qwen3.6-moe

This tiny model is intended for debugging. It is randomly initialized using the configuration adapted from Qwen/Qwen3.6-35B-A3B.

Example usage:

• vLLM

# Multi-token prediction is supported
model_id=tiny-random/qwen3.6-moe
vllm serve $model_id \
     --tensor-parallel-size 2 \
     --speculative-config.method qwen3_next_mtp \
     --speculative-config.num_speculative_tokens 2 \
     --reasoning-parser qwen3 \
     --tool-call-parser qwen3_coder \
     --enable-auto-tool-choice \
     --max-cudagraph-capture-size 8

• SGLang

# Multi-token prediction is supported
model_id=tiny-random/qwen3.6-moe
python3 -m sglang.launch_server \
  --model-path $model_id \
  --tp-size 2 \
  --tool-call-parser qwen3_coder  \
  --reasoning-parser qwen3 \
  --speculative-algo NEXTN \
  --speculative-num-steps 3 \
  --speculative-eagle-topk 1 \
  --speculative-num-draft-tokens 4

• Transformers

import numpy as np
import torch
import transformers
from PIL import Image
from transformers import (
    AutoModel,
    AutoModelForCausalLM,
    AutoProcessor,
    AutoTokenizer,
    Qwen3_5MoeForConditionalGeneration,
)

model_id = "tiny-random/qwen3.6-moe"
model = Qwen3_5MoeForConditionalGeneration.from_pretrained(
    model_id, dtype=torch.bfloat16, device_map="auto",
)
processor = AutoProcessor.from_pretrained(model_id)
messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg",
            },
            {"type": "text", "text": "Describe this image."},
        ],
    }
]

inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt"
).to(model.device)

generated_ids = model.generate(**inputs, max_new_tokens=32)
output_text = processor.batch_decode(generated_ids[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True)
print(output_text)

Codes to create this repo:

<details> <summary>Click to expand</summary>

import json
from copy import deepcopy
from pathlib import Path

import torch
from huggingface_hub import file_exists, hf_hub_download
from transformers import (
    AutoConfig,
    AutoModelForCausalLM,
    AutoProcessor,
    GenerationConfig,
    Qwen3_5MoeForConditionalGeneration,
    set_seed,
)

source_model_id = "Qwen/Qwen3.6-35B-A3B"
save_folder = "/tmp/tiny-random/qwen36-moe"

processor = AutoProcessor.from_pretrained(source_model_id, trust_remote_code=True)
processor.save_pretrained(save_folder)

with open(hf_hub_download(source_model_id, filename='config.json', repo_type='model'), 'r', encoding='utf-8') as f:
    config_json = json.load(f)

config_json['text_config'].update({
    'head_dim': 32,
    'hidden_size': 8,
    "layer_types": ['linear_attention'] * 3 + ['full_attention'],
    'intermediate_size': 32,
    'moe_intermediate_size': 32,
    'num_hidden_layers': 4,
    'num_attention_heads': 8,
    'num_key_value_heads': 4,
    'shared_expert_intermediate_size': 32,
    # 3.6 has higher dim for linear attention vs 3.5
})
config_json['text_config']['rope_parameters']['mrope_section'] = [1, 1, 2]
config_json["tie_word_embeddings"] = False
config_json['vision_config'].update(
    {
        'hidden_size': 64,
        'intermediate_size': 128,
        'num_heads': 2,
        'out_hidden_size': 8,
        'depth': 2,
    }
)
with open(f"{save_folder}/config.json", "w", encoding='utf-8') as f:
    json.dump(config_json, f, indent=2)

config = AutoConfig.from_pretrained(
    save_folder,
    trust_remote_code=True,
)
print(config)
torch.set_default_dtype(torch.bfloat16)
model = Qwen3_5MoeForConditionalGeneration(config)
# with torch.no_grad():
#     for i in range(3):
#         attn = model.model.language_model.layers[i].linear_attn
#         attn.A_log = torch.nn.Parameter(attn.A_log.float())
#         attn.norm.float()

print(model.state_dict()['model.language_model.layers.0.linear_attn.A_log'].dtype)
print(model.state_dict()['model.language_model.layers.0.linear_attn.norm.weight'].dtype)

