OBLITERATUS/docs/checkpoints/support-matrix-v1.json

{
  "schema_id": "obliteratus.checkpoint-support-matrix",
  "schema_version": "1.0.0",
  "generated_from": "5cc43c6e52903497574d80e08dff856028bc47f7",
  "status_vocabulary": ["supported", "conditional", "deferred", "out_of_scope"],
  "rows": [
    {
      "id": "hf-safetensors-existing-loader",
      "subject": "Existing Hugging Face-compatible safetensors input",
      "format": "hf_safetensors",
      "producer_versions": ["Transformers/Hugging Face compatible; exact model-family behavior varies"],
      "adapter": null,
      "model_mapping": "Existing Transformers model class and OBLITERATUS architecture gates",
      "state_scopes": ["model_weights"],
      "safety_level": "ordinary_hf_load",
      "optional_extras": [],
      "capabilities": {
        "detect": {"value": "conditional", "basis": "The standalone inspector classifies bounded HF safetensors structure without loading payloads; model compatibility is separate."},
        "safe_inspect": {"value": "conditional", "basis": "Current bounded JSON and safetensors-header inspection emits structural evidence only."},
        "trusted_inspect": {"value": "out_of_scope", "basis": "Ordinary supported HF safetensors do not need the planned vendor metadata path."},
        "weights_canonicalize": {"value": "out_of_scope", "basis": "Input is already the planned canonical format."},
        "topology_reshard": {"value": "out_of_scope", "basis": "HF file indexes do not encode rank-fragment topology."},
        "surgery": {"value": "conditional", "basis": "Current behavior depends on model architecture, dtype, quantization, kernels, memory, and quality gates."},
        "exact_resume": {"value": "out_of_scope", "basis": "OBLITERATUS is not a training-resume system."},
        "live_multi_node": {"value": "out_of_scope", "basis": "The current loader/runner has no multi-host process topology."}
      },
      "canonical_output": "Hugging Face-compatible safetensors from the current save path",
      "evidence": {
        "references": ["R11", "R12", "R13", "R21"],
        "candidate_commit": "5cc43c6e52903497574d80e08dff856028bc47f7",
        "fixture_digest": null,
        "environment": "Local baseline source and existing repository tests; not a universal model-family qualification",
        "topology": "One OBLITERATUS process; visible devices may receive complete modules",
        "retained_result": null
      },
      "limits": [
        "A registry entry is not proof of a successful model run.",
        "Architecture-specific restrictions, including Qwen hybrid complete-model placement, still apply.",
        "Structural inspection does not establish model-load, surgery, or producer compatibility."
      ]
    },
    {
      "id": "accelerate-device-map-process-local",
      "subject": "Accelerate device_map placement and CPU/disk offload",
      "format": "runtime_placement_not_checkpoint_format",
      "producer_versions": ["Existing project dependency lock"],
      "adapter": null,
      "model_mapping": "Complete modules placed across devices visible to one process",
      "state_scopes": ["runtime_model_placement"],
      "safety_level": "not_applicable",
      "optional_extras": [],
      "capabilities": {
        "detect": {"value": "out_of_scope", "basis": "This is runtime placement, not an input format."},
        "safe_inspect": {"value": "out_of_scope", "basis": "This is runtime placement, not checkpoint inspection."},
        "trusted_inspect": {"value": "out_of_scope", "basis": "This is runtime placement, not checkpoint inspection."},
        "weights_canonicalize": {"value": "out_of_scope", "basis": "Placement does not reconstruct rank fragments."},
        "topology_reshard": {"value": "out_of_scope", "basis": "device_map has no saved rank-fragment contract."},
        "surgery": {"value": "conditional", "basis": "Current process-local path depends on model-family compatibility and resource headroom."},
        "exact_resume": {"value": "out_of_scope", "basis": "No training state is restored."},
        "live_multi_node": {"value": "out_of_scope", "basis": "One process and one host; device placement is not a launcher."}
      },
      "canonical_output": null,
      "evidence": {
        "references": ["R14", "R21"],
        "candidate_commit": "5cc43c6e52903497574d80e08dff856028bc47f7",
        "fixture_digest": null,
        "environment": "Existing OBLITERATUS single-process implementation",
        "topology": "Single host, single OBLITERATUS process, one or more visible devices",
        "retained_result": null
      },
      "limits": [
        "Not rank-based pipeline parallelism.",
        "Not distributed checkpoint intake.",
        "Qwen hybrid models reject generic layer placement across devices."
