Repository navigation

refactor(llm)!: migrate to modular TypeScript runners and reexport models with expanded context lengths - #1508

Merged

barhanc merged 83 commits into

mainfrom

@bh/llm-runner-refactor

Oct 22, 2026

Merged

refactor(llm)!: migrate to modular TypeScript runners and reexport models with expanded context lengths#1508
barhanc merged 83 commits into
mainfrom
@bh/llm-runner-refactor

Commits on Oct 21, 2026
- refactor(llm): remove native C++ LLM runner and tests
  barhanc
  committed
  992c77c
- refactor(bridge): remove llm namespace from rnexecutorchJsi docstring
  barhanc
  committed
  e8a47f6
- style
  barhanc
  committed
  9fe43d3
- feat(model): support returning primitive lists from execute
  barhanc
  committed
  2ae8fa4
- refactor(llm): split runner into runners/ directory and update stats types
  barhanc
  committed
  ec532a6
- feat(llm): add runners/ directory with textRunner, multimodalRunner, gemmaRunner
  barhanc
  committed
  87f4439
- style
  barhanc
  committed
  731edfd
- style
  barhanc
  committed
  4ec801d
- feat(llm): add sampling pipeline (wip)
  barhanc
  committed
  7961f06
- feat(llm): add token sampling pipeline and wire into textRunner
  barhanc
  committed
  d9eaa1e
- wip
  barhanc
  committed
  6d4b0ed
- revert(model): revert primitive list support from execute and bridge docstring
  barhanc
  committed
  3b87789
- style
  barhanc
  committed
  3d27e96
- feat(llm): add native C++ sampler and install map
  barhanc
  committed
  16d3d16
- style(llm): drop churn from sampler wiring
  barhanc
  committed
  fecf0fd
- fix(tests): drop et_llm_runner and LlmRunnerTest from the test build
  barhanc
  committed
  d2563c6
- refactor(llm): prefix runner modules and drop math multinomial
  barhanc
  committed
  afc7ded
- refactor(llm): clean up default sampling options formatting
  barhanc
  committed
  a69ea58
- docs(llm): emphasize in-place logits modification in sample JSDoc
  barhanc
  committed
  0182052
- refactor(llm): remove profiling code from llmTextRunner
  barhanc
  committed
  e868dfa
- style
  barhanc
  committed
  5a040bb
- feat(llm): implement multimodal runner and clean up text runner
  barhanc
  committed
  5fdec2d
- fix(llm): allow rank-4 targetShape in chatPreprocessor
  barhanc
  committed
  07695a3
- refactor(llm): decouple specialized runner specs and polymorphically dispatch from llmRunner
  barhanc
  committed
  5368533
- fix(llm): add descriptive error message for unsupported media segment kind
  barhanc
  committed
  ab82eb1
- refactor(llm): simplify segment kind access in prefill error
  barhanc
  committed
  e7cdb69
- refactor(llm): flatten specialized runner specs into model specs
  barhanc
  committed
  2c63dd0
- refactor(llm): self-validate runner specs inside specialized runner constructors
  barhanc
  committed
  fc0ba83
- style
  barhanc
  committed
  964eb57
- style
  barhanc
  committed
  2ce17cc
- style
  barhanc
  committed
  f13dc53
- update
  barhanc
  committed
  7c74aa9
- update
  barhanc
  committed
  005a09d
- test(llm): drive the chat session through the TypeScript LLM runner
  
  barhanc
  and
  claude
  committed
  24ce001
- feat(nlp-app): add a backend picker to the LLM screen
  
  barhanc
  and
  claude
  committed
  65a31a4
- chore(core): attribute the input shape-check TODO in model.cpp
  
  barhanc
  and
  claude
  committed
  58e92e0
- update
  barhanc
  committed
  a41def8
- test(cpp): cover the native LLM sampler
  
  barhanc
  and
  claude
  committed
  73bb49d
- style
  barhanc
  committed
  f99ffe5
- docs(llm): document the TypeScript LLM runner and the native sampler
  
  barhanc
  and
  claude
  committed
  e4669fe
- feat(models): restore the LLM variants pending a re-export
  
  barhanc
  and
  claude
  committed
  8d7feb6
- fix(models): point LLM tokenizers back at the shared repo-root files
  