model.mtp = torch.nn.ModuleDict({
    "pre_fc_norm_embedding": torch.nn.RMSNorm(config.text_config.hidden_size),
    "fc": torch.nn.Linear(config.text_config.hidden_size * 2, config.text_config.hidden_size, bias=False),
    "layers": torch.nn.ModuleList([deepcopy(model.model.language_model.layers[3])]),
    "norm": torch.nn.RMSNorm(config.text_config.hidden_size),
    "pre_fc_norm_hidden": torch.nn.RMSNorm(config.text_config.hidden_size),
})
torch.set_default_dtype(torch.float32)
if file_exists(filename="generation_config.json", repo_id=source_model_id, repo_type='model'):
    model.generation_config = GenerationConfig.from_pretrained(
        source_model_id, trust_remote_code=True,
    )
    model.generation_config.do_sample = True
    print(model.generation_config)
model = model.cpu()
with torch.no_grad():
    for name, p in sorted(model.named_parameters()):
        torch.nn.init.normal_(p, 0, 0.2)
        print(name, p.shape)
model.save_pretrained(save_folder)

</details>

Printing the model:

<details><summary>Click to expand</summary>

Qwen3_5MoeForConditionalGeneration(
  (model): Qwen3_5MoeModel(
    (visual): Qwen3_5MoeVisionModel(
      (patch_embed): Qwen3_5MoeVisionPatchEmbed(
        (proj): Conv3d(3, 64, kernel_size=(2, 16, 16), stride=(2, 16, 16))
      )
      (pos_embed): Embedding(2304, 64)
      (rotary_pos_emb): Qwen3_5MoeVisionRotaryEmbedding()
      (blocks): ModuleList(
        (0-1): 2 x Qwen3_5MoeVisionBlock(
          (norm1): LayerNorm((64,), eps=1e-06, elementwise_affine=True)
          (norm2): LayerNorm((64,), eps=1e-06, elementwise_affine=True)
          (attn): Qwen3_5MoeVisionAttention(
            (qkv): Linear(in_features=64, out_features=192, bias=True)
            (proj): Linear(in_features=64, out_features=64, bias=True)
          )
          (mlp): Qwen3_5MoeVisionMLP(
            (linear_fc1): Linear(in_features=64, out_features=128, bias=True)
            (linear_fc2): Linear(in_features=128, out_features=64, bias=True)
            (act_fn): GELUTanh()
          )
        )
      )
      (merger): Qwen3_5MoeVisionPatchMerger(
        (norm): LayerNorm((64,), eps=1e-06, elementwise_affine=True)
        (linear_fc1): Linear(in_features=256, out_features=256, bias=True)
        (act_fn): GELU(approximate='none')
        (linear_fc2): Linear(in_features=256, out_features=8, bias=True)
      )
    )
    (language_model): Qwen3_5MoeTextModel(
      (embed_tokens): Embedding(248320, 8)
      (layers): ModuleList(
        (0-2): 3 x Qwen3_5MoeDecoderLayer(
          (linear_attn): Qwen3_5MoeGatedDeltaNet(
            (act): SiLUActivation()
            (conv1d): Conv1d(8192, 8192, kernel_size=(4,), stride=(1,), padding=(3,), groups=8192, bias=False)
            (norm): Qwen3_5MoeRMSNormGated()
            (out_proj): Linear(in_features=4096, out_features=8, bias=False)
            (in_proj_qkv): Linear(in_features=8, out_features=8192, bias=False)
            (in_proj_z): Linear(in_features=8, out_features=4096, bias=False)
            (in_proj_b): Linear(in_features=8, out_features=32, bias=False)
            (in_proj_a): Linear(in_features=8, out_features=32, bias=False)
          )
          (mlp): Qwen3_5MoeSparseMoeBlock(
            (gate): Qwen3_5MoeTopKRouter()
            (experts): Qwen3_5MoeExperts(
              (act_fn): SiLUActivation()
            )
            (shared_expert): Qwen3_5MoeMLP(
              (gate_proj): Linear(in_features=8, out_features=32, bias=False)
