| { |
| "capture": { |
| "capture_content_digest": "ba15dfe4de3e9b8dfeb9b23461aa894b541dbd19f1560ec2d49cb0f8e7989d40", |
| "dtype": "BF16", |
| "dtype_lossless": true, |
| "form": "hidden", |
| "head_not_separable_reason": null, |
| "head_separable": true, |
| "hidden_width": 6144, |
| "lossy_codec": null, |
| "manifest_file": "capture/manifest.json", |
| "manifest_file_sha256": "8234e9f154d04812714fb58d8d05d91b200545c16ff1ad7904a96e171221b0bc", |
| "records_count": 25, |
| "scored_rows_total": 51175, |
| "semantic_point": "after_final_rmsnorm_before_lm_head", |
| "tensor_key": "hidden_states", |
| "total_size_bytes": 628845100, |
| "vocab_size": 154880 |
| }, |
| "coverage": { |
| "complete": true, |
| "declared_records": 25, |
| "index_range": [ |
| 0, |
| 24 |
| ], |
| "missing_indices_sample": [], |
| "present_records": 25, |
| "shard_of": null, |
| "subset_detail": null |
| }, |
| "created_utc": "2026-09-05T12:21:34Z", |
| "dataset": { |
| "author": { |
| "handle": null, |
| "is_registry_maintainer": false, |
| "name": "malaiwah", |
| "role": "capture-author", |
| "url": null |
| }, |
| "base_capture": null, |
| "id": "fidelity--glm53.malaiwah.quant.nvfp4-incoai", |
| "license": "mit", |
| "name": "fidelity--glm53.malaiwah.quant.nvfp4-incoai", |
| "qualification": null, |
| "repository": "malaiwah/glm53-fidelity-nvfp4-incoai-v1", |
| "revision": null, |
| "role": "quant", |
| "structural_status": "sealed" |
| }, |
| "dataset_sha256": "a241052c7d6ffede08b03ea28d9e92851fad05d95ec2c4d0c8e04bef08113f7b", |
| "determinism": { |
| "cold_start_per_run": true, |
| "distinct_evidence_hash_count": 1, |
| "evidence_hashes": [ |
| "ba15dfe4de3e9b8dfeb9b23461aa894b541dbd19f1560ec2d49cb0f8e7989d40" |
| ], |
| "evidence_kind": "hidden_state_tensor_sha256", |
| "identical_across_runs": null, |
| "note": "a single cold capture; cross-run determinism is NOT asserted here. Capture a second cold run and run `fidelity-dataset compare --self-compare` -- an exactly 0.0 result is the SC-1 reproduction confirmation.", |
| "repeat_noise": null, |
| "repeats": [], |
| "run_count": 1 |
| }, |
| "disclosures": [ |
| { |
| "affects_comparability": false, |
| "code": "no_known_deviations", |
| "detail": "captured by engines/tools/hf_capture.py", |
| "severity": "info" |
| }, |
| { |
| "affects_comparability": false, |
| "code": "checkpoint_tensors_intentionally_unused", |
| "detail": "791 checkpoint tensor(s) were present but intentionally unused by this architecture. Their complete set exactly matched pinned allowlist 969cee605deba10dd82afbaa9b1b7a35d2339fdb122efece14530c9030aa1436 (canonical sorted-name SHA-256 61e5f26aed8bca408c5de5347d8e1668b0c5716237dad1fc98c47bc108f4ae57).", |
| "severity": "caveat" |
| }, |
| { |
| "affects_comparability": false, |
| "code": "activation_scales_not_applied", |
| "detail": "the artifact ships a static per-tensor `input_scale` beside each of its 57600 routed-expert modules (a W4A4 serving kernel would quantize activations with it); this is a weights-only decode-and-run measurement, 57600 such tensors were read and NOT applied (activation_scheme static-nvfp4-not-applied), so activation quantization is not captured by the measured KLD -- the same caveat family as the official FP8 release's dynamic activation scheme.", |
| "severity": "caveat" |
| }, |
| { |
| "affects_comparability": false, |
| "code": "reduced_run_count", |
| "detail": "this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus `fidelity-dataset compare --self-compare`, whose exactly-0.0 result is the SC-1 reproduction confirmation.", |
| "severity": "caveat" |
| } |
| ], |
| "format_version": 1, |
| "generation_sanity_probe": { |
| "degenerate": false, |
