Qwen3 30b Plateau 2skill 0617 Iter207

--- library_name: transformers license: apache-2.0 license_link: https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507/blob/main/LICENSE pipeline_tag: text-generation --- # Qwen3-30B-A3B-Instruct-2507 ## Highlights We introduce the updated version of the **Qwen3-30B-A3B non-thinking mode**, named **Qwen3-30B-A3B-Instruct-2507**, featuring the following key enhancements: - **Significant improvements** in general capabilities, including **instruction following, logical reasoning, text comprehension, mathematics, science, coding and tool usage**. - **Substantial gains** in long-tail knowledge coverage across **multiple languages**. - **Markedly better alignment** with user preferences in **subjective and open-ended tasks**, enabling more helpful responses and higher-quality text generation. - **Enhanced capabilities** in **256K long-context understanding**.  ## Model Overview **Qwen3-30B-A3B-Instruct-2507** has the following features: - Type: Causal Language Models - Training Stage: Pretraining & Post-training - Number of Parameters: 30.5B in total and 3.3B activated - Number of Paramaters (Non-Embedding): 29.9B - Number of Layers: 48 - Number of Attention Heads (GQA): 32 for Q and 4 for KV - Number of Experts: 128 - Number of Activated Experts: 8 - Context Length: **262,144 natively**. **NOTE: This model supports only non-thinking mode and does not generate `` `` blocks in its output. Meanwhile, specifying `enable_thinking=False` is no longer required.** For more details, including benchmark evaluation, hardware requirements, and inference performance, please refer to our [blog](https://qwenlm.github.io/blog/qwen3/), [GitHub](https://github.com/QwenLM/Qwen3), and [Documentation](https://qwen.readthedocs.io/en/latest/). ## Performance | | Deepseek-V3-0324 | GPT-4o-0327 | Gemini-2.5-Flash Non-Thinking | Qwen3-235B-A22B Non-Thinking | Qwen3-30B-A3B Non-Thinking | Qwen3-30B-A3B-Instruct-2507 | |--- | --- | --- | --- | --- | --- | --- | | **Knowledge** | | | | | | | | MMLU-Pro | **81.2** | 79.8 | 81.1 | 75.2 | 69.1 | 78.4 | | MMLU-Redux | 90.4 | **91.3** | 90.6 | 89.2 | 84.1 | 89.3 | | GPQA | 68.4 | 66.9 | **78.3** | 62.9 | 54.8 | 70.4 | | SuperGPQA | **57.3** | 51.0 | 54.6 | 48.2 | 42.2 | 53.4 | | **Reasoning** | | | | | | | | AIME25 | 46.6 | 26.7 | **61.6** | 24.7 | 21.6 | 61.3 | | HMMT25 | 27.5 | 7.9 | **45.8** | 10.0 | 12.0 | 43.0 | | ZebraLogic | 83.4 | 52.6 | 57.9 | 37.7 | 33.2 | **90.0** | | LiveBench 20241125 | 66.9 | 63.7 | **69.1** | 62.5 | 59.4 | 69.0 | | **Coding** | | | | | | | | LiveCodeBench v6 (25.02-25.05) | **45.2** | 35.8 | 40.1 | 32.9 | 29.0 | 43.2 | | MultiPL-E | 82.2 | 82.7 | 77.7 | 79.3 | 74.6 | **83.8** | | Aider-Polyglot | 55.1 | 45.3 | 44.0 | **59.6** | 24.4 | 35.6 | | **Alignment** | | | | | | | | IFEval | 82.3 | 83.9 | 84.3 | 83.2 | 83.7 | **84.7** | | Arena-Hard v2* | 45.6 | 61.9 | 58.3 | 52.0 | 24.8 | **69.0** | | Creative Writing v3 | 81.6 | 84.9 | 84.6 | 80.4 | 68.1 | **86.0** | | WritingBench | 74.5 | 75.5 | 80.5 | 77.0 | 72.2 | **85.5** | | **Agent** | | | | | | | | BFCL-v3 | 64.7 | 66.5 | 66.1 | **68.0** | 58.6 | 65.1 | | TAU1-Retail | 49.6 | 60.3# | **65.2** | 65.2 | 38.3 | 59.1 | | TAU1-Airline | 32.0 | 42.8# | **48.0** | 32.0 | 18.0 | 40.0 | | TAU2-Retail | **71.1** | 66.7# | 64.3 | 64.9 | 31.6 | 57.0 | | TAU2-Airline | 36.0 | 42.0# | **42.5** | 36.0 | 18.0 | 38.0 | | TAU2-Telecom | **34.0** | 29.8# | 16.9 | 24.6 | 18.4 | 12.3 | | **Multilingualism** | | | | | | | | MultiIF | 66.5 | 70.4 | 69.4 | 70.2 | **70.8** | 67.9 | | MMLU-ProX | 75.8 | 76.2 | **78.3** | 73.2 | 65.1 | 72.0 | | INCLUDE | 80.1 | 82.1 | **83.8** | 75.6 | 67.8 | 71.9 | | PolyMATH | 32.2 | 25.5 | 41.9 | 27.0 | 23.3 | **43.1** | *: For reproducibility, we report the win rates evaluated by GPT-4.1. \#: Results were generated using GPT-4o-20241120, as access to the native function calling API of GPT-4o-0327 was unavailable. ## Quickstart The code of Qwen3-MoE has been in the latest Hugging Face `transformers` and we advise you to use the latest version of `transformers`. With `transformers =0.4.6.post1` or `vllm>=0.8.5` or to create an OpenAI-compatible API endpoint: - SGLang: ```shell python -m sglang.launch_server --model-path Qwen/Qwen3-30B-A3B-Instruct-2507 --context-length 262144 ``` - vLLM: ```shell vllm serve Qwen/Qwen3-30B-A3B-Instruct-2507 --max-model-len 262144 ``` **Note: If you encounter out-of-memory (OOM) issues, consider reducing the context length to a shorter value, such as `32,768`.