inference_metadata.annotated.yaml

# Annotated review companion for justinchuby/deepseek-r1-distill-qwen-1.5b-onnx-genai.
# Model form: DeepSeek-R1 distilled decoder.
# Canonical authority: inference_metadata.yaml at revision 1427c4896f798893e58ffec91aef65c34de4503a.
# Comments explain authored semantics, derived structure, and fail-closed omissions.
# After YAML parsing this file is mechanically equal to the canonical metadata; comments do not create a second contract.
# Backend implementation, recorded execution, and performance evidence remain separate from this portable metadata.
# Schema spelling `1.0` normalizes under the current reader; versioning gates unknown fields before typed parsing. No component authors batch_capacity, so independent requests must execute per item at every component.
schema_version: 'v1.2'
# The pipeline contains executable package semantics. Deployment/QoS policy and backend-derived execution plans live outside this portable contract.
pipeline:
  # This typed SSA workflow is the authoritative execution contract for the package.
  workflow:
    # Manifest vocabulary declares interpreter/adapter ABI requirements. It is not ORT/native readiness or performance evidence.
    manifest:
      # Current v1 capability strings mirror typed workflow structure; the structure remains the semantic source of truth.
      capabilities:
      - workflow_ssa
      - linear_effects
      - nested_control_flow
      - typed_emit
      - emit_valid_length
      - loop_induction_values
      - serving_service_contract
      - bounded_state_recurrence
    # Workflow inputs are the complete typed boundary. Roles and sources are authored; absent required values must fail before execution.
    inputs:
      # Input `request.input_ids` is required, provided by the request through a typed runtime role, and carries runtime role prompt_tokens@1.0. Unsupported required roles must fail closed.
      request.input_ids:
        # Authored tensor ABI: dtype int64, rank 2, shape [batch_size, sequence_length]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: int64
          rank: 2
          shape:
          - batch_size
          - sequence_length
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # Role identifies runtime semantics without relying on a model-family name.
        role:
          kind: runtime
          version: '1.0'
          role: prompt_tokens
        # Source states who must provide this value; the runtime must not invent application data.
        source:
          kind: request
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.max_iterations` is required, provided by the request through a typed runtime role, and carries runtime role max_output_tokens@1.0. Unsupported required roles must fail closed.
      request.max_iterations:
        # Authored tensor ABI: dtype int64, rank 1, shape [1]. No grouping permission is implied by this shape.
        contract:
          dtype: int64
          rank: 1
          shape:
          - 1
        # Role identifies runtime semantics without relying on a model-family name.
        role:
          kind: runtime
          version: '1.0'
          role: max_output_tokens
        # Source states who must provide this value; the runtime must not invent application data.
        source:
          kind: request
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      package.one_token:
        # Authored tensor ABI: dtype int64, rank 1, shape [batch_size]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: int64
          rank: 1
          shape:
          - batch_size
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # Role identifies runtime semantics without relying on a model-family name.
        role:
          kind: opaque
        # Source states who must provide this value; the runtime must not invent application data.
        source:
          kind: literal
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: false
        # This is a semantic tensor default used only when the optional input is omitted.
        default: 1
      # Input `package.one_step` is optional; omission uses semantic default 1, a package-authored literal/default, and carries opaque. Unsupported required roles must fail closed.
      package.one_step:
        # Authored tensor ABI: dtype int64, rank 1, shape [1]. No grouping permission is implied by this shape.
        contract:
          dtype: int64
          rank: 1
          shape:
          - 1
        # Role identifies runtime semantics without relying on a model-family name.
        role:
          kind: opaque
        # Source states who must provide this value; the runtime must not invent application data.
        source:
          kind: literal
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: false
        # This is a semantic tensor default used only when the optional input is omitted.
        default: 1
      # Input `package.max_context` is optional; omission uses semantic default 131072, a package-authored literal/default, and carries opaque. Unsupported required roles must fail closed.
      package.max_context:
        # Authored tensor ABI: dtype int64, rank 1, shape [1]. No grouping permission is implied by this shape.
        contract:
          dtype: int64
          rank: 1
          shape:
          - 1
        # Role identifies runtime semantics without relying on a model-family name.
        role:
          kind: opaque
        # Source states who must provide this value; the runtime must not invent application data.
        source:
          kind: literal
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: false
        # This is a semantic tensor default used only when the optional input is omitted.
        default: 131072
      # Input `request.prompt_lengths` is optional; omission uses semantic default -1, provided by the application as prompt_lengths, and carries opaque. Unsupported required roles must fail closed.
      request.prompt_lengths:
        # Authored tensor ABI: dtype int64, rank 1, shape [batch_size]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: int64
          rank: 1
          shape:
          - batch_size
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # Role identifies runtime semantics without relying on a model-family name.
        role:
          kind: opaque
        # Source states who must provide this value; the runtime must not invent application data.
        source:
          kind: application
          name: prompt_lengths
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: false
        # This is a semantic tensor default used only when the optional input is omitted.
        default: -1
      # Input `request.eos_ids` is a required resolved runtime value with role eos_token_ids@1.0. An explicit request replaces the package EOS default; otherwise the runtime materializes the package fact. Unsupported roles fail closed.
      request.eos_ids:
        # The tensor contract validates request shape and dtype at the workflow boundary; it does not duplicate an ONNX component port ABI.
        contract:
          # Element type required from the request adapter.
          dtype: int64
          # Tensor rank required before workflow execution.
          rank: 2
          # Symbolic dimensions checked and unified across workflow values.
          shape:
          - batch_size
          - num_eos
          # Request-row identity layout; not independent batching permission.
          batch_layout:
            kind: request_aligned
            axis: 0
        # Typed runtime meaning; unknown required role versions fail admission.
        role:
          # Role namespace.
          kind: runtime
          # Role ABI version.
          version: '1.0'
          # Runtime EOS tensor meaning.
          role: eos_token_ids
        # The request adapter supplies this value; the package does not author it.
        source:
          # Value provider class.
          kind: request
        # Missing required request inputs fail before component execution.
        required: true
      # Input `request.eos_lengths` is a required resolved runtime value with role eos_token_lengths@1.0. An explicit request replaces the package EOS default; otherwise the runtime materializes the package fact. Unsupported roles fail closed.
      request.eos_lengths:
        # The tensor contract validates request shape and dtype at the workflow boundary; it does not duplicate an ONNX component port ABI.
        contract:
          # Element type required from the request adapter.
          dtype: int64
          # Tensor rank required before workflow execution.
          rank: 1
          # Symbolic dimensions checked and unified across workflow values.
          shape:
          - batch_size
          # Request-row identity layout; not independent batching permission.
          batch_layout:
            kind: request_aligned
            axis: 0
        # Typed runtime meaning; unknown required role versions fail admission.
        role:
          # Role namespace.
          kind: runtime
          # Role ABI version.
          version: '1.0'
          # Runtime EOS tensor meaning.
          role: eos_token_lengths
        # The request adapter supplies this value; the package does not author it.
        source:
          # Value provider class.
          kind: request
        # Missing required request inputs fail before component execution.
        required: true
      request.row_max_iterations:
        # Authored tensor ABI: dtype int64, rank 1, shape [batch_size]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: int64
          rank: 1
          shape:
          - batch_size
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # Role identifies runtime semantics without relying on a model-family name.
        role:
          kind: opaque
        # Source states who must provide this value; the runtime must not invent application data.
        source:
          kind: application
          name: row_max_iterations
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: false
        # This is a semantic tensor default used only when the optional input is omitted.
        default: -1
      # Input `request.temperature` is optional; omission uses semantic default 1.0, provided by the request through a typed runtime role, and carries runtime role sampling_temperature@1.0. Unsupported required roles must fail closed.
      request.temperature:
        # Authored tensor ABI: dtype float32, rank 1, shape [batch_size]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float32
          rank: 1
          shape:
          - batch_size
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # Role identifies runtime semantics without relying on a model-family name.
        role:
          kind: runtime
          version: '1.0'
          role: sampling_temperature
        # Source states who must provide this value; the runtime must not invent application data.
        source:
          kind: request
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: false
        # This is a semantic tensor default used only when the optional input is omitted.
        default: 1.0
      # Input `request.top_k` is optional; omission uses semantic default 0, provided by the request through a typed runtime role, and carries runtime role sampling_top_k@1.0. Unsupported required roles must fail closed.
      request.top_k:
        # Authored tensor ABI: dtype int64, rank 1, shape [batch_size]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: int64
          rank: 1
          shape:
          - batch_size
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # Role identifies runtime semantics without relying on a model-family name.
        role:
          kind: runtime
          version: '1.0'
          role: sampling_top_k
        # Source states who must provide this value; the runtime must not invent application data.
        source:
          kind: request
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: false
        # This is a semantic tensor default used only when the optional input is omitted.
        default: 0
      # Input `request.top_p` is optional; omission uses semantic default 1.0, provided by the request through a typed runtime role, and carries runtime role sampling_top_p@1.0. Unsupported required roles must fail closed.
      request.top_p:
        # Authored tensor ABI: dtype float32, rank 1, shape [batch_size]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float32
          rank: 1
          shape:
          - batch_size
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # Role identifies runtime semantics without relying on a model-family name.
