "GteNewForSequenceClassification": (
+        "bert_with_rope",
+        "GteNewForSequenceClassification",
+    ),
+    "JinaVLForRanking": ("jina_vl", "JinaVLForSequenceClassification"),
+    "LlamaBidirectionalForSequenceClassification": (
+        "llama",
+        "LlamaBidirectionalForSequenceClassification",
+    ),
+    "LlamaNemotronVLForSequenceClassification": (
+        "nemotron_vl",
+        "LlamaNemotronVLForSequenceClassification",
+    ),
+    "ModernBertForSequenceClassification": (
+        "modernbert",
+        "ModernBertForSequenceClassification",
+    ),
+    "ModernBertForTokenClassification": (
+        "modernbert",
+        "ModernBertForTokenClassification",
+    ),
+    "RobertaForSequenceClassification": ("roberta", "RobertaForSequenceClassification"),
+    "XLMRobertaForSequenceClassification": (
+        "roberta",
+        "RobertaForSequenceClassification",
+    ),
+}
+_MULTIMODAL_MODELS = {
+    # [Decoder-only]
+    "AriaForConditionalGeneration": ("aria", "AriaForConditionalGeneration"),
+    "AudioFlamingo3ForConditionalGeneration": (
+        "audioflamingo3",
+        "AudioFlamingo3ForConditionalGeneration",
+    ),
+    "MusicFlamingoForConditionalGeneration": (
+        "musicflamingo",
+        "MusicFlamingoForConditionalGeneration",
+    ),
+    "AyaVisionForConditionalGeneration": (
+        "aya_vision",
+        "AyaVisionForConditionalGeneration",
+    ),
+    "BagelForConditionalGeneration": ("bagel", "BagelForConditionalGeneration"),
+    "BeeForConditionalGeneration": ("bee", "BeeForConditionalGeneration"),
+    "Blip2ForConditionalGeneration": ("blip2", "Blip2ForConditionalGeneration"),
+    "ChameleonForConditionalGeneration": (
+        "chameleon",
+        "ChameleonForConditionalGeneration",
+    ),
+    "Cohere2VisionForConditionalGeneration": (
+        "cohere2_vision",
+        "Cohere2VisionForConditionalGeneration",
+    ),
+    "DeepseekVLV2ForCausalLM": ("deepseek_vl2", "DeepseekVLV2ForCausalLM"),
+    "DeepseekOCRForCausalLM": ("deepseek_ocr", "DeepseekOCRForCausalLM"),
+    "DeepseekOCR2ForCausalLM": ("deepseek_ocr2", "DeepseekOCR2ForCausalLM"),
+    "DotsOCRForCausalLM": ("dots_ocr", "DotsOCRForCausalLM"),
+    "Eagle2_5_VLForConditionalGeneration": (
+        "eagle2_5_vl",
+        "Eagle2_5_VLForConditionalGeneration",
+    ),
+    "Ernie4_5_VLMoeForConditionalGeneration": (
+        "ernie45_vl",
+        "Ernie4_5_VLMoeForConditionalGeneration",
+    ),
+    "FireRedASR2ForConditionalGeneration": (
+        "fireredasr2",
+        "FireRedASR2ForConditionalGeneration",
+    ),
+    "FunASRForConditionalGeneration": ("funasr", "FunASRForConditionalGeneration"),  # noqa: E501
+    "FunAudioChatForConditionalGeneration": (
+        "funaudiochat",
+        "FunAudioChatForConditionalGeneration",
+    ),
+    "FuyuForCausalLM": ("fuyu", "FuyuForCausalLM"),
+    "Gemma3ForConditionalGeneration": ("gemma3_mm", "Gemma3ForConditionalGeneration"),  # noqa: E501
+    "Gemma3nForConditionalGeneration": (
+        "gemma3n_mm",
+        "Gemma3nForConditionalGeneration",
+    ),
+    "GlmAsrForConditionalGeneration": ("glmasr", "GlmAsrForConditionalGeneration"),
+    "GLM4VForCausalLM": ("glm4v", "GLM4VForCausalLM"),
+    "Glm4vForConditionalGeneration": ("glm4_1v", "Glm4vForConditionalGeneration"),
+    "Glm4vMoeForConditionalGeneration": ("glm4_1v", "Glm4vMoeForConditionalGeneration"),
+    "GlmOcrForConditionalGeneration": ("glm_ocr", "GlmOcrForConditionalGeneration"),  # noqa: E501
+    "GraniteSpeechForConditionalGeneration": (
+        "granite_speech",
+        "GraniteSpeechForConditionalGeneration",
+    ),
+    "H2OVLChatModel": ("h2ovl", "H2OVLChatModel"),
+    "HunYuanVLForConditionalGeneration": (
+        "hunyuan_vision",
+        "HunYuanVLForConditionalGeneration",
+    ),
+    "StepVLForConditionalGeneration": ("step_vl", "StepVLForConditionalGeneration"),
+    "InternVLChatModel": ("internvl", "InternVLChatModel"),
+    "NemotronH_Nano_VL_V2": ("nano_nemotron_vl", "NemotronH_Nano_VL_V2"),
+    "OpenCUAForConditionalGeneration": (
+        "opencua",
+        "OpenCUAForConditionalGeneration",
+    ),
+    "InternS1ForConditionalGeneration": (
+        "interns1",
+        "InternS1ForConditionalGeneration",
+    ),
+    "InternVLForConditionalGeneration": (
+        "interns1",
+        "InternS1ForConditionalGeneration",
+    ),
+    "InternS1ProForConditionalGeneration": (
+        "interns1_pro",
+        "InternS1ProForConditionalGeneration",
+    ),
+    "Idefics3ForConditionalGeneration": (
+        "idefics3",
+        "Idefics3ForConditionalGeneration",
+    ),
+    "IsaacForConditionalGeneration": ("isaac", "IsaacForConditionalGeneration"),
+    "SmolVLMForConditionalGeneration": ("smolvlm", "SmolVLMForConditionalGeneration"),  # noqa: E501
+    "KananaVForConditionalGeneration": ("kanana_v", "KananaVForConditionalGeneration"),
+    "KeyeForConditionalGeneration": ("keye", "KeyeForConditionalGeneration"),
+    "KeyeVL1_5ForConditionalGeneration": (
+        "keye_vl1_5",
+        "KeyeVL1_5ForConditionalGeneration",
+    ),
+    "RForConditionalGeneration": ("rvl", "RForConditionalGeneration"),
+    "KimiVLForConditionalGeneration": ("kimi_vl", "KimiVLForConditionalGeneration"),  # noqa: E501
+    "KimiK25ForConditionalGeneration": ("kimi_k25", "KimiK25ForConditionalGeneration"),  # noqa: E501
+    "LightOnOCRForConditionalGeneration": (
+        "lightonocr",
+        "LightOnOCRForConditionalGeneration",
+    ),
+    "Lfm2VlForConditionalGeneration": ("lfm2_vl", "Lfm2VLForConditionalGeneration"),
+    "Llama_Nemotron_Nano_VL": ("nemotron_vl", "LlamaNemotronVLChatModel"),
+    "Llama4ForConditionalGeneration": ("mllama4", "Llama4ForConditionalGeneration"),  # noqa: E501
+    "LlavaForConditionalGeneration": ("llava", "LlavaForConditionalGeneration"),
+    "LlavaNextForConditionalGeneration": (
+        "llava_next",
+        "LlavaNextForConditionalGeneration",
