one-surface(parakeet): fold Parakeet ASR onto the C ABI (ROW 1) by localai-bot · Pull Request #121 · mudler/vllm.cpp

localai-bot · 2026-08-07T18:51:11Z

ROW 1 of the ARCH-ONE-SURFACE fold plan (fold #4, audio transcription): grow the ABI -> rewrite the example -> delete the parallel implementation. This PR is the helper claim (operator-helper protocol); head is ONE records-coupled commit for the squash-merge flow.

What landed (order was binding)

Correctness anchor FIRST: deterministic tiny CTC+RNNT checkpoints + clip committed (tests/vllm/models/fixtures/parakeet_e2e, generator in scripts/mm/), transcript goldens captured from the PRE-refactor binary at main@f98e1e48 before any change. The fold gate (test_parakeet_transcription_fold) holds library seam == replicated old pipeline == committed goldens, per head.
W1 tokenizer + library: vllm::Tokenizer implements Metaspace split:true (encode pre-split, MergedWithNext) and the bare Metaspace DECODER (HF decode_chain; the pre-refactor DecodeIds semantics) instead of refusing; Mistral/Gemma Sequence chain regression-pinned; metaspace.rs cases ported. vllm::multimodal::ParakeetTranscriber absorbs WAV ingest (DecodeWavPcm16Mono), head dispatch and id->text; the example's private ReadWav16BitMono/LoadVocab/DecodeIds are deleted.
W2 registry: ParakeetForCTC/ForRNNT/ForTDT registered with the SupportsTranscription mirror on ModelInfo (interfaces.py:1110-1118) - BEYOND-PIN breadth, recorded (pinned vLLM has Parakeet only as the NemotronH audio component). Refuse-by-task everywhere: VT_CHECK(false) refuse-stub forward (runner-routing REFUSE contract), loader refusal via new non-throwing PeekHfArchitectures BEFORE the text HfConfig parse. Registry pins 30 -> 33.
W3 C ABI: vllm_transcribe + params/result structs, ABI v10 -> 11; vllm_engine_load task-dispatches; text and ASR handles refuse each other actionably. First REAL-checkpoint load gated through the public ABI (closes the audit's severity note). FEATURES abi-capability row -> reachable; scripts/abi-capability-allowlist.txt shrinks.
W4 server: task-conditional /v1/audio/transcriptions (multipart file, response_format json/text) through the SAME seam; serving-less ApiServer ctor mirrors vLLM's supported_tasks-conditional registration (generate routes 404 on an ASR server); verified live with curl.
W5 example: examples/parakeet_transcribe is a THIN vllm.h client (vllm::shared); stdout byte-identical to the pre-fold binary on both fixtures; scripts/example-abi-allowlist.txt shrinks and MAX_INTERNAL_REACHING ratchets 12 -> 11 (equality pin + spec claims moved in the same change).

Gates

Full CPU build clean (-Werror, 0 warnings); ctest 349/349.
scripts/agent-preflight.sh EXIT=0 ("All gates green"), including the surface guard now enforcing THIS shrink; all record checkers green unpiped.
Every new test mutation-verified red/green (tokenizer decode rule + split gate, CTC collapse, transcription-only flag, ABI dispatch + input validation, server response_format + route gate).

Residuals (recorded in state.md)

run_batch transcription line; verbose_json/srt/vtt response formats (400 naming the residual); no pretrained checkpoint re-downloaded (safe defaults - fold rests on the committed synthetic pair + P4/P6 pretrained evidence); Whisper/Voxtral remain off-registry (fold #9/#10).

