training, inference, registry

Author: Shaw

ELIZA_1_GGUF_READINESS.md

@@ -47,9 +47,6 @@ Required files:
 - `quantization/polarquant_config.json`
 
 Missing files/evidence:
-- `asr/eliza-1-asr-mmproj.gguf`
-- `asr/eliza-1-asr.gguf`
-- `cache/voice-preset-default.bin`
 - `checksums/SHA256SUMS`
 - `dflash/drafter-0_6b.gguf`
 - `dflash/target-meta.json`
@@ -71,20 +68,14 @@ Missing files/evidence:
 - `evidence/platform/windows-x64-cpu.json`
 - `evidence/platform/windows-x64-vulkan.json`
 - `evidence/release.json`
-- `licenses/LICENSE.asr`
 - `licenses/LICENSE.dflash`
 - `licenses/LICENSE.eliza-1`
 - `licenses/LICENSE.text`
-- `licenses/LICENSE.vad`
-- `licenses/LICENSE.voice`
 - `quantization/fused_turboquant.json`
 - `quantization/polarquant_config.json`
 - `quantization/qjl_config.json`
 - `quantization/turboquant.json`
 - `text/eliza-1-0_6b-32k.gguf`
-- `tts/omnivoice-base-Q4_K_M.gguf`
-- `tts/omnivoice-tokenizer-Q4_K_M.gguf`
-- `vad/silero-vad-int8.onnx`
 
 ## 1_7b
 
@@ -125,45 +116,21 @@ Required files:
 - `quantization/polarquant_config.json`
 
 Missing files/evidence:
-- `asr/eliza-1-asr-mmproj.gguf`
-- `asr/eliza-1-asr.gguf`
-- `cache/voice-preset-default.bin`
-- `checksums/SHA256SUMS`
-- `dflash/drafter-1_7b.gguf`
-- `dflash/target-meta.json`
-- `evals/aggregate.json`
-- `evals/cpu_dispatch.json`
 - `evals/cpu_reference.json`
-- `evals/metal_dispatch.json`
-- `evals/metal_verify.json`
-- `evals/vulkan_dispatch.json`
-- `evals/vulkan_verify.json`
 - `evidence/platform/android-adreno-vulkan.json`
 - `evidence/platform/android-mali-vulkan.json`
-- `evidence/platform/darwin-arm64-metal.json`
 - `evidence/platform/ios-arm64-metal.json`
 - `evidence/platform/linux-x64-cpu.json`
 - `evidence/platform/linux-x64-vulkan.json`
 - `evidence/platform/windows-arm64-cpu.json`
 - `evidence/platform/windows-arm64-vulkan.json`
 - `evidence/platform/windows-x64-cpu.json`
 - `evidence/platform/windows-x64-vulkan.json`
-- `evidence/release.json`
-- `licenses/LICENSE.asr`
-- `licenses/LICENSE.dflash`
-- `licenses/LICENSE.eliza-1`
-- `licenses/LICENSE.text`
-- `licenses/LICENSE.vad`
-- `licenses/LICENSE.voice`
 - `quantization/fused_turboquant.json`
 - `quantization/polarquant_config.json`
 - `quantization/qjl_config.json`
 - `quantization/turboquant.json`
 - `text/eliza-1-1_7b-32k.gguf`
-- `text/eliza-1-1_7b-64k.gguf`
-- `tts/omnivoice-base-Q4_K_M.gguf`
-- `tts/omnivoice-tokenizer-Q4_K_M.gguf`
-- `vad/silero-vad-int8.onnx`
 