+    ),
+    "LlavaNextVideoForConditionalGeneration": (
+        "llava_next_video",
+        "LlavaNextVideoForConditionalGeneration",
+    ),
+    "LlavaOnevisionForConditionalGeneration": (
+        "llava_onevision",
+        "LlavaOnevisionForConditionalGeneration",
+    ),
+    "MantisForConditionalGeneration": ("llava", "MantisForConditionalGeneration"),  # noqa: E501
+    "MiDashengLMModel": ("midashenglm", "MiDashengLMModel"),
+    "MiniMaxVL01ForConditionalGeneration": (
+        "minimax_vl_01",
+        "MiniMaxVL01ForConditionalGeneration",
+    ),
+    "MiniCPMO": ("minicpmo", "MiniCPMO"),
+    "MiniCPMV": ("minicpmv", "MiniCPMV"),
+    "Mistral3ForConditionalGeneration": (
+        "mistral3",
+        "Mistral3ForConditionalGeneration",
+    ),
+    "MolmoForCausalLM": ("molmo", "MolmoForCausalLM"),
+    "Molmo2ForConditionalGeneration": ("molmo2", "Molmo2ForConditionalGeneration"),
+    "NVLM_D": ("nvlm_d", "NVLM_D_Model"),
+    "OpenPanguVLForConditionalGeneration": (
+        "openpangu_vl",
+        "OpenPanguVLForConditionalGeneration",
+    ),
+    "Ovis": ("ovis", "Ovis"),
+    "Ovis2_5": ("ovis2_5", "Ovis2_5"),
+    "Ovis2_6ForCausalLM": ("ovis2_5", "Ovis2_5"),
+    "Ovis2_6_MoeForCausalLM": ("ovis2_5", "Ovis2_5"),
+    "PaddleOCRVLForConditionalGeneration": (
+        "paddleocr_vl",
+        "PaddleOCRVLForConditionalGeneration",
+    ),
+    "PaliGemmaForConditionalGeneration": (
+        "paligemma",
+        "PaliGemmaForConditionalGeneration",
+    ),
+    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
+    "Phi4MMForCausalLM": ("phi4mm", "Phi4MMForCausalLM"),
+    "PixtralForConditionalGeneration": ("pixtral", "PixtralForConditionalGeneration"),  # noqa: E501
+    "QwenVLForConditionalGeneration": ("qwen_vl", "QwenVLForConditionalGeneration"),  # noqa: E501
+    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
+    "Qwen2_5_VLForConditionalGeneration": (
+        "qwen2_5_vl",
+        "Qwen2_5_VLForConditionalGeneration",
+    ),
+    "Qwen2AudioForConditionalGeneration": (
+        "qwen2_audio",
+        "Qwen2AudioForConditionalGeneration",
+    ),
+    "Qwen2_5OmniModel": (
+        "qwen2_5_omni_thinker",
+        "Qwen2_5OmniThinkerForConditionalGeneration",
+    ),
+    "Qwen2_5OmniForConditionalGeneration": (
+        "qwen2_5_omni_thinker",
+        "Qwen2_5OmniThinkerForConditionalGeneration",
+    ),
+    "Qwen3OmniMoeForConditionalGeneration": (
+        "qwen3_omni_moe_thinker",
+        "Qwen3OmniMoeThinkerForConditionalGeneration",
+    ),
+    "Qwen3ASRForConditionalGeneration": (
+        "qwen3_asr",
+        "Qwen3ASRForConditionalGeneration",
+    ),
+    "Qwen3ASRRealtimeGeneration": (
+        "qwen3_asr_realtime",
+        "Qwen3ASRRealtimeGeneration",
+    ),
+    "Qwen3VLForConditionalGeneration": ("qwen3_vl", "Qwen3VLForConditionalGeneration"),  # noqa: E501
+    "Qwen3VLMoeForConditionalGeneration": (
+        "qwen3_vl_moe",
+        "Qwen3VLMoeForConditionalGeneration",
+    ),
+    "Qwen3_5ForConditionalGeneration": (
+        "qwen3_5",
+        "Qwen3_5ForConditionalGeneration",
+    ),
+    "Qwen3_5MoeForConditionalGeneration": (
+        "qwen3_5",
+        "Qwen3_5MoeForConditionalGeneration",
+    ),
+    "SkyworkR1VChatModel": ("skyworkr1v", "SkyworkR1VChatModel"),
+    "Step3VLForConditionalGeneration": ("step3_vl", "Step3VLForConditionalGeneration"),  # noqa: E501
+    "TarsierForConditionalGeneration": ("tarsier", "TarsierForConditionalGeneration"),  # noqa: E501
+    "Tarsier2ForConditionalGeneration": (
+        "qwen2_vl",
+        "Tarsier2ForConditionalGeneration",
+    ),
+    "UltravoxModel": ("ultravox", "UltravoxModel"),
+    "VoxtralForConditionalGeneration": ("voxtral", "VoxtralForConditionalGeneration"),  # noqa: E501
+    "VoxtralRealtimeGeneration": ("voxtral_realtime", "VoxtralRealtimeGeneration"),  # noqa: E501
+    # [Encoder-decoder]
+    "NemotronParseForConditionalGeneration": (
+        "nemotron_parse",
+        "NemotronParseForConditionalGeneration",
+    ),
+    "WhisperForConditionalGeneration": ("whisper", "WhisperForConditionalGeneration"),  # noqa: E501
+}
+_SPECULATIVE_DECODING_MODELS = {
+    "ExtractHiddenStatesModel": ("extract_hidden_states", "ExtractHiddenStatesModel"),
+    "MiMoMTPModel": ("mimo_mtp", "MiMoMTP"),
+    "EagleLlamaForCausalLM": ("llama_eagle", "EagleLlamaForCausalLM"),
+    "EagleLlama4ForCausalLM": ("llama4_eagle", "EagleLlama4ForCausalLM"),
+    "EagleMiniCPMForCausalLM": ("minicpm_eagle", "EagleMiniCPMForCausalLM"),
+    "Eagle3LlamaForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
+    "LlamaForCausalLMEagle3": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
+    "Eagle3Qwen2_5vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
+    "Eagle3Qwen3vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
+    "EagleMistralLarge3ForCausalLM": (
+        "mistral_large_3_eagle",
+        "EagleMistralLarge3ForCausalLM",
+    ),
+    "EagleDeepSeekMTPModel": ("deepseek_eagle", "EagleDeepseekV3ForCausalLM"),
+    "DeepSeekMTPModel": ("deepseek_mtp", "DeepSeekMTP"),
+    "ErnieMTPModel": ("ernie_mtp", "ErnieMTP"),
+    "ExaoneMoeMTP": ("exaone_moe_mtp", "ExaoneMoeMTP"),
+    "NemotronHMTPModel": ("nemotron_h_mtp", "NemotronHMTP"),
+    "LongCatFlashMTPModel": ("longcat_flash_mtp", "LongCatFlashMTP"),
+    "Glm4MoeMTPModel": ("glm4_moe_mtp", "Glm4MoeMTP"),
+    "Glm4MoeLiteMTPModel": ("glm4_moe_lite_mtp", "Glm4MoeLiteMTP"),
+    "GlmOcrMTPModel": ("glm_ocr_mtp", "GlmOcrMTP"),
+    "MedusaModel": ("medusa", "Medusa"),
+    "OpenPanguMTPModel": ("openpangu_mtp", "OpenPanguMTP"),
+    "Qwen3NextMTP": ("qwen3_next_mtp", "Qwen3NextMTP"),
+    "Step3p5MTP": ("step3p5_mtp", "Step3p5MTP"),
+    "Qwen3_5MTP": ("qwen3_5_mtp", "Qwen3_5MTP"),
+    "Qwen3_5MoeMTP": ("qwen3_5_mtp", "Qwen3_5MoeMTP"),
+    # Temporarily disabled.
+    # # TODO(woosuk): Re-enable this once the MLP Speculator is supported in V1.