| "enforced": true, |
| "entropy_nats": 3.6383363063160545, |
| "expect": "Paris", |
| "expected_first_ids": [ |
| 1732, |
| 12089, |
| 40631, |
| 59190 |
| ], |
| "logit_spread": 26.375, |
| "note": "one extra window through the same schedule; its hidden state is discarded and never enters the dataset.", |
| "prompt": "The capital of France is", |
| "schema": "malaiwah.generation-sanity-probe.v1", |
| "status": "pass", |
| "top1_id": 12089, |
| "top1_probability": 0.28770518644792514, |
| "top1_text": " Paris", |
| "top_k": [ |
| { |
| "id": 12089, |
| "probability": 0.28770518644792514, |
| "text": " Paris" |
| }, |
| { |
| "id": 264, |
| "probability": 0.09340419865211294, |
| "text": " a" |
| }, |
| { |
| "id": 825, |
| "probability": 0.09340419865211294, |
| "text": " one" |
| }, |
| { |
| "id": 279, |
| "probability": 0.06419570432765939, |
| "text": " the" |
| }, |
| { |
| "id": 2114, |
| "probability": 0.03893666289657242, |
| "text": " home" |
| } |
| ], |
| "uniform_entropy_nats": 11.950405902510404, |
| "vocab_size": 154880 |
| }, |
| "head": { |
| "applied_in_capture": false, |
| "bias": null, |
| "bits": 16, |
| "compat_tensor_key": "weight", |
| "dtype": "BF16", |
| "equality_receipt": null, |
| "file": "head/weight.safetensors", |
| "file_sha256": "f722d591417749cee9de8ceb977b5f11cf5a1254e2dbf6409df9d170ac3be9b5", |
| "final_norm": { |
| "applied_at_replay": false, |
| "applied_in_capture": true, |
| "dtype": null, |
| "file": null, |
| "file_sha256": null, |
| "shape": null, |
| "tensor_content_sha256": null, |
| "tensor_key": null |
| }, |
| "head_json": "head/head.json", |
| "present": true, |
| "quantized": false, |
| "raw_tensor_sha256": "864f488a0074d236062a4f24800940df7ba611dba76238d4f9f847d9659e11cd", |
| "shape": [ |
| 154880, |
| 6144 |
| ], |
| "source": "native", |
| "tensor_content_sha256": "864f488a0074d236062a4f24800940df7ba611dba76238d4f9f847d9659e11cd", |
| "tensor_key": "lm_head.weight" |
| }, |
| "interop": { |
| "adapted_from": null, |
| "compatible_with": [ |
| "kimi-k3-distribution-fidelity/1" |
| ], |
| "divergences": [ |
| { |
| "field": "head", |
| "id": "D-1", |
| "reason": "every capture declares its OWN head identity by tensor content digest; the comparator refuses a hidden-form comparison across differing heads." |
| }, |
| { |
| "field": "runtime.lane", |
| "id": "D-2", |
| "reason": "lane is required and gated at compare time so registry BIAS-006 is checked before a number exists, not at submission." |
| }, |
| { |
| "field": "panel.scoring_window", |
| "id": "D-3", |
| "reason": "scoring_window is part of panel identity, not a comparator flag: a score_from=1024 number is a different panel, not a variant." |
| }, |
| { |
| "field": "comparison_kind", |
| "id": "D-4", |
| "reason": "self-compare is a first-class mode asserting exactly 0.0." |
| }, |
| { |
| "field": "estimator.accumulation_dtype", |
| "id": "D-5", |
| "reason": "required, fp64 default; an fp32 estimator lands in a different comparability class on the same data and that must be visible." |
| }, |
| { |
| "field": "coverage", |
| "id": "D-6", |
| "reason": "declared vs present record counts with shard_of; fixes our own manifest that claimed 5,120 captures in a 512-file repository." |
| }, |
| { |
| "field": "panel", |
| "id": "D-7", |
| "reason": "the panel is a separately referenceable object, which is what makes a candidate capture publishable standalone." |
| }, |
| { |
| "field": "capture.lossy_codec", |
| "id": "D-8", |
| "reason": "lets a llama.cpp .kld be ingested without laundering a lossy capture as exact." |
| }, |
| { |
| "field": "capture.records[].attention_mask_sha256", |
| "id": "D-9", |
