Simplify composite ONNX GenAI metadata

[justinchuby]

Summary

Align the metadata producer with a normalized composite schema:

• keep each graph contract only at pipeline.models.<component>.io;
• retain top-level model.io only for bare decoder-only packages;
• make pipeline.phases the sole source of component lifecycle scheduling and presence conditions;
• remove duplicated run_on fields from strategy stages;
• explicitly classify diffusion denoisers as every_step.

This gives the two execution sections distinct responsibilities: strategy describes control structure/order/loop semantics, while phases describes when each named model runs.

Validation

• 113 passed, 2 skipped in src/mobius/integrations/onnx_genai/
• Ruff check and format check passed
• git diff --check passed

This is intentionally coordinated with the ONNX GenAI schema/runtime PR and does not preserve the duplicated composite representation.

Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com

[github-actions]

🏗️ Architecture Diff

Comparing 01008cf → 62c69f5

Model	Sub-model	Changes	Status
bert (feature-extraction)	model	0	⚪
falcon	model	0	⚪
gemma2	model	0	⚪
gemma4 (gemma4)	decoder	0	⚪
gemma4 (gemma4)	embedding	0	⚪
gemma4 (gemma4)	vision_encoder	0	⚪
gemma4_text	model	0	⚪
gpt2	model	0	⚪
llama	model	0	⚪
llama (static-cache)	model	0	⚪
mamba (ssm-text-generation)	model	0	⚪
phi3	model	0	⚪
phi3 (static-cache)	model	0	⚪
qwen	model	0	⚪
qwen (static-cache)	model	0	⚪
qwen2	model	0	⚪
qwen2 (static-cache)	model	0	⚪
qwen2_moe	model	0	⚪
qwen2_moe (static-cache)	model	0	⚪
qwen3	model	0	⚪
qwen3 (static-cache)	model	0	⚪
qwen3_5_moe (hybrid-text-generation)	model	0	⚪
qwen3_5_text (hybrid-text-generation)	model	0	⚪
qwen3_5_vl (hybrid-qwen-vl)	decoder	0	⚪
qwen3_5_vl (hybrid-qwen-vl)	embedding	0	⚪
qwen3_5_vl (hybrid-qwen-vl)	vision_encoder	0	⚪
qwen3_moe	model	0	⚪
qwen3_moe (static-cache)	model	0	⚪
qwen3_next (hybrid-text-generation)	model	0	⚪
t5 (seq2seq)	decoder	0	⚪
t5 (seq2seq)	encoder	0	⚪
whisper (speech-to-text)	decoder	0	⚪
whisper (speech-to-text)	encoder	0	⚪

No architecture changes detected. ✅

---

Legend: ⚪ No change · 🔵 Minor (attrs/inits) · 🟡 Moderate (nodes added/removed) · 🔴 Major (interface changed)

[github-actions]