+    # "MLPSpeculatorPreTrainedModel": ("mlp_speculator", "MLPSpeculator"),
+}
+_TRANSFORMERS_SUPPORTED_MODELS = {
+    # Text generation models
+    "SmolLM3ForCausalLM": ("transformers", "TransformersForCausalLM"),
+    # Multimodal models
+    "Emu3ForConditionalGeneration": (
+        "transformers",
+        "TransformersMultiModalForCausalLM",
+    ),
+}
+_TRANSFORMERS_BACKEND_MODELS = {
+    # Text generation models
+    "TransformersForCausalLM": ("transformers", "TransformersForCausalLM"),
+    "TransformersMoEForCausalLM": ("transformers", "TransformersMoEForCausalLM"),
+    # Multimodal models
+    "TransformersMultiModalForCausalLM": (
+        "transformers",
+        "TransformersMultiModalForCausalLM",
+    ),
+    "TransformersMultiModalMoEForCausalLM": (
+        "transformers",
+        "TransformersMultiModalMoEForCausalLM",
+    ),
+    # Embedding models
+    "TransformersEmbeddingModel": ("transformers", "TransformersEmbeddingModel"),
+    "TransformersMoEEmbeddingModel": ("transformers", "TransformersMoEEmbeddingModel"),
+    "TransformersMultiModalEmbeddingModel": (
+        "transformers",
+        "TransformersMultiModalEmbeddingModel",
+    ),
+    # Sequence classification models
+    "TransformersForSequenceClassification": (
+        "transformers",
+        "TransformersForSequenceClassification",
+    ),
+    "TransformersMoEForSequenceClassification": (
+        "transformers",
+        "TransformersMoEForSequenceClassification",
+    ),
+    "TransformersMultiModalForSequenceClassification": (
+        "transformers",
+        "TransformersMultiModalForSequenceClassification",
+    ),
+}
+_VLLM_MODELS = {
+    **_TEXT_GENERATION_MODELS,
+    **_EMBEDDING_MODELS,
+    **_CROSS_ENCODER_MODELS,
+    **_MULTIMODAL_MODELS,
+    **_SPECULATIVE_DECODING_MODELS,
+    **_TRANSFORMERS_SUPPORTED_MODELS,
+    **_TRANSFORMERS_BACKEND_MODELS,
+}
+# This variable is used as the args for subprocess.run(). We
+# can modify  this variable to alter the args if needed. e.g.
+# when we use par format to pack things together, sys.executable
+# might not be the target we want to run.
+_SUBPROCESS_COMMAND = [sys.executable, "-m", "vllm.model_executor.models.registry"]
+_PREVIOUSLY_SUPPORTED_MODELS = {
+    "MotifForCausalLM": "0.10.2",
+    "Phi3SmallForCausalLM": "0.9.2",
+    "Phi4FlashForCausalLM": "0.10.2",
+    "Phi4MultimodalForCausalLM": "0.12.0",
+    # encoder-decoder models except whisper
+    # have been removed for V0 deprecation.
+    "BartModel": "0.10.2",
+    "BartForConditionalGeneration": "0.10.2",
+    "DonutForConditionalGeneration": "0.10.2",
+    "Florence2ForConditionalGeneration": "0.10.2",
+    "MBartForConditionalGeneration": "0.10.2",
+    "MllamaForConditionalGeneration": "0.10.2",
+}
+@dataclass(frozen=True)
+class _ModelInfo:
+    architecture: str
+    is_text_generation_model: bool
+    is_pooling_model: bool
+    attn_type: AttnTypeStr
+    default_seq_pooling_type: SequencePoolingType
+    default_tok_pooling_type: TokenPoolingType
+    supports_cross_encoding: bool
+    supports_late_interaction: bool
+    supports_multimodal: bool
+    supports_multimodal_raw_input_only: bool
+    requires_raw_input_tokens: bool
+    supports_multimodal_encoder_tp_data: bool
+    supports_pp: bool
+    has_inner_state: bool
+    is_attention_free: bool
+    is_hybrid: bool
+    has_noops: bool
+    supports_mamba_prefix_caching: bool
+    supports_transcription: bool
+    supports_transcription_only: bool
+    @staticmethod
+    def from_model_cls(model: type[nn.Module]) -> "_ModelInfo":
+        return _ModelInfo(
+            architecture=model.__name__,
+            is_text_generation_model=is_text_generation_model(model),
+            is_pooling_model=is_pooling_model(model),
+            default_seq_pooling_type=get_default_seq_pooling_type(model),
+            default_tok_pooling_type=get_default_tok_pooling_type(model),
+            attn_type=get_attn_type(model),
+            supports_cross_encoding=supports_cross_encoding(model),
+            supports_late_interaction=supports_late_interaction(model),
+            supports_multimodal=supports_multimodal(model),
+            supports_multimodal_raw_input_only=supports_multimodal_raw_input_only(
+                model
+            ),
+            requires_raw_input_tokens=requires_raw_input_tokens(model),
+            supports_multimodal_encoder_tp_data=supports_multimodal_encoder_tp_data(
+                model
+            ),
+            supports_pp=supports_pp(model),
+            has_inner_state=has_inner_state(model),
+            is_attention_free=is_attention_free(model),
+            is_hybrid=is_hybrid(model),
+            supports_mamba_prefix_caching=supports_mamba_prefix_caching(model),
+            supports_transcription=supports_transcription(model),
+            supports_transcription_only=(
+                supports_transcription(model) and model.supports_transcription_only
+            ),
+            has_noops=has_noops(model),
+        )
+class _BaseRegisteredModel(ABC):
+    @abstractmethod
+    def inspect_model_cls(self) -> _ModelInfo:
+        raise NotImplementedError
+    @abstractmethod
+    def load_model_cls(self) -> type[nn.Module]:
+        raise NotImplementedError
+@dataclass(frozen=True)
+class _RegisteredModel(_BaseRegisteredModel):
+    """
+    Represents a model that has already been imported in the main process.
+    """
+    interfaces: _ModelInfo
+    model_cls: type[nn.Module]
+    @staticmethod
+    def from_model_cls(model_cls: type[nn.Module]):
+        return _RegisteredModel(
+            interfaces=_ModelInfo.from_model_cls(model_cls),
+            model_cls=model_cls,
+        )
+    def inspect_model_cls(self) -> _ModelInfo:
+        return self.interfaces
+    def load_model_cls(self) -> type[nn.Module]:
+        return self.model_cls
+@dataclass(frozen=True)
+class _LazyRegisteredModel(_BaseRegisteredModel):
+    """
+    Represents a model that has not been imported in the main process.
+    """
+    module_name: str
+    class_name: str
+    @staticmethod
+    def _get_cache_dir() -> Path:
+        return Path(envs.VLLM_CACHE_ROOT) / "modelinfos"
+    def _get_cache_filename(self) -> str:
+        cls_name = f"{self.module_name}-{self.class_name}".replace(".", "-")
+        return f"{cls_name}.json"
+    def _load_modelinfo_from_cache(self, module_hash: str) -> _ModelInfo | None:
+        try:
+            try:
+                modelinfo_path = self._get_cache_dir() / self._get_cache_filename()
+                with open(modelinfo_path, encoding="utf-8") as file:
+                    mi_dict = json.load(file)
+            except FileNotFoundError:
+                logger.debug(
+                    "Cached model info file for class %s.%s not found",
+                    self.module_name,
+                    self.class_name,
+                )
+                return None
+            if mi_dict["hash"] != module_hash:
+                logger.debug(
+                    "Cached model info file for class %s.%s is stale",
+                    self.module_name,
+                    self.class_name,
+                )
+                return None
+            # file not changed, use cached _ModelInfo properties
+            return _ModelInfo(**mi_dict["modelinfo"])
+        except Exception:
+            logger.debug(
+                "Cached model info for class %s.%s error. ",
+                self.module_name,
+                self.class_name,
+            )
+            return None
+    def _save_modelinfo_to_cache(self, mi: _ModelInfo, module_hash: str) -> None:
+        """save dictionary json file to cache"""
+        from vllm.model_executor.model_loader.weight_utils import atomic_writer
+        try:
+            modelinfo_dict = {
+                "hash": module_hash,
+                "modelinfo": asdict(mi),
+            }
+            cache_dir = self._get_cache_dir()
+            cache_dir.mkdir(parents=True, exist_ok=True)
+            modelinfo_path = cache_dir / self._get_cache_filename()
+            with atomic_writer(modelinfo_path, encoding="utf-8") as f:
+                json.dump(modelinfo_dict, f, indent=2)
+        except Exception:
+            logger.exception("Error saving model info cache.")