| "reason": "our packed and streaming lanes vary mask construction; a single-request capture path makes it invariant and has no equivalent field." |
| }, |
| { |
| "field": "capture.head_separable", |
| "id": "D-10", |
| "reason": "hidden form is the default; logit form must declare head_separable:false with a reason." |
| } |
| ], |
| "k3_compat_emitted": false, |
| "note": null |
| }, |
| "panel": { |
| "contamination": { |
| "checked": false, |
| "hits": null, |
| "method": "not established", |
| "receipt": null |
| }, |
| "context_length": 2048, |
| "contexts": 25, |
| "panel_file": "panel/panel.json", |
| "panel_file_sha256": "a71d495140a7fc53387547d4cc87d610232c57303a059b57a62fd4f9333f4a4f", |
| "panel_id": "panel--glm53.malaiwah.corpus5x5-v1", |
| "panel_receipt_file": "panel/panel-receipt.json", |
| "panel_receipt_sha256": "9c3bc4f59e8825ac78b366ca0e2988a48ecdbcaf26fdaeaedd3251edb6f9a828", |
| "panel_token_sha256_legacy": "d177a85308683012b4289e35770ef08e4bdc8e17adf59c5d26de3e5da9479734", |
| "remap_file": null, |
| "repository": null, |
| "revision": null, |
| "scored_positions_total": 51175, |
| "scoring_window": { |
| "dropped_positions_total": 0, |
| "min_left_context_tokens": 1, |
| "policy": "every causal prediction position of every window is scored; nothing is dropped", |
| "score_from": 0, |
| "windowed": false |
| }, |
| "suite_token_hash_sha256": "f09ee395f635225a077695ed193d61f7d1e70650cebe1b68f82b16f59d399f86", |
| "tokenizer": { |
| "files": [ |
| { |
| "bytes": 4263, |
| "name": "LICENSE", |
| "sha256": "96e1622099fc9d6b70c9760f007d99e66d7497eec636b63c60fe208401e9170c" |
| }, |
| { |
| "bytes": 10465, |
| "name": "chat_template.jinja", |
| "sha256": "69bb3ab52067898e2466b855407636de559568947f367945842aabcb7fcc1705" |
| }, |
| { |
| "bytes": 3732, |
| "name": "config.json", |
| "sha256": "ca8f2f47b07919a514c0ca223dc2ea2bc7445afaa5ac76c013a3784e096426ca" |
| }, |
| { |
| "bytes": 194, |
| "name": "generation_config.json", |
| "sha256": "ac76b43d8683d3b930126870fc8be73d8679308fe752fa1f381096d8354f6a55" |
| }, |
| { |
| "bytes": 20217442, |
| "name": "tokenizer.json", |
| "sha256": "19e773648cb4e65de8660ea6365e10acca112d42a854923df93db4a6f333a82d" |
| }, |
| { |
| "bytes": 761, |
| "name": "tokenizer_config.json", |
| "sha256": "98b1271574f41abf89427ae2dda030d94dc9478f0edc5a8bd240db213c6fd5fc" |
| } |
| ], |
| "files_verified": true, |
| "id": "zai-org/GLM-5.3-BF16", |
| "identity_sha256": "b18e4d378be7e75fd2b323f7c81cc65640f6a08e1842d3bf4e025d88a9d78bf9", |
| "maximum_token_id_exclusive": 154820, |
| "receipt": null, |
| "repository": "zai-org/GLM-5.3-BF16", |
| "revision": "304b8051cfb2b260b61ce0cbe330e02a98e73639", |
| "vocab_size": 154820 |
| } |
| }, |
| "runtime": { |
| "backend_identity_sha256": null, |
| "file": "runtime/capture-runtime.json", |
| "file_sha256": "6bbba771d9a1a75d3b4ca262a3fdd9fe61f6f317de65351d3d16c397f58404b8", |
| "lane": "streaming", |
| "lane_identity_sha256": "025ae83c1cb0e2c518baa8bd0c46965f0c17b6aaca5b1b875ca5fcea71f565d5", |
| "lane_inferred": false, |
| "runtime_reader_sha256": null, |
| "source": "native", |
| "stack_fingerprint_sha256": "e7ddf6b280474c1709836e227b55dda1247b32148313bc87fd88a2ada704f0e0" |
| }, |
| "schema": "malaiwah.fidelity-dataset.v1", |
| "scope": { |
| "assignments": [ |
| { |
| "bits_per_weight": 16, |
| "format": "bf16", |
| "layer_range": "all", |