🤖 Generated with Claude Code

https://claude.ai/code/session_014fZAcg1WcU8V629k6HWKys

ARCH-ONE-SURFACE ROW 1 on row/PARAKEET-ONE-SURFACE: grow the ABI, rewrite the example, delete the parallel implementation - audio transcription becomes the first remediated lane of the ONE SURFACE program, gated BYTE-IDENTICAL to the pre-fold example. Correctness anchor FIRST: deterministic tiny CTC+RNNT fixtures + clip (tests/vllm/models/fixtures/parakeet_e2e, scripts/mm/parakeet_e2e_fixture_gen .py) with transcript goldens captured from the PRE-refactor binary at main@f98e1e48 before any change; the fold gate (tests/vllm/models/test_parakeet_transcription_fold.cpp) holds library seam == replicated old pipeline == committed goldens for both heads. W1a tokenizer: Metaspace split=true implemented instead of refused (MergedWithNext pre-split; a bare Metaspace DECODER node selects HF's decode_chain rule - the pre-refactor DecodeIds semantics); Mistral/Gemma Sequence chain regression-pinned; metaspace.rs cases ported (tests/vllm/test_tokenizer_metaspace_split.cpp). W1b/c library: vllm::multimodal::ParakeetTranscriber absorbs WAV ingest (DecodeWavPcm16Mono), the model_type head dispatch and tokenizer decode; the example's private ReadWav16BitMono/LoadVocab/DecodeIds are deleted. W2 registry: ParakeetForCTC/ForRNNT/ForTDT registered with the SupportsTranscription mirror on ModelInfo (interfaces.py:1110-1118); BEYOND-PIN breadth recorded (pinned vLLM: Parakeet = NemotronH audio component, registry.py:511-513). Refuse-by-task everywhere: the forward is a VT_CHECK(false) refuse stub (runner-routing REFUSE contract), the loader refuses via the new non-throwing PeekHfArchitectures BEFORE the text HfConfig parse, and registry count pins move 30 -> 33. W3 C ABI: vllm_transcribe + vllm_transcription_params(_default) + vllm_transcription(_free), ABI v10 -> 11; vllm_engine_load task-dispatches a transcription-only dir onto the same seam; text entrypoints and vllm_transcribe refuse each other's handles actionably. First real-checkpoint load gated through the public ABI (test_capi v11 cases). FEATURES abi-capability row flips to reachable; abi-capability-allowlist SHRINKS. W4 server: task-conditional /v1/audio/transcriptions (multipart file, response_format json/text; verbose_json/srt/vtt named residuals) through the SAME seam; serving-less ApiServer construction = vLLM's supported_tasks- conditional registration (generate routes 404 on an ASR server); examples/server dispatches by resolved arch; verified live with curl. run_batch transcription line stays a named residual. W5 example: examples/parakeet_transcribe is a THIN vllm.h client (vllm::shared), stdout byte-identical to the pre-fold binary on both fixtures; example-abi-allowlist SHRINKS and MAX_INTERNAL_REACHING ratchets 12 -> 11 with the equality pin + spec claims moved in the same change. Gates: full CPU build clean (-Werror, 0 warnings); ctest 349/349; surface guard + mutation suites green enforcing the shrink; every new test mutation-verified (tokenizer decode rule + split gate, CTC collapse, transcription-only flag, ABI dispatch + input validation, server response_format + route gate). Records: STATUS/BENCHMARKS/FEATURES keyed rows, model-matrix rows, state.md entry + NOW refresh, both ONE-SURFACE specs mark ROW 1 landed. FOLLOWING_AGENTS_PROTOCOL Assisted-by: Claude Code:claude-fable-5 [ClaudeCode]

localai-bot · 2026-08-07T20:45:25Z

Landed on main as fc636c7 (mudler-authored squash via commit-tree). Two-round review per protocol (round-1 found 2 test gaps, kill-confirmed fixed, scoped re-review MERGE-CLEAN). Operator-run gates on the merged tree pre-push: keyed-record auto-merge verified byte-identical to the deterministic rebuild (main wholesale + branch delta, all 5 files); fold/metaspace/capi/api_server suites 4/4 SUCCESS on a fresh build; eleven record/doc/guard gates unpiped 00000000000. The C ABI is now v11 with vllm_transcribe — the first non-text capability on the one surface; registry 33 archs; guard ratchet 12→11. Fold ROW 2 (MiniMax-H3 video ABI slice) dispatches next.

