hipfire

Port hipfire compute kernels to a new RDNA / CDNA architecture (gfx1201/gfx1200/gfx94x/gfx1150/etc.). Use when adding support for a new GPU arch, fixing arch-specific kernel codegen failures (e.g. "Cannot select intrinsic %llvm.amdgcn.wmma..."), or refactoring dispatch.rs's arch-conditional branches. Captures the WMMA operand-shape matrix, builtin name table per arch, dispatch routing convention, validation procedure (channel-test / coherence-gate / speed-gate), contributor onboarding workflow, and known correctness traps. Triggers on phrases like "port to gfx12", "9070 XT support", "R9700 support", "WMMA gfx12", "Cannot select intrinsic wmma", "amdgcn.wmma", "new arch port", "cross-arch kernel".

Cleanly stop, free the port, and restart `hipfire serve`. Use when serve "Failed to start (port in use)", a stale daemon holds VRAM, an os-error-2/JSON-parse pre-warm crash left a zombie singleton, or you just want a guaranteed-fresh daemon. Kills both bun CLI serve and the spawned target/release/examples/daemon, reaps stale ~/.hipfire/daemon.pid + serve.pid, fuser-frees the port, then relaunches.

Optimize hipfire HIP/compute kernels — pick a tuning lever (multi-row, K-tile depth, prefetch, wave-size port, WMMA/MFMA, fused projections, ISA flags) and validate the win across the supported RDNA arch matrix. Use when you've identified a hot kernel, want to land a real perf win, and need to NOT regress on archs you don't have hardware for. Codifies the methodology from this repo's actual perf history — wave64 CDNA3 port (commit 4105035, 2× decode), nontemporal-load revert (34eb024, -13% caught only by clean-baseline bisect), gfx12 WMMA port (PR

Use for hipfire quant calibration, imatrix-driven experiments, KLD/PPL quality evaluation, k-map/format selection, MQ/HFQ/HFP/MFP tradeoff work, ParoQuant-style weight transform planning, and KV policy planning. Use when deciding whether a calibrated model candidate should be promoted, rejected, packaged, or sent through Atlas for AR/DFlash perf validation.

Triage and repair hipfire runtime failures such as daemon hangs, stale serve.pid, port 11435 conflicts, ROCm include-path problems, missing precompiled kernels, VRAM OOM, kernel JIT failures, and multi-turn recall regressions. Use after diagnostics identify a likely runtime issue or when the user asks to fix a broken hipfire serve/run flow.

Run and interpret hipfire GPU diagnostics for ROCm/HIP bring-up, missing kernels, test_kernels failures, inference smoke failures, and install/runtime environment problems. Use when a user asks to diagnose hipfire, check GPU readiness, run baseline tests, or explain diagnostic output.

Use Kernel Atlas to collect phase-aware hipfire measurements and render ISA Fit View visualizations for AMD GPU kernels, quant formats, and architectures. Use when a user asks how MQ/HFQ/HFP/Q8 quants occupy hardware, asks for an ASCII ISA visualization, wants to compare gfx1010/gfx1030/gfx11/gfx12 kernel fit, or wants an agent-readable "left on table" summary from Atlas rows.

Guide a tester through hipfire bring-up, smoke tests, DFlash opt-in checks, MQ format sanity, multi-turn recall, CLI surface checks, and benchmark reporting on AMD RDNA/CDNA GPUs. Use when the user wants a standard test matrix or upstream-ready tester report.

Use when porting a hipfire feature/fix branch authored against pre-0.1.20 master onto post-modular master. Walks through the engine→hipfire-runtime + per-arch-crate split mechanically, then surfaces semantic conflicts that need human judgment.

Extract VGPR/SGPR/LDS/spill counts and AMDGPU notes from a compiled HIP kernel `.hsaco` for any AMD GPU arch (CDNA wave64 — gfx906/908/90a/942 — and RDNA wave32 — gfx1010 through gfx12). Lets you compute theoretical occupancy and identify register-pressure / LDS-pressure constraints.