inference engine for rdna3. codex slop.

[rishavsanjay]

Radeon Forge

Oracle-guided, benchmark-tuned local inference engineering agent for AMD Radeon / ROCm.

Implemented

• workload/objective/correctness/resource contracts
• hard numerical, no-spill, stability, and resource gates
• append-only experiment evidence ledger and local memory
• successive-halving autotuner
• AMD kernel metadata parsing
• gfx1100 HIP compilation backend
• executable JSON benchmark protocol
• native gfx1100 assembly-GEMM family using alternate valid VOPD/FMAC schedules and occupancy limits
• gfx1100 workload adapter with tinygrad numerical oracle
• provenance-tagged corpus of tinygrad fused-kernel patterns
• deterministic local retrieval with file-and-line citations
• loopback-only OpenAI-compatible planner client; remote planner endpoints are rejected
• deny-by-default scoped permission controller
• multi-turn proposal, revision, approval, authorization, completion, and failure state machine
• unit coverage for metadata, hard gates, permissions, command protocol, local endpoint restrictions, retrieval, and approval state

Honest current boundary

• No W7900 speed result is claimed yet.
• extra/llama_kernels implementations currently compiled for gfx950 are architectural patterns until individually compiled and validated on gfx1100.
• The native extra/gemm/amd_asm_matmul.py path is the first real gfx1100 hardware family.
• The GEMM family validates the search, permission, correctness, and evidence infrastructure; the final submission metric remains P95 end-to-end private-agent task latency under quality, correctness, privacy, memory, and stability constraints.
• No local planner model/runtime has yet been selected and benchmarked on the W7900.
• The interactive terminal/web surface, frozen private-agent workload, held-out evaluation, and Radeon Cloud evidence remain in progress.

Validation status

The initial isolated core unit suite passed. A fresh repository clone/full-tree run could not be performed from the current execution environment because outbound GitHub DNS resolution is unavailable. No GitHub Actions workflow was triggered for this fork PR. W7900 compilation, correctness, and benchmark evidence must be produced on the assigned Radeon Cloud host before this draft is made ready.

See extra/radeon_forge/README.md and extra/radeon_forge/KERNEL_CORPUS.md for architecture, commands, metrics, provenance, and scope.