…I v11, /v1/audio/transcriptions, thin-client example (#121) ONE-SURFACE fold ROW 1. The library absorbs everything the example privately owned: Metaspace split:true decode lands in vllm::Tokenizer (replacing the FromHfJson refusal), WAV ingest reuses the existing DecodeWavPcm16Mono, and a ParakeetTranscriber seam composes processor -> conformer encoder -> CTC/RNNT/TDT head -> tokenizer decode. Three Parakeet head archs register via REGISTER_VLLM_MODEL (registry 30->33) with transcription-only task refusal both directions. include/vllm.h gains vllm_transcribe + params/default/free idiom, VLLM_ABI_VERSION 10->11 (floor-pinned >= 11 in test_capi). The server routes /v1/audio/transcriptions through the same seam, gated on a transcriber engine (socket-level 404 pin on text servers, /v1/videos-precedent style). examples/parakeet_transcribe is rewritten as a vllm.h-only thin client - stdout byte-identical to the pre-fold binary - and the surface guard ratchet shrinks 12->11 with its equality pin. Correctness: three-arm fold gate (library seam == replicated old pipeline == committed goldens captured at f98e1e4 pre-refactor) on synthetic CTC+RNNT fixtures; Mistral/DeepSeek real-checkpoint tokenizer parity unchanged. Two-round review per protocol: round-1 mutation review found 2 test gaps (route gate on text servers untested; ABI floor unpinned), fixed and kill-confirmed by a scoped re-review, MERGE-CLEAN. Operator re-ran the fold/tokenizer/capi/server suites on this merged tree: 4/4 SUCCESS. FOLLOWING_AGENTS_PROTOCOL Assisted-by: Claude Code:claude-fable-5 [ClaudeCode]

…binary at branch base The three-arm fold gate's arm C, captured BEFORE any refactor (the #121 discipline): minimax_h3_video_fold_fixture.h writes a deterministic tiny checkpoint set (857KB F32 ComfyUI-GGUF DiT at the :3786 reduced geometry, the reduced ViT3D video VAE + BigVGAN audio VAE as on-disk safetensors+config.json, prompt embeds), and the committed goldens are what the PRE-fold minimax-h3-gen binary at this branch base rendered on it: minimax-h3-gen --dit dit.gguf --keep-quant --video-vae ... --audio-vae ... --prompt-embeds prompt_embeds.f32 --partition fl2va --steps 3 --frames 5 --height 32 --width 32 --device cpu --ffmpeg /bin/true 8 PPM frames (32x32) + audio.wav (stereo, 32000 Hz), byte-stable across two runs (diff -r clean), plus the pre-fold minimax-h3-mux --print-only argv for both the audio and silent forms. The fold gate to come holds the library seam, the replicated old pipeline, and the rewritten thin clients to these bytes. Also captured (for the refuse-both-directions pin): vllm_engine_load on this H3 fixture dir today returns status 2 with 'hf_config: cannot open .../config.json'. FOLLOWING_AGENTS_PROTOCOL Assisted-by: Claude Code:claude-fable-5 [ClaudeCode]