      ]
    },
    {
      "id": "pytorch-dcp-fsdp-planned",
      "subject": "PyTorch DCP and FSDP model-weight intake",
      "format": "pytorch_dcp_or_fsdp_state",
      "producer_versions": ["Unselected; must be exact-version qualified"],
      "adapter": "Deferred exact-version PyTorch adapter",
      "model_mapping": "Requires a predeclared target model state schema",
      "state_scopes": ["model_weights"],
      "safety_level": "safe_structure",
      "optional_extras": ["checkpoint-pytorch-planned"],
      "capabilities": {
        "detect": {"value": "conditional", "basis": "Current bounded inventory recognizes DCP/FSDP marker structure without reading DCP metadata."},
        "safe_inspect": {"value": "conditional", "basis": "Current inventory-only classification keeps DCP metadata opaque and trust-gated."},
        "trusted_inspect": {"value": "deferred", "basis": "Blocked on security-owner acceptance, exact-profile controls, and an exact PyTorch version."},
        "weights_canonicalize": {"value": "deferred", "basis": "No exact-version adapter has been qualified."},
        "topology_reshard": {"value": "deferred", "basis": "Only exact model-weight cases with retained fixture evidence may qualify."},
        "surgery": {"value": "deferred", "basis": "Requires successful canonical output then existing OBLITERATUS gates."},
        "exact_resume": {"value": "out_of_scope", "basis": "Version 1 emits weights only."},
        "live_multi_node": {"value": "out_of_scope", "basis": "Checkpoint intake is not distributed execution."}
      },
      "canonical_output": "Planned HF safetensors plus conversion manifest",
      "evidence": {
        "references": ["R01", "R02", "R03", "R17", "R18", "R27", "R28", "R29"],
        "candidate_commit": null,
        "fixture_digest": null,
        "environment": null,
        "topology": null,
        "retained_result": null
      },
      "limits": ["No producer reader, adapter, conversion implementation, or qualified version band exists.", "Structural classification alone is not payload compatibility.", "Exact resume and arbitrary stateful/planner objects are excluded."]
    },
    {
      "id": "megatron-bridge-planned",
      "subject": "Megatron distributed model-weight intake through a model-aware Bridge",
      "format": "megatron_torch_dist_with_newer_forms_explicitly_deferred",
      "producer_versions": ["Unselected; must be exact-version and format qualified"],
      "adapter": "Deferred model-aware Megatron Bridge adapter",
      "model_mapping": "Requires supported Megatron Bridge/provider mapping and reference configuration",
      "state_scopes": ["model_weights"],
      "safety_level": "safe_structure",
      "optional_extras": ["checkpoint-megatron-planned"],
      "capabilities": {
        "detect": {"value": "conditional", "basis": "Current bounded inventory recognizes declared Megatron torch_dist marker structure without a vendor import."},
        "safe_inspect": {"value": "conditional", "basis": "Current inspection reports bounded structural facts only and does not interpret model semantics."},
        "trusted_inspect": {"value": "deferred", "basis": "Blocked on security-owner acceptance, exact-profile controls, and an exact vendor stack."},
        "weights_canonicalize": {"value": "deferred", "basis": "Blocked on common infrastructure and model-aware mapping."},
        "topology_reshard": {"value": "deferred", "basis": "Must be proven for exact model-weight topology and format."},
        "surgery": {"value": "deferred", "basis": "Requires canonical output and existing model-family gates."},
        "exact_resume": {"value": "out_of_scope", "basis": "Optimizer/resume compatibility is version/format-specific and outside v1."},
        "live_multi_node": {"value": "out_of_scope", "basis": "Bridge conversion is not an OBLITERATUS distributed runtime."}
      },
      "canonical_output": "Planned HF safetensors plus conversion manifest",
      "evidence": {
        "references": ["R04", "R05", "R06", "R07", "R27", "R28"],
        "candidate_commit": null,
        "fixture_digest": null,
        "environment": null,
        "topology": null,
        "retained_result": null
      },
      "limits": ["Offsets alone never authorize generic QKV/MLP/expert reconstruction.", "No producer reader, adapter, conversion implementation, or qualified model/version exists.", "Structural classification alone is not payload compatibility."]
    },
    {
      "id": "deepspeed-zero-universal-planned",
      "subject": "DeepSpeed ZeRO or Universal model-weight intake",
      "format": "deepspeed_zero_or_universal",
      "producer_versions": ["Unselected; must be exact-version qualified"],
      "adapter": "Deferred exact-version DeepSpeed adapter",
      "model_mapping": "Official consolidation or Universal mapping for an exact compatible case",
      "state_scopes": ["model_weights"],
      "safety_level": "safe_structure",
      "optional_extras": ["checkpoint-deepspeed-planned"],
      "capabilities": {
        "detect": {"value": "conditional", "basis": "Current bounded inventory recognizes declared ZeRO or Universal marker structure without a DeepSpeed import."},
        "safe_inspect": {"value": "conditional", "basis": "Current inspection reports bounded structural facts only and keeps framework serialization opaque."},
        "trusted_inspect": {"value": "deferred", "basis": "Blocked on security-owner acceptance and exact-profile controls; official consolidation reads framework serialization."},
        "weights_canonicalize": {"value": "deferred", "basis": "No adapter has passed the resource and trust gates."},
        "topology_reshard": {"value": "deferred", "basis": "Universal behavior must be proven for the exact mapping/version."},
        "surgery": {"value": "deferred", "basis": "Requires canonical output and existing model gates."},
        "exact_resume": {"value": "out_of_scope", "basis": "Version 1 emits weights only."},
        "live_multi_node": {"value": "out_of_scope", "basis": "Checkpoint consolidation is not live OBLITERATUS execution."}
      },
      "canonical_output": "Planned HF safetensors plus conversion manifest",
      "evidence": {
        "references": ["R08", "R09", "R10", "R17", "R27", "R28"],
        "candidate_commit": null,
        "fixture_digest": null,
        "environment": null,
        "topology": null,
        "retained_result": null
      },
      "limits": ["No producer reader, adapter, conversion implementation, or qualified version band exists.", "Structural classification alone is not payload compatibility.", "Input is trust-gated and peak CPU RAM/disk must be admitted."]