        role:
          kind: runtime
          version: '1.0'
          role: sampling_top_p
        # Source states who must provide this value; the runtime must not invent application data.
        source:
          kind: request
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: false
        # This is a semantic tensor default used only when the optional input is omitted.
        default: 1.0
      # Input `request.min_p` is optional; omission uses semantic default 0.0, provided by the request through a typed runtime role, and carries runtime role sampling_min_p@1.0. Unsupported required roles must fail closed.
      request.min_p:
        # Authored tensor ABI: dtype float32, rank 1, shape [batch_size]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float32
          rank: 1
          shape:
          - batch_size
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # Role identifies runtime semantics without relying on a model-family name.
        role:
          kind: runtime
          version: '1.0'
          role: sampling_min_p
        # Source states who must provide this value; the runtime must not invent application data.
        source:
          kind: request
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: false
        # This is a semantic tensor default used only when the optional input is omitted.
        default: 0.0
      # Input `request.seed` is optional; omission uses semantic default 0, provided by the request through a typed runtime role, and carries runtime role seed@1.0. Unsupported required roles must fail closed.
      request.seed:
        # Authored tensor ABI: dtype int64, rank 1, shape [batch_size]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: int64
          rank: 1
          shape:
          - batch_size
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # Role identifies runtime semantics without relying on a model-family name.
        role:
          kind: runtime
          version: '1.0'
          role: seed
        # Source states who must provide this value; the runtime must not invent application data.
        source:
          kind: request
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: false
        # This is a semantic tensor default used only when the optional input is omitted.
        default: 0
      # Input `request.rng_counter` is optional; omission uses semantic default 0, provided by the application as rng_counter, and carries opaque. Unsupported required roles must fail closed.
      request.rng_counter:
        # Authored tensor ABI: dtype int64, rank 1, shape [batch_size]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: int64
          rank: 1
          shape:
          - batch_size
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # Role identifies runtime semantics without relying on a model-family name.
        role:
          kind: opaque
        # Source states who must provide this value; the runtime must not invent application data.
        source:
          kind: application
          name: rng_counter
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: false
        # This is a semantic tensor default used only when the optional input is omitted.
        default: 0
      # Input `package.active` is optional; omission uses semantic default True, a package-authored literal/default, and carries opaque. Unsupported required roles must fail closed.
      package.active:
        # Authored tensor ABI: dtype bool, rank 1, shape [batch_size]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: bool
          rank: 1
          shape:
          - batch_size
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # Role identifies runtime semantics without relying on a model-family name.
        role:
          kind: opaque
        # Source states who must provide this value; the runtime must not invent application data.
        source:
          kind: literal
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: false
        # This is a semantic tensor default used only when the optional input is omitted.
        default: true
      # Input `package.not_done` is optional; omission uses semantic default False, a package-authored literal/default, and carries opaque. Unsupported required roles must fail closed.
      package.not_done:
        # Authored tensor ABI: dtype bool, rank 1, shape [batch_size]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: bool
          rank: 1
          shape:
          - batch_size
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # Role identifies runtime semantics without relying on a model-family name.
        role:
          kind: opaque
        # Source states who must provide this value; the runtime must not invent application data.
        source:
          kind: literal
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: false
        # This is a semantic tensor default used only when the optional input is omitted.
        default: false
      # Input `package.cache_lengths` is optional; omission uses semantic default 0, a package-authored literal/default, and carries opaque. Unsupported required roles must fail closed.
      package.cache_lengths:
        # Authored tensor ABI: dtype int64, rank 1, shape [batch_size]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: int64
          rank: 1
          shape:
          - batch_size
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # Role identifies runtime semantics without relying on a model-family name.
        role:
          kind: opaque
        # Source states who must provide this value; the runtime must not invent application data.
        source:
          kind: literal
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: false
        # This is a semantic tensor default used only when the optional input is omitted.
        default: 0
      # Input `package.zero_batch` is optional; omission uses semantic default 0, a package-authored literal/default, and carries opaque. Unsupported required roles must fail closed.
      package.zero_batch:
        # Authored tensor ABI: dtype int64, rank 1, shape [batch_size]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: int64
          rank: 1
          shape:
          - batch_size
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # Role identifies runtime semantics without relying on a model-family name.
        role:
          kind: opaque
        # Source states who must provide this value; the runtime must not invent application data.
        source:
          kind: literal
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: false
        # This is a semantic tensor default used only when the optional input is omitted.
        default: 0
    # Workflow outputs are the complete public result boundary and retain explicit tensor contracts.
    outputs:
      # Output `tokens` publishes the authored tokens result at the pre_adapter stage; it is not backend-readiness evidence.
      tokens:
        # Authored tensor ABI: dtype int64, rank 2, shape [batch_size, generated_sequence]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: int64
          rank: 2
          shape:
          - batch_size
          - generated_sequence
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        role: tokens
        # Stage records whether adapters/postprocessing have run before this value is exposed.
        stage: pre_adapter
    # Components are the executable artifacts and semantic policy programs. Their presence does not prove ORT/native readiness or performance.
    components:
      # Component `model` is a primary model compute graph using onnx from `model.onnx`. No batch_capacity is authored, so the runtime must not group independent requests at this component.
      model:
        # Selects a packaged implementation/artifact only; provider/backend readiness requires separate recorded evidence.
        implementation:
          kind: onnx
          artifact: model.onnx
        # Ports bind workflow names to the artifact's exact graph ABI; model-family inference is forbidden.
        ports:
          # Semantic port roles identify tokens, logits, positions, masks, caches, or media without architecture-name conditionals.
          roles:
            input_ids: token_ids
            attention_mask: attention_mask
            logits: logits
      # Component `token_sampler` is an executable semantic policy component implementing `onnx-genai.token-sampler` using onnx from `policies/token_sampler.onnx`. Policy graphs compute tensors and semantics; they are not deployment/QoS knobs or backend readiness claims. No batch_capacity is authored, so the runtime must not group independent requests at this component.
      token_sampler:
        # Selects a packaged implementation/artifact only; provider/backend readiness requires separate recorded evidence.
        implementation:
          kind: onnx
          artifact: policies/token_sampler.onnx
        # Ports bind workflow names to the artifact's exact graph ABI; model-family inference is forbidden.
        ports:
          # Exact graph inputs; each tensor shape/dtype is authored.
          inputs:
            # Port `logits`: Authored tensor ABI: dtype float32, rank 2, shape [batch, vocabulary]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            logits:
              dtype: float32
              rank: 2
              shape:
              - batch
              - vocabulary
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `temperature`: Authored tensor ABI: dtype float32, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            temperature:
              dtype: float32
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `top_k`: Authored tensor ABI: dtype int64, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            top_k:
              dtype: int64
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `top_p`: Authored tensor ABI: dtype float32, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            top_p:
              dtype: float32
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `min_p`: Authored tensor ABI: dtype float32, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            min_p:
              dtype: float32
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `seed`: Authored tensor ABI: dtype int64, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            seed:
              dtype: int64
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `counter`: Authored tensor ABI: dtype int64, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            counter:
              dtype: int64
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `active`: Authored tensor ABI: dtype bool, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            active:
              dtype: bool
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `done`: Authored tensor ABI: dtype bool, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            done:
              dtype: bool
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
          # Exact graph outputs; each tensor shape/dtype is authored.
          outputs:
            # Port `token`: Authored tensor ABI: dtype int64, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            token:
              dtype: int64
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `next_counter`: Authored tensor ABI: dtype int64, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            next_counter:
              dtype: int64
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
        # Semantic component ABI `onnx-genai.token-sampler`; its bindings/parameters affect output correctness.
        contract:
          id: onnx-genai.token-sampler
          version: '2'
          bindings:
            logits: logits
            token: token
            temperature: temperature
            top_k: top_k
            top_p: top_p
            min_p: min_p
            active: active
            done: done
            seed: seed
            counter: counter
            next_counter: next_counter
          parameters:
            mode: seeded_stochastic
            batching: per_row
            inactive_rows: preserve
        # An override may replace this component only through the same validated semantic contract.
        # Application replacement is allowed only through the same typed component contract.
        application_overridable: true
      # Component `termination` is an executable semantic policy component implementing `onnx-genai.termination-predicate` using onnx from `policies/termination.onnx`. Policy graphs compute tensors and semantics; they are not deployment/QoS knobs or backend readiness claims. No batch_capacity is authored, so the runtime must not group independent requests at this component.
      termination:
        # Selects a packaged implementation/artifact only; provider/backend readiness requires separate recorded evidence.
        implementation:
          kind: onnx
          artifact: policies/termination.onnx
        # Ports bind workflow names to the artifact's exact graph ABI; model-family inference is forbidden.
        ports:
          # Exact graph inputs; each tensor shape/dtype is authored.
          inputs:
            # Port `tokens`: Authored tensor ABI: dtype int64, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            tokens:
              dtype: int64
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `eos_ids`: Authored tensor ABI: dtype int64, rank 2, shape [batch, num_eos]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            eos_ids:
              dtype: int64
              rank: 2
              shape:
              - batch
              - num_eos
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `eos_lengths`: Authored tensor ABI: dtype int64, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            eos_lengths:
              dtype: int64
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `iteration`: Authored tensor ABI: dtype int64, rank 1, shape [1]. No grouping permission is implied by this shape.
