# Annotated review companion for justinchuby/onnx-genai-example-gemma4-e2b.
# Model form: hybrid full/sliding-attention text decoder.
# Canonical authority: inference_metadata.yaml at revision 79ca25afe326719e4daab79430c90195dfd28f3b.
# 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.
# 
# Canonical file's existing review notes:
# Faithful onnx-genai v1 inference_metadata for google/gemma-4-E2B-it text decoder.
# Authored to the onnx-genai #1716 schema (example 23 shape) using the REAL exported
# graph port names (past_key_values.N.key / present.N.key) and REAL owner layers:
# full_attention owners (head_dim 512): layers [4, 9, 14]
# sliding_attention owners (head_dim 256): layers [0, 1, 2, 3, 5, 6, 7, 8, 10, 11, 12, 13]
# The 20 shared-KV layers borrow an owner's buffer inside the graph and expose no ports.
# MoE is DISABLED in this checkpoint (enable_moe_block=false) -> dense MLP, not invented.
# final_logit_softcapping / tie_word_embeddings are graph-internal. Validates against
# onnx-genai PR #1716 schema/inference_metadata.schema.json.
# Schema spelling `v1` 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
# 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, typed_emit, serving_service_contract]
    # Workflow inputs are the complete typed boundary. Roles and sources are authored; absent required values must fail before execution.
    inputs:
      # Input `request.active` is required, provided by the application as active, and carries opaque. Unsupported required roles must fail closed.
      request.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.
        contract:
          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}
        # 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: active}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.done` is required, provided by the application as done, and carries opaque. Unsupported required roles must fail closed.
      request.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.
        contract:
          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}
        # 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: done}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.accepted_len` is required, provided by the application as accepted_len, and carries opaque. Unsupported required roles must fail closed.
      request.accepted_len:
        # 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.
        contract:
          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}
        # 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: accepted_len}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # 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, 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, sequence]
          # 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.attention_mask` is required, provided by the application as attention_mask, and carries opaque. Unsupported required roles must fail closed.
      request.attention_mask:
        # Authored tensor ABI: dtype int64, rank 2, shape [batch, 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, sequence]
          # 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: attention_mask}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.0.key` is required, provided by the application as past_key_values.0.key, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.0.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: past_key_values.0.key}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.0.value` is required, provided by the application as past_key_values.0.value, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.0.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: past_key_values.0.value}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.1.key` is required, provided by the application as past_key_values.1.key, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.1.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: past_key_values.1.key}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.1.value` is required, provided by the application as past_key_values.1.value, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.1.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: past_key_values.1.value}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.2.key` is required, provided by the application as past_key_values.2.key, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.2.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: past_key_values.2.key}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.2.value` is required, provided by the application as past_key_values.2.value, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.2.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: past_key_values.2.value}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.3.key` is required, provided by the application as past_key_values.3.key, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.3.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: past_key_values.3.key}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.3.value` is required, provided by the application as past_key_values.3.value, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.3.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: past_key_values.3.value}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.4.key` is required, provided by the application as past_key_values.4.key, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.4.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, full_kv_heads, past_sequence, full_head_dim]. 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, full_kv_heads, past_sequence, full_head_dim]
          # 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: past_key_values.4.key}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.4.value` is required, provided by the application as past_key_values.4.value, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.4.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, full_kv_heads, past_sequence, full_head_dim]. 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, full_kv_heads, past_sequence, full_head_dim]