…le, allowlist 11->10 ROW 7 ONE-SURFACE B4 (§20.3, task #281), rebased on #121 (Parakeet took v11). - include/vllm.h + src/capi/vllm_c.cpp: vllm_complete_tokens (ABI v12) — blocking completion from a PRE-TOKENIZED prompt (vLLM TokensPrompt, the AsyncLLM token-id add_request overload), returning the generated token ids and optionally the detokenized vllm_completion. Purely additive; capi gate test_capi 35/35·290 with null-contract + truncation + string-leg-identity cases (mutation-verified). - examples/kimi_linear_gen: REWRITTEN as a thin public-ABI client (vllm.h + vllm::shared only) — vllm_engine_load + vllm_complete_tokens over the §12 golden's pre-tokenized prompts, two-length-diff steady-rate report. The private harness (internal bf16-resident loader + KimiDecodeCache incremental drive) is deleted from examples; the CLI-incremental REFERENCE leg moves to the env-gated tests/vllm/models/test_kimi_linear_fold_gate.cpp (VT_KIMI_MODEL_DIR/VT_KIMI_GOLDEN_DIR; skips in the CPU suite), where internal headers are legitimate. - scripts/example-abi-allowlist.txt: kimi_linear_gen row REMOVED; MAX_INTERNAL_REACHING 11->10; equality pin + spec claims moved in the same change (test_check_surface_coverage 46 passed; surface-coverage OK: 10 internal-reachers, kimi is the third clean ABI client). - docs/FEATURES.md capability table: Pre-tokenized completion row (reachable). FOLLOWING_AGENTS_PROTOCOL Assisted-by: Claude Code:claude-fable-5 [ClaudeCode]

…-> 12 ARCH-ONE-SURFACE ROW 2, the C surface over the W1 seam — the ratified shapes (.agents/specs/one-surface-abi.md:92-108) as shipped: vllm_video_engine (an opaque handle SEPARATE from vllm_engine, because H3 loads a checkpoint SET, not a config.json directory), vllm_video_model_params/_default, vllm_video_params/_default, vllm_video_result, vllm_video_engine_load/free, vllm_video_generate, vllm_video_result_free — plus the engine-free vllm_video_mux_argv(+_params_default,+_free) composer the minimax-h3-mux thin client needs (a disclosed append beyond the ratified proposal; the encoding contract stays the library's, the caller execs). Deviations from the proposal, each argued in the spec: + prompt_embeds_path (without an encoder there is NO conditioning path), + partition (the #77 guard refuses every render without it), + fp4_resident and + output_dir; - task/duration (derivable / expressible via num_frames); ref_image stays the ratified SINGLE image (multi-image ref2va remains C++-seam-reachable, named residual). mux_argv is NULL-terminated (execvp-ready). Zero values preserve behaviour (noise_aug <= 0 resolves to the 1.0 pin engine-side). Refuse-both-directions: the seam now refuses a config.json-style directory LOUDLY naming vllm_engine_load; vllm_engine_load on the H3 fixture dir keeps failing byte-for-byte as captured at v11 (status 2, missing config.json). Gates: vllm_capi_c_check (strict-C11 compile) green; test_capi 40/40 (373 assertions) including the v12 section — the FULL golden e2e through the C marshalling (frames+WAV byte-identical to the pre-fold binary's committed goldens, mux argv pinned), the zero-value contract, both refusal directions, and the argument contract; ABI floor pin advanced >= 11 -> >= 12 (the #121 lesson — the == checks move with the macro). FOLLOWING_AGENTS_PROTOCOL Assisted-by: Claude Code:claude-fable-5 [ClaudeCode]