            iteration:
              dtype: int64
              rank: 1
              shape:
              - 1
            # Port `max_iterations`: Authored tensor ABI: dtype int64, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            max_iterations:
              dtype: int64
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `active`: Authored tensor ABI: dtype bool, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            active:
              dtype: bool
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
          # Exact graph outputs; each tensor shape/dtype is authored.
          outputs:
            # Port `done`: Authored tensor ABI: dtype bool, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            done:
              dtype: bool
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `next_active`: Authored tensor ABI: dtype bool, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            next_active:
              dtype: bool
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `continue`: Authored tensor ABI: dtype bool, rank 1, shape [1]. No grouping permission is implied by this shape.
            continue:
              dtype: bool
              rank: 1
              shape:
              - 1
        # Semantic component ABI `onnx-genai.termination-predicate`; its bindings/parameters affect output correctness.
        contract:
          id: onnx-genai.termination-predicate
          version: '2'
          bindings:
            tokens: tokens
            eos_ids: eos_ids
            iteration: iteration
            max_iterations: max_iterations
            eos_lengths: eos_lengths
            active: active
            done: done
            next_active: next_active
            continue: continue
          parameters:
            batching: per_row
            inactive_rows: preserve
      # Component `token_state_update` is an executable semantic policy component implementing `onnx-genai.state-update` using onnx from `policies/token_state_update.onnx`. Policy graphs compute tensors and semantics; they are not deployment/QoS knobs or backend readiness claims. No batch_capacity is authored, so the runtime must not group independent requests at this component.
      token_state_update:
        # Selects a packaged implementation/artifact only; provider/backend readiness requires separate recorded evidence.
        implementation:
          kind: onnx
          artifact: policies/token_state_update.onnx
        # Ports bind workflow names to the artifact's exact graph ABI; model-family inference is forbidden.
        ports:
          # Exact graph inputs; each tensor shape/dtype is authored.
          inputs:
            # Port `current`: Authored tensor ABI: dtype int64, rank 2, shape [batch, 1]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            current:
              dtype: int64
              rank: 2
              shape:
              - batch
              - 1
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `update`: Authored tensor ABI: dtype int64, rank 2, shape [batch, 1]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            update:
              dtype: int64
              rank: 2
              shape:
              - batch
              - 1
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `active`: Authored tensor ABI: dtype bool, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            active:
              dtype: bool
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `done`: Authored tensor ABI: dtype bool, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            done:
              dtype: bool
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
          # Exact graph outputs; each tensor shape/dtype is authored.
          outputs:
            # Port `next`: Authored tensor ABI: dtype int64, rank 2, shape [batch, 1]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            next:
              dtype: int64
              rank: 2
              shape:
              - batch
              - 1
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
        # Semantic component ABI `onnx-genai.state-update`; its bindings/parameters affect output correctness.
        contract:
          id: onnx-genai.state-update
          version: '2'
          bindings:
            current: current
            update: update
            active: active
            done: done
            next: next
          parameters:
            batching: per_row
            inactive_rows: preserve
      # Component `last_token_logits` is a logit-selection policy graph using onnx from `policies/last_token_logits.onnx`. Policy graphs compute tensors and semantics; they are not deployment/QoS knobs or backend readiness claims. No batch_capacity is authored, so the runtime must not group independent requests at this component.
      last_token_logits:
        # Selects a packaged implementation/artifact only; provider/backend readiness requires separate recorded evidence.
        implementation:
          kind: onnx
          artifact: policies/last_token_logits.onnx
        # Ports bind workflow names to the artifact's exact graph ABI; model-family inference is forbidden.
        ports:
          # Exact graph inputs; each tensor shape/dtype is authored.
          inputs:
            # Port `logits`: Authored tensor ABI: dtype float16, rank 3, shape [batch, sequence, vocabulary]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            logits:
              dtype: float16
              rank: 3
              shape:
              - batch
              - sequence
              - vocabulary
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
          # Exact graph outputs; each tensor shape/dtype is authored.
          outputs:
            # Port `last_logits`: Authored tensor ABI: dtype float32, rank 2, shape [batch, vocabulary]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            last_logits:
              dtype: float32
              rank: 2
              shape:
              - batch
              - vocabulary
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
      # Component `decoder_state_initializer` is a primary model compute graph using onnx from `policies/decoder_state_initializer.onnx`. No batch_capacity is authored, so the runtime must not group independent requests at this component.
      decoder_state_initializer:
        # Selects a packaged implementation/artifact only; provider/backend readiness requires separate recorded evidence.
        implementation:
          kind: onnx
          artifact: policies/decoder_state_initializer.onnx
        # Ports bind workflow names to the artifact's exact graph ABI; model-family inference is forbidden.
        ports:
          # Exact graph inputs; each tensor shape/dtype is authored.
          inputs:
            # Port `prompt_tokens`: Authored tensor ABI: dtype int64, rank 2, shape [batch, prompt_sequence]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            prompt_tokens:
              dtype: int64
              rank: 2
              shape:
              - batch
              - prompt_sequence
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `prompt_lengths`: Authored tensor ABI: dtype int64, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            prompt_lengths:
              dtype: int64
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `max_iterations`: Authored tensor ABI: dtype int64, rank 1, shape [1]. No grouping permission is implied by this shape.
            max_iterations:
              dtype: int64
              rank: 1
              shape:
              - 1
          # Exact graph outputs; each tensor shape/dtype is authored.
          outputs:
            # Port `attention_mask`: Authored tensor ABI: dtype int64, rank 2, shape [batch, capacity]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            attention_mask:
              dtype: int64
              rank: 2
              shape:
              - batch
              - capacity
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `body_attention_mask`: Authored tensor ABI: dtype int64, rank 2, shape [batch, capacity]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            body_attention_mask:
              dtype: int64
              rank: 2
              shape:
              - batch
              - capacity
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `token_slot`: Authored tensor ABI: dtype int64, rank 2, shape [batch, 1]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            token_slot:
              dtype: int64
              rank: 2
              shape:
              - batch
              - 1
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `generated_lengths`: Authored tensor ABI: dtype int64, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            generated_lengths:
              dtype: int64
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `cache_lengths`: Authored tensor ABI: dtype int64, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            cache_lengths:
              dtype: int64
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.0.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.0.key:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.0.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.0.value:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.1.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.1.key:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.1.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.1.value:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.10.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.10.key:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.10.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.10.value:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.11.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.11.key:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.11.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.11.value:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.12.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.12.key:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.12.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.12.value:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.13.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.13.key:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.13.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.13.value:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.14.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.14.key:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.14.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.14.value:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.15.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.15.key:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.15.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.15.value:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.16.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.16.key:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.16.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.16.value:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.17.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.17.key:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.17.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.17.value:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.18.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.18.key:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.18.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.18.value:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.19.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.19.key:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.19.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.19.value:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.2.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.2.key:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.2.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.2.value:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.20.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.20.key:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.20.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.20.value:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.21.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.21.key:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.21.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.21.value:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.22.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.22.key:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.22.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.22.value:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.23.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.23.key:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.23.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.23.value:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.24.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.24.key:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.24.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.24.value:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.25.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.25.key:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.25.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.25.value:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.26.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.26.key:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.26.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.26.value:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.27.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.27.key:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.27.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.27.value:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.3.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.3.key:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.3.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.3.value:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.4.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.4.key:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.4.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.4.value:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.5.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.5.key:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.5.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.5.value:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.6.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.6.key:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.6.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.6.value:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.7.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.7.key:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.7.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.7.value:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.8.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.8.key:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.8.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.8.value:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.9.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.9.key:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `past_key_values.9.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, capacity, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            past_key_values.9.value:
              dtype: float16
              rank: 4
              shape:
              - batch_size
              - 2
              - capacity
              - 128
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
      # Component `decoder_step_update` is a primary model compute graph using onnx from `policies/decoder_step_update.onnx`. No batch_capacity is authored, so the runtime must not group independent requests at this component.
      decoder_step_update:
        # Selects a packaged implementation/artifact only; provider/backend readiness requires separate recorded evidence.
        implementation:
          kind: onnx
          artifact: policies/decoder_step_update.onnx
        # Ports bind workflow names to the artifact's exact graph ABI; model-family inference is forbidden.
        ports:
          # Exact graph inputs; each tensor shape/dtype is authored.
          inputs:
            # Port `attention_mask`: Authored tensor ABI: dtype int64, rank 2, shape [batch, context]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            attention_mask:
              dtype: int64
              rank: 2
              shape:
              - batch
              - context
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `logical_length`: Authored tensor ABI: dtype int64, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            logical_length:
              dtype: int64
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
          # Exact graph outputs; each tensor shape/dtype is authored.
          outputs:
            # Port `next_attention_mask`: Authored tensor ABI: dtype int64, rank 2, shape [batch, context]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            next_attention_mask:
              dtype: int64
              rank: 2
              shape:
              - batch
              - context
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
      # Component `cache_length_update` is a state-transition policy graph using onnx from `policies/cache_length_update.onnx`. Policy graphs compute tensors and semantics; they are not deployment/QoS knobs or backend readiness claims. No batch_capacity is authored, so the runtime must not group independent requests at this component.