          # 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: past_key_values.4.value}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.5.key` is required, provided by the application as past_key_values.5.key, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.5.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: past_key_values.5.key}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.5.value` is required, provided by the application as past_key_values.5.value, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.5.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: past_key_values.5.value}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.6.key` is required, provided by the application as past_key_values.6.key, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.6.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: past_key_values.6.key}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.6.value` is required, provided by the application as past_key_values.6.value, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.6.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: past_key_values.6.value}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.7.key` is required, provided by the application as past_key_values.7.key, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.7.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: past_key_values.7.key}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.7.value` is required, provided by the application as past_key_values.7.value, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.7.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: past_key_values.7.value}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.8.key` is required, provided by the application as past_key_values.8.key, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.8.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: past_key_values.8.key}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.8.value` is required, provided by the application as past_key_values.8.value, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.8.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: past_key_values.8.value}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.9.key` is required, provided by the application as past_key_values.9.key, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.9.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, full_kv_heads, past_sequence, full_head_dim]. 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, full_kv_heads, past_sequence, full_head_dim]
          # 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: past_key_values.9.key}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.9.value` is required, provided by the application as past_key_values.9.value, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.9.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, full_kv_heads, past_sequence, full_head_dim]. 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, full_kv_heads, past_sequence, full_head_dim]
          # 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: past_key_values.9.value}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.10.key` is required, provided by the application as past_key_values.10.key, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.10.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: past_key_values.10.key}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.10.value` is required, provided by the application as past_key_values.10.value, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.10.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: past_key_values.10.value}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.11.key` is required, provided by the application as past_key_values.11.key, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.11.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: past_key_values.11.key}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.11.value` is required, provided by the application as past_key_values.11.value, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.11.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: past_key_values.11.value}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.12.key` is required, provided by the application as past_key_values.12.key, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.12.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: past_key_values.12.key}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.12.value` is required, provided by the application as past_key_values.12.value, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.12.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: past_key_values.12.value}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.13.key` is required, provided by the application as past_key_values.13.key, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.13.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: past_key_values.13.key}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.13.value` is required, provided by the application as past_key_values.13.value, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.13.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: past_key_values.13.value}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.14.key` is required, provided by the application as past_key_values.14.key, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.14.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, full_kv_heads, past_sequence, full_head_dim]. 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, full_kv_heads, past_sequence, full_head_dim]
          # 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: past_key_values.14.key}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
      # Input `request.past_key_values.14.value` is required, provided by the application as past_key_values.14.value, and carries opaque. Unsupported required roles must fail closed.
      request.past_key_values.14.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, full_kv_heads, past_sequence, full_head_dim]. 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, full_kv_heads, past_sequence, full_head_dim]
          # 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: past_key_values.14.value}
        # Required values fail closed when absent; optional values may use only their authored default/presence rule.
        required: true
    # Workflow outputs are the complete public result boundary and retain explicit tensor contracts.
    outputs:
      # Output `logits` publishes the authored tensor result at the pre_adapter stage; it is not backend-readiness evidence.
      logits:
        # Authored tensor ABI: dtype float16, rank 3, shape [batch, sequence, vocab]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
        contract:
          dtype: float16
          rank: 3
          shape: [batch, sequence, vocab]
          # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
          batch_layout: {kind: request_aligned, axis: 0}
        role: tensor
        # 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 `decoder` 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.
      decoder:
        # 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:
          # Exact graph inputs; each tensor shape/dtype is authored.
          inputs:
            # Port `input_ids`: Authored tensor ABI: dtype int64, rank 2, shape [batch, sequence]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            input_ids:
              dtype: int64
              rank: 2
              shape: [batch, sequence]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `attention_mask`: Authored tensor ABI: dtype int64, rank 2, shape [batch, sequence]. 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, sequence]
              # 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, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
              # 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, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
              # 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, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
              # 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, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
              # 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, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
              # 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, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
              # 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, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
              # 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, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
              # 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, full_kv_heads, past_sequence, full_head_dim]. 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, full_kv_heads, past_sequence, full_head_dim]
              # 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, full_kv_heads, past_sequence, full_head_dim]. 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, full_kv_heads, past_sequence, full_head_dim]
              # 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, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
              # 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, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
              # 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, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
              # 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, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
              # 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, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
              # 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, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
              # 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, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
              # 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, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
              # 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, full_kv_heads, past_sequence, full_head_dim]. 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, full_kv_heads, past_sequence, full_head_dim]
              # 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, full_kv_heads, past_sequence, full_head_dim]. 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, full_kv_heads, past_sequence, full_head_dim]
              # 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, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
              # 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, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
              # 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, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
              # 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, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
              # 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, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
              # 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, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
              # 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, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
              # 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, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
              # 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, full_kv_heads, past_sequence, full_head_dim]. 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, full_kv_heads, past_sequence, full_head_dim]
              # 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, full_kv_heads, past_sequence, full_head_dim]. 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, full_kv_heads, past_sequence, full_head_dim]
              # 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 `logits`: Authored tensor ABI: dtype float16, rank 3, shape [batch, sequence, vocab]. 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, vocab]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.0.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, total_sequence, sliding_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.0.key:
              dtype: float16
              rank: 4
              shape: [batch, sliding_kv_heads, total_sequence, sliding_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.0.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, total_sequence, sliding_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.0.value:
              dtype: float16
              rank: 4
              shape: [batch, sliding_kv_heads, total_sequence, sliding_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.1.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, total_sequence, sliding_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.1.key:
              dtype: float16
              rank: 4
              shape: [batch, sliding_kv_heads, total_sequence, sliding_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.1.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, total_sequence, sliding_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.1.value:
              dtype: float16
              rank: 4
              shape: [batch, sliding_kv_heads, total_sequence, sliding_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.2.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, total_sequence, sliding_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.2.key:
              dtype: float16
              rank: 4
              shape: [batch, sliding_kv_heads, total_sequence, sliding_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.2.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, total_sequence, sliding_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.2.value:
              dtype: float16
              rank: 4
              shape: [batch, sliding_kv_heads, total_sequence, sliding_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.3.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, total_sequence, sliding_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.3.key:
              dtype: float16
              rank: 4
              shape: [batch, sliding_kv_heads, total_sequence, sliding_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.3.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, total_sequence, sliding_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.3.value:
              dtype: float16
              rank: 4
              shape: [batch, sliding_kv_heads, total_sequence, sliding_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.4.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch, full_kv_heads, total_sequence, full_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.4.key:
              dtype: float16
              rank: 4
              shape: [batch, full_kv_heads, total_sequence, full_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.4.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch, full_kv_heads, total_sequence, full_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.4.value:
              dtype: float16
              rank: 4
              shape: [batch, full_kv_heads, total_sequence, full_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.5.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, total_sequence, sliding_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.5.key:
              dtype: float16
              rank: 4
              shape: [batch, sliding_kv_heads, total_sequence, sliding_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.5.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, total_sequence, sliding_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.5.value:
              dtype: float16
              rank: 4
              shape: [batch, sliding_kv_heads, total_sequence, sliding_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.6.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, total_sequence, sliding_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.6.key:
              dtype: float16
              rank: 4
              shape: [batch, sliding_kv_heads, total_sequence, sliding_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.6.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, total_sequence, sliding_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.6.value:
              dtype: float16
              rank: 4
              shape: [batch, sliding_kv_heads, total_sequence, sliding_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.7.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, total_sequence, sliding_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.7.key:
              dtype: float16
              rank: 4
              shape: [batch, sliding_kv_heads, total_sequence, sliding_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.7.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, total_sequence, sliding_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.7.value:
              dtype: float16
              rank: 4
              shape: [batch, sliding_kv_heads, total_sequence, sliding_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.8.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, total_sequence, sliding_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.8.key:
              dtype: float16
              rank: 4
              shape: [batch, sliding_kv_heads, total_sequence, sliding_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.8.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, total_sequence, sliding_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.8.value:
              dtype: float16
              rank: 4
              shape: [batch, sliding_kv_heads, total_sequence, sliding_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.9.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch, full_kv_heads, total_sequence, full_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.9.key:
              dtype: float16
              rank: 4
              shape: [batch, full_kv_heads, total_sequence, full_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.9.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch, full_kv_heads, total_sequence, full_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.9.value:
              dtype: float16
              rank: 4
              shape: [batch, full_kv_heads, total_sequence, full_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.10.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, total_sequence, sliding_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.10.key:
              dtype: float16
              rank: 4
              shape: [batch, sliding_kv_heads, total_sequence, sliding_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.10.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, total_sequence, sliding_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.10.value:
              dtype: float16
              rank: 4
              shape: [batch, sliding_kv_heads, total_sequence, sliding_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.11.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, total_sequence, sliding_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.11.key:
              dtype: float16
              rank: 4
              shape: [batch, sliding_kv_heads, total_sequence, sliding_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.11.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, total_sequence, sliding_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.11.value:
              dtype: float16
              rank: 4
              shape: [batch, sliding_kv_heads, total_sequence, sliding_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.12.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, total_sequence, sliding_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.12.key:
              dtype: float16
              rank: 4
              shape: [batch, sliding_kv_heads, total_sequence, sliding_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.12.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, total_sequence, sliding_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.12.value:
              dtype: float16
              rank: 4
              shape: [batch, sliding_kv_heads, total_sequence, sliding_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.13.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, total_sequence, sliding_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.13.key:
              dtype: float16
              rank: 4
              shape: [batch, sliding_kv_heads, total_sequence, sliding_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.13.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, total_sequence, sliding_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.13.value:
              dtype: float16
              rank: 4
              shape: [batch, sliding_kv_heads, total_sequence, sliding_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.14.key`: Authored tensor ABI: dtype float16, rank 4, shape [batch, full_kv_heads, total_sequence, full_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.14.key:
              dtype: float16
              rank: 4
              shape: [batch, full_kv_heads, total_sequence, full_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Port `present.14.value`: Authored tensor ABI: dtype float16, rank 4, shape [batch, full_kv_heads, total_sequence, full_head_dim]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            present.14.value:
              dtype: float16
              rank: 4
              shape: [batch, full_kv_heads, total_sequence, full_head_dim]
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
            # Post-final-norm hidden state (the lm_head input == HF hidden_states[-1]).
            # Declared for truthful graph contract; consumed only by a borrowed-KV
            # speculative drafter as its folded_carry_seed (see the assistant/speculative
            # packages). Standalone generation does not route it, so the port is
            # optional: this workflow leaves it unbound while the speculative
            # workflow binds it.
            # Port `hidden_states.34`: Authored tensor ABI: dtype float16, rank 3, shape [batch, sequence, hidden]. Request rows stay aligned on axis 0; this preserves identity but does not by itself authorize grouping independent requests.
            hidden_states.34:
              dtype: float16
              rank: 3
              shape: [batch, sequence, hidden]
              optional: true
              # Tensor row layout is structural. Only component batch_capacity can authorize grouping independent requests.
              batch_layout: {kind: request_aligned, axis: 0}
          # Semantic port roles identify tokens, logits, positions, masks, caches, or media without architecture-name conditionals.
          roles: {input_ids: token_ids, logits: logits}
    # State cells make recurrence, cache mutation, lifetime, and ownership explicit. Physical paging/tiering/placement are not encoded by tensor names.