…— vllm_video_*, ABI v12, /v1/videos via the library seam, thin-client examples ONE-SURFACE fold ROW 2 (row/H3-VIDEO-ABI, task #283; the #121 precedent's shape: three-arm fold gate, guard-ratchet shrink). The library absorbs everything the examples privately owned. GOLDENS FIRST: minimax_h3_video_fold_fixture.h writes a deterministic tiny checkpoint set (857KB F32 ComfyUI-GGUF DiT at the test_minimax_h3.cpp:3786 reduced geometry, on-disk reduced ViT3D/BigVGAN VAEs + config.json, prompt embeds); the PRE-fold minimax-h3-gen binary AT THE BRANCH BASE (fc636c7) rendered it (--partition fl2va --steps 3 --frames 5 --height/width 32, CPU, keep-quant; run-to-run deterministic, diff -r clean) and its 8 frames + WAV + both minimax-h3-mux --print-only argv lines are the committed goldens in fixtures/minimax_h3_video_fold/. W1 seam: vllm::multimodal::MiniMaxH3VideoEngine absorbs the 1293-line example driver AND the server's 354-line /v1/videos twin — 4 DiT loader arms (GGUF keep-quant / dequant-bf16 host+streamed / bf16-shard stream / NVFP4 bf16+fp4-resident), VAE decoder + lazy encoder halves, the H3-Encoder tower staged once (GGUF or bf16 shards), the #77 partition guard, fl2va/ref2va conditioning, the byte-exact splitmix64 Box-Muller noise streams (VT_H3_GAUSSIAN_NOISE honored; seeded requests derive the audio stream via splitmix64), artifact writing and mux-argv assembly. The library SPAWNS NOTHING (mkdir -p became std::filesystem; ffmpeg stays caller-side — the ratified 2026-08-03 boundary). MiniMaxH3VideoGenParamsFromRequest is the ONE /v1/videos mapping so HTTP and FFI cannot drift. W2 ABI: the ratified vllm_video_* slice (one-surface-abi.md:92-108) — vllm_video_engine_load/free, vllm_video_model_params/_default, vllm_video_params/_default, vllm_video_generate, vllm_video_result/_free — plus the engine-free vllm_video_mux_argv(+_params_default,+_free) composer the mux thin client needs. VLLM_ABI_VERSION 11 -> 12; test_capi floor pin >= 12 (the == checks move with the macro). As-shipped deltas vs the proposal argued in the spec (+prompt_embeds_path/+partition/+fp4_resident/+output_dir; -task/-duration; ONE ref_image). Zero values preserve behaviour. Refusals both directions: vllm_video_engine_load on a text dir names vllm_engine_load; vllm_engine_load on the H3 dir fails byte-for-byte as captured at v11. W3 server: the VideoState block is DELETED; /v1/videos = seam Load + a runner lambda (FromRequest -> Generate -> fork/execvp(mux_argv)). Direct MiniMaxH3 refs 54 -> 7, all seam type names. DISCLOSED numeric deltas on the server arm (no goldens existed; the drift WAS the defect): legacy single-stream uniform noise -> the shared ratified recipe; host-f32 GGUF default -> keep-quant; new --video-dequant-bf16 flag. W4 examples: minimax_h3_gen (1293 -> 216 lines) + minimax_h3_mux are vllm.h+vllm::shared thin clients, BOTH byte-identical to the pre-fold binaries on the fixture (frames+WAV cmp clean; mux argv diff clean; the pre-fold diagnostic modes died with the private pipeline — named residual). Ratchet: both example-abi-allowlist rows removed, MAX_INTERNAL_REACHING 11 -> 9 with the equality pin + spec claims moved; the abi-capability video row closes; FEATURES video row -> reachable naming the four symbols. Gates: three-arm fold gate test_minimax_h3_video_fold 4 cases / 131 assertions GREEN (seam == replicated pre-fold pipeline == committed goldens, byte-identical; partition-guard + input refusals; FromRequest field-complete); test_capi 40/373 incl. the v12 golden e2e THROUGH the C marshalling; vllm_capi_c_check strict-C11 green; test_minimax_h3 75/75 unchanged; test_openai_api_server 45/566 incl. the socket-level routes-do-not-exist-without-a-runner 404 pin; test_video_api 14/14; check-surface-coverage green with the SHRINK enforced + its 46-test mutation suite. GB10 real-video re-verification through the v12 ABI is a NAMED RESIDUAL (box on the Kimi campaign); records updated (STATUS/BENCHMARKS/FEATURES/NOW/state/model-matrix + both specs). FOLLOWING_AGENTS_PROTOCOL Assisted-by: Claude Code:claude-fable-5 [ClaudeCode]