      cache_length_update:
        # Selects a packaged implementation/artifact only; provider/backend readiness requires separate recorded evidence.
        implementation:
          kind: onnx
          artifact: policies/cache_length_update.onnx
        # Ports bind workflow names to the artifact's exact graph ABI; model-family inference is forbidden.
        ports:
          # Exact graph inputs; each tensor shape/dtype is authored.
          inputs:
            # Port `left`: Authored tensor ABI: dtype int64, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            left:
              dtype: int64
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `right`: Authored tensor ABI: dtype int64, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            right:
              dtype: int64
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `active`: Authored tensor ABI: dtype bool, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            active:
              dtype: bool
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `done`: Authored tensor ABI: dtype bool, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            done:
              dtype: bool
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
          # Exact graph outputs; each tensor shape/dtype is authored.
          outputs:
            # Port `total`: Authored tensor ABI: dtype int64, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            total:
              dtype: int64
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
      # Component `termination_batch_initializer` is a termination policy graph using onnx from `policies/termination_batch_initializer.onnx`. Policy graphs compute tensors and semantics; they are not deployment/QoS knobs or backend readiness claims. No batch_capacity is authored, so the runtime must not group independent requests at this component.
      termination_batch_initializer:
        # Selects a packaged implementation/artifact only; provider/backend readiness requires separate recorded evidence.
        implementation:
          kind: onnx
          artifact: policies/termination_batch_initializer.onnx
        # Ports bind workflow names to the artifact's exact graph ABI; model-family inference is forbidden.
        ports:
          # Exact graph inputs; each tensor shape/dtype is authored.
          inputs:
            # Port `input_eos_ids`: Authored tensor ABI: dtype int64, rank 2, shape [batch, num_eos]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            input_eos_ids:
              dtype: int64
              rank: 2
              shape:
              - batch
              - num_eos
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `input_eos_lengths`: Authored tensor ABI: dtype int64, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            input_eos_lengths:
              dtype: int64
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `input_max_iterations`: Authored tensor ABI: dtype int64, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            input_max_iterations:
              dtype: int64
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `fallback_max_iterations`: Authored tensor ABI: dtype int64, rank 1, shape [1]. No grouping permission is implied by this shape.
            fallback_max_iterations:
              dtype: int64
              rank: 1
              shape:
              - 1
            # Port `active`: Authored tensor ABI: dtype bool, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            active:
              dtype: bool
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
          # Exact graph outputs; each tensor shape/dtype is authored.
          outputs:
            # Port `row_eos_ids`: Authored tensor ABI: dtype int64, rank 2, shape [batch, num_eos]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            row_eos_ids:
              dtype: int64
              rank: 2
              shape:
              - batch
              - num_eos
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `eos_lengths`: Authored tensor ABI: dtype int64, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            eos_lengths:
              dtype: int64
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `max_iterations`: Authored tensor ABI: dtype int64, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            max_iterations:
              dtype: int64
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
      # Component `token_to_slot` is an authored workflow compute component using onnx from `policies/token_to_slot.onnx`. No batch_capacity is authored, so the runtime must not group independent requests at this component.
      token_to_slot:
        # Selects a packaged implementation/artifact only; provider/backend readiness requires separate recorded evidence.
        implementation:
          kind: onnx
          artifact: policies/token_to_slot.onnx
        # Ports bind workflow names to the artifact's exact graph ABI; model-family inference is forbidden.
        ports:
          # Exact graph inputs; each tensor shape/dtype is authored.
          inputs:
            # Port `token`: Authored tensor ABI: dtype int64, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            token:
              dtype: int64
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
          # Exact graph outputs; each tensor shape/dtype is authored.
          outputs:
            # Port `slot`: Authored tensor ABI: dtype int64, rank 2, shape [batch, 1]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            slot:
              dtype: int64
              rank: 2
              shape:
              - batch
              - 1
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
      # Component `generated_length_update` is a state-transition policy graph using onnx from `policies/generated_length_update.onnx`. Policy graphs compute tensors and semantics; they are not deployment/QoS knobs or backend readiness claims. No batch_capacity is authored, so the runtime must not group independent requests at this component.
      generated_length_update:
        # Selects a packaged implementation/artifact only; provider/backend readiness requires separate recorded evidence.
        implementation:
          kind: onnx
          artifact: policies/generated_length_update.onnx
        # Ports bind workflow names to the artifact's exact graph ABI; model-family inference is forbidden.
        ports:
          # Exact graph inputs; each tensor shape/dtype is authored.
          inputs:
            # Port `left`: Authored tensor ABI: dtype int64, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            left:
              dtype: int64
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `right`: Authored tensor ABI: dtype int64, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            right:
              dtype: int64
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `active`: Authored tensor ABI: dtype bool, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            active:
              dtype: bool
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
            # Port `done`: Authored tensor ABI: dtype bool, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            done:
              dtype: bool
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
          # Exact graph outputs; each tensor shape/dtype is authored.
          outputs:
            # Port `total`: Authored tensor ABI: dtype int64, rank 1, shape [batch]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            total:
              dtype: int64
              rank: 1
              shape:
              - batch
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout:
                kind: request_aligned
                axis: 0
    # State cells make recurrence, cache mutation, lifetime, and ownership explicit. Physical paging/tiering/placement are not encoded by tensor names.
    state:
      # State `token` is semantic tensor carry between workflow steps. Scope=invocation, recurrence=invariant. The authored recurrence determines how it evolves.
      token:
        # Authored tensor ABI: dtype int64, rank 2, shape [batch_size, 1]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: int64
          rank: 2
          shape:
          - batch_size
          - 1
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: initializer.token_slot
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence:
          kind: invariant
      # State `logits` is semantic tensor carry between workflow steps. Scope=invocation, recurrence=invariant. The authored recurrence determines how it evolves.
      logits:
        # Authored tensor ABI: dtype float32, rank 2, shape [batch_size, 151936]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float32
          rank: 2
          shape:
          - batch_size
          - 151936
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.last_logits
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence:
          kind: invariant
      # State `generated_lengths` is loop/scheduler control state. Scope=invocation, recurrence=invariant. The values affect correctness, not merely performance.
      generated_lengths:
        # Authored tensor ABI: dtype int64, rank 1, shape [batch_size]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: int64
          rank: 1
          shape:
          - batch_size
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # State class distinguishes semantic state from advisory/runtime-only bookkeeping.
        class: semantic
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: initializer.generated_lengths
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence:
          kind: invariant
      # State `active` is loop/scheduler control state. Scope=invocation, recurrence=invariant. The values affect correctness, not merely performance.
      active:
        # Authored tensor ABI: dtype bool, rank 1, shape [batch_size]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: bool
          rank: 1
          shape:
          - batch_size
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # State class distinguishes semantic state from advisory/runtime-only bookkeeping.
        class: semantic
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: package.active
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence:
          kind: invariant
      # State `done` is loop/scheduler control state. Scope=invocation, recurrence=invariant. The values affect correctness, not merely performance.
      done:
        # Authored tensor ABI: dtype bool, rank 1, shape [batch_size]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: bool
          rank: 1
          shape:
          - batch_size
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # State class distinguishes semantic state from advisory/runtime-only bookkeeping.
        class: semantic
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: package.not_done
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence:
          kind: invariant
      # State `accepted_len` is loop/scheduler control state. Scope=invocation, recurrence=invariant. The values affect correctness, not merely performance.
      accepted_len:
        # Authored tensor ABI: dtype int64, rank 1, shape [batch_size]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: int64
          rank: 1
          shape:
          - batch_size
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # State class distinguishes semantic state from advisory/runtime-only bookkeeping.
        class: semantic
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: package.zero_batch
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence:
          kind: invariant
      # State `cache_lengths` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_lengths:
        # Authored tensor ABI: dtype int64, rank 1, shape [batch_size]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: int64
          rank: 1
          shape:
          - batch_size
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # State class distinguishes semantic state from advisory/runtime-only bookkeeping.
        class: semantic
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: initializer.cache_lengths
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence:
          kind: invariant
      # State `rng_counter` is deterministic random-stream state. Scope=invocation, recurrence=invariant. Its recurrence preserves reproducibility across loop steps.
      rng_counter:
        # Authored tensor ABI: dtype int64, rank 1, shape [batch_size]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: int64
          rank: 1
          shape:
          - batch_size
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # State class distinguishes semantic state from advisory/runtime-only bookkeeping.
        class: semantic
        # Initializer names the exact first SSA value for this state cell.
        initializer: request.rng_counter
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence:
          kind: invariant
      # State `attention_mask` is semantic tensor carry between workflow steps. Scope=invocation, recurrence=invariant. The authored recurrence determines how it evolves.
