feat: Gemma 4 cutover — eliza-1 Qwen→Gemma, eliza-1-only stack, cloud HF-proxy, dynamic-fit memory (#9033)#9060
Merged
Merged
Conversation
Gemma 4 E2B+E4B text gen + E2B vision (mmproj) + assembled-bundle load all verified through the fork on CPU (stock-q8_0 KV path). Device/CUDA/Metal lanes remain hardware-gated. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…a auto-provision) Make Feed deployable as a single always-on Railway web service (reusing the existing Steward + cloud inference), with no external cron and no migrate step: - startInternalCronLoop(): in-process scheduler (gated on ENABLE_INTERNAL_CRON_SCHEDULER=true) that fires the game loop's entry crons (game-tick fans out the rest) every 60s against the local server with the CRON_SECRET — so one web container runs the live game. Wired at boot in instrumentation.ts. - railway.json: Nixpacks build (bun install + feed web build) + healthcheck. - scripts/railway-start.sh: ensure schema via drizzle-kit push (the migration history has parallel 0000 baselines that cannot apply to a fresh DB), then next start — auto-provisioned 'just works' boot. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…oud inference) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ernel contract (#9033) The M6 schema change (REQUIRED_KERNELS_BY_TIER → turboquant_q4 only; QJL/Polar/ turbo3_tcq optional for Gemma's stock-q8_0 KV) left the validator + tests on the old Qwen contract: - validator.ts: drop the 'ctx>64k requires turbo3_tcq' hard rule (Gemma handles long context via native windowed-SWA + shared-KV; turbo3_tcq is now optional). - manifest.test.ts: assert the Gemma required set (turboquant_q4 only), turbo3_tcq optional-when-long-context is now accepted; retarget rejection triggers. - delete obsolete generic-gguf tests (backend-runtime-class.test.ts, assignment-not-servable-route.test.ts — M9 removed those code paths); trim the removed canServeRuntimeClassOnHost suite from assignment-validation.test.ts (its setAssignment boundary tests still own the non-eliza-1-rejection contract). Gates: manifest 49/49, the manifest+assignment+catalog combo 72/72, typecheck 0. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…d, idempotent) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…t drift (#9033) - memory-benchmark: plannedKvQuant qjl1_256→q8_0 (Gemma stock KV). - fused-eliza1-no-regression: decideBackend no longer carries a runtimeClass field or a generic-gguf backend (eliza-1-only); assert backend==llama-cpp for both known + unknown catalog entries. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…9033) - eot-scorer: <|im_end|>→<end_of_turn>, <|im_start|>user→<start_of_turn>user (M2 EOT). - text-bundle + mmproj-routing: same-file NextN MTP → separate official Gemma drafter (mtp/drafter-<slug>.gguf component + runtime.mtp.drafterFile, draftMax 4). Also fixed a real test-bundle drafter-path mismatch the catalog change exposed. - vision-describe: cache family qwen3-vl→gemma-vl (M2 vision default). - downloader: HF bearer → Eliza Cloud API-key via cloud HF-proxy (M10a). Full plugin-local-inference suite via vitest: 179 files, 1869 pass, 0 fail. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…r develop merge (#9033) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Your free trial has ended. If you'd like to continue receiving code reviews, you can add a payment method here.
Contributor
|
Important Review skippedAuto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the ⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Run ID: You can disable this status message by setting the Use the checkbox below for a quick retry:
✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
Contributor
|
Claude encountered an error —— View job I'll analyze this and get back to you. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Cuts the eliza-1 local model line from the Qwen3.5/3.6 backbone to Gemma 4 end-to-end, and consolidates the local stack to eliza-1-only. Tracking: #9033. Supersedes #8794 (closed); resolves #8808 (closed, generic-GGUF removed); addresses #8807/#8809 (M10).
What's in this PR (all CPU/test-verified)
memory_calcvocab 248320→262144; abliterate dense surgery; catalog/types/device-fit (tokenizerFamilygemma4, separate-drafter MTP, stock-q8_0 KV); voice EOT<|im_end|>→<end_of_turn>; vision familyqwen3-vl→gemma-vl; AGENTS/native contracts. ~250 files swept (mechanical + load-bearing), strict KEEP of frozen Qwen3-ASR/OmniVoice/Embedding/turn-detector lineage + banned-name guards.RuntimeClasscollapsed tofused-eliza1.packages/cloud-api/v1/hf-proxy+shared/hf-proxy.ts— HF downloads route through Eliza Cloud (no local HF keys).estimatedMbfix, desktop bench harness.assemble_local_gemma_bundle.pybuilds a verified-loadable E2B bundle; legacy-tier purge helper.libelizainferenceFFI seam + LiteRT/MLX/CoreML scaffolds (gated off; compile-verified).Verification
@elizaos/plugin-local-inference: 179 files, 1869 tests pass, 0 fail (vitest).@elizaos/shared: 65 files, 947 pass, 0 fail. cloud-api/cloud-shared/ui/benchmarks typecheck 0 errors.Deliberately NOT in this PR (gated, tracked in #9033)
elizaos/eliza-1Gemma bundle upload — gated on the non-standard MTP-drafter safetensors→GGUF conversion (base ships MTP-disabled meanwhile).🤖 Generated with Claude Code