Performance Comparison

Comparing 01008cf → 62c69f5

Model	Metric	Baseline	Current	Delta
bert (feature-extraction)	model_size_bytes	359 KB	359 KB	+0.0%	⚪
bert (feature-extraction)	num_nodes	60	60	+0.0%	⚪
falcon	model_size_bytes	364 KB	364 KB	+0.0%	⚪
falcon	num_nodes	66	66	+0.0%	⚪
gemma2	model_size_bytes	428 KB	428 KB	+0.0%	⚪
gemma2	num_nodes	105	105	+0.0%	⚪
gpt2	model_size_bytes	388 KB	388 KB	+0.0%	⚪
gpt2	num_nodes	54	54	+0.0%	⚪
llama	model_size_bytes	425 KB	425 KB	+0.0%	⚪
llama	num_nodes	60	60	+0.0%	⚪
llama (static-cache)	model_size_bytes	425 KB	425 KB	+0.0%	⚪
llama (static-cache)	num_nodes	56	56	+0.0%	⚪
mamba (ssm-text-generation)	model_size_bytes	296 KB	296 KB	+0.0%	⚪
mamba (ssm-text-generation)	num_nodes	94	94	+0.0%	⚪
phi3	model_size_bytes	421 KB	421 KB	+0.0%	⚪
phi3	num_nodes	58	58	+0.0%	⚪
phi3 (static-cache)	model_size_bytes	421 KB	421 KB	+0.0%	⚪
phi3 (static-cache)	num_nodes	54	54	+0.0%	⚪
qwen2	model_size_bytes	425 KB	425 KB	+0.0%	⚪
qwen2	num_nodes	60	60	+0.0%	⚪
qwen2 (static-cache)	model_size_bytes	425 KB	425 KB	+0.0%	⚪
qwen2 (static-cache)	num_nodes	56	56	+0.0%	⚪
qwen3_5_moe (hybrid-text-generation)	model_size_bytes	506 KB	506 KB	+0.0%	⚪
qwen3_5_moe (hybrid-text-generation)	num_nodes	264	264	+0.0%	⚪
qwen3_5_text (hybrid-text-generation)	model_size_bytes	458 KB	458 KB	+0.0%	⚪
qwen3_5_text (hybrid-text-generation)	num_nodes	126	126	+0.0%	⚪
qwen3_5_vl (hybrid-qwen-vl)	model_size_bytes	977 KB	977 KB	+0.0%	⚪
qwen3_5_vl (hybrid-qwen-vl)	num_nodes	428	428	+0.0%	⚪
t5 (seq2seq)	model_size_bytes	836 KB	836 KB	+0.0%	⚪
t5 (seq2seq)	num_nodes	166	166	+0.0%	⚪
whisper (speech-to-text)	model_size_bytes	1008 KB	1008 KB	+0.0%	⚪
whisper (speech-to-text)	num_nodes	128	128	+0.0%	⚪

No performance regressions.

[Copilot]

Pull request overview

This PR updates the ONNX GenAI inference metadata emitter to produce a normalized composite schema, reducing duplicated/ambiguous scheduling and I/O contract representation across model, pipeline.strategy, and pipeline.phases.

Changes:

• Removes duplicated run_on scheduling fields from pipeline.strategy.stages, making pipeline.phases the single scheduling/presence source of truth.
• Removes top-level model.io emission for composite packages (e.g., native VLM), keeping I/O contracts scoped to pipeline.models.<component>.io (while preserving model.io for bare single-model packages via the non-pipeline path).
• Explicitly classifies diffusion denoisers as run_on: every_step in pipeline.phases.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
src/mobius/integrations/onnx_genai/inference_metadata.py	Normalizes composite metadata by deduplicating stage scheduling fields and scoping I/O contracts to per-component locations; sets diffusion denoiser phase scheduling explicitly.
src/mobius/integrations/onnx_genai/inference_metadata_test.py	Updates tests to assert the new normalized schema (phases-based scheduling; decoder I/O read from pipeline.models.decoder.io; no composite model.io).

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[justinchuby]

Cross-repo contract audit against onnx-genai 4c3c4b6 found these concrete mismatches in 487d6ca:

1. Metadata emits fp32 while the workflow schema/validator uses float32 (also float16/bfloat16). Please emit canonical workflow dtype names.
2. Seeded sampler currently uses scalar temperature, seed, counter, and next_counter; the published contract requires batched temperature[B] and counter RNG seed[B], offset[B], next_offset[B].
3. EOS graph has only token_ids/eos_token_ids -> terminated; the current termination role also requires iteration and max_iterations. I will split/relax the runtime contract so EOS predicate and loop max remain independently composable rather than forcing limit math into EOS.
4. Euler graph exposes sample, derivative, scalar sigma, sigma_next; the current role's step/schedule mapping cannot describe both sigma values. I will revise the semantic role to explicit current/next schedule values.
5. Masked update needs semantic mappings for confidence and threshold; the current role's integer step is not equivalent. I will revise this contract.
6. Speculative artifact consumes probabilities/uniform and returns acceptance mask/count, whereas the current verifier role expects score tensors, accepted tokens, and done. I will revise the role to primitive acceptance math; token correction/state update remains separate workflow components.
7. Token state update grows [B,T] -> [B,T+1]; current validation incorrectly requires identical current/next shapes. I will change it to dtype/rank compatibility and rely on state recurrence bounds.
8. add_policy_components_to_workflow currently emits ports only, without each component's required policy semantic mapping/effect declaration. Those must be emitted once the revised fields land.

