quark-torch-model-intake

Model validation for Quark PTQ checks architecture, layer counts, and quantization readiness without loading full model weights. It inspects model configurations to identify transformer compatibility constraints, custom loading needs like trust_remote_code, and risk factors affecting post-training quantization.

Architecture analysis detects model family, hidden layer counts, and MoE status by inspecting the model path. This identifies the quantizable layer structure and determines whether MoE layers require extra preparation before applying post-training quantization.

Compatibility checks identify transformers version constraints and whether custom loading needs like trust_remote_code are required. This validation assesses risk factors and special template handling needs to ensure the model architecture is ready for post-training quantization.

Analyzing local model directories for quantization faces limitations from missing config details and unsupported loading requirements. The assessment extracts structural facts without loading full model weights, so incomplete configurations can prevent accurate risk evaluation and quantization readiness checks.

Risk assessment for quantization validates transformer compatibility and checks configuration details without loading full model weights. It extracts structural facts like architecture and MoE status to identify risk factors that could cause downstream failures during post-training quantization.