+    @logtime(logger=logger, msg="Registry inspect model class")
+    def inspect_model_cls(self) -> _ModelInfo:
+        model_path = Path(__file__).parent / f"{self.module_name.split('.')[-1]}.py"
+        module_hash = None
+        if model_path.exists():
+            with open(model_path, "rb") as f:
+                module_hash = safe_hash(f.read(), usedforsecurity=False).hexdigest()
+            mi = self._load_modelinfo_from_cache(module_hash)
+            if mi is not None:
+                logger.debug(
+                    "Loaded model info for class %s.%s from cache",
+                    self.module_name,
+                    self.class_name,
+                )
+                return mi
+            else:
+                logger.debug(
+                    "Cache model info for class %s.%s miss. Loading model instead.",
+                    self.module_name,
+                    self.class_name,
+                )
+        # Performed in another process to avoid initializing CUDA
+        mi = _run_in_subprocess(
+            lambda: _ModelInfo.from_model_cls(self.load_model_cls())
+        )
+        logger.debug(
+            "Loaded model info for class %s.%s", self.module_name, self.class_name
+        )
+        # save cache file
+        if module_hash is not None:
+            self._save_modelinfo_to_cache(mi, module_hash)
+        return mi
+    def load_model_cls(self) -> type[nn.Module]:
+        mod = importlib.import_module(self.module_name)
+        return getattr(mod, self.class_name)
+@lru_cache(maxsize=128)
+def _try_load_model_cls(
+    model_arch: str,
+    model: _BaseRegisteredModel,
+) -> type[nn.Module] | None:
+    from vllm.platforms import current_platform
+    current_platform.verify_model_arch(model_arch)
+    try:
+        return model.load_model_cls()
+    except Exception:
+        logger.exception("Error in loading model architecture '%s'", model_arch)
+        return None
+@lru_cache(maxsize=128)
+def _try_inspect_model_cls(
+    model_arch: str,
+    model: _BaseRegisteredModel,
+) -> _ModelInfo | None:
+    try:
+        return model.inspect_model_cls()
+    except Exception:
+        logger.exception("Error in inspecting model architecture '%s'", model_arch)
+        return None
+@dataclass
+class _ModelRegistry:
+    # Keyed by model_arch
+    models: dict[str, _BaseRegisteredModel] = field(default_factory=dict)
+    def get_supported_archs(self) -> Set[str]:
+        return self.models.keys()
+    def register_model(
+        self,
+        model_arch: str,
+        model_cls: type[nn.Module] | str,
+    ) -> None:
+        """
+        Register an external model to be used in vLLM.
+        `model_cls` can be either:
+        - A [`torch.nn.Module`][] class directly referencing the model.
+        - A string in the format `<module>:<class>` which can be used to
+          lazily import the model. This is useful to avoid initializing CUDA
+          when importing the model and thus the related error
+          `RuntimeError: Cannot re-initialize CUDA in forked subprocess`.
+        """
+        if not isinstance(model_arch, str):
+            msg = f"`model_arch` should be a string, not a {type(model_arch)}"
+            raise TypeError(msg)
+        if model_arch in self.models:
+            logger.warning(
+                "Model architecture %s is already registered, and will be "
+                "overwritten by the new model class %s.",
+                model_arch,
+                model_cls,
+            )
+        if isinstance(model_cls, str):
+            split_str = model_cls.split(":")
+            if len(split_str) != 2:
+                msg = "Expected a string in the format `<module>:<class>`"
+                raise ValueError(msg)
+            model = _LazyRegisteredModel(*split_str)
+        elif isinstance(model_cls, type) and issubclass(model_cls, nn.Module):
+            model = _RegisteredModel.from_model_cls(model_cls)
+        else:
+            msg = (
+                "`model_cls` should be a string or PyTorch model class, "
+                f"not a {type(model_arch)}"
+            )
+            raise TypeError(msg)
+        self.models[model_arch] = model
+    def _raise_for_unsupported(self, architectures: list[str]):
+        all_supported_archs = self.get_supported_archs()
+        if any(arch in all_supported_archs for arch in architectures):
+            raise ValueError(
+                f"Model architectures {architectures} failed "
+                "to be inspected. Please check the logs for more details."
+            )
+        for arch in architectures:
+            if arch in _PREVIOUSLY_SUPPORTED_MODELS:
+                previous_version = _PREVIOUSLY_SUPPORTED_MODELS[arch]
+                raise ValueError(
+                    f"Model architecture {arch} was supported in vLLM until "
+                    f"v{previous_version}, and is not supported anymore. "
+                    "Please use an older version of vLLM if you want to "
+                    "use this model architecture."
+                )
+        raise ValueError(
+            f"Model architectures {architectures} are not supported for now. "
+            f"Supported architectures: {all_supported_archs}"
+        )
+    def _try_load_model_cls(self, model_arch: str) -> type[nn.Module] | None:
+        if model_arch not in self.models:
+            return None
+        return _try_load_model_cls(model_arch, self.models[model_arch])
+    def _try_inspect_model_cls(self, model_arch: str) -> _ModelInfo | None:
+        if model_arch not in self.models:
+            return None
+        return _try_inspect_model_cls(model_arch, self.models[model_arch])
+    def _try_resolve_transformers(
+        self,
+        architecture: str,
+        model_config: ModelConfig,
+    ) -> str | None:
+        if architecture in _TRANSFORMERS_BACKEND_MODELS:
+            return architecture
+        auto_map: dict[str, str] = (
+            getattr(model_config.hf_config, "auto_map", None) or dict()
+        )
+        # Make sure that config class is always initialized before model class,
+        # otherwise the model class won't be able to access the config class,
+        # the expected auto_map should have correct order like:
+        # "auto_map": {
+        #     "AutoConfig": "<your-repo-name>--<config-name>",
+        #     "AutoModel": "<your-repo-name>--<config-name>",
+        #     "AutoModelFor<Task>": "<your-repo-name>--<config-name>",
+        # },
+        for prefix in ("AutoConfig", "AutoModel"):
+            for name, module in auto_map.items():
+                if name.startswith(prefix):
+                    try_get_class_from_dynamic_module(
+                        module,
+                        model_config.model,
+                        revision=model_config.revision,
+                        trust_remote_code=model_config.trust_remote_code,
+                        warn_on_fail=False,
+                    )
+        model_module = getattr(transformers, architecture, None)
+        if model_module is None:
+            for name, module in auto_map.items():
+                if name.startswith("AutoModel"):
+                    model_module = try_get_class_from_dynamic_module(
+                        module,
+                        model_config.model,
+                        revision=model_config.revision,
+                        trust_remote_code=model_config.trust_remote_code,
+                        warn_on_fail=True,
+                    )
+                    if model_module is not None:
+                        break
+            else:
+                if model_config.model_impl != "transformers":
+                    return None
+                raise ValueError(
+                    f"Cannot find model module. {architecture!r} is not a "
+                    "registered model in the Transformers library (only "
+                    "relevant if the model is meant to be in Transformers) "
+                    "and 'AutoModel' is not present in the model config's "
+                    "'auto_map' (relevant if the model is custom)."
+                )
+        if not model_module.is_backend_compatible():
+            if model_config.model_impl != "transformers":
+                return None
+            raise ValueError(
+                f"The Transformers implementation of {architecture!r} "
+                "is not compatible with vLLM."