Please continue making sampler RNG batched/counter-based and canonicalizing dtypes. I will push the contract corrections and runtime E2E fixtures to #828, then post the new SHA here for exact producer alignment.

[justinchuby]

Runtime E2E review of dd1cb838 found three remaining blocking mismatches in build_decoder_workflow_metadata:

1. Decoder logits is rank 3 [B,T,V], but both sampler artifacts declare rank 2 [B,V]. onnx-genai now validates SSA edge contracts and will reject this package. Please ship/invoke an ONNX final-position selector (or make the sampler artifact explicitly accept rank-3 logits and select the last position); the runtime will not add host slicing.
2. Workflow loop semantics are continue while condition is true. The generated graph binds condition: loop.done, which stops after the first ordinary token and continues after EOS. Please invoke an ONNX boolean inverter and bind a continue value, or change the termination artifact to expose continuation explicitly.
3. loop.iteration is a fixed application input and is never loop-carried/updated. The termination graph's max-iteration predicate therefore does not advance. The host loop bound still caps execution, but the policy output is wrong. Please add an ONNX counter state-update component/carry or remove the redundant limit predicate from the termination artifact in a coordinated contract revision.

Also, emitted tokens is declared rank-1 [B] with mode: append; that only represents a batch-1 flat stream. For batched output the boundary must define [B,T] accumulation (or use per-row event emission with the serving contract). I am making the runtime fail rather than silently inspect only batch row 0.

[justinchuby]

ONNX GenAI branch phi/effect-merge blocker is fixed at ad8e5e4f1a766eb95787c97266a6ed554e1b28ec on PR #828.

Emit branches with:

outputs:
  selected.tokens:
    cases: { "true": accepted.tokens, "false": corrected.tokens }
effects:
  rng:
    incoming: rng.0
    cases: { "true": rng.accepted, "false": rng.corrected }
    produces: rng.joined

Include default in each output/effect mapping whenever the branch has a default node. All phi contracts must unify. Case-local successor tokens may differ (or match each other), but produces is a distinct joined token. Only phi outputs and explicit emits escape branch scope.

Exact contract/docs: crates/onnx-genai-metadata/src/schema/ir.rs and docs/WORKFLOW_POLICY_COMPONENTS.md; generated schema is updated. Speculative token/KV/RNG/emit E2E passes.

[justinchuby]

ONNX GenAI preprocessing-to-workflow SSA blocker is fixed at justinchuby/onnx-genai@f72782c64fbe04ac6e12593a69ecff47781f023b (draft PR #828).

Producer target:

• declare adapter component { kind: adapter, abi: onnx-genai.image-preprocess, version: "1" } and pin manifest.adapter_abis["onnx-genai.image-preprocess"]: "1"
• adapter input encoded: uint8 rank 1 [encoded_bytes]
• add exact TensorContract to every preprocessing.image.outputs[]; source is processor-local and required; name is the workflow SSA value
• invoke with explicit port maps, e.g. outputs: { pixel_values: image.pixel_values, grid: image.grid }
• output cannot be optional in workflow metadata
• feed those SSA values to vision/embedding/decoder invokes; do not emit legacy component.input names

Canonical documentation: docs/WORKFLOW_POLICY_COMPONENTS.md (Versioned adapter invocation). Schema: schema/inference_metadata.schema.json. Runtime E2E covers encoded request image -> preprocessing adapter -> pixel/grid SSA -> vision -> embedding -> decoder -> post-adapter emit. Branch phi/effect support remains at prior commit ad8e5e4f1a766eb95787c97266a6ed554e1b28ec.

