Issues

Qwen export harness (StableHLO + parameters) for the IREE path — none exists; FunctionGemma is the only tool-calling model with a compiled leg

#411

· michalharakal opened

on Sep 3, 2026

iree-android: stateful prefill/step contract with a reusable prompt-prefix KV snapshot (and an embeddings-input variant for Vulkan)

#410

· michalharakal opened

on Sep 3, 2026

llm-agent: QwenChatTemplate(enableThinking = false) pre-fills an empty <think> block, which Qwen2.5-Instruct never saw — tool calling drops to 0–1/8 vs 4/8 with the official Qwen2.5 template

#409

· michalharakal opened

on Sep 3, 2026

FunctionGemma redecode contract: LM head + argmax run over all SEQ positions (36 % of every step, 1 GB scratch at SEQ 1024 in a 32-bit process); parameter archives are per-SEQ

#406

· michalharakal opened

on Sep 3, 2026

[gemma3n] Split into its own family module + conformance (postponed)

#377

· michalharakal opened

on Sep 1, 2026

Retire kllama-cli's legacy dense-FP32 LlamaRuntime path; route Llama/Mistral GGUF through the DSL like Qwen already does

#354

· michalharakal opened

on Aug 29, 2026

Two independent GGUF tokenizer implementations silently disagree (GGUFTokenizer vs TokenizerFactory)

#330

· michalharakal opened

on Aug 21, 2026

Antora docs site has no version selector — pin per release instead of a single unversioned component

#317

· michalharakal opened

on Aug 13, 2026

Validate kllama Apple native-kernel fix (PR #315) on real iOS/macOS hardware

help wanted

#316

· michalharakal opened

on Aug 13, 2026

SmolLM2-135M compiled leg: DSL → StableHLO → IREE showcase (gemma-iree precedent)

#305

· michalharakal opened

on Aug 12, 2026

Register the native cinterop kernels on Apple targets in the runtime facades

#298

· michalharakal opened

on Aug 11, 2026

Evaluate the existing BERT/MiniLM embedding model (feature/bert-dsl-embeddings): quality + on-device numbers, and a merge path

#273

· michalharakal opened

on Aug 9, 2026

Navigation Menu

Uh oh!

All issues

Qwen export harness (StableHLO + parameters) for the IREE path — none exists; FunctionGemma is the only tool-calling model with a compiled leg

iree-android: stateful prefill/step contract with a reusable prompt-prefix KV snapshot (and an embeddings-input variant for Vulkan)

llm-agent: QwenChatTemplate(enableThinking = false) pre-fills an empty <think> block, which Qwen2.5-Instruct never saw — tool calling drops to 0–1/8 vs 4/8 with the official Qwen2.5 template

FunctionGemma redecode contract: LM head + argmax run over all SEQ positions (36 % of every step, 1 GB scratch at SEQ 1024 in a 32-bit process); parameter archives are per-SEQ

[gemma3n] Split into its own family module + conformance (postponed)

Retire kllama-cli's legacy dense-FP32 LlamaRuntime path; route Llama/Mistral GGUF through the DSL like Qwen already does

Two independent GGUF tokenizer implementations silently disagree (GGUFTokenizer vs TokenizerFactory)

Antora docs site has no version selector — pin per release instead of a single unversioned component

Validate kllama Apple native-kernel fix (PR #315) on real iOS/macOS hardware

SmolLM2-135M compiled leg: DSL → StableHLO → IREE showcase (gemma-iree precedent)

Register the native cinterop kernels on Apple targets in the runtime facades

Evaluate the existing BERT/MiniLM embedding model (feature/bert-dsl-embeddings): quality + on-device numbers, and a merge path

Uh oh!

Search results