…odel_params.device, ABI v14, --device on server + cli ONE-SURFACE fold ROW 8 (row/DEVICE-KNOB, task #284), the smallest: an embedder can now ASK for a device instead of inheriting the accelerator-first probe. Mirrors vLLM's DeviceConfig.device names (vllm/config/device.py:13 @ 555967922; an explicit device is assigned VERBATIM and never substituted, device.py:61-66 - upstream's own --device flag was removed in v0.10, bc8a8ce5e, so the field mirrors the surviving config surface). include/vllm.h gains vllm_model_params.device (int32, 0=auto/1=cpu/ 2=cuda; 0 MUST be auto so zero-initialized structs keep the pre-v14 probe byte-identical; cpu-before-cuda follows the shipped v12 vllm_video_model_params.device precedent shifted by the auto slot). VLLM_ABI_VERSION -> 14 per the operator's version-collision heads-up (#122 reconciles the Kimi vllm_complete_tokens slice to v13 at its landing; the changelog names the gap and this branch rebases after). Floor pin advanced to >= 14 (the #121 == lesson). New vllm/config/device.h (+.cpp): vllm::Device, DeviceFromString, DeviceName. Plumb: EngineParams::device -> SelectQueue explicit arms via the pure LoadedEngine::ResolveExplicitDeviceType (explicit cpu NEVER probes; explicit cuda requires the registered kCUDA platform and fails LOUD otherwise; explicit-arm queue-creation failures propagate, unlike auto's byte-identical catch-and-fall-back). FromModelDir resolves an explicit device BEFORE any path/config I/O (DeviceConfig resolves at config time, arg_utils.py:1878), so a device error is never masked by a path error. capi validates 0/1/2 and refuses explicit cuda on the CPU-hosted transcription stack. Thin clients: --device auto|cpu|cuda on vllm-cli (ABI field only) and the server (DeviceFromString; --video-device untouched, separate engine). Also folds the #123 review's three MINOR findings (tests/docs only): c_header_compile.c actually references the v11+v12 surface + the v14 field (its "every entry point" claim went stale at v10); the v12 changelog block moved to chronological position; the H3 fold fixture flag list gained --keep-quant. Gates: full CPU build -Werror clean; test_capi (4 new v14 cases + floor), test_loaded_engine_dense (+3 device cases incl. the pure policy matrix that pins "explicit cpu beats a REGISTERED accelerator" CPU-side), test_dlopen, test_openai_api_server (+explicit-cpu serving smoke), test_minimax_h3_video_fold, test_model_loader_gguf, test_platform all green; surface-coverage green, ratchet UNTOUCHED at 9; STATUS ratchet paid by removing a stale back-to-back merge duplicate (279200 -> 279150). Ten mutations killed and reverted (defaults, zero-map, range, both plumb sites, ctor arm, unconditional fallback, policy flip, parse flip, forgotten bump); kills recorded in .agents/state.md. Residuals: CUDA-build A/B (no dgx in scope) and the pending #122 rebase, both named in the state entry. FOLLOWING_AGENTS_PROTOCOL Assisted-by: Claude Code:claude-fable-5 [ClaudeCode]