+            )
+        return model_config._get_transformers_backend_cls()
+    def _normalize_arch(
+        self,
+        architecture: str,
+        model_config: ModelConfig,
+    ) -> str:
+        if architecture in self.models:
+            return architecture
+        # This may be called in order to resolve runner_type and convert_type
+        # in the first place, in which case we consider the default match
+        match = try_match_architecture_defaults(
+            architecture,
+            runner_type=getattr(model_config, "runner_type", None),
+            convert_type=getattr(model_config, "convert_type", None),
+        )
+        if match:
+            suffix, _ = match
+            # Get the name of the base model to convert
+            for repl_suffix, _ in iter_architecture_defaults():
+                base_arch = architecture.replace(suffix, repl_suffix)
+                if base_arch in self.models:
+                    return base_arch
+        return architecture
+    def inspect_model_cls(
+        self,
+        architectures: str | list[str],
+        model_config: ModelConfig,
+    ) -> tuple[_ModelInfo, str]:
+        if isinstance(architectures, str):
+            architectures = [architectures]
+        if not architectures:
+            raise ValueError("No model architectures are specified")
+        # Require transformers impl
+        if model_config.model_impl == "transformers":
+            arch = self._try_resolve_transformers(architectures[0], model_config)
+            if arch is not None:
+                model_info = self._try_inspect_model_cls(arch)
+                if model_info is not None:
+                    return (model_info, arch)
+        elif model_config.model_impl == "terratorch":
+            model_info = self._try_inspect_model_cls("Terratorch")
+            return (model_info, "Terratorch")
+        # Fallback to transformers impl (after resolving convert_type)
+        if (
+            all(arch not in self.models for arch in architectures)
+            and model_config.model_impl == "auto"
+            and getattr(model_config, "convert_type", "none") == "none"
+        ):
+            arch = self._try_resolve_transformers(architectures[0], model_config)
+            if arch is not None:
+                model_info = self._try_inspect_model_cls(arch)
+                if model_info is not None:
+                    return (model_info, arch)
+        for arch in architectures:
+            normalized_arch = self._normalize_arch(arch, model_config)
+            model_info = self._try_inspect_model_cls(normalized_arch)
+            if model_info is not None:
+                return (model_info, arch)
+        # Fallback to transformers impl (before resolving runner_type)
+        if (
+            all(arch not in self.models for arch in architectures)
+            and model_config.model_impl == "auto"
+        ):
+            arch = self._try_resolve_transformers(architectures[0], model_config)
+            if arch is not None:
+                model_info = self._try_inspect_model_cls(arch)
+                if model_info is not None:
+                    return (model_info, arch)
+        return self._raise_for_unsupported(architectures)
+    def resolve_model_cls(
+        self,
+        architectures: str | list[str],
+        model_config: ModelConfig,
+    ) -> tuple[type[nn.Module], str]:
+        if isinstance(architectures, str):
+            architectures = [architectures]
+        if not architectures:
+            raise ValueError("No model architectures are specified")
+        # Require transformers impl
+        if model_config.model_impl == "transformers":
+            arch = self._try_resolve_transformers(architectures[0], model_config)
+            if arch is not None:
+                model_cls = self._try_load_model_cls(arch)
+                if model_cls is not None:
+                    return (model_cls, arch)
+        elif model_config.model_impl == "terratorch":
+            arch = "Terratorch"
+            model_cls = self._try_load_model_cls(arch)
+            if model_cls is not None:
+                return (model_cls, arch)
+        # Fallback to transformers impl (after resolving convert_type)
+        if (
+            all(arch not in self.models for arch in architectures)
+            and model_config.model_impl == "auto"
+            and getattr(model_config, "convert_type", "none") == "none"
+        ):
+            arch = self._try_resolve_transformers(architectures[0], model_config)
+            if arch is not None:
+                model_cls = self._try_load_model_cls(arch)
+                if model_cls is not None:
+                    return (model_cls, arch)
+        for arch in architectures:
+            normalized_arch = self._normalize_arch(arch, model_config)
+            model_cls = self._try_load_model_cls(normalized_arch)
+            if model_cls is not None:
+                return (model_cls, arch)
+        # Fallback to transformers impl (before resolving runner_type)
+        if (
+            all(arch not in self.models for arch in architectures)
+            and model_config.model_impl == "auto"
+        ):
+            arch = self._try_resolve_transformers(architectures[0], model_config)
+            if arch is not None:
+                model_cls = self._try_load_model_cls(arch)
+                if model_cls is not None:
+                    return (model_cls, arch)
+        return self._raise_for_unsupported(architectures)
+    def is_text_generation_model(
+        self,
+        architectures: str | list[str],
+        model_config: ModelConfig,
+    ) -> bool:
+        model_cls, _ = self.inspect_model_cls(architectures, model_config)
+        return model_cls.is_text_generation_model
+    def is_pooling_model(
+        self,
+        architectures: str | list[str],
+        model_config: ModelConfig,
+    ) -> bool:
+        model_cls, _ = self.inspect_model_cls(architectures, model_config)
+        return model_cls.is_pooling_model
+    def is_cross_encoder_model(
+        self,
+        architectures: str | list[str],
+        model_config: ModelConfig,
+    ) -> bool:
+        model_cls, _ = self.inspect_model_cls(architectures, model_config)
+        return model_cls.supports_cross_encoding
+    def is_multimodal_model(
+        self,
+        architectures: str | list[str],
+        model_config: ModelConfig,
+    ) -> bool:
+        model_cls, _ = self.inspect_model_cls(architectures, model_config)
+        return model_cls.supports_multimodal
+    def is_multimodal_raw_input_only_model(
+        self,
+        architectures: str | list[str],
+        model_config: ModelConfig,
+    ) -> bool:
+        model_cls, _ = self.inspect_model_cls(architectures, model_config)
+        return model_cls.supports_multimodal_raw_input_only
+    def is_pp_supported_model(
+        self,
+        architectures: str | list[str],
+        model_config: ModelConfig,
+    ) -> bool:
+        model_cls, _ = self.inspect_model_cls(architectures, model_config)
+        return model_cls.supports_pp
+    def model_has_inner_state(
+        self,
+        architectures: str | list[str],
+        model_config: ModelConfig,
+    ) -> bool:
+        model_cls, _ = self.inspect_model_cls(architectures, model_config)
+        return model_cls.has_inner_state
+    def is_attention_free_model(
+        self,
+        architectures: str | list[str],
+        model_config: ModelConfig,
+    ) -> bool:
+        model_cls, _ = self.inspect_model_cls(architectures, model_config)
+        return model_cls.is_attention_free
+    def is_hybrid_model(
+        self,
+        architectures: str | list[str],
+        model_config: ModelConfig,
+    ) -> bool:
+        model_cls, _ = self.inspect_model_cls(architectures, model_config)
+        return model_cls.is_hybrid
+    def is_noops_model(
+        self,
+        architectures: str | list[str],
+        model_config: ModelConfig,
+    ) -> bool:
+        model_cls, _ = self.inspect_model_cls(architectures, model_config)
+        return model_cls.has_noops
+    def is_transcription_model(
+        self,
+        architectures: str | list[str],
+        model_config: ModelConfig,
+    ) -> bool:
+        model_cls, _ = self.inspect_model_cls(architectures, model_config)
+        return model_cls.supports_transcription
+    def is_transcription_only_model(
+        self,
+        architectures: str | list[str],
+        model_config: ModelConfig,
+    ) -> bool:
+        model_cls, _ = self.inspect_model_cls(architectures, model_config)
+        return model_cls.supports_transcription_only
+ModelRegistry = _ModelRegistry(
+    {
+        model_arch: _LazyRegisteredModel(
+            module_name=f"vllm.model_executor.models.{mod_relname}",
+            class_name=cls_name,
+        )
+        for model_arch, (mod_relname, cls_name) in _VLLM_MODELS.items()
+    }
+)
+_T = TypeVar("_T")
+def _run_in_subprocess(fn: Callable[[], _T]) -> _T:
+    # NOTE: We use a temporary directory instead of a temporary file to avoid
+    # issues like https://stackoverflow.com/questions/23212435/permission-denied-to-write-to-my-temporary-file
+    with tempfile.TemporaryDirectory() as tempdir:
+        output_filepath = os.path.join(tempdir, "registry_output.tmp")
+        # `cloudpickle` allows pickling lambda functions directly
+        import cloudpickle
+        input_bytes = cloudpickle.dumps((fn, output_filepath))
+        # cannot use `sys.executable __file__` here because the script
+        # contains relative imports
+        returned = subprocess.run(
+            _SUBPROCESS_COMMAND, input=input_bytes, capture_output=True
+        )
+        # check if the subprocess is successful
+        try:
+            returned.check_returncode()
+        except Exception as e:
+            # wrap raised exception to provide more information
+            raise RuntimeError(
+                f"Error raised in subprocess:\n{returned.stderr.decode()}"
+            ) from e
+        with open(output_filepath, "rb") as f:
+            return pickle.load(f)
+def _run() -> None:
+    # Setup plugins
+    from vllm.plugins import load_general_plugins
+    load_general_plugins()
+    fn, output_file = pickle.loads(sys.stdin.buffer.read())
+    result = fn()
+    with open(output_file, "wb") as f:
+        f.write(pickle.dumps(result))
+if __name__ == "__main__":
+    _run()

patches/solar_open.py ADDED Viewed

	@@ -0,0 +1,693 @@

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+#
+# Adopted from
+# https://github.com/huggingface/transformers/blob/v5.0.0/src/transformers/models/solar_open/modeling_solar_open.py
+#
+#
+# Copyright 2026 Upstage Inc.