      attention_mask:
        # Authored tensor ABI: dtype int64, rank 2, shape [batch_size, context]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: int64
          rank: 2
          shape:
          - batch_size
          - context
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: initializer.attention_mask
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence:
          kind: invariant
      # State `cache_0` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_0:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.0.key
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_1` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_1:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.1.key
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_2` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_2:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.2.key
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_3` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_3:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.3.key
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_4` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_4:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.4.key
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_5` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_5:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.5.key
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_6` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_6:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.6.key
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_7` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_7:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.7.key
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_8` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_8:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.8.key
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_9` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_9:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.9.key
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_10` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_10:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.10.key
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_11` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_11:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.11.key
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_12` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_12:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.12.key
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_13` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_13:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.13.key
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_14` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_14:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.14.key
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_15` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_15:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.15.key
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_16` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_16:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.16.key
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_17` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_17:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.17.key
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_18` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_18:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.18.key
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_19` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_19:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.19.key
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_20` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_20:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.20.key
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_21` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_21:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.21.key
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_22` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_22:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.22.key
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_23` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_23:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.23.key
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_24` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_24:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.24.key
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_25` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_25:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.25.key
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_26` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_26:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.26.key
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_27` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_27:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.27.key
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_28` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_28:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.0.value
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_29` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_29:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.1.value
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_30` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_30:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.2.value
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_31` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_31:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.3.value
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_32` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_32:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.4.value
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_33` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_33:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.5.value
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_34` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_34:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.6.value
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_35` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_35:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.7.value
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_36` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_36:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.8.value
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_37` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_37:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.9.value
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_38` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_38:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.10.value
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_39` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_39:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.11.value
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_40` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_40:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.12.value
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_41` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_41:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.13.value
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_42` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_42:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.14.value
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_43` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_43:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.15.value
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_44` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_44:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.16.value
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_45` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_45:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.17.value
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_46` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_46:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.18.value
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_47` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_47:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.19.value
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_48` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_48:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.20.value
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_49` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_49:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.21.value
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_50` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_50:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.22.value
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_51` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_51:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.23.value
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_52` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_52:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.24.value
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_53` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_53:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.25.value
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_54` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_54:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.26.value
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
      # State `cache_55` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=bounded. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      cache_55:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch_size, 2, past_sequence_length, 128]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 4
          shape:
          - batch_size
          - 2
          - past_sequence_length
          - 128
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout:
            kind: request_aligned
            axis: 0
        # `invocation` defines the semantic lifetime boundary; it is not a worker-placement policy.
        scope: invocation
        # Initializer names the exact first SSA value for this state cell.
        initializer: decoder.setup.present.27.value
        # Recurrence defines the next-value rule: {kind=bounded, axis=2, max=package.max_context}.
        recurrence:
          kind: bounded
          axis: 2
          max: package.max_context
        # Runtime management means the engine preserves this semantic state; storage layout remains backend policy.
        management: runtime
        # The state may be released only at this correctness boundary.
        release_boundary: invocation
        # Service-group membership links this cell to the typed cache/state service below.
        service_group: decoder_cache
    # Serving binds workflow tensors to scheduler/session semantics. It does not choose worker count, provider, QoS, or claim thread safety.
    serving:
      # Active-row binding tells the scheduler which request rows still execute.
      active: active
      # Done-row binding is the semantic termination signal.
      done: done
      # Accepted-length binding controls cache/state advancement after verification.
      accepted_len: accepted_len
      # State service groups connect semantic cells to cache/recurrent update contracts used by serving admission.
      state_service:
        # Each group declares semantic update/rollback behavior; physical cache storage remains deployment/runtime policy.
        groups:
          # Group `decoder_cache` is semantic `full_attention` state in `bnsh` layout with update None; rollback/fork/snapshot are correctness capabilities.
          decoder_cache:
            kind: full_attention
            # Sequence axis identifies logical history growth for append/rollback validation.
            sequence_axis: 2
            # Layout names tensor axis order, not physical paging or device placement.
            layout: bnsh
            logical_lengths: cache_lengths
            aliasing: permitted
            reuse:
              prefix_reusable: true
              evictable_prefix: false
            # Port bindings connect service state to exact component input/output names.
            ports:
              model:
                cache_0:
                  input: past_key_values.0.key
                  output: present.0.key
                  role: key
                  layer: 0
                cache_1:
                  input: past_key_values.1.key
                  output: present.1.key
                  role: key
                  layer: 1
                cache_2:
                  input: past_key_values.2.key
                  output: present.2.key
                  role: key
                  layer: 2
                cache_3:
                  input: past_key_values.3.key
                  output: present.3.key
                  role: key
                  layer: 3
                cache_4:
                  input: past_key_values.4.key
                  output: present.4.key
                  role: key
                  layer: 4
                cache_5:
                  input: past_key_values.5.key
                  output: present.5.key
                  role: key
                  layer: 5
                cache_6:
                  input: past_key_values.6.key
                  output: present.6.key
                  role: key
                  layer: 6
                cache_7:
                  input: past_key_values.7.key
                  output: present.7.key
                  role: key
                  layer: 7
                cache_8:
                  input: past_key_values.8.key
                  output: present.8.key
                  role: key
                  layer: 8
                cache_9:
                  input: past_key_values.9.key
                  output: present.9.key
                  role: key
                  layer: 9
                cache_10:
                  input: past_key_values.10.key
                  output: present.10.key
                  role: key
                  layer: 10
                cache_11:
                  input: past_key_values.11.key
                  output: present.11.key
                  role: key
                  layer: 11
                cache_12:
                  input: past_key_values.12.key
                  output: present.12.key
                  role: key
                  layer: 12
                cache_13:
                  input: past_key_values.13.key
                  output: present.13.key
                  role: key
                  layer: 13
                cache_14:
                  input: past_key_values.14.key
                  output: present.14.key
                  role: key
                  layer: 14
                cache_15:
                  input: past_key_values.15.key
                  output: present.15.key
                  role: key
                  layer: 15
                cache_16:
                  input: past_key_values.16.key
                  output: present.16.key
                  role: key
                  layer: 16
                cache_17:
                  input: past_key_values.17.key
                  output: present.17.key
                  role: key
                  layer: 17
                cache_18:
                  input: past_key_values.18.key
                  output: present.18.key
                  role: key
                  layer: 18
                cache_19:
                  input: past_key_values.19.key
                  output: present.19.key
                  role: key
                  layer: 19
                cache_20:
                  input: past_key_values.20.key
                  output: present.20.key
                  role: key
                  layer: 20
                cache_21:
                  input: past_key_values.21.key
                  output: present.21.key
                  role: key
                  layer: 21
                cache_22:
                  input: past_key_values.22.key
                  output: present.22.key
                  role: key
                  layer: 22
                cache_23:
                  input: past_key_values.23.key
                  output: present.23.key
                  role: key
                  layer: 23
                cache_24:
                  input: past_key_values.24.key
                  output: present.24.key
                  role: key
                  layer: 24
                cache_25:
                  input: past_key_values.25.key
                  output: present.25.key
                  role: key
                  layer: 25
                cache_26:
                  input: past_key_values.26.key
                  output: present.26.key
                  role: key
                  layer: 26
                cache_27:
                  input: past_key_values.27.key
                  output: present.27.key
                  role: key
                  layer: 27
                cache_28:
                  input: past_key_values.0.value
                  output: present.0.value
                  role: value
                  layer: 0
                cache_29:
                  input: past_key_values.1.value
                  output: present.1.value
                  role: value
                  layer: 1
                cache_30:
                  input: past_key_values.2.value
                  output: present.2.value
                  role: value
                  layer: 2
                cache_31:
                  input: past_key_values.3.value
                  output: present.3.value
                  role: value
                  layer: 3
                cache_32:
                  input: past_key_values.4.value
                  output: present.4.value
                  role: value
                  layer: 4
                cache_33:
                  input: past_key_values.5.value
                  output: present.5.value
                  role: value
                  layer: 5
                cache_34:
                  input: past_key_values.6.value
                  output: present.6.value
                  role: value
                  layer: 6
                cache_35:
                  input: past_key_values.7.value
                  output: present.7.value
                  role: value
                  layer: 7
                cache_36:
                  input: past_key_values.8.value
                  output: present.8.value
                  role: value
                  layer: 8
                cache_37:
                  input: past_key_values.9.value
                  output: present.9.value
                  role: value
                  layer: 9
                cache_38:
                  input: past_key_values.10.value
                  output: present.10.value
                  role: value
                  layer: 10
                cache_39:
                  input: past_key_values.11.value
                  output: present.11.value
                  role: value
                  layer: 11
                cache_40:
                  input: past_key_values.12.value
                  output: present.12.value
                  role: value
                  layer: 12
                cache_41:
                  input: past_key_values.13.value
                  output: present.13.value
                  role: value
                  layer: 13
                cache_42:
                  input: past_key_values.14.value
                  output: present.14.value
                  role: value
                  layer: 14
                cache_43:
                  input: past_key_values.15.value
                  output: present.15.value
                  role: value
                  layer: 15
                cache_44:
                  input: past_key_values.16.value
                  output: present.16.value
                  role: value
                  layer: 16
                cache_45:
                  input: past_key_values.17.value
                  output: present.17.value
                  role: value
                  layer: 17
                cache_46:
                  input: past_key_values.18.value
                  output: present.18.value
                  role: value
                  layer: 18
                cache_47:
                  input: past_key_values.19.value
                  output: present.19.value
                  role: value
                  layer: 19
                cache_48:
                  input: past_key_values.20.value
                  output: present.20.value
                  role: value
                  layer: 20
                cache_49:
                  input: past_key_values.21.value
                  output: present.21.value
                  role: value
                  layer: 21
                cache_50:
                  input: past_key_values.22.value
                  output: present.22.value
                  role: value
                  layer: 22
                cache_51:
                  input: past_key_values.23.value
                  output: present.23.value
                  role: value
                  layer: 23
                cache_52:
                  input: past_key_values.24.value
                  output: present.24.value
                  role: value
                  layer: 24
                cache_53:
                  input: past_key_values.25.value
                  output: present.25.value
                  role: value
                  layer: 25
                cache_54:
                  input: past_key_values.26.value
                  output: present.26.value
                  role: value
                  layer: 26
                cache_55:
                  input: past_key_values.27.value
                  output: present.27.value
                  role: value
                  layer: 27
    # The ordered SSA program defines actual execution and dataflow; capability strings are not a substitute for these steps.
    steps:
      # Run a bounded semantic loop (request.max_iterations); carried state and termination values make progress explicit and fail closed if unsupported.