[justinchuby]

Follow-up audit of current Mobius head 151014eb found remaining generated-workflow blockers:

1. Decoder state declaration still says state.token.initializer: request.<token_input> ([B,T]) while the state contract/current value is [B,1]. ONNX GenAI now validates declared state initializers statically and against the concrete value at loop entry (21f6994d). Set the initializer to the setup-produced [B,1] token value (token.setup) or restructure setup.
2. loop.iteration is still a literal zero and is not updated/carried. The termination max-iteration input therefore remains semantically wrong. Use an ONNX counter/state-update component and carry its [B] output; no host loop-index fallback will be added.
3. Decoder setup samples/emits one token, then the loop runs max_iterations body executions. That can emit max_output_tokens + 1; setup EOS also cannot prevent the first body execution because the declared loop condition is produced in the body. Either make setup initialization-only and perform all output steps in the body, or ship ONNX policy components/dataflow that compute the remaining body budget and gate the first body.
4. The same setup-plus-body accounting should be checked in masked diffusion: setup performs one masked update, then the body may run the full requested iteration count.

TTS does not need a new host induction primitive: per the component-centric contract, initialize an integer state cell and invoke/carry a generic ONNX counter update inside the nested loop; use that SSA value for step_index/position selection.

Please generate one concrete package/YAML at the fixed head so I can run ONNX GenAI load+execution cross-repo.

[justinchuby]

ONNX GenAI loop induction blocker is fixed in justinchuby/onnx-genai@e877270123d38d49012e70ecbf7254e720cd3431 (PR #828).

Producers may now declare:

iteration:
  value: loop.i
  contract: { dtype: int64, rank: 1, shape: [batch] }

and bind loop.i directly to solver/masked-update/termination step, RNG counters, schedule selectors, or TTS step_index. It is zero-based, body/condition-scoped, deterministic, and lexically distinct for nested loops. No ONNX counter component is needed for the control-flow induction variable; reverse/remaining values still require an ordinary ONNX component.

Please migrate decoder/diffusion/masked/TTS loop iteration inputs to this field. The separate decoder setup/body output-count and initializer issues from the prior comment still need producer fixes before cross-repo execution.

[justinchuby]

Cross-repo execution against ONNX GenAI #828 found one remaining producer-side contract mismatch (still present at current head 8f96651c0031ed082229ee404210c50de4a55216):

• Decoder/VLM workflows carry decoder.setup.logits as invariant state.logits, then update it from decoder.body.logits (workflow_metadata.py around lines 1150-1195 and 2041-2105). A normal decoder produces setup logits [B, prompt_sequence, V] and body logits [B, 1, V], so the generic runtime correctly rejects the first recurrence as an invariant shape change.
• Please carry last-token logits instead: invoke last_token_logits after the setup decoder, initialize a rank-2 [B,V] state from that result, invoke it after each body decoder, and carry the rank-2 result; alternatively restructure the loop so full rank-3 logits are not a carried invariant.

Other exact 774448b generated-package results:

• masked diffusion executes end-to-end with the actual Mobius policy artifacts;
• codec executes end-to-end;
• decoder executes when the synthetic decoder is constrained to [B,1,V], confirming request binding, invoke, loop, KV growth, effects, and emit paths.

ONNX GenAI fixes discovered by this run: scalar defaults now materialize unbound symbolic axes as singleton tensors, and component shape symbols are invocation-local (with adapter allocation retaining package-scope symbols). Permanent conformance/regression tests are being pushed to #828.

[justinchuby]

Fixed the decoder/VLM logits recurrence in 87bd5b4: setup and body decoder outputs are normalized through the generated last_token_logits component, and only rank-2 [B,V] values are carried invariantly. I generated a decoder with setup [B,T,V] and body [B,1,V] and ran ONNX GenAI conformance at 2ad2e0d; mobius_decoder_workflow_executes passed for 3 tokens. Diffusion also executed end-to-end for two Euler iterations on the same runtime. Muse metadata was regenerated and validates semantically; local SHA-256 is f1a33d7f162c350e79a37d11d8be5be0690a3c8c6cc0e71ce2b428c368dccabf.