| "note": "78 tensors: model.layers.N.self_attn.o_proj.weight. read from incoai/GLM-5.3-NVFP4@54e52520606f96b3d9fc84088ad22882a61648ac model.safetensors.index.json + every shard header (ranged HTTP): a routed expert projection is nvfp4 (e2m1, group 16 along the input axis, weight U8 [out, in/2] + weight_scale F8_E4M3 [out, in/16] + weight_scale_2 F32 scalar; 57600 such modules) when it ships the modelopt component set, native at its stored dtype when it ships whole; every other tensor native at its stored dtype. Declared by quantization_config: quant_method modelopt, quant_algo NVFP4, config_groups.group_0.weights num_bits=4 group_size=16, producer modelopt 0.45.0. 57600 modules also ship a per-tensor F32 input_scale (a static activation quantity: recorded, never applied by the weights-only decode).", |
| "tensor_class": "attn.o", |
| "treatment": "native" |
| }, |
| { |
| "bits_per_weight": 16, |
| "format": "bf16", |
| "layer_range": "all", |
| "note": "261 tensors: model.layers.N.self_attn.indexer.k_norm.bias, model.layers.N.self_attn.indexer.k_norm.weight, model.layers.N.self_attn.indexer.weights_proj.weight, model.layers.N.self_attn.indexer.wk.weight, model.layers.N.self_attn.indexer.wq_b.weight, model.layers.N.self_attn.kv_a_layernorm.weight. read from incoai/GLM-5.3-NVFP4@54e52520606f96b3d9fc84088ad22882a61648ac model.safetensors.index.json + every shard header (ranged HTTP): a routed expert projection is nvfp4 (e2m1, group 16 along the input axis, weight U8 [out, in/2] + weight_scale F8_E4M3 [out, in/16] + weight_scale_2 F32 scalar; 57600 such modules) when it ships the modelopt component set, native at its stored dtype when it ships whole; every other tensor native at its stored dtype. Declared by quantization_config: quant_method modelopt, quant_algo NVFP4, config_groups.group_0.weights num_bits=4 group_size=16, producer modelopt 0.45.0. 57600 modules also ship a per-tensor F32 input_scale (a static activation quantity: recorded, never applied by the weights-only decode).", |
| "tensor_class": "attn.other", |
| "treatment": "native" |
| }, |
| { |
| "bits_per_weight": 16, |
| "format": "bf16", |
| "layer_range": "all", |
| "note": "312 tensors: model.layers.N.self_attn.kv_a_proj_with_mqa.weight, model.layers.N.self_attn.kv_b_proj.weight, model.layers.N.self_attn.q_a_proj.weight, model.layers.N.self_attn.q_b_proj.weight. read from incoai/GLM-5.3-NVFP4@54e52520606f96b3d9fc84088ad22882a61648ac model.safetensors.index.json + every shard header (ranged HTTP): a routed expert projection is nvfp4 (e2m1, group 16 along the input axis, weight U8 [out, in/2] + weight_scale F8_E4M3 [out, in/16] + weight_scale_2 F32 scalar; 57600 such modules) when it ships the modelopt component set, native at its stored dtype when it ships whole; every other tensor native at its stored dtype. Declared by quantization_config: quant_method modelopt, quant_algo NVFP4, config_groups.group_0.weights num_bits=4 group_size=16, producer modelopt 0.45.0. 57600 modules also ship a per-tensor F32 input_scale (a static activation quantity: recorded, never applied by the weights-only decode).", |
| "tensor_class": "attn.qkv", |
| "treatment": "native" |
| }, |
| { |
| "bits_per_weight": 16, |
| "format": "bf16", |
| "layer_range": "all", |
| "note": "1 tensors: model.embed_tokens.weight. read from incoai/GLM-5.3-NVFP4@54e52520606f96b3d9fc84088ad22882a61648ac model.safetensors.index.json + every shard header (ranged HTTP): a routed expert projection is nvfp4 (e2m1, group 16 along the input axis, weight U8 [out, in/2] + weight_scale F8_E4M3 [out, in/16] + weight_scale_2 F32 scalar; 57600 such modules) when it ships the modelopt component set, native at its stored dtype when it ships whole; every other tensor native at its stored dtype. Declared by quantization_config: quant_method modelopt, quant_algo NVFP4, config_groups.group_0.weights num_bits=4 group_size=16, producer modelopt 0.45.0. 57600 modules also ship a per-tensor F32 input_scale (a static activation quantity: recorded, never applied by the weights-only decode).", |
