[inureyes]

Summary

• Add the distinct optional prefill_embeddings_deepstack.main entry with compact [Nv], [K, Nv, hidden], and [K] side tensors and post-language-layer sparse residual injection.
• Validate config pairing, descriptors, byte counts, expanded-sequence positions, target layers, finite features, and static bounds in Rust and again in the C ABI.
• Wire optional VMFB compilation/loading, artifact identity, single-session and ragged batch execution, and request-owned cancellation/error cleanup without changing decode state.
• Preserve token prefill, ordinary embeddings prefill, and decode IR byte-for-byte when DeepStack is declared.
• Add a diagnostic-only graph that reuses the production shared layer loop and exposes every post-hook hidden state for the fixed MLX Qwen3-VL oracle; it is never loaded into serving sessions.

Validation

• cargo fmt --all -- --check
• cargo test -p mlxcel-xla --lib (193 passed, 6 opt-in ignored, 0 failed)
• cargo test -p mlxcel-xla --lib deepstack (9 passed)
• cargo clippy -p mlxcel-xla --lib --tests -- -D warnings
• IREE_CUDA_HOME=... IREE_CUDA_COMPILE=... cargo clippy -p mlxcel-xla --lib --features iree -- -D warnings
• CARGO_TARGET_DIR=... python3 spike/openxla/deepstack_prefill_check.py compiled and ran both production and diagnostic Qwen3-shaped VMFBs on IREE local-task. The fixture ports Qwen3VLModel::deepstack_process semantics and exactly checks the full first/middle/last [3, 256, 4] post-hook residual tensor against visual-row values 11/21/31, final logits, and every K/V element for both graphs.

Closes #867

[inureyes]

Addressed the acceptance-coverage gap in 0adb6ab:

• Added a diagnostic-only DeepStack prefill graph that reuses the production shared transformer loop and captures every first/middle/last post-hook residual state.
• Ported the fixed MLX Qwen3VLModel::deepstack_process fixture: visual positions 1..=3, unit hidden states, feature value 10, and exact post-hook visual rows 11/21/31.
• Ran both production and diagnostic VMFBs under IREE local-task. The oracle now compares the complete [3, 256, 4] post-hook tensor, final logits, and every K/V element; no output is ignored.
• Revalidated 193 passing library tests, 9 focused DeepStack tests, fmt, and default/IREE Clippy with warnings denied.

No CSV files were added or changed.

[inureyes]

Addressed the selected-capacity validation ordering gap in 12bf0b6:

• JSON parsing now validates only intrinsic DeepStack schema properties (positive maximum, sorted/unique/in-range target layers).
• with_context_capacity and emission validate the visual maximum against the caller-selected static capacity.
• Regression coverage proves context capacity 512 with visual maximum 300 is accepted, while selected capacity 256 with maximum 300 is rejected.

Validation:

• cargo test -p mlxcel-xla --lib deepstack -- --nocapture: 9 passed
• cargo test -p mlxcel-xla --lib: 193 passed, 6 ignored
• default Clippy with -D warnings: passed
• diagnostics/IREE-feature Clippy with -D warnings: passed
• real IREE local-task production and diagnostic entries: both matched all expected outputs
• git diff --check: clean
• CSV changes: none

[inureyes]

Rebased PR #893 onto origin/main@e739ca1 (merged PR #894) and preserved both contracts:

• M-RoPE keeps its explicit 1D/3D position mode, signed per-request rope_delta, fail-closed config validation, and ragged slot lifecycle.
• DeepStack keeps its distinct optional VMFB, compact owned payload, post-layer hook, Rust/C validation, single/ragged execution, diagnostic graph, and prefill-only ownership.
• Rebase integration commit 09f1aba forwards the embedded M-RoPE delta for DeepStackPrepared batch requests and exercises that combination through cancellation and slot reuse.
• Both internal runtime config variants are boxed after the combined schemas crossed Clippy's large-enum threshold; this is load-time indirection only and does not change emitted or native contracts.

Validation:

• focused DeepStack suite: 9 passed
• focused M-RoPE suite: 8 passed
• mixed text/M-RoPE/DeepStack ragged slot lifecycle: 1 passed
• full mlxcel-xla library suite: 203 passed, 6 ignored
• default Clippy with -D warnings: passed
• IREE-feature Clippy with the pinned CUDA IREE source/build tree and -D warnings: passed
• real IREE local-task DeepStack production and diagnostic VMFBs: both matched all expected outputs, including the full Qwen3-VL post-hook state tensor, final logits, and every K/V element
• cargo fmt --all -- --check and git diff --check: passed
• CSV changes: none

Issue #867's six acceptance criteria were checked against this evidence. The branch was updated with --force-with-lease; no merge was performed.

[inureyes]

Rebased the DeepStack branch onto origin/main at 09641945 after #895 merged, then force-pushed the rewritten commits with --force-with-lease.

