generated: '2026-08-16'
method: searched
source: >-
  https://github.com/furiosa-ai/furiosa-sdk/blob/main/python/furiosa-server/README.md,
  https://developer.furiosa.ai/latest/en/furiosa_llm/furiosa-llm-serve.html,
  https://developer.furiosa.ai/latest/en/furiosa_llm/responses-api.html,
  https://developer.furiosa.ai/latest/en/whatsnew/release-2026.3.0.html,
  openapi/furiosa-predict-v2.yaml
summary: >-
  FuriosaAI's conformance story is unusually strong on INTEROPERABILITY standards and empty on
  REGULATORY ones. It implements four borrowed API contracts more or less verbatim - which is
  exactly the right posture for an inference-accelerator vendor, because it means an existing
  OpenAI or KServe client works against RNGD unchanged. What it does not publish is any
  certification, audit report or compliance program, so no Compliance pointer is emitted.
conformance:
  - id: kserve-predict-v2
    name: KServe / KFServing Predict Protocol v2 (V2 Dataplane)
    conforms: true
    evidence: >-
      The furiosa-server README states the API is "compliant with KFServing's V2 Dataplane
      specification", and FuriosaAI publishes the protocol's OpenAPI and .proto verbatim in
      python/furiosa-server/openapi/ and /proto/. Paths, operationIds and schemas match the
      upstream required_api.md.
    artifacts:
      - openapi/furiosa-predict-v2.yaml
      - grpc/furiosa-predict.proto
  - id: triton-model-repository
    name: Triton Inference Server Model Repository extension (HTTP/REST + gRPC)
    conforms: true
    evidence: >-
      README names the Triton Model Repository specification; /v2/repository/index,
      /load and /unload are implemented with the upstream request/response schemas.
    artifacts:
      - openapi/furiosa-model-repository-v2.yaml
      - grpc/furiosa-model-repository.proto
  - id: openai-api
    name: OpenAI API compatibility (Completions, Chat Completions, Embeddings, Models)
    conforms: partial
    evidence: >-
      `furiosa-llm serve` exposes /v1/completions, /v1/chat/completions, /v1/embeddings,
      /v1/models and /v1/models/{model_id}, and the docs say every undescribed parameter
      "inherits its behavior from the corresponding parameter in the OpenAI API".
    deviations:
      - n is currently limited to 1.
      - logprobs / top_logprobs are marked experimental.
      - The `model` field is required by clients but ignored by the server (one model per process).
      - Adds non-OpenAI fields - reasoning, return_token_ids, artifact_id, max_prompt_len, max_context_len, runtime_config.
  - id: openresponses
    name: OpenResponses specification (Responses API)
    conforms: partial
    evidence: >-
      "Furiosa-LLM implements the OpenResponses specification"; POST /v1/responses,
      GET /v1/responses/{id}, POST /v1/responses/{id}/cancel.
    deviations:
      - Multimodal input types (input_image, input_audio, input_file) accepted but silently ignored.
      - Built-in tools (web_search, file_search, code_interpreter, computer_use, mcp) not supported - custom function tools only.
      - background=true accepted but has no effect.
      - truncation="auto" accepted; only "disabled" implemented.
      - presence_penalty and frequency_penalty accepted but not functional.
      - Adds top_k, which is not in the specification.
  - id: vllm-extensions
    name: vLLM Score and Rerank API extensions
    conforms: true
    evidence: >-
      /score, /v1/score, /rerank, /v1/rerank and /v2/rerank are implemented against the vLLM
      specification, which the docs cite directly. Currently limited to Qwen3-Rerank models.
  - id: prometheus-exposition
    name: Prometheus text exposition on /metrics
    conforms: true
    evidence: >-
      GET /metrics serves vLLM-compatible metrics plus furiosa_llm_* collectors (gauges,
      counters and histograms with model_name/engine labels). GET-only as of 2026.3.0.
  - id: opentelemetry
    name: OpenTelemetry
    conforms: true
    evidence: Roadmap records "✅ Enhanced observability with OpenTelemetry and per-device metrics" delivered in 2026 Q1-Q2.
  - id: json-schema
    name: JSON Schema structured output
    conforms: true
    evidence: >-
      Structured output constrains generation to a caller-supplied JSON Schema via
      response_format / text.format type json_schema, backed by libguidance and xgrammar.
  - id: kubernetes-dra
    name: Kubernetes Dynamic Resource Allocation (DRA) and Device Plugin
    conforms: true
    evidence: >-
      furiosa-device-plugin, furiosa-dra-driver, furiosa-feature-discovery, furiosa-npu-operator
      and furiosa-metrics-exporter are published as multi-arch images; the NPU operator images
      are Red Hat OpenShift-certified and UBI variants ship on quay.io.
  - id: cdi
    name: Container Device Interface (CDI)
    conforms: true
    evidence: Roadmap - "✅ Container Runtime and Container Interface Device (CDI) support" (2025 Q1-Q2); furiosa-cdi package ships in the APT/YUM trains.
  - id: rfc9457
    name: RFC 9457 Problem Details for HTTP APIs
    conforms: false
    evidence: >-
      No application/problem+json anywhere; the KServe {"error": string} envelope is used
      instead. See errors/furiosa-problem-types.yml.
  - id: oauth2
    name: OAuth 2.0
    conforms: false
    evidence: No authorization server, no scopes; /.well-known/oauth-authorization-server 404 on every host (well-known/furiosa-well-known.yml).
  - id: oidc
    name: OpenID Connect
    conforms: false
    evidence: /.well-known/openid-configuration 404 on every host.
certifications: []
certifications_note: >-
  No SOC 2, ISO 27001, PCI DSS, HIPAA or FedRAMP claim is published anywhere on furiosa.ai or the
  developer center, and there is no trust center. The only third-party certification FuriosaAI
  publishes is a PRODUCT one - Red Hat OpenShift certification of the cloud-native operator images -
  which is a platform-compatibility certification, not a security or compliance attestation, and is
  recorded above under kubernetes-dra rather than here. NO Compliance pointer is emitted.
conforms_count: 9