    },
    {
      "id": "peft-lora-safe-artifacts",
      "subject": "PEFT LoRA safetensors inspection and exact-base artifact export",
      "format": "peft_lora_safetensors",
      "producer_versions": ["PEFT format contract; live compatibility remains version-conditional"],
      "adapter": null,
      "model_mapping": "Exact base revision, weights/tokenizer digests, architecture, vocabulary, ties, and full target-module paths are required",
      "state_scopes": ["adapter_weights"],
      "safety_level": "safe_structure",
      "optional_extras": [],
      "capabilities": {
        "detect": {"value": "conditional", "basis": "Current inspector recognizes adapter_config.json with safetensors structure without loading payloads."},
        "safe_inspect": {"value": "conditional", "basis": "Current bounded JSON/header inspection preserves adapter and base identity evidence."},
        "trusted_inspect": {"value": "out_of_scope", "basis": "Canonical PEFT safetensors artifacts do not require a vendor metadata reader."},
        "weights_canonicalize": {"value": "out_of_scope", "basis": "LoRA artifact export is not distributed rank-fragment canonicalization."},
        "topology_reshard": {"value": "out_of_scope", "basis": "PEFT adapter tensors do not establish distributed checkpoint topology."},
        "surgery": {"value": "conditional", "basis": "The Python exporter writes canonical PEFT files only when exact base identity is supplied; the default pipeline otherwise writes a truthfully unsupported safe artifact."},
        "exact_resume": {"value": "out_of_scope", "basis": "Adapter artifacts do not contain complete training-resume state."},
        "live_multi_node": {"value": "out_of_scope", "basis": "Adapter export is an offline artifact operation."}
      },
      "canonical_output": "adapter_model.safetensors, adapter_config.json, adapter_manifest.json, provenance, and model card",
      "evidence": {
        "references": ["R16", "R27"],
        "candidate_commit": null,
        "fixture_digest": null,
        "environment": "Mandatory CPU artifact tests plus a conditional upstream PEFT loader check; not a universal base-model qualification",
        "topology": "Offline single-process adapter artifact export",
        "retained_result": null
      },
      "limits": [
        "Exact base and tokenizer identity are mandatory for a canonical claim.",
        "The default pipeline does not invent missing identity and emits a safely serialized unsupported artifact instead.",
        "Live base-model behavior remains conditional on the exact PEFT, Transformers, model, and runtime versions."
      ]
    },
    {
      "id": "live-multi-node-surgery-research",
      "subject": "Live multi-node OBLITERATUS surgery",
      "format": "runtime_not_checkpoint_format",
      "producer_versions": [],
      "adapter": null,
      "model_mapping": "Fixed-membership preflight only; model execution mapping is unimplemented",
      "state_scopes": ["runtime_execution"],
      "safety_level": "not_applicable",
      "optional_extras": [],
      "capabilities": {
        "detect": {"value": "out_of_scope", "basis": "Runtime capability is not a source format."},
        "safe_inspect": {"value": "out_of_scope", "basis": "Runtime capability is not checkpoint inspection."},
        "trusted_inspect": {"value": "out_of_scope", "basis": "Runtime capability is not checkpoint inspection."},
        "weights_canonicalize": {"value": "out_of_scope", "basis": "Offline conversion is a separate subsystem."},
        "topology_reshard": {"value": "out_of_scope", "basis": "Runtime ownership does not establish checkpoint conversion."},
        "surgery": {"value": "deferred", "basis": "Preflight and CPU/Gloo protocol contracts are available; model loading and surgery payloads are not."},
        "exact_resume": {"value": "out_of_scope", "basis": "No training-resume product is planned."},
        "live_multi_node": {"value": "deferred", "basis": "Candidate preflight only; no exact physical profile, multi-host qualification, or supported model workflow exists."}
      },
      "canonical_output": null,
      "evidence": {
        "references": ["R06", "R15", "R21"],
        "candidate_commit": null,
        "fixture_digest": null,
        "environment": null,
        "topology": null,
        "retained_result": null
      },
      "limits": ["Current --remote runs one OBLITERATUS process on one SSH host.", "Canonicalizing a distributed checkpoint does not make surgery distributed.", "The distributed command validates fixed membership and exits before model allocation."]
    }
  ]
}