Kiln Documentation

ROCm/HF next-token result schema

Interpret the bound source, process, model, full logits, retained artifact, and candidate attribution for one ROCm oracle run.

Result of one independent eager Hugging Face next-token comparison on ROCm. The record binds the pushed source tree, implementation bytes, request, model, process containment, full-vocabulary logits artifact, and verdict; it attributes this exact divergent state rather than proving general model parity.

Schema identity

Property Value
Title Kiln ROCm Hugging Face Next-Token Oracle Result v2
$id https://ericflo.github.io/kiln/schemas/rocm-hf-next-token-oracle-v2.schema.json
Dialect https://json-schema.org/draft/2020-12/schema
Root type object
Root object closed

Root fields

Field Required Type Constraints and default Description
schema yes any const "kiln.rocm-hf-next-token-oracle.v2" Schema identifier for this oracle-result format.
result_sha256 yes sha256 - Canonical SHA-256 identity of the complete result after removing this self-hash field.
created_at_utc yes string format date-time UTC time when the result record was created.
duration_seconds yes number exclusive minimum 0 Total wall-clock duration of model fingerprinting, contained oracle execution, verification, and result assembly.
source yes object closed object Clean pushed repository identity used to build and execute the oracle.
implementation yes object closed object Content identities of the worker, interpreter, and containment runner that produced the oracle evidence.
request yes object closed object Repository-local attribution request, its canonical identity, and the paired candidate-source receipts it binds.
model_identity yes modelIdentity - Recomputed content identity of the exact model, tokenizer, chat template, configuration, and weight shards.
model_fingerprint yes modelFingerprintEvidence - Implementation and process evidence for the separate contained model-fingerprinting step.
containment yes object closed object Host-memory preflight, network isolation, and process-exit evidence for the oracle worker.
oracle yes oracleEvidence - Full-vocabulary Hugging Face output, candidate ranks, runtime identity, and resource evidence for the bound input state.
reference_artifact yes object closed object Ignored local safetensors file retaining the complete full-vocabulary logits used by later attribution work.
verdict yes object closed object Attribution of the Hugging Face argmax to one disputed candidate, or to neither candidate.

Definitions

sha256Lowercase SHA-256 digest with the `sha256:` algorithm prefix.0 fields

Type: string. Constraints: pattern ^sha256:[0-9a-f]{64}$.

gitObjectFull lowercase SHA-1 Git object identifier.0 fields

Type: string. Constraints: pattern ^[0-9a-f]{40}$.

tokenIdTokenizer vocabulary ID for the bound Qwen3.5-4B model identity.0 fields

Type: integer. Constraints: minimum 0; maximum 248319.

receiptReferenceRepository-relative qualification receipt with hashes of both its file bytes and canonical content.3 fields

Type: object. Constraints: closed object.

Field Required Type Constraints and default Description
path yes string minimum length 1 Repository-relative path to the retained receipt.
file_sha256 yes sha256 - SHA-256 identity of the receipt file's exact bytes.
content_sha256 yes sha256 - SHA-256 identity of the receipt's canonical content.
sourcePairSource revision and qualification receipts for the Kiln and vLLM candidate-producing runs.3 fields

Type: object. Constraints: closed object.

Field Required Type Constraints and default Description
source_commit yes gitObject - Full Git commit shared by the paired candidate runs.
kiln_receipt yes receiptReference - Retained Kiln qualification receipt and its content identities.
vllm_receipt yes receiptReference - Retained vLLM qualification receipt and its content identities.
weightFileOne model weight shard and its exact size and content identity.3 fields

Type: object. Constraints: closed object.

Field Required Type Constraints and default Description
path yes string minimum length 1 Path or stable shard name within the bound model source.
bytes yes integer minimum 1 Exact weight-shard size in bytes.
sha256 yes sha256 - SHA-256 identity of the shard's exact bytes.
modelIdentityComplete content identity for the model inputs used by candidate and oracle processes.6 fields

Type: object. Constraints: closed object.

Field Required Type Constraints and default Description
id yes string minimum length 1 Human-readable model identifier recorded alongside the content hashes.
content_sha256 yes sha256 - Aggregate content identity for the complete model source.
config_hash yes sha256 - SHA-256 identity of the effective model configuration.
tokenizer_hash yes sha256 - SHA-256 identity of the tokenizer inputs.
chat_template_hash yes sha256 - SHA-256 identity of the chat template used to render the prompt.
weight_files yes array<weightFile> minimum items 1 Complete ordered inventory of model weight shards.
modelFingerprintEvidenceSource identity and bounded process evidence for the separate model-provenance hashing process.3 fields

Type: object. Constraints: closed object.

Field Required Type Constraints and default Description
implementation_sha256 yes sha256 - SHA-256 identity of the model-fingerprint implementation.
python_sha256 yes sha256 - SHA-256 identity of the Python executable used for fingerprinting.
process yes processEvidence - Bounded process evidence for the separate fingerprint worker.
topTokenOne of the ten highest Hugging Face logits, ordered from highest to lowest.3 fields

Type: object. Constraints: closed object.

