REPOMIND

REPOMIND is an open-source repo-scale coding agent that ingests an entire git repository at 256K context on a single AMD MI300X and reasons across the whole codebase with multi-step tool use. MIT licensed. Built for the AMD Developer Hackathon 2026. Why MI300X: Qwen3-Coder-Next-FP8 weights (~80 GB) + 256K KV cache @ FP8 (~38 GB) + activations (~25 GB) ≈ 143 GB total. H100 80 GB cannot accommodate this on a single card by VRAM accounting; MI300X 192 GB has the headroom. AMD Day-0 ROCm 7 post (Feb 2026) positioned this exact workload — REPOMIND is the first open-source proof shipped. Verified on real hardware (2026-05-05/06, 124-min stress test, 2 sessions, $4.12 total): • Memory: 77.29 GiB weights + 94.58 GiB KV cache + 92% VRAM peak. max_model_len=262144 confirmed. • Concurrency (24-cell matrix, default Triton): 31/31 at 8K, 16K, 32K AND 64K. 25/31 at 128K. 6-8 at 256K within 15-min window. • Long-context: 3/3 needle-in-haystack pass at 200K. Model recovers sentinel function + constant from middle of 199K-token prompt. • End-to-end repo Q&A: 9/9 correct across REPOMIND self (68K), Flask (408K), pytorch/vision (1.3M tokens — 5× larger than any context window). Priority-aware chunker fits to 180K. • Tuning A/B: tried --attention-backend ROCM_AITER_FA. Throughput 2-4× higher BUT output degenerates to repeating punctuation on FP8 KV cache (137/144 cells broken). Default Triton production-safe; filed for AMD upstream investigation. Stack: Qwen3-Coder-Next-FP8 + vLLM 0.17.1 + ROCm 7.2 + SC-TIR agent loop + 5 tools (read_file, grep_codebase, execute_code, run_tests, git_log). Market unlock: regulated industries (banks, defense, pharma, Apple iOS) cannot legally use SaaS coding agents. $1.99/hr cloud, 70-140 dev seats per MI300X, breaks even vs Cursor in 3-6 months. REPOMIND is the first open-source option for compliance-locked enterprises. Evidence pack: 7 JSON results + 5 plots + raw outputs + rocm-smi + benchmark scripts, all reproducible.