Fetching metadata from the HF Docker repository...
-
benchmark: MI300X results — 603 TFLOPS FP16, 4.6 TB/s bandwidth
-
fix: compatibility with PyTorch 2.9 (total_mem -> total_memory)
-
feat: AST-level Python analysis + MI300X benchmark suite
- AST Transformer: Uses Python's ast module for tree-level analysis
- Detects device assignments, CUDA call sites, embedded kernels
- Identifies env mutations and inline kernel strings structurally
- Not regex — real compiler-level understanding of code structure
- MI300X Benchmark Suite: Ready to run on AMD Developer Cloud
- Device info, memory bandwidth, and GEMM TFLOPS tests
- Saves JSON results as proof of AMD GPU usage
- Wired AST findings into analyzer, API, and orchestrator