Field Required Type Constraints and default Description
token_id yes tokenId - Vocabulary ID of this ranked token.
text yes string - Tokenizer-decoded text for this token.
logit yes number - F32 logit produced by the eager Hugging Face reference.
candidateTokenOne disputed engine candidate with its Hugging Face logit and one-based full-vocabulary rank.5 fields

Type: object. Constraints: closed object.

Field Required Type Constraints and default Description
engine yes any enum "kiln", "vllm" Candidate-producing inference engine.
token_id yes tokenId - Vocabulary ID predicted by the named engine.
text yes string - Tokenizer-decoded text for the predicted token.
logit yes number - Hugging Face logit at this candidate's vocabulary ID.
rank yes integer minimum 1; maximum 248320 One-based rank of this candidate in the complete Hugging Face vocabulary, with 1 denoting the argmax.
oracleEvidenceEager Hugging Face next-token evidence for one exact request and model state.31 fields

Type: object. Constraints: closed object.

Field Required Type Constraints and default Description
attention_implementation yes any const "eager" Transformers full-attention implementation selected for the oracle; fixed to eager execution.
argmax yes tokenId - Vocabulary ID with the highest Hugging Face logit.
argmax_text yes string minimum length 1 Tokenizer-decoded text for the Hugging Face argmax token.
candidate_tokens yes array minimum items 2; maximum items 2 The two disputed candidates, ordered as Kiln then vLLM, annotated with their Hugging Face logits and ranks.
configuration_sha256 yes sha256 - SHA-256 identity of the fixed oracle execution configuration enforced by the worker.
deterministic_algorithms yes any const true Whether PyTorch deterministic algorithms were required; fixed to true.
device yes string minimum length 1 Diagnostic accelerator name reported by PyTorch. It identifies this run but is not a dispatch-policy input.
dtype yes any const "bfloat16" Model execution dtype used by the oracle.
duration_seconds yes number exclusive minimum 0 Wall-clock duration of the eager Hugging Face worker itself.
input_token_count yes integer minimum 1 Number of tokens in the complete prompt-plus-continuation input.
input_token_ids_sha256 yes sha256 - Canonical SHA-256 identity of the complete input-token sequence.
linear_attention_implementation yes any const "transformers_torch_fallback" Transformers implementation selected for Gated DeltaNet linear-attention layers.
logits_sha256 yes sha256 - SHA-256 identity of the contiguous F32 full-vocabulary logit bytes after the bound continuation state.
memory_high_events yes integer minimum 0 Number of cgroup memory.events high-threshold events observed during the contained worker.
memory_max_events yes integer minimum 0 Number of cgroup memory.events max-limit events observed during the contained worker.
memory_oom_events yes integer minimum 0 Number of cgroup out-of-memory events observed during the contained worker.
memory_oom_kill_events yes integer minimum 0 Number of cgroup out-of-memory kill events observed during the contained worker.
memory_peak_bytes yes integer minimum 1 Peak cgroup memory consumption in bytes during the contained worker.
memory_swap_bytes yes integer minimum 0 Cgroup swap usage in bytes reported for the contained worker.
modeling_sha256 yes sha256 - SHA-256 identity of the pinned Transformers Qwen3.5 modeling implementation expected by the worker.
output_bytes yes integer minimum 1 Exact size in bytes of the retained full-vocabulary logits artifact.
request_id yes string minimum length 1 Identifier copied from the bound next-token request.
request_sha256 yes sha256 - Canonical SHA-256 identity of the bound next-token request.
tf32_allowed yes any const false Whether TF32 matrix operations were permitted; fixed to false for this comparison.
top_logit_margin yes number exclusive minimum 0 Positive difference between the highest and second-highest Hugging Face logits.
top_tokens yes array<topToken> minimum items 10; maximum items 10 Exactly the ten highest Hugging Face tokens and logits, ordered from highest to lowest.
torch_commit yes gitObject - Pinned PyTorch source commit used by the qualification environment.
torch_hip_version yes string minimum length 1 HIP runtime version reported by the pinned PyTorch build.
torch_version yes any const "2.13.0+rocm7.2" Pinned PyTorch distribution version required by this result schema.
transformers_version yes any const "5.13.1" Pinned Transformers distribution version required by this result schema.
vocab yes any const 248320 Vocabulary size used for logits, ranks, and token-ID bounds.
processEvidenceBounded worker execution record emitted by the shared process-containment runner.4 fields

Type: object. Constraints: closed object.

Field Required Type Constraints and default Description
elapsed_seconds yes number minimum 0 Wall-clock duration observed by the containment runner.
schema yes any const "kiln.hf-process-containment.v1" Schema identifier for the shared process-containment evidence.
timeout_seconds yes number exclusive minimum 0 Maximum worker duration enforced by the containment runner.
worker_exit_code yes any const 0 Contained worker exit code; zero is required for retained oracle evidence.