…odel_params.device, ABI v14, --device on server + cli ONE-SURFACE fold ROW 8 (row/DEVICE-KNOB, task #284), the smallest: an embedder can now ASK for a device instead of inheriting the accelerator-first probe. Mirrors vLLM's DeviceConfig.device names (vllm/config/device.py:13 @ 555967922; an explicit device is assigned VERBATIM and never substituted, device.py:61-66 - upstream's own --device flag was removed in v0.10, bc8a8ce5e, so the field mirrors the surviving config surface). include/vllm.h gains vllm_model_params.device (int32, 0=auto/1=cpu/ 2=cuda; 0 MUST be auto so zero-initialized structs keep the pre-v14 probe byte-identical; cpu-before-cuda follows the shipped v12 vllm_video_model_params.device precedent shifted by the auto slot). VLLM_ABI_VERSION 13 -> 14 on top of #122's vllm_complete_tokens v13 (rebased onto 9f772ca; the changelog now reads v10..v14 chronological - the v12-after-v11 move is the #123-review minor kept through the rebase). Floor pin advanced to >= 14 (the #121 == lesson). New vllm/config/device.h (+.cpp): vllm::Device, DeviceFromString, DeviceName. Plumb: EngineParams::device -> SelectQueue explicit arms via the pure LoadedEngine::ResolveExplicitDeviceType (explicit cpu NEVER probes; explicit cuda requires the registered kCUDA platform and fails LOUD otherwise; explicit-arm queue-creation failures propagate, unlike auto's byte-identical catch-and-fall-back). FromModelDir resolves an explicit device BEFORE any path/config I/O (DeviceConfig resolves at config time, arg_utils.py:1878), so a device error is never masked by a path error. capi validates 0/1/2 and refuses explicit cuda on the CPU-hosted transcription stack. Thin clients: --device auto|cpu|cuda on vllm-cli (ABI field only) and the server (DeviceFromString; --video-device untouched, separate engine). Also folds the #123 review's three MINOR findings (tests/docs only): c_header_compile.c actually references the v11+v12 surface + the v14 field (composed with #122's vllm_complete_tokens reference; its "every entry point" claim went stale at v10); the v12 changelog block moved to chronological position; the H3 fold fixture flag list gained --keep-quant. Gates: full CPU build -Werror clean; test_capi (4 new v14 cases + floor >= 14, #122's v13 cases green alongside), test_loaded_engine_dense (+3 device cases incl. the pure policy matrix that pins "explicit cpu beats a REGISTERED accelerator" CPU-side), test_dlopen, test_openai_api_server (+explicit-cpu serving smoke), test_minimax_h3_video_fold, test_model_loader_gguf, test_platform all green; surface-coverage green, allowlist ratchet UNTOUCHED at 8 (post #122); STATUS ratchet paid by removing a stale back-to-back merge duplicate (279200 -> 279150, measured 279111 post-rebase). Ten mutations killed and reverted (defaults, zero-map, range, both plumb sites, ctor arm, unconditional fallback, policy flip, parse flip, forgotten bump); kills recorded in .agents/state.md. Residual: CUDA-build A/B (no dgx in scope), named in the state entry. FOLLOWING_AGENTS_PROTOCOL Assisted-by: Claude Code:claude-fable-5 [ClaudeCode]

mudler force-pushed the row/PARAKEET-ONE-SURFACE branch from a26a225 to 162a21e Compare August 7, 2026 20:06

mudler force-pushed the row/PARAKEET-ONE-SURFACE branch from 162a21e to 5984be3 Compare August 7, 2026 20:33

localai-bot closed this Aug 7, 2026

localai-bot mentioned this pull request Aug 7, 2026

feat(abi): explicit device selection through the ONE surface - vllm_model_params.device, ABI v14, --device on server + cli #136

Closed

localai-bot mentioned this pull request Aug 8, 2026

feat(abi): embeddings/pooling through the ONE surface — LlamaModel arch, PoolingRunner in the engine step, vllm_embed (ABI v15), live /v1/embeddings #137

Closed

Navigation Menu

one-surface(parakeet): fold Parakeet ASR onto the C ABI (ROW 1) - #121

one-surface(parakeet): fold Parakeet ASR onto the C ABI (ROW 1)#121
localai-bot wants to merge 1 commit into
mainfrom
row/PARAKEET-ONE-SURFACE

localai-bot commented Aug 7, 2026 •

edited

Loading

Uh oh!

localai-bot commented Aug 7, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

localai-bot commented Aug 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What landed (order was binding)

Gates

Residuals (recorded in state.md)

Uh oh!

localai-bot commented Aug 7, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

localai-bot commented Aug 7, 2026 •

edited

Loading