vllm - 💡(How to fix) Fix [Bug]: EagleMistralLarge3Model crashes with AttributeError: 'use_mha' on vLLM 0.21.0 [1 pull requests]

Your current environment

Environment

• vLLM version: 0.21.0
• Model: mistralai/Mistral-Large-3-675B-Instruct-2512 with Eagle speculative decoding
• Hardware: 8x H200 NVL, TP=8

(APIServer pid=1) INFO 05-21 08:22:26 [utils.py:240] non-default args: {'model_tag': 'mistralai/Mistral-Large-3-675B-Instruct-2512-NVFP4', 'enable_auto_tool_choice': True, 'tool_call_parser': 'mistral', 'model': 'mistralai/Mistral-Large-3-675B-Instruct-2512-NVFP4', 'tokenizer_mode': 'mistral', 'served_model_name': ['mistral-large-3-675b-instruct-2512'], 'config_format': 'mistral', 'load_format': 'mistral', 'tensor_parallel_size': 8, 'speculative_config': {'model': 'mistralai/Mistral-Large-3-675B-Instruct-2512-Eagle', 'num_speculative_tokens': 3, 'method': 'eagle', 'max_model_len': '16384'}}

🐛 Describe the bug

Description

EagleMistralLarge3Model fails during weight loading with:

File "/usr/local/lib/python3.12/dist-packages/vllm/model_executor/models/deepseek_v2.py", line 1355, in load_weights
    if self.use_mha:
       ^^^^^^^^^^^^
AttributeError: 'EagleMistralLarge3Model' object has no attribute 'use_mha'

This worked in vLLM 0.19.x / 0.20.2 and regressed in 0.21.0.

Root Cause

PR #16383 moved load_weights from DeepseekV2ForCausalLM to DeepseekV2Model, which now references self.use_mha. DeepseekV2Model.__init__ sets this attribute, but EagleMistralLarge3Model.__init__ calls nn.Module.__init__(self) directly instead of super().__init__(), so use_mha is never initialized.

This is the same class of bug fixed in #37232 (missing aux_hidden_state_layers). The underlying fragility — bypassing the parent __init__ — was flagged by reviewers at the time.

How to Reproduce

vllm serve mistralai/Mistral-Large-3-675B-Instruct-2512 \
  --speculative-config '{"model": "mistralai/Mistral-Large-3-675B-Instruct-2512-Eagle", "num_speculative_tokens": 3, "method": "eagle"}'

Suggested Fix

Add self.use_mha = False to EagleMistralLarge3Model.__init__ in vllm/model_executor/models/mistral_large_3_eagle.py. Mistral Large 3 uses MLA, so this is always False.

Longer term, the bypassed super().__init__() pattern will keep causing these regressions whenever DeepseekV2Model.__init__ gains new attributes.

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.