[justinchuby]

Remaining cross-repo blockers after 87bd5b4: (1) real speculative rejection cannot be represented faithfully because WorkflowNode::Emit has no accepted-length/validity operand, while verifier KV state also needs rollback/truncation to the accepted prefix; the current fixed-shape branch fixture is schema-valid but is not a correct rejection execution. (2) Real Qwen3-TTS still needs talker/code-predictor KV recurrence and the trained group-0/predictor embedding transition; the current representative nested-loop fixture validates structure but is not a real-package runtime E2E. Independent decoder, masked diffusion, codec, and Euler diffusion runtime paths pass; VLM/Muse validates against the latest schema.

[justinchuby]

Performance acceptance instrumentation is pushed in e864abe.

The new paired-run gate rejects non-identical model/runtime/EP/device/precision/batch/shape/sampling/RNG/KV/capture/warmup conditions, then checks throughput, TTFT, peak memory, H2D/D2H counts+bytes, device syncs, session/kernel boundaries, device residency, and required island capture/replay. Required plans cover decoder+min-p+termination, speculative accept/reject, and grammar-delimited islands.

Current measured upstream baseline at ONNX GenAI 8bacf8c is not release-passing:

• H200 decoder steady throughput workflow/native: 0.903 (9.7% regression).
• H200 min-p throughput: 0.957 (within 5%), but warm TTFT 4.36/3.64 ms (19.8% regression).
• Cold startup: decoder 467/49 ms, min-p 231/18 ms, rooted in first-run output-extent discovery/stable-binding construction.
• Synthetic decoder/min-p islands captured once/replayed 503 times; speculative verifier/policy also captured/replayed.

Real Mobius package/KV/per-row serving measurements are still absent. Additionally, application_overridable samplers remain excluded by is_fusible_component, so the actual Mobius overrideable sampler path cannot yet demonstrate the required fused island. PR remains draft/not performance-ready.

[justinchuby]

Architecture cleanup progress:

• 66e5281: removed the closed PolicyRole enum. Policy ONNX artifacts now embed versioned namespaced contract IDs, and workflow declarations derive ID/version from artifact metadata.
• ada6303: decoder and masked-diffusion loops now consume lexical loop induction directly; generated iteration state and iteration_increment.onnx are removed. Serialized invokes contain semantic tensor mappings only, with no effect-token fields/read-write chains.
• Full Mobius suite: 3694 passed, 52 skipped; Ruff passed.

Muse concise metadata currently measures 600 lines / 546 mapping fields / 16,339 bytes, down from the reviewed 1,183-line source (583 lines, 49.3% reduction). A pre-lowering flattened-field baseline was not retained, so only the current 546-field count is evidence-backed.

Remaining producer work is blocked on exact current #828 contracts, reported at issuecomment-5274711168: per-row Emit.valid_length[B]/ragged append and logical state lengths, continue_when polarity, generic KV alias/group/sequence-axis semantics, and a shipped semantic-validator CLI. Current 8bacf8c still requires one-element emit length, so removing speculative ReduceMin now would create metadata the shipped validator/executor rejects. Legacy schema fixtures/continue-Not paths will be deleted when those contracts land; PR remains draft.

[justinchuby]

ONNX GenAI producer blockers are resolved at justinchuby/onnx-genai@8215649100a0a27be15b04045fddde777c8248fc on PR #828. Producer migration can resume.

Use schema/inference_metadata.schema.json from that commit and run the shipped package validator:
cargo run -p onnx-genai-metadata --bin validate_metadata -- <generated-package-dir>

Key migration: continue_when is pre-test/zero-trip; emit valid_length: int64[B] without reduction and consume output.row.<row>; declare serving active/done/accepted_len/slot_ids; bind cache state to a KV service_group with sequence_axis, layout, semantic logical_lengths: int64[B], storage, and past/present aliases; request sources no longer duplicate the semantic role.

Full exact checklist and YAML are in docs/WORKFLOW_POLICY_COMPONENTS.md; the detailed coordination comment is on #828.