 ## 9b
 
@@ -210,9 +177,6 @@ Required files:
 - `quantization/polarquant_config.json`
 
 Missing files/evidence:
-- `asr/eliza-1-asr-mmproj.gguf`
-- `asr/eliza-1-asr.gguf`
-- `cache/voice-preset-default.bin`
 - `checksums/SHA256SUMS`
 - `dflash/drafter-9b.gguf`
 - `dflash/target-meta.json`
@@ -239,22 +203,16 @@ Missing files/evidence:
 - `evidence/platform/windows-x64-cuda.json`
 - `evidence/platform/windows-x64-vulkan.json`
 - `evidence/release.json`
-- `licenses/LICENSE.asr`
 - `licenses/LICENSE.dflash`
 - `licenses/LICENSE.eliza-1`
 - `licenses/LICENSE.text`
-- `licenses/LICENSE.vad`
 - `licenses/LICENSE.vision`
-- `licenses/LICENSE.voice`
 - `quantization/fused_turboquant.json`
 - `quantization/polarquant_config.json`
 - `quantization/qjl_config.json`
 - `quantization/turboquant.json`
 - `text/eliza-1-9b-128k.gguf`
 - `text/eliza-1-9b-64k.gguf`
-- `tts/omnivoice-base-Q8_0.gguf`
-- `tts/omnivoice-tokenizer-Q8_0.gguf`
-- `vad/silero-vad-int8.onnx`
 - `vision/mmproj-9b.gguf`
 
 ## 27b
@@ -302,9 +260,6 @@ Required files:
 - `quantization/polarquant_config.json`
 
 Missing files/evidence:
-- `asr/eliza-1-asr-mmproj.gguf`
-- `asr/eliza-1-asr.gguf`
-- `cache/voice-preset-default.bin`
 - `checksums/SHA256SUMS`
 - `dflash/drafter-27b.gguf`
 - `dflash/target-meta.json`
@@ -327,22 +282,16 @@ Missing files/evidence:
 - `evidence/platform/windows-x64-cuda.json`
 - `evidence/platform/windows-x64-vulkan.json`
 - `evidence/release.json`
-- `licenses/LICENSE.asr`
 - `licenses/LICENSE.dflash`
 - `licenses/LICENSE.eliza-1`
 - `licenses/LICENSE.text`
-- `licenses/LICENSE.vad`
 - `licenses/LICENSE.vision`
-- `licenses/LICENSE.voice`
 - `quantization/fused_turboquant.json`
 - `quantization/polarquant_config.json`
 - `quantization/qjl_config.json`
 - `quantization/turboquant.json`
 - `text/eliza-1-27b-128k.gguf`
 - `text/eliza-1-27b-256k.gguf`
-- `tts/omnivoice-base-Q8_0.gguf`
-- `tts/omnivoice-tokenizer-Q8_0.gguf`
-- `vad/silero-vad-int8.onnx`
 - `vision/mmproj-27b.gguf`
 
 ## 27b-256k
@@ -389,9 +338,6 @@ Required files:
 - `quantization/polarquant_config.json`
 
 Missing files/evidence:
-- `asr/eliza-1-asr-mmproj.gguf`
-- `asr/eliza-1-asr.gguf`
-- `cache/voice-preset-default.bin`
 - `checksums/SHA256SUMS`
 - `dflash/drafter-27b-256k.gguf`
 - `dflash/target-meta.json`
@@ -413,19 +359,13 @@ Missing files/evidence:
 - `evidence/platform/linux-x64-rocm.json`
 - `evidence/platform/linux-x64-vulkan.json`
 - `evidence/release.json`
-- `licenses/LICENSE.asr`
 - `licenses/LICENSE.dflash`
 - `licenses/LICENSE.eliza-1`
 - `licenses/LICENSE.text`
-- `licenses/LICENSE.vad`
 - `licenses/LICENSE.vision`
-- `licenses/LICENSE.voice`
 - `quantization/fused_turboquant.json`
 - `quantization/polarquant_config.json`
 - `quantization/qjl_config.json`
 - `quantization/turboquant.json`
 - `text/eliza-1-27b-256k.gguf`
-- `tts/omnivoice-base-Q8_0.gguf`
-- `tts/omnivoice-tokenizer-Q8_0.gguf`
-- `vad/silero-vad-int8.onnx`
 - `vision/mmproj-27b-256k.gguf`