| "tensor_class": "embed_tokens", |
| "treatment": "native" |
| }, |
| { |
| "bits_per_weight": 16, |
| "format": "bf16", |
| "layer_range": "all", |
| "note": "1 tensors: lm_head.weight. read from incoai/GLM-5.3-NVFP4@54e52520606f96b3d9fc84088ad22882a61648ac model.safetensors.index.json + every shard header (ranged HTTP): a routed expert projection is nvfp4 (e2m1, group 16 along the input axis, weight U8 [out, in/2] + weight_scale F8_E4M3 [out, in/16] + weight_scale_2 F32 scalar; 57600 such modules) when it ships the modelopt component set, native at its stored dtype when it ships whole; every other tensor native at its stored dtype. Declared by quantization_config: quant_method modelopt, quant_algo NVFP4, config_groups.group_0.weights num_bits=4 group_size=16, producer modelopt 0.45.0. 57600 modules also ship a per-tensor F32 input_scale (a static activation quantity: recorded, never applied by the weights-only decode).", |
| "tensor_class": "lm_head", |
| "treatment": "native" |
| }, |
| { |
| "bits_per_weight": 16, |
| "format": "bf16", |
| "layer_range": "all", |
| "note": "3 tensors: model.layers.N.mlp.down_proj.weight. read from incoai/GLM-5.3-NVFP4@54e52520606f96b3d9fc84088ad22882a61648ac model.safetensors.index.json + every shard header (ranged HTTP): a routed expert projection is nvfp4 (e2m1, group 16 along the input axis, weight U8 [out, in/2] + weight_scale F8_E4M3 [out, in/16] + weight_scale_2 F32 scalar; 57600 such modules) when it ships the modelopt component set, native at its stored dtype when it ships whole; every other tensor native at its stored dtype. Declared by quantization_config: quant_method modelopt, quant_algo NVFP4, config_groups.group_0.weights num_bits=4 group_size=16, producer modelopt 0.45.0. 57600 modules also ship a per-tensor F32 input_scale (a static activation quantity: recorded, never applied by the weights-only decode).", |
| "tensor_class": "mlp.down", |
| "treatment": "native" |
| }, |
| { |
| "bits_per_weight": 16, |
| "format": "bf16", |
| "layer_range": "all", |
| "note": "3 tensors: model.layers.N.mlp.gate_proj.weight. read from incoai/GLM-5.3-NVFP4@54e52520606f96b3d9fc84088ad22882a61648ac model.safetensors.index.json + every shard header (ranged HTTP): a routed expert projection is nvfp4 (e2m1, group 16 along the input axis, weight U8 [out, in/2] + weight_scale F8_E4M3 [out, in/16] + weight_scale_2 F32 scalar; 57600 such modules) when it ships the modelopt component set, native at its stored dtype when it ships whole; every other tensor native at its stored dtype. Declared by quantization_config: quant_method modelopt, quant_algo NVFP4, config_groups.group_0.weights num_bits=4 group_size=16, producer modelopt 0.45.0. 57600 modules also ship a per-tensor F32 input_scale (a static activation quantity: recorded, never applied by the weights-only decode).", |
| "tensor_class": "mlp.gate", |
| "treatment": "native" |
| }, |
| { |
| "bits_per_weight": 16, |
| "format": "bf16", |
| "layer_range": "all", |
| "note": "3 tensors: model.layers.N.mlp.up_proj.weight. read from incoai/GLM-5.3-NVFP4@54e52520606f96b3d9fc84088ad22882a61648ac model.safetensors.index.json + every shard header (ranged HTTP): a routed expert projection is nvfp4 (e2m1, group 16 along the input axis, weight U8 [out, in/2] + weight_scale F8_E4M3 [out, in/16] + weight_scale_2 F32 scalar; 57600 such modules) when it ships the modelopt component set, native at its stored dtype when it ships whole; every other tensor native at its stored dtype. Declared by quantization_config: quant_method modelopt, quant_algo NVFP4, config_groups.group_0.weights num_bits=4 group_size=16, producer modelopt 0.45.0. 57600 modules also ship a per-tensor F32 input_scale (a static activation quantity: recorded, never applied by the weights-only decode).", |