    state:
      # State `past_key_values.0.key` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.0.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: request.past_key_values.0.key
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: sliding_attention
      # State `past_key_values.0.value` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.0.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: request.past_key_values.0.value
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: sliding_attention
      # State `past_key_values.1.key` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.1.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: request.past_key_values.1.key
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: sliding_attention
      # State `past_key_values.1.value` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.1.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: request.past_key_values.1.value
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: sliding_attention
      # State `past_key_values.2.key` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.2.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: request.past_key_values.2.key
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: sliding_attention
      # State `past_key_values.2.value` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.2.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: request.past_key_values.2.value
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: sliding_attention
      # State `past_key_values.3.key` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.3.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: request.past_key_values.3.key
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: sliding_attention
      # State `past_key_values.3.value` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.3.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: request.past_key_values.3.value
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: sliding_attention
      # State `past_key_values.4.key` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.4.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, full_kv_heads, past_sequence, full_head_dim]. 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, full_kv_heads, past_sequence, full_head_dim]
          # 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: request.past_key_values.4.key
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: full_attention
      # State `past_key_values.4.value` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.4.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, full_kv_heads, past_sequence, full_head_dim]. 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, full_kv_heads, past_sequence, full_head_dim]
          # 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: request.past_key_values.4.value
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: full_attention
      # State `past_key_values.5.key` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.5.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: request.past_key_values.5.key
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: sliding_attention
      # State `past_key_values.5.value` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.5.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: request.past_key_values.5.value
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: sliding_attention
      # State `past_key_values.6.key` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.6.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: request.past_key_values.6.key
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: sliding_attention
      # State `past_key_values.6.value` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.6.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: request.past_key_values.6.value
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: sliding_attention
      # State `past_key_values.7.key` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.7.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: request.past_key_values.7.key
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: sliding_attention
      # State `past_key_values.7.value` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.7.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: request.past_key_values.7.value
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: sliding_attention
      # State `past_key_values.8.key` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.8.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: request.past_key_values.8.key
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: sliding_attention
      # State `past_key_values.8.value` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.8.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: request.past_key_values.8.value
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: sliding_attention
      # State `past_key_values.9.key` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.9.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, full_kv_heads, past_sequence, full_head_dim]. 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, full_kv_heads, past_sequence, full_head_dim]
          # 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: request.past_key_values.9.key
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: full_attention
      # State `past_key_values.9.value` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.9.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, full_kv_heads, past_sequence, full_head_dim]. 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, full_kv_heads, past_sequence, full_head_dim]
          # 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: request.past_key_values.9.value
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: full_attention
      # State `past_key_values.10.key` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.10.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: request.past_key_values.10.key
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: sliding_attention
      # State `past_key_values.10.value` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.10.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: request.past_key_values.10.value
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: sliding_attention
      # State `past_key_values.11.key` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.11.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: request.past_key_values.11.key
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: sliding_attention
      # State `past_key_values.11.value` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.11.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: request.past_key_values.11.value
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: sliding_attention
      # State `past_key_values.12.key` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.12.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: request.past_key_values.12.key
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: sliding_attention
      # State `past_key_values.12.value` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.12.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: request.past_key_values.12.value
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: sliding_attention
      # State `past_key_values.13.key` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.13.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: request.past_key_values.13.key
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: sliding_attention
      # State `past_key_values.13.value` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.13.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, sliding_kv_heads, past_sequence, sliding_head_dim]. 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, sliding_kv_heads, past_sequence, sliding_head_dim]
          # 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: request.past_key_values.13.value
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: sliding_attention
      # State `past_key_values.14.key` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.14.key:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, full_kv_heads, past_sequence, full_head_dim]. 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, full_kv_heads, past_sequence, full_head_dim]
          # 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: request.past_key_values.14.key
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: full_attention
      # State `past_key_values.14.value` is semantic key/value or recurrent cache state. Scope=invocation, recurrence=invariant. Physical paging, tiering, prefix storage, and mirror dtype remain runtime policy.
      past_key_values.14.value:
        # Authored tensor ABI: dtype float16, rank 4, shape [batch, full_kv_heads, past_sequence, full_head_dim]. 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, full_kv_heads, past_sequence, full_head_dim]
          # 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: request.past_key_values.14.value
        # Recurrence defines the next-value rule: {kind=invariant}.
        recurrence: {kind: invariant}
        # 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: full_attention
    # The ordered SSA program defines actual execution and dataflow; capability strings are not a substitute for these steps.
    steps:
      # Invoke component `decoder` with explicit SSA input/output bindings; this is executable tensor computation, not a host-side capability flag.