+# Copyright 2026 The vLLM team.
+#
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Inference-only Solar Open model compatible with HuggingFace weights."""
+import typing
+from collections.abc import Callable, Iterable
+from itertools import islice
+import torch
+from torch import nn
+from vllm.compilation.decorators import support_torch_compile
+from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config
+from vllm.distributed import (
+    get_ep_group,
+    get_pp_group,
+    get_tensor_model_parallel_world_size,
+)
+from vllm.logger import init_logger
+from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.attention.attention import Attention
+from vllm.model_executor.layers.fused_moe import FusedMoE, SharedFusedMoE
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.linear import (
+    MergedColumnParallelLinear,
+    QKVParallelLinear,
+    RowParallelLinear,
+)
+from vllm.model_executor.layers.logits_processor import LogitsProcessor
+from vllm.model_executor.layers.quantization import QuantizationConfig
+from vllm.model_executor.layers.rotary_embedding import get_rope
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
+from vllm.model_executor.model_loader.weight_utils import (
+    default_weight_loader,
+    maybe_remap_kv_scale_name,
+)
+from vllm.model_executor.models.interfaces import SupportsLoRA, SupportsPP
+from vllm.model_executor.models.utils import (
+    AutoWeightsLoader,
+    PPMissingLayer,
+    is_pp_missing_parameter,
+    make_empty_intermediate_tensors_factory,
+    make_layers,
+    maybe_prefix,
+)
+from vllm.sequence import IntermediateTensors
+logger = init_logger(__name__)
+class SolarOpenDecoderLayer(nn.Module):
+    def __init__(
+        self,
+        hf_config,
+        cache_config: CacheConfig | None = None,
+        quant_config: QuantizationConfig | None = None,
+        prefix: str = "",
+        enable_eplb: bool = False,
+    ) -> None:
+        super().__init__()
+        self.hidden_size = hf_config.hidden_size
+        rope_theta = getattr(hf_config, "rope_theta", 1000000)
+        max_position_embeddings = getattr(hf_config, "max_position_embeddings", 131072)
+        # DecoderLayers are created with `make_layers` which passes the prefix
+        # with the layer's index.
+        layer_idx = int(prefix.split(sep=".")[-1])
+        self.layer_idx = layer_idx
+        self.self_attn = SolarOpenAttention(
+            hf_config=hf_config,
+            hidden_size=self.hidden_size,
+            num_heads=hf_config.num_attention_heads,
+            num_kv_heads=hf_config.num_key_value_heads,
+            rope_theta=rope_theta,
+            max_position_embeddings=max_position_embeddings,
+            head_dim=hf_config.head_dim,
+            qkv_bias=hf_config.attention_bias,
+            cache_config=cache_config,
+            quant_config=quant_config,
+            prefix=f"{prefix}.self_attn",
+        )
+        self.mlp = SolarOpenMoE(
+            hf_config=hf_config,
+            quant_config=quant_config,
+            prefix=f"{prefix}.mlp",
+            enable_eplb=enable_eplb,
+        )
+        self.input_layernorm = RMSNorm(
+            hf_config.hidden_size, eps=hf_config.rms_norm_eps
+        )
+        self.post_attention_layernorm = RMSNorm(
+            hf_config.hidden_size, eps=hf_config.rms_norm_eps
+        )
+        self.routed_scaling_factor = hf_config.routed_scaling_factor
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        residual: torch.Tensor | None,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        if residual is None:
+            residual = hidden_states
+            hidden_states = self.input_layernorm(hidden_states)
+        else:
+            hidden_states, residual = self.input_layernorm(hidden_states, residual)
+        hidden_states = self.self_attn(positions=positions, hidden_states=hidden_states)
+        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
+        hidden_states = self.mlp(hidden_states)
+        return hidden_states, residual
+class SolarOpenMLP(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
+        quant_config: QuantizationConfig | None = None,
+        reduce_results: bool = True,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.gate_up_proj = MergedColumnParallelLinear(
+            hidden_size,
+            [intermediate_size] * 2,
+            bias=False,
+            quant_config=quant_config,
+            prefix=f"{prefix}.gate_up_proj",
+        )
+        self.down_proj = RowParallelLinear(
+            intermediate_size,
+            hidden_size,
+            bias=False,
+            quant_config=quant_config,
+            reduce_results=reduce_results,
+            prefix=f"{prefix}.down_proj",
+        )
+        if hidden_act != "silu":
+            raise ValueError(
+                f"Unsupported activation: {hidden_act}. Only silu is supported for now."
+            )
+        self.act_fn = SiluAndMul()
+    def forward(self, x):
+        gate_up, _ = self.gate_up_proj(x)
+        x = self.act_fn(gate_up)
+        x, _ = self.down_proj(x)
+        return x
+class SolarOpenMoE(nn.Module):
+    def __init__(
+        self,
+        hf_config,
+        quant_config: QuantizationConfig | None = None,
+        prefix: str = "",
+        enable_eplb: bool = False,
+    ):
+        super().__init__()
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.routed_scaling_factor = hf_config.routed_scaling_factor
+        self.ep_group = get_ep_group().device_group
+        self.ep_rank = self.ep_group.rank()
+        self.ep_size = self.ep_group.size()
+        self.n_routed_experts: int = hf_config.n_routed_experts
+        self.n_shared_experts: int = hf_config.n_shared_experts
+        if hf_config.hidden_act != "silu":
+            raise ValueError(
+                f"Unsupported activation: {hf_config.hidden_act}. "
+                "Only silu is supported for now."
+            )
+        self.gate = nn.Linear(
+            hf_config.hidden_size,
+            hf_config.n_routed_experts,
+            bias=False,
+            dtype=torch.float32,
+        )
+        self.gate.e_score_correction_bias = nn.Parameter(
+            torch.empty(hf_config.n_routed_experts, dtype=torch.float32)
+        )
+        # Load balancing settings.