[justinchuby]

Cross-repo execution against onnx-genai #828 imported 3b0445a immutably. Decoder/masked/codec execute; four packages fail before workflow runtime:

• VLM embedding/model.onnx: inputs_embeds is declared float but produced from int64 (emit_inputs_embeds).
• Diffusion text_encoder/model.onnx: encoder_hidden_states is declared float but produced from int64.
• Speculative proposer/model.onnx: graph outputs proposed_tokens/scores have no producers.
• TTS code_predictor/model.onnx: missing initializer producer for code_predictor.layers.0.input_layernorm.weight.

All seven metadata files pass the callable validator, demonstrating why runtime E2E is now mandatory. ONNX CI/test wiring is ready and will pin the corrected producer commit once these graph defects are regenerated.

[justinchuby]

Confirmed corrected producer head 92bc47ef: all seven checked-in packages now execute under #828 (7/7), including VLM preprocessing, Euler loop, nested TTS, and speculative grammar forced-token rejection/correction. One remaining producer diagnostic: the TTS E2E emits ORT memory-pattern shape-reuse warnings as sequence shapes grow ({1,1,6} != {1,1,2..4}, {1,6} != {1,2..4}, later {1,1,7}). Execution is correct and ORT falls back, but shared dim_param names appear to claim equality across changing sequence dimensions. Please tighten those graph shape symbols in a follow-up to preserve allocator/capture quality.

[justinchuby]

Final validation at 9c0cc54b: ONNX GenAI metadata/schema validation passed; all 7/7 checked workflow packages execute (decoder, VLM, diffusion, masked diffusion, codec, TTS, speculative); deterministic fixture regeneration passed; Ruff passed; Mobius suite passed (3696 passed, 52 skipped). Muse metadata: 764 lines, 720 fields, 20,545 bytes, SHA-256 5f4019a492f2278728743e9e2658be00455c0952076098b3cbfae5d0514c0191. The Integration (fast) rerun remains queued without a runner; its prior failures were isolated DeepSeek numerical-parity tests unrelated to this metadata/serialization diff. No merge or HF upload performed.

[justinchuby]

CI follow-up: the Integration (fast) rerun completed with the identical two pre-existing DeepSeek numerical-parity failures (test_deepseek_v2_lite_prefill_logits_match, test_deepseek_non_mla_decoder_prefill_logits_match). All other checks passed again, including ONNX GenAI metadata and L1/L3/L4/L5. The failures do not exercise package save, workflow metadata, TTS, or regenerated fixtures.

[justinchuby]

Real Muse H200 follow-up is pushed at c332bec (producer dtype fix d7927c3).

Paired native result on ORT 1.28/H200, exact 68-token prompt, 128 new tokens, 1 warmup + 3 runs, CUDA Graph/shared KV: 63.3653 tok/s median, 49.0217 ms TTFT (2.60% above the historical 61.76 tok/s baseline).

The real metadata workflow currently cannot complete in frozen ONNX GenAI a341c463; exact runtime blockers found while executing the 52-layer package:

1. text-only VLM cannot omit request.image (optional tensor defaults unsupported);
2. image adapter rejects derived total_patches before preprocessing;
3. CUDA loop carries/continue masks cannot be cloned/inspected without host materialization;
4. cuDNN fused GQA fails for the real prefill shape, requiring unfused fallback.

Mobius also fixed BF16 decoder logits → float32 sampler ABI and added the exact runnable native/workflow harness. I returned the PR to draft and have not replaced the HF metadata, because the currently frozen runtime cannot execute the published workflow without runtime changes.