    - kind: invoke
      component: decoder
      # SSA bindings supply exact values to this step; no implicit tensor discovery occurs.
      inputs: {input_ids: request.input_ids, attention_mask: request.attention_mask, past_key_values.0.key: request.past_key_values.0.key, past_key_values.0.value: request.past_key_values.0.value, past_key_values.1.key: request.past_key_values.1.key, past_key_values.1.value: request.past_key_values.1.value, past_key_values.2.key: request.past_key_values.2.key, past_key_values.2.value: request.past_key_values.2.value, past_key_values.3.key: request.past_key_values.3.key, past_key_values.3.value: request.past_key_values.3.value, past_key_values.4.key: request.past_key_values.4.key, past_key_values.4.value: request.past_key_values.4.value, past_key_values.5.key: request.past_key_values.5.key, past_key_values.5.value: request.past_key_values.5.value, past_key_values.6.key: request.past_key_values.6.key, past_key_values.6.value: request.past_key_values.6.value, past_key_values.7.key: request.past_key_values.7.key, past_key_values.7.value: request.past_key_values.7.value, past_key_values.8.key: request.past_key_values.8.key,
        past_key_values.8.value: request.past_key_values.8.value, past_key_values.9.key: request.past_key_values.9.key, past_key_values.9.value: request.past_key_values.9.value, past_key_values.10.key: request.past_key_values.10.key, past_key_values.10.value: request.past_key_values.10.value, past_key_values.11.key: request.past_key_values.11.key, past_key_values.11.value: request.past_key_values.11.value, past_key_values.12.key: request.past_key_values.12.key, past_key_values.12.value: request.past_key_values.12.value, past_key_values.13.key: request.past_key_values.13.key, past_key_values.13.value: request.past_key_values.13.value, past_key_values.14.key: request.past_key_values.14.key, past_key_values.14.value: request.past_key_values.14.value}
      # SSA output names become the only values subsequent steps may consume.
      outputs: {logits: decoder.logits, present.0.key: decoder.past_key_values.0.key, present.0.value: decoder.past_key_values.0.value, present.1.key: decoder.past_key_values.1.key, present.1.value: decoder.past_key_values.1.value, present.2.key: decoder.past_key_values.2.key, present.2.value: decoder.past_key_values.2.value, present.3.key: decoder.past_key_values.3.key, present.3.value: decoder.past_key_values.3.value, present.4.key: decoder.past_key_values.4.key, present.4.value: decoder.past_key_values.4.value, present.5.key: decoder.past_key_values.5.key, present.5.value: decoder.past_key_values.5.value, present.6.key: decoder.past_key_values.6.key, present.6.value: decoder.past_key_values.6.value, present.7.key: decoder.past_key_values.7.key, present.7.value: decoder.past_key_values.7.value, present.8.key: decoder.past_key_values.8.key, present.8.value: decoder.past_key_values.8.value, present.9.key: decoder.past_key_values.9.key, present.9.value: decoder.past_key_values.9.value, present.10.key: decoder.past_key_values.10.key,
        present.10.value: decoder.past_key_values.10.value, present.11.key: decoder.past_key_values.11.key, present.11.value: decoder.past_key_values.11.value, present.12.key: decoder.past_key_values.12.key, present.12.value: decoder.past_key_values.12.value, present.13.key: decoder.past_key_values.13.key, present.13.value: decoder.past_key_values.13.value, present.14.key: decoder.past_key_values.14.key, present.14.value: decoder.past_key_values.14.value}
      # Publish `decoder.logits` as output `logits` using replace semantics.
    - {kind: emit, value: decoder.logits, output: logits, mode: replace}
    # 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: request.active
      # Done-row binding is the semantic termination signal.
      done: request.done
      # Accepted-length binding controls cache/state advancement after verification.
      accepted_len: request.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 `full_attention` is semantic `full_attention` state in `bnsh` layout with update {kind=append}; rollback/fork/snapshot are correctness capabilities.