    - kind: loop
      setup:
        # Invoke component `decoder_state_initializer` with explicit SSA input/output bindings; this is executable tensor computation, not a host-side capability flag.
      - kind: invoke
        component: decoder_state_initializer
        # SSA bindings supply exact values to this step; no implicit tensor discovery occurs.
        inputs:
          prompt_tokens: request.input_ids
          prompt_lengths: request.prompt_lengths
          max_iterations: request.max_iterations
        # SSA output names become the only values subsequent steps may consume.
        outputs:
          attention_mask: initializer.attention_mask
          body_attention_mask: initializer.body_attention_mask
          token_slot: initializer.token_slot
          generated_lengths: initializer.generated_lengths
          cache_lengths: initializer.cache_lengths
          past_key_values.0.key: initializer.past_key_values.0.key
          past_key_values.0.value: initializer.past_key_values.0.value
          past_key_values.1.key: initializer.past_key_values.1.key
          past_key_values.1.value: initializer.past_key_values.1.value
          past_key_values.10.key: initializer.past_key_values.10.key
          past_key_values.10.value: initializer.past_key_values.10.value
          past_key_values.11.key: initializer.past_key_values.11.key
          past_key_values.11.value: initializer.past_key_values.11.value
          past_key_values.12.key: initializer.past_key_values.12.key
          past_key_values.12.value: initializer.past_key_values.12.value
          past_key_values.13.key: initializer.past_key_values.13.key
          past_key_values.13.value: initializer.past_key_values.13.value
          past_key_values.14.key: initializer.past_key_values.14.key
          past_key_values.14.value: initializer.past_key_values.14.value
          past_key_values.15.key: initializer.past_key_values.15.key
          past_key_values.15.value: initializer.past_key_values.15.value
          past_key_values.16.key: initializer.past_key_values.16.key
          past_key_values.16.value: initializer.past_key_values.16.value
          past_key_values.17.key: initializer.past_key_values.17.key
          past_key_values.17.value: initializer.past_key_values.17.value
          past_key_values.18.key: initializer.past_key_values.18.key
          past_key_values.18.value: initializer.past_key_values.18.value
          past_key_values.19.key: initializer.past_key_values.19.key
          past_key_values.19.value: initializer.past_key_values.19.value
          past_key_values.2.key: initializer.past_key_values.2.key
          past_key_values.2.value: initializer.past_key_values.2.value
          past_key_values.20.key: initializer.past_key_values.20.key
          past_key_values.20.value: initializer.past_key_values.20.value
          past_key_values.21.key: initializer.past_key_values.21.key
          past_key_values.21.value: initializer.past_key_values.21.value
          past_key_values.22.key: initializer.past_key_values.22.key
          past_key_values.22.value: initializer.past_key_values.22.value
          past_key_values.23.key: initializer.past_key_values.23.key
          past_key_values.23.value: initializer.past_key_values.23.value
          past_key_values.24.key: initializer.past_key_values.24.key
          past_key_values.24.value: initializer.past_key_values.24.value
          past_key_values.25.key: initializer.past_key_values.25.key
          past_key_values.25.value: initializer.past_key_values.25.value
          past_key_values.26.key: initializer.past_key_values.26.key
          past_key_values.26.value: initializer.past_key_values.26.value
          past_key_values.27.key: initializer.past_key_values.27.key
          past_key_values.27.value: initializer.past_key_values.27.value
          past_key_values.3.key: initializer.past_key_values.3.key
          past_key_values.3.value: initializer.past_key_values.3.value
          past_key_values.4.key: initializer.past_key_values.4.key
          past_key_values.4.value: initializer.past_key_values.4.value
          past_key_values.5.key: initializer.past_key_values.5.key
          past_key_values.5.value: initializer.past_key_values.5.value
          past_key_values.6.key: initializer.past_key_values.6.key
          past_key_values.6.value: initializer.past_key_values.6.value
          past_key_values.7.key: initializer.past_key_values.7.key
          past_key_values.7.value: initializer.past_key_values.7.value
          past_key_values.8.key: initializer.past_key_values.8.key
          past_key_values.8.value: initializer.past_key_values.8.value
          past_key_values.9.key: initializer.past_key_values.9.key
          past_key_values.9.value: initializer.past_key_values.9.value
        # Invoke component `model` with explicit SSA input/output bindings; this is executable tensor computation, not a host-side capability flag.
      - kind: invoke
        component: model
        # SSA bindings supply exact values to this step; no implicit tensor discovery occurs.
        inputs:
          input_ids: request.input_ids
          past_key_values.0.key: initializer.past_key_values.0.key
          past_key_values.1.key: initializer.past_key_values.1.key
          past_key_values.2.key: initializer.past_key_values.2.key
          past_key_values.3.key: initializer.past_key_values.3.key
          past_key_values.4.key: initializer.past_key_values.4.key
          past_key_values.5.key: initializer.past_key_values.5.key
          past_key_values.6.key: initializer.past_key_values.6.key
          past_key_values.7.key: initializer.past_key_values.7.key
          past_key_values.8.key: initializer.past_key_values.8.key
          past_key_values.9.key: initializer.past_key_values.9.key
          past_key_values.10.key: initializer.past_key_values.10.key
          past_key_values.11.key: initializer.past_key_values.11.key
          past_key_values.12.key: initializer.past_key_values.12.key
          past_key_values.13.key: initializer.past_key_values.13.key
          past_key_values.14.key: initializer.past_key_values.14.key
          past_key_values.15.key: initializer.past_key_values.15.key
          past_key_values.16.key: initializer.past_key_values.16.key
          past_key_values.17.key: initializer.past_key_values.17.key
          past_key_values.18.key: initializer.past_key_values.18.key
          past_key_values.19.key: initializer.past_key_values.19.key
          past_key_values.20.key: initializer.past_key_values.20.key
          past_key_values.21.key: initializer.past_key_values.21.key
          past_key_values.22.key: initializer.past_key_values.22.key
          past_key_values.23.key: initializer.past_key_values.23.key
          past_key_values.24.key: initializer.past_key_values.24.key
          past_key_values.25.key: initializer.past_key_values.25.key
          past_key_values.26.key: initializer.past_key_values.26.key
          past_key_values.27.key: initializer.past_key_values.27.key
          past_key_values.0.value: initializer.past_key_values.0.value
          past_key_values.1.value: initializer.past_key_values.1.value
          past_key_values.2.value: initializer.past_key_values.2.value
          past_key_values.3.value: initializer.past_key_values.3.value
          past_key_values.4.value: initializer.past_key_values.4.value
          past_key_values.5.value: initializer.past_key_values.5.value
          past_key_values.6.value: initializer.past_key_values.6.value
          past_key_values.7.value: initializer.past_key_values.7.value
          past_key_values.8.value: initializer.past_key_values.8.value
          past_key_values.9.value: initializer.past_key_values.9.value
          past_key_values.10.value: initializer.past_key_values.10.value
          past_key_values.11.value: initializer.past_key_values.11.value
          past_key_values.12.value: initializer.past_key_values.12.value
          past_key_values.13.value: initializer.past_key_values.13.value
          past_key_values.14.value: initializer.past_key_values.14.value
          past_key_values.15.value: initializer.past_key_values.15.value
          past_key_values.16.value: initializer.past_key_values.16.value
          past_key_values.17.value: initializer.past_key_values.17.value
          past_key_values.18.value: initializer.past_key_values.18.value
          past_key_values.19.value: initializer.past_key_values.19.value
          past_key_values.20.value: initializer.past_key_values.20.value
          past_key_values.21.value: initializer.past_key_values.21.value
          past_key_values.22.value: initializer.past_key_values.22.value
          past_key_values.23.value: initializer.past_key_values.23.value
          past_key_values.24.value: initializer.past_key_values.24.value
          past_key_values.25.value: initializer.past_key_values.25.value
          past_key_values.26.value: initializer.past_key_values.26.value
          past_key_values.27.value: initializer.past_key_values.27.value
          attention_mask: initializer.attention_mask
        # SSA output names become the only values subsequent steps may consume.
        outputs:
          logits: decoder.setup.logits
          present.0.key: decoder.setup.present.0.key
          present.1.key: decoder.setup.present.1.key
          present.2.key: decoder.setup.present.2.key
          present.3.key: decoder.setup.present.3.key
          present.4.key: decoder.setup.present.4.key
          present.5.key: decoder.setup.present.5.key
          present.6.key: decoder.setup.present.6.key
          present.7.key: decoder.setup.present.7.key
          present.8.key: decoder.setup.present.8.key
          present.9.key: decoder.setup.present.9.key
          present.10.key: decoder.setup.present.10.key
          present.11.key: decoder.setup.present.11.key
          present.12.key: decoder.setup.present.12.key
          present.13.key: decoder.setup.present.13.key
          present.14.key: decoder.setup.present.14.key
          present.15.key: decoder.setup.present.15.key
          present.16.key: decoder.setup.present.16.key
          present.17.key: decoder.setup.present.17.key
          present.18.key: decoder.setup.present.18.key
          present.19.key: decoder.setup.present.19.key
          present.20.key: decoder.setup.present.20.key
          present.21.key: decoder.setup.present.21.key
          present.22.key: decoder.setup.present.22.key
          present.23.key: decoder.setup.present.23.key
          present.24.key: decoder.setup.present.24.key
          present.25.key: decoder.setup.present.25.key
          present.26.key: decoder.setup.present.26.key
          present.27.key: decoder.setup.present.27.key
          present.0.value: decoder.setup.present.0.value
          present.1.value: decoder.setup.present.1.value
          present.2.value: decoder.setup.present.2.value
          present.3.value: decoder.setup.present.3.value
          present.4.value: decoder.setup.present.4.value
          present.5.value: decoder.setup.present.5.value
          present.6.value: decoder.setup.present.6.value
          present.7.value: decoder.setup.present.7.value
          present.8.value: decoder.setup.present.8.value
          present.9.value: decoder.setup.present.9.value
          present.10.value: decoder.setup.present.10.value
          present.11.value: decoder.setup.present.11.value
          present.12.value: decoder.setup.present.12.value
          present.13.value: decoder.setup.present.13.value
          present.14.value: decoder.setup.present.14.value
          present.15.value: decoder.setup.present.15.value
          present.16.value: decoder.setup.present.16.value
          present.17.value: decoder.setup.present.17.value
          present.18.value: decoder.setup.present.18.value
          present.19.value: decoder.setup.present.19.value
          present.20.value: decoder.setup.present.20.value
          present.21.value: decoder.setup.present.21.value
          present.22.value: decoder.setup.present.22.value
          present.23.value: decoder.setup.present.23.value
          present.24.value: decoder.setup.present.24.value
          present.25.value: decoder.setup.present.25.value
          present.26.value: decoder.setup.present.26.value
          present.27.value: decoder.setup.present.27.value
        # Invoke component `termination_batch_initializer` with explicit SSA input/output bindings; this is executable tensor computation, not a host-side capability flag.
      - kind: invoke
        component: termination_batch_initializer
        # SSA bindings supply exact values to this step; no implicit tensor discovery occurs.
        inputs:
          input_eos_ids: request.eos_ids
          input_eos_lengths: request.eos_lengths
          input_max_iterations: request.row_max_iterations
          fallback_max_iterations: request.max_iterations
          active: package.active
        # SSA output names become the only values subsequent steps may consume.
        outputs:
          row_eos_ids: termination.eos_ids
          eos_lengths: termination.eos_lengths
          max_iterations: termination.max_iterations
        # Invoke component `last_token_logits` with explicit SSA input/output bindings; this is executable tensor computation, not a host-side capability flag.
      - kind: invoke
        component: last_token_logits
        # SSA bindings supply exact values to this step; no implicit tensor discovery occurs.
        inputs:
          logits: decoder.setup.logits
        # SSA output names become the only values subsequent steps may consume.
        outputs:
          last_logits: decoder.setup.last_logits
      # Nested steps remain part of the typed executable workflow.
      steps:
        # Invoke component `token_sampler` with explicit SSA input/output bindings; this is executable tensor computation, not a host-side capability flag.
      - kind: invoke
        component: token_sampler
        # SSA bindings supply exact values to this step; no implicit tensor discovery occurs.
        inputs:
          logits: logits
          temperature: request.temperature
          top_k: request.top_k
          top_p: request.top_p
          min_p: request.min_p
          seed: request.seed
          counter: rng_counter
          active: active
          done: done
        # SSA output names become the only values subsequent steps may consume.
        outputs:
          token: sample.body
          next_counter: sample.next_counter
        # Invoke component `token_to_slot` with explicit SSA input/output bindings; this is executable tensor computation, not a host-side capability flag.
      - kind: invoke
        component: token_to_slot
        # SSA bindings supply exact values to this step; no implicit tensor discovery occurs.
        inputs:
          token: sample.body
        # SSA output names become the only values subsequent steps may consume.
        outputs:
          slot: sample.slot
        # Invoke component `generated_length_update` with explicit SSA input/output bindings; this is executable tensor computation, not a host-side capability flag.
      - kind: invoke
        component: generated_length_update
        # SSA bindings supply exact values to this step; no implicit tensor discovery occurs.
        inputs:
          left: generated_lengths
          right: package.one_token
          active: active
          done: done
        # SSA output names become the only values subsequent steps may consume.
        outputs:
          total: token.next_lengths
        # Invoke component `generated_length_update` with explicit SSA input/output bindings; this is executable tensor computation, not a host-side capability flag.
      - kind: invoke
        component: generated_length_update
        # SSA bindings supply exact values to this step; no implicit tensor discovery occurs.
        inputs:
          left: package.zero_batch
          right: package.one_token
          active: active
          done: done
        # SSA output names become the only values subsequent steps may consume.
        outputs:
          total: token.emitted_length
        # Invoke component `token_state_update` with explicit SSA input/output bindings; this is executable tensor computation, not a host-side capability flag.
      - kind: invoke
        component: token_state_update
        # SSA bindings supply exact values to this step; no implicit tensor discovery occurs.
        inputs:
          current: token
          update: sample.slot
          active: active
          done: done
        # SSA output names become the only values subsequent steps may consume.
        outputs:
          next: token.body
        # Invoke component `termination` with explicit SSA input/output bindings; this is executable tensor computation, not a host-side capability flag.
      - kind: invoke
        component: termination
        # SSA bindings supply exact values to this step; no implicit tensor discovery occurs.
        inputs:
          tokens: sample.body
          eos_ids: termination.eos_ids
          eos_lengths: termination.eos_lengths
          iteration: loop.iteration
          max_iterations: termination.max_iterations
          active: active
        # SSA output names become the only values subsequent steps may consume.
        outputs:
          done: loop.done
          continue: loop.continue
          next_active: loop.next_active
        # Invoke component `cache_length_update` with explicit SSA input/output bindings; this is executable tensor computation, not a host-side capability flag.
      - kind: invoke
        component: cache_length_update
        # SSA bindings supply exact values to this step; no implicit tensor discovery occurs.
        inputs:
          left: cache_lengths
          right: package.one_token
          active: active
          done: done
        # SSA output names become the only values subsequent steps may consume.
        outputs:
          total: cache_lengths.next
        # Invoke component `cache_length_update` with explicit SSA input/output bindings; this is executable tensor computation, not a host-side capability flag.
      - kind: invoke
        component: cache_length_update
        # SSA bindings supply exact values to this step; no implicit tensor discovery occurs.
        inputs:
          left: package.zero_batch
          right: package.one_token
          active: active
          done: done
        # SSA output names become the only values subsequent steps may consume.
        outputs:
          total: accepted_len.next
        # Publish `token.body` as output `tokens` using append semantics.
      - kind: emit
        value: token.body
        output: tokens
        # Emit/update mode is semantic replacement/append behavior, not a buffering optimization.
        mode: append
        # Valid length prevents padded/unused elements from becoming observable output.
        valid_length: token.emitted_length
        when: active
        # Invoke component `decoder_step_update` with explicit SSA input/output bindings; this is executable tensor computation, not a host-side capability flag.
      - kind: invoke
        component: decoder_step_update
        # SSA bindings supply exact values to this step; no implicit tensor discovery occurs.
        inputs:
          attention_mask: attention_mask
          logical_length: cache_lengths
        # SSA output names become the only values subsequent steps may consume.
        outputs:
          next_attention_mask: decoder_step.body_attention_mask
        # Invoke component `model` with explicit SSA input/output bindings; this is executable tensor computation, not a host-side capability flag.