  barhanc
  and
  claude
  committed
  f3c231b
- style
  barhanc
  committed
  a34884f
- feat(models): point the LLMs at the re-exported scratch builds
  
  barhanc
  and
  claude
  committed
  d6f3119
- feat(nlp-app): list every LLM variant on the LLM screen
  
  barhanc
  and
  claude
  committed
  7b2c7c8
- fix
  barhanc
  committed
  bfc91c1
- fix
  barhanc
  committed
  3f5196c
- update
  barhanc
  committed
  ea5b543
- update
  barhanc
  committed
  45c9c0f
- refactor(llm): use tensor array destructuring pattern in multimodal runner
  barhanc
  committed
  2b35466
- feat(nlp-app): set tailored generation configs for each LLM model family
  barhanc
  committed
  e74754b
- refactor(llm): align runner reset bounds and clean up gemma runner TODOs
  barhanc
  committed
  253391a
- fix: remove llmBrand
  barhanc
  committed
  6f03765
- test(llm): add tests for multimodal runner, variant dispatch, sampler options and construction failure
  barhanc
  committed
  7ad4a1d
- refactor(llm): initialize prefill totStats to zero stats and omit non-null assertion
  barhanc
  committed
  4462dac
- fix(core): reject non-finite numbers in double and float conversions
  barhanc
  committed
  1d4b3b4
- feat(llm): add penaltyLastN window and seed to sampler with full-chain pipeline tests
  barhanc
  committed
  893c08f
- fix(build): run prepare before workspace typecheck
  barhanc
  committed
  6d84a23
- fix(llm): prevent EOS tokens reaching onToken and fix Hammer/Llama tool parsing
  barhanc
  committed
  5312189
- feat(llm): support Gemma Omni model and audio prefill chunking
  barhanc
  committed
  5b55ea8
- test(llm): add test suite for createLLMGemmaRunner and remove TODO in runner
  barhanc
  committed
  c75e696
- test(llm): allow arbitrary shapes in sample and test flattening
  barhanc
  committed
  4376330
- fix
  barhanc
  committed
  87fc502
- fix(llm): mark reset and getKVCacheState as worklet in runners
  barhanc
  committed
  a32fce1
- feat(llm): add Gemma Omni Vulkan model variant
  barhanc
  committed
  1dbf87f
- feat(models): wire Gemma 4 Omni E2B MLX variant in models registry and NLP example app
  barhanc
  committed
  37d28ac
- feat(models): update Qwen 3 0.6B MLX variant to int8 in models registry and NLP app
  barhanc
  committed
  a9eddf1
- feat(models): rename Gemma MLX variant to gemma_4 in models registry
  barhanc
  committed
  f20cb48
- fix(llm): match HF Gemma PLE for media prefill and drop padded audio tokens
  barhanc
  committed
  d54f17f
- chore(nlp): update react-native-audio-api to 0.13.6
  barhanc
  committed
  4b9a932
- fix(llm): initialize top-k heap for constrained sampling
  barhanc
  committed
  bce7bb9
- feat(nlp): add constrained decoding Potion Lab demo
  barhanc
  committed
  aab6fd3
- fix(nlp): explain constrained decoding validation results
  barhanc
  committed
  3cd3bc2
- docs: await createLLMRunner directly in generation example
  barhanc
  committed
  966fb0e
- fix: reject overlapping LLM chat session turns
  barhanc
  committed
  0b07db7
- fix: preserve split UTF-8 characters in LLM streaming
  barhanc
  committed
  2534cd2
- docs: clarify local model filesystem path requirements
  barhanc
  committed
  a715834
- fix: roll back cancelled LLM chat turns
  barhanc
  committed
  2711a7e
- docs: clarify LLM chat rollback behavior
  barhanc
  committed
  bc79589
- feat: add LFM2.5 2.6B models and example tool calling
  barhanc
  committed
  de5d4d1
- docs: simplify LFM2.5 2.6B model description
  barhanc
  committed
  f88bf5a
- fix: pass chat special tokens and advance generation seeds
  barhanc
  committed
  4fdaf13
Commits on Oct 22, 2026
- feat: support configurable LLM prefill length
  barhanc
  committed
  e7cebc1

Commits

Commits on Oct 21, 2026

Commits on Oct 22, 2026