{
  "kind": "quatrain_qwen38_ple_refinement_export",
  "schema_version": 1,
  "source_model": "Qwen/Qwen3.8-Flash-Next",
  "source_revision": "de4b8e4d43b917e7706784d8bb445c9af86a3540",
  "step": 1000,
  "tensor_count": 299817,
  "tensor_bytes": 110091566076,
  "shards": 41,
  "precision": {
    "routed_experts": "nvfp4_group16",
    "ple_tables": "nvfp4_group16",
    "shared_experts": "mxfp8_block32",
    "trained_norms": "float32",
    "attention_ple_projections_convolution_embeddings_lm_head": "source_bf16",
    "vision": "source_precision_unmodified",
    "mtp_experts": "mxfp8_block32",
    "mtp_non_expert": "source_bf16"
  },
  "activation_calibration": {
    "progress": {
      "documents": 59753,
      "seconds": 3365.1721622999994,
      "step": 863,
      "steps": 863,
      "tokens": 100134710
    },
    "policy": {
      "input_scale_divisor": 2688.0,
      "observed_layer_experts": 24570,
      "statistic": "absolute_maximum_over_all_observed_rows_v1",
      "w13": "layer_max_measured_scale_repeated_for_every_expert_v1",
      "w2_fallback_layer_experts": [
        {
          "expert": 181,
          "exported_scale": 0.01422991044819355,
          "layer": 0
        },
        {
          "expert": 193,
          "exported_scale": 0.01422991044819355,
          "layer": 0
        },
        {
          "expert": 244,
          "exported_scale": 0.01422991044819355,
          "layer": 0
        },
        {
          "expert": 271,
          "exported_scale": 0.01422991044819355,
          "layer": 0
        },
        {
          "expert": 413,
          "exported_scale": 0.01422991044819355,
          "layer": 0
        },
        {
          "expert": 424,
          "exported_scale": 0.01422991044819355,
          "layer": 0
        }
      ],
      "w2_observed": "preserve_measured_per_expert_scale_v1",
      "w2_unobserved": "same_layer_max_measured_w2_scale_v1"
    },
    "complete_expert_observation_coverage": false,
    "source_trunk_step": 5500,
    "recalibrated_after_ple_refinement": false
  },
  "validation": {
    "inventory": "exact names/shapes/dtypes",
    "packing": "every refined PLE table decoded and checked against training NVFP4 QDQ",
    "write_readback": "all emitted tensors compared byte-for-byte",
    "payload_hashes": false,
    "full_model_serving": "not run",
    "mtp_packing": "every decoded matrix checked against the block32 MXFP8 reference/QDQ",
    "unmodified_shards": "independent byte copies; unchanged tensor inventory",
    "mtp_acceptance_rate": "not measured",
    "replicated_ple_norms": "exact agreement across all checkpoint owners",
    "base_preservation": "unchanged shards copied independently; mixed shards retain exact non-PLE values"
  },
  "serving_requirements": [
    "group16 E2M1/E4M3 PLE table loader and gather with one FP32 global scale",
    "FP32 trained PLE and hyperconnection normalization weights",
    "MXFP8 shared and MTP routed experts with E4M3 values and UE8M0 block32 scales"
  ],
  "mtp_quantization": {
    "method": "post_training",
    "experts": 512,
    "block_shape": [
      1,
      32
    ],
    "values": "float8_e4m3fn",
    "scales": "uint8_ue8m0",
    "tensor_bytes_saved": 2437939200
  },
  "trunk_step": 5500,
  "ple_refinement": {
    "step": 1000,
    "base_export": "Qwen3.8-Flash-Next-NVFP4-QAD-step5500-mxfp8-mtp-v1",
    "trainable": "PLE table masters and FP32 PLE norms only",
    "global_scale": 3.324236240587197e-05,
    "global_scale_policy": {
      "quantization": "nvfp4_group16_checkpoint_global_amax_v1",
      "scale_source": "initial_checkpoint_ple_table_masters",
      "global_scale_divisor": 2688.0
    }
  }
}