      - kind: invoke
        component: model
        # SSA bindings supply exact values to this step; no implicit tensor discovery occurs.
        inputs:
          input_ids: token.body
          past_key_values.0.key: cache_0
          past_key_values.1.key: cache_1
          past_key_values.2.key: cache_2
          past_key_values.3.key: cache_3
          past_key_values.4.key: cache_4
          past_key_values.5.key: cache_5
          past_key_values.6.key: cache_6
          past_key_values.7.key: cache_7
          past_key_values.8.key: cache_8
          past_key_values.9.key: cache_9
          past_key_values.10.key: cache_10
          past_key_values.11.key: cache_11
          past_key_values.12.key: cache_12
          past_key_values.13.key: cache_13
          past_key_values.14.key: cache_14
          past_key_values.15.key: cache_15
          past_key_values.16.key: cache_16
          past_key_values.17.key: cache_17
          past_key_values.18.key: cache_18
          past_key_values.19.key: cache_19
          past_key_values.20.key: cache_20
          past_key_values.21.key: cache_21
          past_key_values.22.key: cache_22
          past_key_values.23.key: cache_23
          past_key_values.24.key: cache_24
          past_key_values.25.key: cache_25
          past_key_values.26.key: cache_26
          past_key_values.27.key: cache_27
          past_key_values.0.value: cache_28
          past_key_values.1.value: cache_29
          past_key_values.2.value: cache_30
          past_key_values.3.value: cache_31
          past_key_values.4.value: cache_32
          past_key_values.5.value: cache_33
          past_key_values.6.value: cache_34
          past_key_values.7.value: cache_35
          past_key_values.8.value: cache_36
          past_key_values.9.value: cache_37
          past_key_values.10.value: cache_38
          past_key_values.11.value: cache_39
          past_key_values.12.value: cache_40
          past_key_values.13.value: cache_41
          past_key_values.14.value: cache_42
          past_key_values.15.value: cache_43
          past_key_values.16.value: cache_44
          past_key_values.17.value: cache_45
          past_key_values.18.value: cache_46
          past_key_values.19.value: cache_47
          past_key_values.20.value: cache_48
          past_key_values.21.value: cache_49
          past_key_values.22.value: cache_50
          past_key_values.23.value: cache_51
          past_key_values.24.value: cache_52
          past_key_values.25.value: cache_53
          past_key_values.26.value: cache_54
          past_key_values.27.value: cache_55
          attention_mask: decoder_step.body_attention_mask
        # SSA output names become the only values subsequent steps may consume.
        outputs:
          logits: decoder.body.logits
          present.0.key: decoder.body.present.0.key
          present.1.key: decoder.body.present.1.key
          present.2.key: decoder.body.present.2.key
          present.3.key: decoder.body.present.3.key
          present.4.key: decoder.body.present.4.key
          present.5.key: decoder.body.present.5.key
          present.6.key: decoder.body.present.6.key
          present.7.key: decoder.body.present.7.key
          present.8.key: decoder.body.present.8.key
          present.9.key: decoder.body.present.9.key
          present.10.key: decoder.body.present.10.key
          present.11.key: decoder.body.present.11.key
          present.12.key: decoder.body.present.12.key
          present.13.key: decoder.body.present.13.key
          present.14.key: decoder.body.present.14.key
          present.15.key: decoder.body.present.15.key
          present.16.key: decoder.body.present.16.key
          present.17.key: decoder.body.present.17.key
          present.18.key: decoder.body.present.18.key
          present.19.key: decoder.body.present.19.key
          present.20.key: decoder.body.present.20.key
          present.21.key: decoder.body.present.21.key
          present.22.key: decoder.body.present.22.key
          present.23.key: decoder.body.present.23.key
          present.24.key: decoder.body.present.24.key
          present.25.key: decoder.body.present.25.key
          present.26.key: decoder.body.present.26.key
          present.27.key: decoder.body.present.27.key
          present.0.value: decoder.body.present.0.value
          present.1.value: decoder.body.present.1.value
          present.2.value: decoder.body.present.2.value
          present.3.value: decoder.body.present.3.value
          present.4.value: decoder.body.present.4.value
          present.5.value: decoder.body.present.5.value
          present.6.value: decoder.body.present.6.value
          present.7.value: decoder.body.present.7.value
          present.8.value: decoder.body.present.8.value
          present.9.value: decoder.body.present.9.value
          present.10.value: decoder.body.present.10.value
          present.11.value: decoder.body.present.11.value
          present.12.value: decoder.body.present.12.value
          present.13.value: decoder.body.present.13.value
          present.14.value: decoder.body.present.14.value
          present.15.value: decoder.body.present.15.value
          present.16.value: decoder.body.present.16.value
          present.17.value: decoder.body.present.17.value
          present.18.value: decoder.body.present.18.value
          present.19.value: decoder.body.present.19.value
          present.20.value: decoder.body.present.20.value
          present.21.value: decoder.body.present.21.value
          present.22.value: decoder.body.present.22.value
          present.23.value: decoder.body.present.23.value
          present.24.value: decoder.body.present.24.value
          present.25.value: decoder.body.present.25.value
          present.26.value: decoder.body.present.26.value
          present.27.value: decoder.body.present.27.value
        # Invoke component `last_token_logits` with explicit SSA input/output bindings; this is executable tensor computation, not a host-side capability flag.
      - kind: invoke
        component: last_token_logits
        # SSA bindings supply exact values to this step; no implicit tensor discovery occurs.
        inputs:
          logits: decoder.body.logits
        # SSA output names become the only values subsequent steps may consume.
        outputs:
          last_logits: decoder.body.last_logits
      continue_when: active
      max_iterations: request.max_iterations
      carried:
      - cell: token
        next: token.body
      - cell: logits
        next: decoder.body.last_logits
      - cell: generated_lengths
        next: token.next_lengths
      - cell: active
        next: loop.next_active
      - cell: done
        next: loop.done
      - cell: cache_lengths
        next: cache_lengths.next
      - cell: accepted_len
        next: accepted_len.next
      - cell: rng_counter
        next: sample.next_counter
      - cell: attention_mask
        next: decoder_step.body_attention_mask
      - cell: cache_0
        next: decoder.body.present.0.key
      - cell: cache_1
        next: decoder.body.present.1.key
      - cell: cache_2
        next: decoder.body.present.2.key
      - cell: cache_3
        next: decoder.body.present.3.key
      - cell: cache_4
        next: decoder.body.present.4.key
      - cell: cache_5
        next: decoder.body.present.5.key
      - cell: cache_6
        next: decoder.body.present.6.key
      - cell: cache_7
        next: decoder.body.present.7.key
      - cell: cache_8
        next: decoder.body.present.8.key
      - cell: cache_9
        next: decoder.body.present.9.key
      - cell: cache_10
        next: decoder.body.present.10.key
      - cell: cache_11
        next: decoder.body.present.11.key
      - cell: cache_12
        next: decoder.body.present.12.key
      - cell: cache_13
        next: decoder.body.present.13.key
      - cell: cache_14
        next: decoder.body.present.14.key
      - cell: cache_15
        next: decoder.body.present.15.key
      - cell: cache_16
        next: decoder.body.present.16.key
      - cell: cache_17
        next: decoder.body.present.17.key
      - cell: cache_18
        next: decoder.body.present.18.key
      - cell: cache_19
        next: decoder.body.present.19.key
      - cell: cache_20
        next: decoder.body.present.20.key
      - cell: cache_21
        next: decoder.body.present.21.key
      - cell: cache_22
        next: decoder.body.present.22.key
      - cell: cache_23
        next: decoder.body.present.23.key
      - cell: cache_24
        next: decoder.body.present.24.key
      - cell: cache_25
        next: decoder.body.present.25.key
      - cell: cache_26
        next: decoder.body.present.26.key
      - cell: cache_27
        next: decoder.body.present.27.key
      - cell: cache_28
        next: decoder.body.present.0.value
      - cell: cache_29
        next: decoder.body.present.1.value
      - cell: cache_30
        next: decoder.body.present.2.value
      - cell: cache_31
        next: decoder.body.present.3.value
      - cell: cache_32
        next: decoder.body.present.4.value
      - cell: cache_33
        next: decoder.body.present.5.value
      - cell: cache_34
        next: decoder.body.present.6.value
      - cell: cache_35
        next: decoder.body.present.7.value
      - cell: cache_36
        next: decoder.body.present.8.value
      - cell: cache_37
        next: decoder.body.present.9.value
      - cell: cache_38
        next: decoder.body.present.10.value
      - cell: cache_39
        next: decoder.body.present.11.value
      - cell: cache_40
        next: decoder.body.present.12.value
      - cell: cache_41
        next: decoder.body.present.13.value
      - cell: cache_42
        next: decoder.body.present.14.value
      - cell: cache_43
        next: decoder.body.present.15.value
      - cell: cache_44
        next: decoder.body.present.16.value
      - cell: cache_45
        next: decoder.body.present.17.value
      - cell: cache_46
        next: decoder.body.present.18.value
      - cell: cache_47
        next: decoder.body.present.19.value
      - cell: cache_48
        next: decoder.body.present.20.value
      - cell: cache_49
        next: decoder.body.present.21.value
      - cell: cache_50
        next: decoder.body.present.22.value
      - cell: cache_51
        next: decoder.body.present.23.value
      - cell: cache_52
        next: decoder.body.present.24.value
      - cell: cache_53
        next: decoder.body.present.25.value
      - cell: cache_54
        next: decoder.body.present.26.value
      - cell: cache_55
        next: decoder.body.present.27.value
      termination: generation_eos
      iteration:
        value: loop.iteration
        contract:
          dtype: int64
          rank: 1
          shape:
          - 1
# Package facts are portable execution defaults and identities. They are not deployment policy, backend support claims, or request overrides.
package:
  # Tokenizer facts interpret request/model token IDs. Numeric special-token facts are package defaults; token spellings and chat templates remain in the tokenizer assets.
  tokenizer:
    # Execution-relevant token IDs have one authority here. Request EOS values are optional overrides; workflow inputs and termination components do not own another package-default copy.
    special_tokens:
      # Ordered set: generation stops on any listed package-default ID.
      eos_token_id:
      - 151643
      # Padding ID used by package-authored tensor contracts.
      pad_token_id: 151643
      # Beginning-of-sequence ID.
      bos_token_id: 151646