ELIZA_1_RELEASE_ASSET_STATUS.md

@@ -0,0 +1,100 @@
+# Eliza-1 Release Asset Status
+
+Generated on 2026-05-11 after local iOS smoke + asset acquisition.
+
+This is a release-prep ledger, not a release approval. All upstream source
+weights acquired here are staged under `source/` and are not final Eliza-1
+weights. The publishable files must still land at the paths listed in
+`ELIZA_1_GGUF_READINESS.md`.
+
+## Local Bundle Roots
+
+All local bundles are under:
+
+`/Users/shawwalters/.eliza/local-inference/models/`
+
+Staged bundle directories:
+
+- `eliza-1-0_6b.bundle`
+- `eliza-1-1_7b.bundle`
+- `eliza-1-9b.bundle`
+- `eliza-1-27b.bundle`
+- `eliza-1-27b-256k.bundle`
+
+`stage_eliza1_bundle_assets.py --link-mode hardlink` now hardlinks Hub cache
+blobs into tier bundles, so repeated ASR/TTS assets do not consume disk once
+per tier.
+
+## Acquired Non-Text Assets
+
+Every tier has these runtime side assets staged with SHA-256 evidence in
+`evidence/bundle-assets.json`:
+
+- TTS: `Serveurperso/OmniVoice-GGUF`
+- ASR: `ggml-org/Qwen3-ASR-0.6B-GGUF` for 0.6B/1.7B/9B tiers
+- ASR: `ggml-org/Qwen3-ASR-1.7B-GGUF` for 27B tiers
+- VAD: `onnx-community/silero-vad`
+- Cache: deterministic `cache/voice-preset-default.bin`
+
+## Acquired Source Weights
+
+Each tier has `evidence/source-weights.json` and source files under `source/`:
+
+| Tier | Text source | DFlash source | Vision source |
+| --- | --- | --- | --- |
+| `0_6b` | `Qwen/Qwen3-0.6B-GGUF` / `Qwen3-0.6B-Q8_0.gguf` | Missing upstream drafter | n/a |
+| `1_7b` | `Qwen/Qwen3-1.7B-GGUF` / `Qwen3-1.7B-Q8_0.gguf` | Missing upstream drafter | n/a |
+| `9b` | `unsloth/Qwen3.5-9B-GGUF` / `Qwen3.5-9B-Q4_K_M.gguf` | `lym00/Qwen3.5-9B-DFlash-GGUF-Test` / `Qwen3.5-9B-DFlash-q8_0.gguf` | `unsloth/Qwen3.5-9B-GGUF` / `mmproj-F16.gguf` |
+| `27b` | `batiai/Qwen3.6-27B-GGUF` / `Qwen-Qwen3.6-27B-Q4_K_M.gguf` | `spiritbuun/Qwen3.6-27B-DFlash-GGUF` / `dflash-draft-3.6-q8_0.gguf` | `batiai/Qwen3.6-27B-GGUF` / `mmproj-Qwen-Qwen3.6-27B-Q6_K.gguf` |
+| `27b-256k` | same as `27b` | same as `27b` | same as `27b` |
+
+These are conversion/training inputs only. Do not rename them into `text/` or
+`dflash/` release paths without running the final Eliza-1 train, quantize,
+kernel, eval, checksum, license, and upload gates.
+
+## iOS Evidence
+
+XCFramework physical-device smoke now passes:
+
+- Report: `packages/inference/verify/hardware-results/ios-device-smoke-2026-05-11.json`
+- Device: iPhone 15 Pro, iOS 26.3.1, UDID `00008130-001955E91EF8001C`
+- Result: 3/3 XCTest cases passed
+- Voice ABI: required and not skipped
+
+This proves XCFramework structure, Metal availability, runtime symbols, and
+fail-closed ABI calls. It does not prove a final weight-backed Capacitor app
+route yet.
+
+## Publish Blockers
+
+- Final Eliza-1 text GGUFs are not produced for any tier.
+- Final 0.6B and 1.7B DFlash drafter GGUFs are missing.
+- The 9B DFlash source comes from a test repo and is not release-grade.
+- Final long-context variants (`32k`, `64k`, `128k`, `256k`) are not generated
+  and evaluated from trained Eliza-1 checkpoints.
+- Final `dflash/target-meta.json` files are missing for all tiers except any
+  local stand-in artifacts already marked non-publishable.
+- Final evals are missing: text, ASR WER, VAD latency, TTS RTF, expressive voice,
+  DFlash acceptance, first token, first audio, barge-in, 30-turn endurance,
+  mobile RSS, thermal, and backend dispatch.
+- Final release checksums and release-reviewed license attestations are missing
+  for the staged final paths.
+- Final platform evidence is incomplete across native Linux Vulkan, Android
+  graph dispatch, CUDA, ROCm, GH200/H200, native Windows, and weight-backed iOS.
+- `elizalabs` Hugging Face upload evidence is still absent. Existing accessible
+  staging evidence is not a release namespace proof.
+
+## Next Release Actions
+
+1. Train/fine-tune the Eliza-1 text checkpoints for each tier.
+2. Produce or train matching DFlash drafters for 0.6B/1.7B/9B/27B tiers.
+3. Quantize final text and drafters into the exact `text/` and `dflash/` paths
+   required by `ELIZA_1_GGUF_READINESS.md`.
+4. Generate final quantization sidecars for TurboQuant, fused TurboQuant, QJL,
+   and PolarQuant from the exact final bytes.
+5. Run all local eval gates and write final `evals/*.json`.
+6. Run platform hardware dispatch gates and write final `evidence/platform/*.json`.
+7. Generate `checksums/SHA256SUMS` and `evidence/release.json` only after every
+   gate is green.
+8. Upload to `elizalabs/eliza-1-*` and preserve upload logs/Hub URLs in
+   `evidence/release.json`.

