Spaces:
Sleeping
Sleeping
Commit History
-
feat: Aether v2 with RoPE fix -- PyTorch vs Aether side by side
TaylorRoot cause found and fixed: RoPE was using NeoX style (half-split pairing) but SmolLM2/LLaMA needs adjacent pairing (x[i], x[i+1]). This corrupted every attention head in every layer. Also includes: - WASM SIMD for matVec, rmsNorm, softmax, fusedSiluMul - JS fallback for LM head (>100MB) and attention - Type-aware GGUF dequant (Q8_0 with type field, not byte guessing) - Parallel execution (whichever engine finishes first shows first) - Streaming results via Gradio generatorfcac5c7
-
revert: back to working PyTorch for both models
TaylorAether inference has deeper bugs (tokenizer, weight layout, or numerical accumulation). Debug separately in edge-workers. Both models on PyTorch/transformers -- base vs Buleyean comparison that was working before. Streaming results so faster one shows first.0c8d249 -
perf: add WASM SIMD kernels + use Q4_K_M for faster inference
TaylorMajor changes: - Bundle simd-kernels-standalone.wasm (14KB) from Aether - WASM SIMD matVec, rmsNorm, softmax, fusedSiluMul, flashAttention - Switch from Q8_0 (360MB) to Q4_K_M (210MB, half the work) - Reduce max_tokens to 50 for snappier demo - Proper Q4_K dequantization with getScaleMinK4 - Falls back to JS if WASM SIMD unavailable7336fde -
feat: PyTorch vs Aether side-by-side inference
TaylorLeft column: base SmolLM2-360M on PyTorch CPU (the standard) Right column: Buleyean SmolLM2-360M on Aether (our engine) Aether runs the entire inference pipeline in pure TypeScript + JS: GGUF parse -> dequant -> matVec -> RoPE -> SwiGLU -> sampling Zero external ML dependencies. Shows timing for both. Docker SDK with Python 3.11 + Node.js 20.c92238b -
fix: switch to Gradio SDK, drop Docker build
TaylorDocker build compiles llama-cpp-python from source and times out on cpu-basic. Gradio SDK uses pre-built wheels. Kept the polished app.py and aeon.css untouched.854ee42 -
Upload folder using huggingface_hub
Verified signature868752a -
Upload folder using huggingface_hub
Verified signatureb4d8480 -
fix: switch from Docker to Gradio SDK to fix build timeout
Taylorllama-cpp-python compiles from source in Docker, timing out on cpu-basic. Gradio SDK builder has pre-built wheels.103b9da -
Upload folder using huggingface_hub
Verified signatured77d3cd -
Upload folder using huggingface_hub
Verified signatureb56d971 -
Upload folder using huggingface_hub
Verified signature940f26c -
Upload folder using huggingface_hub
Verified signature75a4d80 -
Upload folder using huggingface_hub
Verified signature6cf8d36