          full_attention:
            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
            # Update kind (append/replace/indexed) is model semantics and must match graph state transitions.
            update: {kind: append}
            reuse: {prefix_reusable: true, evictable_prefix: false}
            # Port bindings connect service state to exact component input/output names.
            ports:
              decoder:
                past_key_values.4.key: {input: past_key_values.4.key, output: present.4.key, role: key, layer: 4}
                past_key_values.4.value: {input: past_key_values.4.value, output: present.4.value, role: value, layer: 4}
                past_key_values.9.key: {input: past_key_values.9.key, output: present.9.key, role: key, layer: 9}
                past_key_values.9.value: {input: past_key_values.9.value, output: present.9.value, role: value, layer: 9}
                past_key_values.14.key: {input: past_key_values.14.key, output: present.14.key, role: key, layer: 14}
                past_key_values.14.value: {input: past_key_values.14.value, output: present.14.value, role: value, layer: 14}
          # Group `sliding_attention` is semantic `sliding_attention` state in `bnsh` layout with update {kind=append}; rollback/fork/snapshot are correctness capabilities.
          sliding_attention:
            kind: sliding_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
            # Update kind (append/replace/indexed) is model semantics and must match graph state transitions.
            update: {kind: append}
            reuse: {prefix_reusable: true, evictable_prefix: true}
            # Port bindings connect service state to exact component input/output names.
            ports:
              decoder:
                past_key_values.0.key: {input: past_key_values.0.key, output: present.0.key, role: key, layer: 0}
                past_key_values.0.value: {input: past_key_values.0.value, output: present.0.value, role: value, layer: 0}
                past_key_values.1.key: {input: past_key_values.1.key, output: present.1.key, role: key, layer: 1}
                past_key_values.1.value: {input: past_key_values.1.value, output: present.1.value, role: value, layer: 1}
                past_key_values.2.key: {input: past_key_values.2.key, output: present.2.key, role: key, layer: 2}
                past_key_values.2.value: {input: past_key_values.2.value, output: present.2.value, role: value, layer: 2}
                past_key_values.3.key: {input: past_key_values.3.key, output: present.3.key, role: key, layer: 3}
                past_key_values.3.value: {input: past_key_values.3.value, output: present.3.value, role: value, layer: 3}
                past_key_values.5.key: {input: past_key_values.5.key, output: present.5.key, role: key, layer: 5}
                past_key_values.5.value: {input: past_key_values.5.value, output: present.5.value, role: value, layer: 5}
                past_key_values.6.key: {input: past_key_values.6.key, output: present.6.key, role: key, layer: 6}
                past_key_values.6.value: {input: past_key_values.6.value, output: present.6.value, role: value, layer: 6}
                past_key_values.7.key: {input: past_key_values.7.key, output: present.7.key, role: key, layer: 7}
                past_key_values.7.value: {input: past_key_values.7.value, output: present.7.value, role: value, layer: 7}
                past_key_values.8.key: {input: past_key_values.8.key, output: present.8.key, role: key, layer: 8}
                past_key_values.8.value: {input: past_key_values.8.value, output: present.8.value, role: value, layer: 8}
                past_key_values.10.key: {input: past_key_values.10.key, output: present.10.key, role: key, layer: 10}
                past_key_values.10.value: {input: past_key_values.10.value, output: present.10.value, role: value, layer: 10}
                past_key_values.11.key: {input: past_key_values.11.key, output: present.11.key, role: key, layer: 11}
                past_key_values.11.value: {input: past_key_values.11.value, output: present.11.value, role: value, layer: 11}
                past_key_values.12.key: {input: past_key_values.12.key, output: present.12.key, role: key, layer: 12}
                past_key_values.12.value: {input: past_key_values.12.value, output: present.12.value, role: value, layer: 12}
                past_key_values.13.key: {input: past_key_values.13.key, output: present.13.key, role: key, layer: 13}
                past_key_values.13.value: {input: past_key_values.13.value, output: present.13.value, role: value, layer: 13}