              (up_proj): Linear(in_features=8, out_features=32, bias=False)
              (down_proj): Linear(in_features=32, out_features=8, bias=False)
              (act_fn): SiLUActivation()
            )
            (shared_expert_gate): Linear(in_features=8, out_features=1, bias=False)
          )
          (input_layernorm): Qwen3_5MoeRMSNorm((8,), eps=1e-06)
          (post_attention_layernorm): Qwen3_5MoeRMSNorm((8,), eps=1e-06)
        )
        (3): Qwen3_5MoeDecoderLayer(
          (self_attn): Qwen3_5MoeAttention(
            (q_proj): Linear(in_features=8, out_features=512, bias=False)
            (k_proj): Linear(in_features=8, out_features=128, bias=False)
            (v_proj): Linear(in_features=8, out_features=128, bias=False)
            (o_proj): Linear(in_features=256, out_features=8, bias=False)
            (q_norm): Qwen3_5MoeRMSNorm((32,), eps=1e-06)
            (k_norm): Qwen3_5MoeRMSNorm((32,), eps=1e-06)
          )
          (mlp): Qwen3_5MoeSparseMoeBlock(
            (gate): Qwen3_5MoeTopKRouter()
            (experts): Qwen3_5MoeExperts(
              (act_fn): SiLUActivation()
            )
            (shared_expert): Qwen3_5MoeMLP(
              (gate_proj): Linear(in_features=8, out_features=32, bias=False)
              (up_proj): Linear(in_features=8, out_features=32, bias=False)
              (down_proj): Linear(in_features=32, out_features=8, bias=False)
              (act_fn): SiLUActivation()
            )
            (shared_expert_gate): Linear(in_features=8, out_features=1, bias=False)
          )
          (input_layernorm): Qwen3_5MoeRMSNorm((8,), eps=1e-06)
          (post_attention_layernorm): Qwen3_5MoeRMSNorm((8,), eps=1e-06)
        )
      )
      (norm): Qwen3_5MoeRMSNorm((8,), eps=1e-06)
      (rotary_emb): Qwen3_5MoeTextRotaryEmbedding()
    )
  )
  (lm_head): Linear(in_features=8, out_features=248320, bias=False)
  (mtp): ModuleDict(
    (pre_fc_norm_embedding): RMSNorm((8,), eps=None, elementwise_affine=True)
    (fc): Linear(in_features=16, out_features=8, bias=False)
    (layers): ModuleList(
      (0): Qwen3_5MoeDecoderLayer(
        (self_attn): Qwen3_5MoeAttention(
          (q_proj): Linear(in_features=8, out_features=512, bias=False)
          (k_proj): Linear(in_features=8, out_features=128, bias=False)
          (v_proj): Linear(in_features=8, out_features=128, bias=False)
          (o_proj): Linear(in_features=256, out_features=8, bias=False)
          (q_norm): Qwen3_5MoeRMSNorm((32,), eps=1e-06)
          (k_norm): Qwen3_5MoeRMSNorm((32,), eps=1e-06)
        )
        (mlp): Qwen3_5MoeSparseMoeBlock(
          (gate): Qwen3_5MoeTopKRouter()
          (experts): Qwen3_5MoeExperts(
            (act_fn): SiLUActivation()
          )
          (shared_expert): Qwen3_5MoeMLP(
            (gate_proj): Linear(in_features=8, out_features=32, bias=False)
            (up_proj): Linear(in_features=8, out_features=32, bias=False)
            (down_proj): Linear(in_features=32, out_features=8, bias=False)
            (act_fn): SiLUActivation()
          )
          (shared_expert_gate): Linear(in_features=8, out_features=1, bias=False)
        )
        (input_layernorm): Qwen3_5MoeRMSNorm((8,), eps=1e-06)
        (post_attention_layernorm): Qwen3_5MoeRMSNorm((8,), eps=1e-06)
      )
    )
    (norm): RMSNorm((8,), eps=None, elementwise_affine=True)
    (pre_fc_norm_hidden): RMSNorm((8,), eps=None, elementwise_affine=True)
  )
)

</details>

Test environment:

• torch: 2.11.0
• transformers: 5.5.0