Fix Mistral-Small-3.1-24B contrib: get_neuron_config_cls returns wrong class

[jimburtoft]

Description

This is a fix to an existing contrib model, not a new contribution.

The Mistral3InferenceConfig.get_neuron_config_cls() method returns the base NeuronConfig instead of Mistral3NeuronConfig. This means the custom attn_cls field defined in Mistral3NeuronConfig is never used.

Changes

One-line fix in modeling_mistral3.py: change get_neuron_config_cls to return Mistral3NeuronConfig instead of NeuronConfig.

This is a correctness fix -- the decoder layer currently hardcodes NeuronMistral3Attention directly so the attn_cls field on Mistral3NeuronConfig is effectively dead code, but the config class should still return the correct subclass for consistency and to enable future use of attn_cls.

Validation Results

Metric	Value
Instance	trn2.3xlarge (spot, $0.90/hr)
TP	4 (LNC=2)
Compile time	258s
Throughput	~34.5 tok/s
Token match	88% (4/5 prompts 100% match)

Files Changed

File	Change
contrib/models/Mistral-Small-3.1-24B-Instruct-2503/src/modeling_mistral3.py	Fix get_neuron_config_cls return type

Testing

Compiled and ran on trn2.3xlarge with TP=4, Neuron SDK 2.27, BF16. Model generates coherent text with 88% token match (4 of 5 prompts exact match against HuggingFace CPU reference).