Community contributions: M1-M4 compat, security fixes, docs, benchmarks, and community dashboard by dev-erik · Pull Request #25 · maderix/ANE

[fix] M1/M2/M3 MIL syntax compatibility (upstream PR #6): use program(1.0), ios16 target, tensor types across 18 files

dev-erik

committed

[fix] Token sampling underflow fix (upstream PR #17): prevent size_t wraparound on short datasets in both train_large variants

dev-erik

committed

[fix] Dashboard sudo hang fix (upstream PR #20): prevent blocking when password is required for powermetrics

dev-erik

committed

[fix] Security hardening (upstream PRs #5, #7): stack-protector-strong, format-security flags, NULL guards on ane_compile/fread/fopen, tokenize.py input validation

dev-erik

committed

[feat] ANE ChainingRequest API prototype: baseline measurement for multi-kernel pipelining without recompile overhead

dev-erik

committed

[docs] Developer documentation: architecture diagrams, complete API reference, benchmark guide, M4 Max results, security audit report

dev-erik

committed

[feat] Benchmark runner and mlpackage generator: run_benchmarks.sh for full test suite, gen_mlpackages.py for CoreML model generation

dev-erik

committed

[feat] Community benchmark system: standardized JSON output, auto-submit to dashboard, aggregation script, M4 Max reference result

dev-erik

committed

[docs] Community fork README, CONTRIBUTING guide, issue templates, gitignore: rewritten README with quickstart, env vars, benchmark instructions, dashboard link

dev-erik

committed

[feat] Merge upstream: dynamic weight training, CLI fixes, dashboard v2

dev-erik

committed

[feat] Merge upstream PRs #21, #23, #26: NEON-optimized training (train_opt), double-buffered async ANE training (train_double_buffer), Qwen2.5-0.5B LLM inference (inference/). Added get_path() env…

dev-erik

committed

[feat] Inference server mode: keep ANE kernels loaded between prompts (stdin loop + Unix socket server). Subsequent queries respond in ~0.5s instead of ~6s. run.py auto-connects to socket server wh…

dev-erik

committed

[feat] Optimize inference: vectorize ops (NEON/vDSP), gate debug output, skip unused ANE compilation, add round-trip benchmark timing, pure C HTTP API with tokenizer

dev-erik

committed

[feat][gpu] Q4 quantization, Metal GPU shaders, ANE kernel fusion, memory safety

dev-erik

committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Community contributions: M1-M4 compat, security fixes, docs, benchmarks, and community dashboard#25

Community contributions: M1-M4 compat, security fixes, docs, benchmarks, and community dashboard#25
dev-erik wants to merge 14 commits into
maderix:mainfrom
dev-erik:main

Commits on Mar 3, 2026

[fix] M1/M2/M3 MIL syntax compatibility (upstream PR #6): use program(1.0), ios16 target, tensor types across 18 files

[fix] Token sampling underflow fix (upstream PR #17): prevent size_t wraparound on short datasets in both train_large variants

[fix] Dashboard sudo hang fix (upstream PR #20): prevent blocking when password is required for powermetrics

[fix] Security hardening (upstream PRs #5, #7): stack-protector-strong, format-security flags, NULL guards on ane_compile/fread/fopen, tokenize.py input validation

[feat] ANE ChainingRequest API prototype: baseline measurement for multi-kernel pipelining without recompile overhead

[docs] Developer documentation: architecture diagrams, complete API reference, benchmark guide, M4 Max results, security audit report

[feat] Benchmark runner and mlpackage generator: run_benchmarks.sh for full test suite, gen_mlpackages.py for CoreML model generation

[feat] Community benchmark system: standardized JSON output, auto-submit to dashboard, aggregation script, M4 Max reference result

[docs] Community fork README, CONTRIBUTING guide, issue templates, gitignore: rewritten README with quickstart, env vars, benchmark instructions, dashboard link

[feat] Merge upstream: dynamic weight training, CLI fixes, dashboard v2

[feat] Merge upstream PRs #21, #23, #26: NEON-optimized training (train_opt), double-buffered async ANE training (train_double_buffer), Qwen2.5-0.5B LLM inference (inference/). Added get_path() env…

[feat] Inference server mode: keep ANE kernels loaded between prompts (stdin loop + Unix socket server). Subsequent queries respond in ~0.5s instead of ~6s. run.py auto-connects to socket server wh…

[feat] Optimize inference: vectorize ops (NEON/vDSP), gate debug output, skip unused ANE compilation, add round-trip benchmark timing, pure C HTTP API with tokenizer

[feat][gpu] Q4 quantization, Metal GPU shaders, ANE kernel fusion, memory safety