** For local use, applications such as Ollama, LMStudio, MLX-LM, llama.cpp, and KTransformers have also supported Qwen3. ## Agentic Use Qwen3 excels in tool calling capabilities. We recommend using [Qwen-Agent](https://github.com/QwenLM/Qwen-Agent) to make the best use of agentic ability of Qwen3. Qwen-Agent encapsulates tool-calling templates and tool-calling parsers internally, greatly reducing coding complexity. To define the available tools, you can use the MCP configuration file, use the integrated tool of Qwen-Agent, or integrate other tools by yourself. ```python from qwen_agent.agents import Assistant # Define LLM llm_cfg = { 'model': 'Qwen3-30B-A3B-Instruct-2507', # Use a custom endpoint compatible with OpenAI API: 'model_server': 'http://localhost:8000/v1', # api_base 'api_key': 'EMPTY', } # Define Tools tools = [ {'mcpServers': { # You can specify the MCP configuration file 'time': { 'command': 'uvx', 'args': ['mcp-server-time', '--local-timezone=Asia/Shanghai'] }, "fetch": { "command": "uvx", "args": ["mcp-server-fetch"] } } }, 'code_interpreter', # Built-in tools ] # Define Agent bot = Assistant(llm=llm_cfg, function_list=tools) # Streaming generation messages = [{'role': 'user', 'content': 'https://qwenlm.github.io/blog/ Introduce the latest developments of Qwen'}] for responses in bot.run(messages=messages): pass print(responses) ``` ## Processing Ultra-Long Texts To support **ultra-long context processing** (up to **1 million tokens**), we integrate two key techniques: - **[Dual Chunk Attention](https://arxiv.org/abs/2402.17463) (DCA)**: A length extrapolation method that splits long sequences into manageable chunks while preserving global coherence. - **[MInference](https://arxiv.org/abs/2407.02490)**: A sparse attention mechanism that reduces computational overhead by focusing on critical token interactions. Together, these innovations significantly improve both **generation quality** and **inference efficiency** for sequences beyond 256K tokens. On sequences approaching 1M tokens, the system achieves up to a **3× speedup** compared to standard attention implementations. For full technical details, see the [Qwen2.5-1M Technical Report](https://arxiv.org/abs/2501.15383). ### How to Enable 1M Token Context > [!NOTE] > To effectively process a 1 million token context, users will require approximately **240 GB** of total GPU memory. This accounts for model weights, KV-cache storage, and peak activation memory demands. #### Step 1: Update Configuration File Download the model and replace the content of your `config.json` with `config_1m.json`, which includes the config for length extrapolation and sparse attention. ```bash export MODELNAME=Qwen3-30B-A3B-Instruct-2507 huggingface-cli download Qwen/${MODELNAME} --local-dir ${MODELNAME} mv ${MODELNAME}/config.json ${MODELNAME}/config.json.bak mv ${MODELNAME}/config_1m.json ${MODELNAME}/config.json ``` #### Step 2: Launch Model Server After updating the config, proceed with either **vLLM** or **SGLang** for serving the model. #### Option 1: Using vLLM To run Qwen with 1M context support: ```bash pip install -U vllm \ --torch-backend=auto \ --extra-index-url https://wheels.vllm.ai/nightly ``` Then launch the server with Dual Chunk Flash Attention enabled: ```bash VLLM_ATTENTION_BACKEND=DUAL_CHUNK_FLASH_ATTN VLLM_USE_V1=0 \ vllm serve ./Qwen3-30B-A3B-Instruct-2507 \ --tensor-parallel-size 4 \ --max-model-len 1010000 \ --enable-chunked-prefill \ --max-num-batched-tokens 131072 \ --enforce-eager \ --max-num-seqs 1 \ --gpu-memory-utilization 0.85 ``` ##### Key Parameters | Parameter | Purpose | |--------|--------| | `VLLM_ATTENTION_BACKEND=DUAL_CHUNK_FLASH_ATTN` | Enables the custom attention kernel for long-context efficiency | | `--max-model-len 1010000` | Sets maximum context length to ~1M tokens | | `--enable-chunked-prefill` | Allows chunked prefill for very long inputs (avoids OOM) | | `--max-num-batched-tokens 131072` | Controls batch size during prefill; balances throughput and memory | | `--enforce-eager` | Disables CUDA graph capture (required for dual chunk attention) | | `--max-num-seqs 1` | Limits concurrent sequences due to extreme memory usage | | `--gpu-memory-utilization 0.85` | Set the fraction of GPU memory to be used for the model executor | #### Option 2: Using SGLang First, clone and install the specialized branch: ```bash git clone https://github.com/sgl-project/sglang.git cd sglang pip install -e "python[all]" ``` Launch the server with DCA support: ```bash python3 -m sglang.launch_server \ --model-path ./Qwen3-30B-A3B-Instruct-2507 \ --context-length 1010000 \ --mem-frac 0.75 \ --attention-backend dual_chunk_flash_attn \ --tp 4 \ --chunked-prefill-size 131072 ``` ##### Key Parameters | Parameter | Purpose | |---------|--------| | `--attention-backend dual_chunk_flash_attn` | Activates Dual Chunk Flash Attention | | `--context-length 1010000` | Defines max input length | | `--mem-frac 0.75` | The fraction of the memory used for static allocation (model weights and KV cache memory pool). Use a smaller value if you see out-of-memory errors. | | `--tp 4` | Tensor parallelism size (matches model sharding) | | `--chunked-prefill-size 131072` | Prefill chunk size for handling long inputs without OOM | #### Troubleshooting: 1. Encountering the error: "The model's max sequence length (xxxxx) is larger than the maximum number of tokens that can be stored in the KV cache." or "RuntimeError: Not enough memory. Please try to increase --mem-fraction-static." The VRAM reserved for the KV cache is insufficient. - vLLM: Consider reducing the ``max_model_len`` or increasing the ``tensor_parallel_size`` and ``gpu_memory_utilization``. Alternatively, you can reduce ``max_num_batched_tokens``, although this may significantly slow down inference. - SGLang: Consider reducing the ``context-length`` or increasing the ``tp`` and ``mem-frac``. Alternatively, you can reduce ``chunked-prefill-size``, although this may significantly slow down inference. 2. Encountering the error: "torch.OutOfMemoryError: CUDA out of memory." The VRAM reserved for activation weights is insufficient. You can try lowering ``gpu_memory_utilization`` or ``mem-frac``, but be aware that this might reduce the VRAM available for the KV cache. 3. Encountering the error: "Input prompt (xxxxx tokens) + lookahead slots (0) is too long and exceeds the capacity of the block manager." or "The input (xxx xtokens) is longer than the model's context length (xxx tokens)." The input is too lengthy. Consider using a shorter sequence or increasing the ``max_model_len`` or ``context-length``. #### Long-Context Performance We test the model on an 1M version of the [RULER](https://arxiv.org/abs/2404.06654) benchmark. | Model Name | Acc avg | 4k | 8k | 16k | 32k | 64k | 96k | 128k | 192k | 256k | 384k | 512k | 640k | 768k | 896k | 1000k | |---------------------------------------------|---------|------|------|------|------|------|------|------|------|------|------|------|------|------|------|-------| | Qwen3-30B-A3B (Non-Thinking) | 72.0 | 97.1 | 96.1 | 95.0 | 92.2 | 82.6 | 79.7 | 76.9 | 70.2 | 66.3 | 61.9 | 55.4 | 52.6 | 51.5 | 52.0 | 50.9 | | Qwen3-30B-A3B-Instruct-2507 (Full Attention) | 86.8 | 98.0 | 96.7 | 96.9 | 97.2 | 93.4 | 91.0 | 89.1 | 89.8 | 82.5 | 83.6 | 78.4 | 79.7 | 77.6 | 75.7 | 72.8 | |...