| "tensor_class": "mlp.up", |
| "treatment": "native" |
| }, |
| { |
| "bits_per_weight": 4.0, |
| "format": "nvfp4", |
| "layer_range": "all", |
| "note": "57600 tensors: model.layers.N.mlp.experts.E.down_proj.weight, model.layers.N.mlp.experts.E.gate_proj.weight, model.layers.N.mlp.experts.E.up_proj.weight. read from incoai/GLM-5.3-NVFP4@54e52520606f96b3d9fc84088ad22882a61648ac model.safetensors.index.json + every shard header (ranged HTTP): a routed expert projection is nvfp4 (e2m1, group 16 along the input axis, weight U8 [out, in/2] + weight_scale F8_E4M3 [out, in/16] + weight_scale_2 F32 scalar; 57600 such modules) when it ships the modelopt component set, native at its stored dtype when it ships whole; every other tensor native at its stored dtype. Declared by quantization_config: quant_method modelopt, quant_algo NVFP4, config_groups.group_0.weights num_bits=4 group_size=16, producer modelopt 0.45.0. 57600 modules also ship a per-tensor F32 input_scale (a static activation quantity: recorded, never applied by the weights-only decode).", |
| "tensor_class": "moe.experts", |
| "treatment": "quantized" |
| }, |
| { |
| "bits_per_weight": null, |
| "format": "mixed", |
| "layer_range": "3-77", |
| "note": "class mixes formats on the same layers (SCOPE-004 admits one row per class and layer_range): 75 x native:bf16@16 (model.layers.N.mlp.gate.weight); 75 x native:fp32@32 (model.layers.N.mlp.gate.e_score_correction_bias). read from incoai/GLM-5.3-NVFP4@54e52520606f96b3d9fc84088ad22882a61648ac model.safetensors.index.json + every shard header (ranged HTTP): a routed expert projection is nvfp4 (e2m1, group 16 along the input axis, weight U8 [out, in/2] + weight_scale F8_E4M3 [out, in/16] + weight_scale_2 F32 scalar; 57600 such modules) when it ships the modelopt component set, native at its stored dtype when it ships whole; every other tensor native at its stored dtype. Declared by quantization_config: quant_method modelopt, quant_algo NVFP4, config_groups.group_0.weights num_bits=4 group_size=16, producer modelopt 0.45.0. 57600 modules also ship a per-tensor F32 input_scale (a static activation quantity: recorded, never applied by the weights-only decode).", |
| "tensor_class": "moe.router", |
| "treatment": "native" |
| }, |
| { |
| "bits_per_weight": 16, |
| "format": "bf16", |
| "layer_range": "all", |
| "note": "225 tensors: model.layers.N.mlp.shared_experts.down_proj.weight, model.layers.N.mlp.shared_experts.gate_proj.weight, model.layers.N.mlp.shared_experts.up_proj.weight. read from incoai/GLM-5.3-NVFP4@54e52520606f96b3d9fc84088ad22882a61648ac model.safetensors.index.json + every shard header (ranged HTTP): a routed expert projection is nvfp4 (e2m1, group 16 along the input axis, weight U8 [out, in/2] + weight_scale F8_E4M3 [out, in/16] + weight_scale_2 F32 scalar; 57600 such modules) when it ships the modelopt component set, native at its stored dtype when it ships whole; every other tensor native at its stored dtype. Declared by quantization_config: quant_method modelopt, quant_algo NVFP4, config_groups.group_0.weights num_bits=4 group_size=16, producer modelopt 0.45.0. 57600 modules also ship a per-tensor F32 input_scale (a static activation quantity: recorded, never applied by the weights-only decode).", |
| "tensor_class": "moe.shared_expert", |
| "treatment": "native" |
| }, |
| { |
| "bits_per_weight": null, |
| "format": "mixed", |
| "layer_range": "78", |