+        vllm_config = get_current_vllm_config()
+        eplb_config = vllm_config.parallel_config.eplb_config
+        self.enable_eplb = enable_eplb
+        self.n_redundant_experts = eplb_config.num_redundant_experts
+        self.n_logical_experts = self.n_routed_experts
+        self.n_physical_experts = self.n_logical_experts + self.n_redundant_experts
+        self.n_local_physical_experts = self.n_physical_experts // self.ep_size
+        self.physical_expert_start = self.ep_rank * self.n_local_physical_experts
+        self.physical_expert_end = (
+            self.physical_expert_start + self.n_local_physical_experts
+        )
+        if hf_config.n_shared_experts is not None:
+            intermediate_size = (
+                hf_config.moe_intermediate_size * hf_config.n_shared_experts
+            )
+            self.shared_experts = SolarOpenMLP(
+                hidden_size=hf_config.hidden_size,
+                intermediate_size=intermediate_size,
+                hidden_act=hf_config.hidden_act,
+                quant_config=quant_config,
+                reduce_results=False,
+                prefix=f"{prefix}.shared_experts",
+            )
+            self.experts = SharedFusedMoE(
+                shared_experts=self.shared_experts,
+                num_experts=hf_config.n_routed_experts,
+                top_k=hf_config.num_experts_per_tok,
+                hidden_size=hf_config.hidden_size,
+                intermediate_size=hf_config.moe_intermediate_size,
+                reduce_results=False,
+                renormalize=hf_config.norm_topk_prob,
+                quant_config=quant_config,
+                use_grouped_topk=True,
+                num_expert_group=hf_config.n_group,
+                topk_group=hf_config.topk_group,
+                prefix=f"{prefix}.experts",
+                scoring_func="sigmoid",
+                # we do scaling outside, set factor to 1.0 to avoid double mul
+                routed_scaling_factor=1.0,
+                e_score_correction_bias=self.gate.e_score_correction_bias,
+                enable_eplb=self.enable_eplb,
+                num_redundant_experts=self.n_redundant_experts,
+            )
+        else:
+            self.experts = FusedMoE(
+                num_experts=hf_config.n_routed_experts,
+                top_k=hf_config.num_experts_per_tok,
+                hidden_size=hf_config.hidden_size,
+                intermediate_size=hf_config.moe_intermediate_size,
+                reduce_results=False,
+                renormalize=hf_config.norm_topk_prob,
+                quant_config=quant_config,
+                use_grouped_topk=True,
+                num_expert_group=hf_config.n_group,
+                topk_group=hf_config.topk_group,
+                prefix=f"{prefix}.experts",
+                scoring_func="sigmoid",
+                # we do scaling outside, set factor to 1.0 to avoid double mul
+                routed_scaling_factor=1.0,
+                e_score_correction_bias=self.gate.e_score_correction_bias,
+                enable_eplb=self.enable_eplb,
+                num_redundant_experts=self.n_redundant_experts,
+            )
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        num_tokens, hidden_dim = hidden_states.shape
+        hidden_states = hidden_states.view(-1, hidden_dim)
+        # router_logits: (num_tokens, n_experts)
+        router_logits = self.gate(hidden_states.to(dtype=torch.float32))
+        fused_moe_out = self.experts(
+            hidden_states=hidden_states, router_logits=router_logits
+        )
+        if self.shared_experts is not None:
+            shared_output, final_hidden_states = fused_moe_out
+            assert shared_output is not None
+            final_hidden_states = (
+                final_hidden_states * self.routed_scaling_factor + shared_output
+            )
+        else:
+            final_hidden_states = fused_moe_out * self.routed_scaling_factor
+        if self.tp_size > 1:
+            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(
+                final_hidden_states
+            )
+        return final_hidden_states.view(num_tokens, hidden_dim)
+class SolarOpenAttention(nn.Module):
+    def __init__(
+        self,
+        hf_config,
+        hidden_size: int,
+        num_heads: int,
+        num_kv_heads: int,
+        rope_theta: float = 1000000,
+        max_position_embeddings: int = 131072,
+        head_dim: int | None = 128,
+        qkv_bias: bool = False,
+        cache_config: CacheConfig | None = None,
+        quant_config: QuantizationConfig | None = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.hidden_size = hidden_size
+        tp_size = get_tensor_model_parallel_world_size()
+        self.total_num_heads = num_heads
+        assert self.total_num_heads % tp_size == 0
+        self.num_heads = self.total_num_heads // tp_size
+        self.total_num_kv_heads = num_kv_heads
+        if self.total_num_kv_heads >= tp_size:
+            # Number of KV heads is greater than TP size, so we partition
+            # the KV heads across multiple tensor parallel GPUs.
+            assert self.total_num_kv_heads % tp_size == 0
+        else:
+            # Number of KV heads is less than TP size, so we replicate
+            # the KV heads across multiple tensor parallel GPUs.
+            assert tp_size % self.total_num_kv_heads == 0
+        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
+        self.head_dim = head_dim or (hidden_size // self.total_num_heads)
+        self.q_size = self.num_heads * self.head_dim
+        self.kv_size = self.num_kv_heads * self.head_dim
+        self.scaling = self.head_dim**-0.5
+        self.rope_theta = rope_theta
+        self.max_position_embeddings = max_position_embeddings
+        self.qkv_proj = QKVParallelLinear(
+            hidden_size,
+            self.head_dim,
+            self.total_num_heads,
+            self.total_num_kv_heads,
+            bias=qkv_bias,
+            quant_config=quant_config,
+            prefix=f"{prefix}.qkv_proj",
+        )
+        self.o_proj = RowParallelLinear(
+            self.total_num_heads * self.head_dim,
+            hidden_size,
+            bias=False,
+            quant_config=quant_config,
+            prefix=f"{prefix}.o_proj",
+        )
+        hf_config.rope_parameters.setdefault("partial_rotary_factor", 1.0)
+        self.rotary_emb = get_rope(
+            self.head_dim,
+            max_position=max_position_embeddings,
+            rope_parameters=hf_config.rope_parameters,
+        )
+        self.attn = Attention(
+            self.num_heads,
+            self.head_dim,
+            self.scaling,
+            num_kv_heads=self.num_kv_heads,
+            cache_config=cache_config,
+            quant_config=quant_config,
+            prefix=f"{prefix}.attn",
+        )
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+    ) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(hidden_states)
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+        q, k = self.rotary_emb(positions, q, k)
+        attn_output = self.attn(q, k, v)
+        output, _ = self.o_proj(attn_output)
+        return output
+@support_torch_compile(
+    dynamic_arg_dims={
+        "input_ids": 0,
+        "positions": -1,
+        "intermediate_tensors": 0,
+        "inputs_embeds": 0,
+    }
+)
+class SolarOpenModel(nn.Module):
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+        hf_config = vllm_config.model_config.hf_config
+        cache_config = vllm_config.cache_config
+        quant_config = vllm_config.quant_config
+        enable_eplb = vllm_config.parallel_config.enable_eplb
+        self.hf_config = hf_config
+        self.vocab_size = hf_config.vocab_size
+        if get_pp_group().is_first_rank:
+            self.embed_tokens = VocabParallelEmbedding(
+                hf_config.vocab_size,
+                hf_config.hidden_size,
+                prefix=f"{prefix}.embed_tokens",
+            )
+        else:
+            self.embed_tokens = PPMissingLayer()
+        self.start_layer, self.end_layer, self.layers = make_layers(
+            hf_config.num_hidden_layers,
+            lambda prefix: SolarOpenDecoderLayer(
+                hf_config=hf_config,
+                cache_config=cache_config,
+                quant_config=quant_config,
+                prefix=prefix,
+                enable_eplb=enable_eplb,
+            ),
+            prefix=f"{prefix}.layers",
+        )
+        if get_pp_group().is_last_rank:
+            self.norm = RMSNorm(hf_config.hidden_size, eps=hf_config.rms_norm_eps)
+        else:
+            self.norm = PPMissingLayer()
+        self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
+            ["hidden_states", "residual"], hf_config.hidden_size
+        )
+    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.embed_tokens(input_ids)
+    def forward(
+        self,
+        input_ids: torch.Tensor | None,
+        positions: torch.Tensor,
+        intermediate_tensors: IntermediateTensors | None = None,
+        inputs_embeds: torch.Tensor | None = None,
+    ) -> torch.Tensor | IntermediateTensors:
+        if get_pp_group().is_first_rank:
+            if inputs_embeds is not None:
+                hidden_states = inputs_embeds
+            else:
+                hidden_states = self.embed_input_ids(input_ids)
+            residual = None
+        else:
+            assert intermediate_tensors is not None
+            hidden_states = intermediate_tensors["hidden_states"]
+            residual = intermediate_tensors["residual"]
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
+            hidden_states, residual = layer(positions, hidden_states, residual)
+        if not get_pp_group().is_last_rank:
+            return IntermediateTensors(
+                {"hidden_states": hidden_states, "residual": residual}
+            )
+        hidden_states, _ = self.norm(hidden_states, residual)
+        return hidden_states
+    def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
+        # Params for weights, fp8 weight scales, fp8 activation scales
+        # (param_name, weight_name, expert_id, shard_id)
+        return SharedFusedMoE.make_expert_params_mapping(
+            self,
+            ckpt_gate_proj_name="gate_proj",
+            ckpt_down_proj_name="down_proj",
+            ckpt_up_proj_name="up_proj",
+            num_experts=self.hf_config.n_routed_experts,
+        )
+    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
+        stacked_params_mapping = [
+            # (param_name, shard_name, shard_id)
+            ("qkv_proj", "q_proj", "q"),
+            ("qkv_proj", "k_proj", "k"),
+            ("qkv_proj", "v_proj", "v"),
+            ("gate_up_proj", "gate_proj", 0),
+            ("gate_up_proj", "up_proj", 1),
+        ]
+        params_dict = dict(self.named_parameters())
+        loaded_params: set[str] = set()
+        expert_params_mapping = self.get_expert_mapping()
+        for name, loaded_weight in weights:
+            for param_name, weight_name, shard_id in stacked_params_mapping:
+                # Skip non-stacked layers and experts (experts handled below).
