{ "status": "pass", "loader": "dflash 0.1.0 / Transformers 5.15.0", "architecture": "DFlash2DraftModel", "parameters": 1924404480, "checkpoint": "selective native block-FP8 with BF16 preserved modules", "test_hardware": "NVIDIA RTX 3090 (SM 8.6)", "observed_behavior": "Transformers loaded all 81 source tensors and dequantized serialized FP8 matrices to BF16 because SM 8.6 lacks native W8A8 FP8.", "module_routing": { "mlp_gate_projection_after_load": "bfloat16 (dequantized from FP8)", "q_projection_after_load": "bfloat16 (preserved)", "candidate_selector_projection_after_load": "bfloat16 (preserved)" }, "limitation": "This confirms checkpoint/metadata compatibility, not native FP8 kernel execution." }