| "note": "class mixes formats on the same layers (SCOPE-004 admits one row per class and layer_range): 790 x native:bf16@16 (model.layers.N.eh_proj.weight, model.layers.N.enorm.weight, model.layers.N.hnorm.weight); 1 x native:fp32@32 (model.layers.N.mlp.gate.e_score_correction_bias). read from incoai/GLM-5.3-NVFP4@54e52520606f96b3d9fc84088ad22882a61648ac model.safetensors.index.json + every shard header (ranged HTTP): a routed expert projection is nvfp4 (e2m1, group 16 along the input axis, weight U8 [out, in/2] + weight_scale F8_E4M3 [out, in/16] + weight_scale_2 F32 scalar; 57600 such modules) when it ships the modelopt component set, native at its stored dtype when it ships whole; every other tensor native at its stored dtype. Declared by quantization_config: quant_method modelopt, quant_algo NVFP4, config_groups.group_0.weights num_bits=4 group_size=16, producer modelopt 0.45.0. 57600 modules also ship a per-tensor F32 input_scale (a static activation quantity: recorded, never applied by the weights-only decode).", |
| "tensor_class": "mtp", |
| "treatment": "native" |
| }, |
| { |
| "bits_per_weight": 16, |
| "format": "bf16", |
| "layer_range": "all", |
| "note": "157 tensors: model.layers.N.input_layernorm.weight, model.layers.N.post_attention_layernorm.weight, model.norm.weight. read from incoai/GLM-5.3-NVFP4@54e52520606f96b3d9fc84088ad22882a61648ac model.safetensors.index.json + every shard header (ranged HTTP): a routed expert projection is nvfp4 (e2m1, group 16 along the input axis, weight U8 [out, in/2] + weight_scale F8_E4M3 [out, in/16] + weight_scale_2 F32 scalar; 57600 such modules) when it ships the modelopt component set, native at its stored dtype when it ships whole; every other tensor native at its stored dtype. Declared by quantization_config: quant_method modelopt, quant_algo NVFP4, config_groups.group_0.weights num_bits=4 group_size=16, producer modelopt 0.45.0. 57600 modules also ship a per-tensor F32 input_scale (a static activation quantity: recorded, never applied by the weights-only decode).", |
| "tensor_class": "norm", |
| "treatment": "native" |
| } |
| ], |
| "head_policy": "native", |
| "kv_cache_dtype": "bf16", |
| "policy": "mixed", |
| "scope_digest": "attn.o=native:bf16@16|attn.other=native:bf16@16|attn.qkv=native:bf16@16|embed_tokens=native:bf16@16|lm_head=native:bf16@16|mlp.down=native:bf16@16|mlp.gate=native:bf16@16|mlp.up=native:bf16@16|moe.experts=quantized:nvfp4@4|moe.router=native:mixed|moe.shared_expert=native:bf16@16|mtp=native:mixed|norm=native:bf16@16|head=native|kv=bf16" |
| }, |
| "seal": { |
| "checksums_file": "checksums.txt", |
| "checksums_format": "sha256sum", |
| "checksums_sha256": "3ad756cfd2d44b2fe30b621ce02c95d8b97c2a763b0567034b90207aa99f88f8", |
| "covers": [ |
| "every published file except checksums.txt and fidelity-dataset.json, via checksums.txt", |
| "fidelity-dataset.json itself, via the self-blanked seal", |
| "every capture tensor a second time by CONTENT, via capture.capture_content_digest" |
| ], |
| "excludes": [ |
| "checksums.txt", |
| "fidelity-dataset.json" |
| ], |
| "external_anchor": null, |
| "method": "self_blanked_canonical_json_sha256", |
| "seal_field": "dataset_sha256" |
| }, |
| "upstream_receipts": [], |
| "weights": { |
| "artifact_ref": null, |
| "checkpoint_identity_sha256": "2bafa83a937934c04da9cbc20540babe9dddf945aa3926392497d25f1a8c58b6", |
| "codec": "nvfp4", |
| "config_sha256": "48025a5f559f7a747734be8499810a22d34330577e1566a28810466319f57ef4", |
| "declared_bits": 4.0, |
| "declared_head_bits": 16, |
| "index_sha256": null, |
| "model_ref": null, |
| "model_revision": "54e52520606f96b3d9fc84088ad22882a61648ac", |
| "quantized": true, |
| "repository": "incoai/GLM-5.3-NVFP4", |
| "revision": "54e52520606f96b3d9fc84088ad22882a61648ac" |
| } |
| } |
|
|