[justinchuby]

The generic adapter ABI is now published in ONNX GenAI commit 21a935c241f9fb8bb4b77e4df39b2716b0f70a26 (PR #828). Normative contract: https://github.com/justinchuby/onnx-genai/blob/21a935c241f9fb8bb4b77e4df39b2716b0f70a26/docs/WORKFLOW_POLICY_COMPONENTS.md#parameter-adapters-lora

Please persist/freeze producer artifacts and cross-repo fixtures against this exact commit:

adapters:
  base_model_fingerprint: onnx-genai-targeted-base-v1:sha256:<64 lowercase hex>
  selection:
    row_ids: request.row_ids
    request_epochs: request.request_epochs
    adapter_ids: request.adapter_ids
    adapter_counts: request.adapter_counts
    scales: request.adapter_scales
    active: request.active
    max_adapters: 4
  application_capability: onnx-genai.adapters@1

Canonical files are under adapters/<alias>/, with lowercase SHA-256 of exact bytes. Formats: RFC 8785 float32 JSON (targets.<weight_key>.{a,b}), safetensors (<weight_key>.a/.b), or upstream ORT .onnx_adapter (TORT, v1). Artifact indices are contiguous from zero. The targeted-base fingerprint hashes RFC 8785 canonical JSON over sorted (component,parameter) records containing exact initializer dtype/shape/logical little-endian bytes and sorted direct-consumer identity/attributes.

Selection is immutable SSA: IDs [B,K], counts [B], scales [B,K], row IDs/epochs [B], optional active [B]; unused slots are exactly -1/0. Composition is ordered base + Σ scale*(alpha/rank)*B*(A*x). Compaction carries all selection/state together by (row_id,request_epoch); epoch changes on slot reuse.

Please generate exact executable fixtures for heterogeneous B>1 rows, no-adapter row, composed adapters, compaction/reorder, slot reuse, eviction/reload, capture replay, plus invalid base fingerprint/shape/checksum/selection cases, and report artifact hashes for byte-identity consumption.

[justinchuby]

Supersedes the earlier 21a935c adapter draft. Do not freeze fixtures against that commit.

The canonical ABI integrating native LoRA PR #318 (813a9b53) and #374 (326fddcf) is ONNX GenAI commit 903a2d1aad6e58ecd966db70e8b2fd310fce146a on PR #828. Normative contract and exact reuse/adapt/retire matrix: https://github.com/justinchuby/onnx-genai/blob/903a2d1aad6e58ecd966db70e8b2fd310fce146a/docs/WORKFLOW_POLICY_COMPONENTS.md#parameter-adapters-lora

Key producer changes before fixture freeze:

• emit one top-level InferenceMetadata.adapters; there is no workflow.adapters;
• emit authoritative target_manifest.targets[] with stable id, generic component, exact base parameter, optional output_value/resolved output_slice, activation_dtype, dimensions, and optional Phase-1 graph_inputs.{a,b,scale};
• declare source format/capability: hf_peft pairs checksummed adapter_config.json + safetensors; ort_genai is checksummed upstream TORT v1; Mobius is not required to emit FlatBuffers;
• artifact bindings[] reference manifest target IDs and weight keys, with optional per-target rank/alpha;
• request SSA is slot_ids[B], request_epochs[B], segments[B,K], adapter_counts[B], scales[B,K], optional active[B]; unused slots are exactly -1/0;
• K=1 preserves Phase-2 segment routing; K>1 is ordered additive composition. Compaction carries slot ID, epoch, segments, counts, scales, active flag, and model state together.

Please update persistence and generate exact fixtures against 903a2d1a for PEFT and/or ORT sources, heterogeneous B>1, base-only, ordered composition, compaction/reorder, slot reuse, eviction/reload, capture replay, and invalid fingerprint/shape/checksum/ID cases. Report producer commit plus exact file hashes before ONNX GenAI consumes/finalizes them.

[justinchuby]

Exact producer ABI is now finalized at ONNX GenAI d9482bca0ad6ccec907a2b49faf033c2006a9fa9 (supersedes 903a2d1a for fixture pinning).