+                if weight_name not in name:
+                    continue
+                # We have mlp.experts[0].gate_proj in the checkpoint.
+                # Since we handle the experts below in expert_params_mapping,
+                # we need to skip here BEFORE we update the name, otherwise
+                # name will be updated to mlp.experts[0].gate_up_proj, which
+                # will then be updated below in expert_params_mapping
+                # for mlp.experts[0].gate_gate_up_proj, which breaks load.
+                if ("mlp.experts." in name) and name not in params_dict:
+                    continue
+                name = name.replace(weight_name, param_name)
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+                if is_pp_missing_parameter(name, self):
+                    continue
+                param = params_dict[name]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, shard_id)
+                break
+            else:
+                is_expert_weight = False
+                for mapping in expert_params_mapping:
+                    param_name, weight_name, expert_id, shard_id = mapping
+                    if weight_name not in name:
+                        continue
+                    # Anyway, this is an expert weight and should not be
+                    # attempted to load as other weights later
+                    is_expert_weight = True
+                    # Do not modify `name` since the loop may continue here
+                    # Instead, create a new variable
+                    name_mapped = name.replace(weight_name, param_name)
+                    if is_pp_missing_parameter(name_mapped, self):
+                        continue
+                    param = params_dict[name_mapped]
+                    # We should ask the weight loader to return success or not
+                    # here since otherwise we may skip experts with other
+                    # available replicas.
+                    weight_loader = typing.cast(
+                        Callable[..., bool], param.weight_loader
+                    )
+                    success = weight_loader(
+                        param,
+                        loaded_weight,
+                        name_mapped,
+                        shard_id=shard_id,
+                        expert_id=expert_id,
+                        return_success=True,
+                    )
+                    if success:
+                        name = name_mapped
+                        break
+                else:
+                    if is_expert_weight:
+                        # We've checked that this is an expert weight
+                        # However it's not mapped locally to this rank
+                        # So we simply skip it
+                        continue
+                    # Skip loading extra bias for GPTQ models.
+                    if name.endswith(".bias") and name not in params_dict:
+                        continue
+                    # Remapping the name of FP8 kv-scale.
+                    name = maybe_remap_kv_scale_name(name, params_dict)
+                    if name is None:
+                        continue
+                    if is_pp_missing_parameter(name, self):
+                        continue
+                    param = params_dict[name]
+                    weight_loader = getattr(
+                        param, "weight_loader", default_weight_loader
+                    )
+                    weight_loader(param, loaded_weight)
+            loaded_params.add(name)
+        return loaded_params
+class SolarOpenForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
+    packed_modules_mapping = {
+        "qkv_proj": [
+            "q_proj",
+            "k_proj",
+            "v_proj",
+        ],
+        "gate_up_proj": [
+            "gate_proj",
+            "up_proj",
+        ],
+    }
+    fall_back_to_pt_during_load = False
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+        config = vllm_config.model_config.hf_config
+        quant_config = vllm_config.quant_config
+        self.config = config
+        self.quant_config = quant_config
+        self.model = SolarOpenModel(
+            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
+        )
+        if get_pp_group().is_last_rank:
+            self.lm_head = ParallelLMHead(
+                config.vocab_size,
+                config.hidden_size,
+                quant_config=quant_config,
+                prefix=maybe_prefix(prefix, "lm_head"),
+            )
+        else:
+            self.lm_head = PPMissingLayer()
+        self.logits_processor = LogitsProcessor(config.vocab_size)
+        self.make_empty_intermediate_tensors = (
+            self.model.make_empty_intermediate_tensors
+        )
+        self.expert_weights = []
+        # Set MoE hyperparameters
+        self.num_moe_layers = config.num_hidden_layers - config.first_k_dense_replace
+        self.num_expert_groups = config.n_group
+        self.moe_layers: list[FusedMoE] = []
+        example_moe = None
+        for layer in self.model.layers:
+            if isinstance(layer, PPMissingLayer):
+                continue
+            assert isinstance(layer, SolarOpenDecoderLayer)
+            if isinstance(layer.mlp, SolarOpenMoE):
+                # Pick last one layer since the first ones may be dense layers.
+                example_moe = layer.mlp
+                self.moe_layers.append(layer.mlp.experts)
+        if example_moe is None:
+            raise RuntimeError("No SolarOpenMoE layer found in model.layers.")
+        self.num_logical_experts = example_moe.n_logical_experts
+        self.num_physical_experts = example_moe.n_physical_experts
+        self.num_local_physical_experts = example_moe.n_local_physical_experts
+        self.num_routed_experts = example_moe.n_routed_experts
+        self.num_shared_experts = example_moe.n_shared_experts
+        self.num_redundant_experts = example_moe.n_redundant_experts
+    def set_eplb_state(
+        self,
+        expert_load_view: torch.Tensor,
+        logical_to_physical_map: torch.Tensor,
+        logical_replica_count: torch.Tensor,
+    ) -> None:
+        for layer_idx, layer in enumerate(self.moe_layers):
+            # Register the expert weights.
+            self.expert_weights.append(layer.get_expert_weights())
+            layer.set_eplb_state(
+                moe_layer_idx=layer_idx,
+                expert_load_view=expert_load_view,
+                logical_to_physical_map=logical_to_physical_map,
+                logical_replica_count=logical_replica_count,
+            )
+    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.model.embed_input_ids(input_ids)
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        intermediate_tensors: IntermediateTensors | None = None,
+        inputs_embeds: torch.Tensor | None = None,
+    ) -> torch.Tensor | IntermediateTensors:
+        hidden_states = self.model(
+            input_ids, positions, intermediate_tensors, inputs_embeds
+        )
+        return hidden_states
+    def compute_logits(
+        self,
+        hidden_states: torch.Tensor,
+    ) -> torch.Tensor | None:
+        logits = self.logits_processor(self.lm_head, hidden_states)
+        return logits
+    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
+        loader = AutoWeightsLoader(self)
+        return loader.load_weights(weights)
+    def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
+        return self.model.get_expert_mapping()