Integration review:

• The rebase completed without textual conflicts. The only files touched by both lines of work were emitter/config.rs and iree.rs.
• feat(xla): admit image requests in CLI and continuous-batch serving #895's LLaVA wrapper handling remains intact: root config.json selects nested text_config, including the wrapper-level quantization fallback.
• feat(xla): admit image requests in CLI and continuous-batch serving #895's canonical / language_model.* dense checkpoint namespace resolution remains intact alongside the optional DeepStack artifact/runtime path.
• This PR's final diff does not touch the root MediaRequestMetadata, image preprocessing, XLA serve worker/admission, backend, loading, or CLI files.
• M-RoPE rope_delta remains preserved for DeepStack-prepared batch inputs.

Validation after the rebase:

• cargo fmt --all -- --check — passed
• DeepStack-focused mlxcel-xla tests — 9 passed
• M-RoPE-focused mlxcel-xla tests — 8 passed
• mixed text/M-RoPE slot cancellation and reuse test — 1 passed
• full cargo test -p mlxcel-xla --lib — 204 passed, 6 ignored
• default mlxcel-xla Clippy with -D warnings — passed
• IREE-feature mlxcel-xla Clippy with -D warnings — passed
• real-IREE DeepStack production/diagnostic oracle — passed; both EXEC @main runs matched expected outputs, including fixed Qwen3-VL post-hook states, logits, and every K/V value
• root xla-iree worker/admission tests — 6 passed
• root xla-iree image preprocessing tests — 7 passed
• root XLA backend image-capability test — 1 passed
• root XLA CLI partial-image and output-audio rejection tests — 2 passed
• git diff --check — passed
• tracked/untracked CSV check — no CSV files in the PR diff or worktree

The branch head is now 14a039fbb1e74e96353fb0d3fc111f82ef782fc6.

[inureyes]

Implemented the final HIGH-severity DeepStack/M-RoPE corrections in da7706f.

Key changes:

• DeepStack single and batch preparation now uses the loaded runtime position mode, preserving public PreparedPositions::Mrope3D buffers and signed rope deltas.
• The DeepStack C ABI now carries an explicit position mode, rejects mode disagreement, and validates 1D [S] versus M-RoPE [3,S] descriptors including exact byte lengths.
• Sequential state commits the signed rope delta only after native prefill succeeds.
• Internal scheduler mutation coverage was replaced with public single/batch fixtures covering negative and positive deltas, first-decode parity, failure-state retention, cancellation, slot reuse, pending admission, and wrong-mode rejection.
• The fixed Qwen3-VL IREE oracle now executes both 1D and M-RoPE production and diagnostic entries and checks post-hook states, logits, and every K/V value.

Validation:

• DeepStack focused tests: 11 passed
• M-RoPE focused tests: 8 passed
• Full mlxcel-xla tests: 205 passed, 6 ignored
• Root real-IREE prepared-prefill integration: 2 passed
• Fixed real-IREE oracle: 4 exact runs passed (1D production/diagnostic and M-RoPE production/diagnostic)
• Root feat(xla): admit image requests in CLI and continuous-batch serving #895 admission regressions: worker 6 passed; preprocess 7 passed; capability 1 passed; CLI partial-image 1 passed; CLI audio 1 passed
• Root xla-iree C compilation/no-run: passed
• Default cargo clippy --fix --allow-dirty --allow-staged -p mlxcel-xla --lib --tests -- -D warnings: passed with no rewrite
• IREE library/tests Clippy and root integration-test Clippy: passed
• cargo fmt --all -- --check, git diff --check, and Python syntax validation: passed
• CSV audit: no changed or untracked CSV files

The PR remains open and unmerged for CI/review.

[inureyes]

Closed the final mutation-sensitivity evidence gap in 2ec2a10.

The production public API remains unchanged. Sequential prefill completion and decode-coordinate calculation now share a private transactional DecodePositionState: only an Ok native prefill result commits the next signed delta. The focused public-DTO regression proves a DeepStack request with sequence_len=4, delta=-2 commits first decode coordinate 2, then a failed replacement carrying delta=+5 leaves coordinate 2 unchanged.

Batch admission now shares one private conversion helper between the production submit_deepstack_prepared call site and the scheduler regression. Public DeepStackPreparedPrefill requests carrying -1, +2, and 0 are asserted while pending, after slot admission, and as exact first decode coordinates 3, 6, and 4. Cancelling the positive-delta row and admitting text into the same slot proves reuse restores coordinate 4 with delta zero. No test mutates PreparedIreePositions or scheduler deltas to manufacture the state under test.

Validation:

• DeepStack focused: 13 passed
• M-RoPE focused: 8 passed
• Full mlxcel-xla: 207 passed, 6 ignored
• Root real-IREE public prepared-prefill integration: 2 passed
• Default Clippy fix mode: passed with no rewrite
• IREE library/tests Clippy: passed
• Root real-IREE integration-test Clippy: passed
• cargo fmt --all -- --check and git diff --check: passed
• Changed/untracked CSV audit: empty

PR #893 remains open and unmerged.