adapters:
  base_model_fingerprint: onnx-genai-targeted-base-v1:sha256:<hex>
  target_manifest:
    targets:
      - id: <stable generic target id>
        component: <pipeline component or model>
        parameter: <exact immutable initializer>
        node_name: <exact ONNX projection node>
        output_name: <exact ONNX output value>
        activation_dtype: float16
        input_features: <K>
        output_features: <N>
        output_slice: { role: <producer label>, offset: <u>, width: <u> } # optional
        graph_inputs: { a: <optional A input>, b: <optional B input>, scale: <optional scale input> } # optional Phase-1 seam
  discovery_fallback: disabled # tooling_only only while importer resolves the manifest
  selection:
    slot_ids: <int64[B] SSA>
    request_epochs: <int64[B] SSA>
    segments: <int64[B,K] SSA>
    adapter_counts: <int64[B] SSA>
    scales: <float32[B,K] SSA>
    active: <bool[B] SSA> # optional
    max_adapters: <K>
  application_capability: onnx-genai.adapters@1
  portable_fallback: true
  artifacts:
    <alias>:
      index: <contiguous segment id>
      identity: <stable identity>
      version: <version>
      base_model_fingerprint: <same fingerprint>
      rank: <default rank>
      alpha: <default alpha>
      dtype: <factor dtype>
      provenance: { producer: mobius, source: <optional URI>, revision: <optional immutable revision> }
      weights:
        - format: hf_peft
          loader_capability: onnx-genai.adapters.hf-peft@1
          location: adapters/<alias>/adapter_model.safetensors
          sha256: <exact bytes>
          config_location: adapters/<alias>/adapter_config.json
          config_sha256: <exact bytes>
          scale_encoding: alpha_over_rank
        # ORT alternative: format=ort_genai, loader_capability=onnxruntime.lora-adapter@1,
        # location=...onnx_adapter, sha256=..., scale_encoding=baked
      bindings:
        - target: <manifest target id>
          weight_key: <canonical A/B key>
          rank: <optional override>
          alpha: <optional override>

Important #374 detail now explicit: TORT scale is baked, so ORT sources require scale_encoding: baked and must not receive alpha/rank again. PEFT is primary/provenance-preserving with alpha_over_rank. Exact node/output names and labeled fused slices are retained for graph validation; execution consumes the resolved bindings, never model-family discovery.

Normative docs/matrix: https://github.com/justinchuby/onnx-genai/blob/d9482bca0ad6ccec907a2b49faf033c2006a9fa9/docs/WORKFLOW_POLICY_COMPONENTS.md#parameter-adapters-lora

[justinchuby]

The final ONNX GenAI adapter ABI superseding the earlier d9482bca pin is now published at justinchuby/onnx-genai@793bfe9b5489257a8ae4126ba66b760e3b2cfe19 (PR justinchuby/onnx-genai#828).

Please serialize/freeze producer fixtures against these exact public target names:

• component
• initializer
• optional layer_index
• optional target rank / alpha
• optional output_slice { role, offset, width, rank, alpha }
• optional Phase-1 graph_inputs { a, b, scale }

Effective artifact binding rank/alpha is checked against both target and output-slice policy. component + initializer is also the canonical targeted-fingerprint identity. This matches the authoritative exact-graph manifest direction in producer 7bb3fac; model-family discovery remains producer/import tooling only.

[justinchuby]

Final adapter consumer pin correction: use Mobius 62c69f51cbc9495c1c91a69221424e709880f5ad against ONNX GenAI 2af34dcad0e429604ca66d7ba1388ee2e688756e; do not use d9bf9ffc or 9a3f523.

The final committed YAML is here and uses top-level adapters, target_manifest, SSA selection, artifact bindings, structured provenance, loader capabilities, and scale encoding. SHA-256 is c63c9a3ac59f87383044fe3e47fa7b7b615c823763a0fda8b07e764e4d729e0e (stale d9bf YAML was 8701473391a93707ac86d78a00ba7fd4387109d2576c3cabb09d88ee32e7bec4).

The final conformance test supplies slot_ids, request_epochs, adapter_segments, adapter_counts, adapter_scales, and active via .with_input; SHA-256 15ae3864e7003e0989cf76a060fdcff0472897ceafe633a92e808d9afde3b17d (stale d9bf test was 8297818c0166ce338c6a70a2efa865bf54787370586c84d7dd164220b7096446). AdapterSelection exists only as a test-side lowering helper; it is never passed out-of-band to the runtime.