ELIZA_1_TESTING_TODO.md

@@ -6,7 +6,7 @@ Status as of 2026-05-11 on this workspace:
 - Standalone Vulkan SPIR-V fixtures pass on this Apple Silicon host through MoltenVK, including TurboQuant, QJL, PolarQuant, and Polar+QJL residual.
 - Built-fork Vulkan graph dispatch source wiring now exists for `GGML_OP_ATTN_SCORE_QJL`, `GGML_OP_ATTN_SCORE_TBQ` (`turbo3`, `turbo4`, `turbo3_tcq`), and `GGML_OP_ATTN_SCORE_POLAR` (`use_qjl=0/1`), but runtime-ready capability bits stay false until native Vulkan graph smoke passes on physical hardware.
 - `adb devices -l` currently shows only `emulator-5554`; emulator Vulkan is diagnostic only and is not recordable Eliza-1 hardware evidence.
-- `xcrun xctrace list devices` currently shows `Shaw's iPhone (26.3.1)` offline even though `xcrun devicectl list devices` sees the iPhone 15 Pro as paired/available; simulator results are not physical iOS evidence.
+- Physical iOS XCFramework smoke now passes on `Shaw's iPhone (26.3.1)` / iPhone 15 Pro UDID `00008130-001955E91EF8001C`; simulator results remain non-recordable for release bundle evidence.
 - CUDA, ROCm, GH200, and native Windows runners are present and fail closed, but this Mac cannot provide recordable target hardware evidence.
 - No final Eliza-1 release bundles exist yet with final weights, hashes, eval outputs, license manifests, and Hugging Face upload evidence.
 
@@ -201,10 +201,16 @@ node packages/app-core/scripts/ios-xcframework/run-physical-device-smoke.mjs
 
 Current iOS blockers:
 
-- The physical iPhone is visible to CoreDevice as paired/available, but `xctrace` still lists UDID `00008130-001955E91EF8001C` as offline.
-- Retrying with the CoreDevice identifier reached an interactive `Password:` prompt before XCTest output; do not enter credentials inside the runner.
+- XCFramework physical-device smoke is PASS as of 2026-05-11:
+  `packages/inference/verify/hardware-results/ios-device-smoke-2026-05-11.json`
+  reports 3/3 XCTest cases passing without `--skip-voice-abi`.
+- The fixed runner treats CoreDevice `connected` state as connected, and
+  `build-xcframework.mjs` refreshes the runtime ABI shim before packaging so
+  stale shim archives cannot reintroduce the old TTS ABI crash.
+- Remaining iOS blocker: run a real Eliza-1 bundle smoke from the Capacitor app
+  shell with final text + DFlash + TTS + ASR payloads and record first token,
+  first audio, peak RSS, thermal state, and at least one local voice route.
 - Simulator runs do not count as physical iOS evidence.
-- Physical smoke must validate the embedded Metal library, Capacitor bridge load, and at least one real local-inference route from the app shell.
 
 ## Release Bundle Evidence