Skip to content

AGENTS.md

Latest commit

 

History

History
26 lines (21 loc) · 2.49 KB

AGENTS.md

File metadata and controls

26 lines (21 loc) · 2.49 KB

EriDiffusion Trainer – Agent Guardrails

Hard Rules

  • GPU BF16 end-to-end. With STRICT_BF16=1 there is zero tolerance for FP32 device tensors or host staging. Do not revive tensor.to_vec() / to_vec_f32() or any CPU mirror of BF16 buffers.
  • Broadcast/repeat/index stay on device. The helpers are already GPU-only; do not sneak in to_vec() or host iterators “for debugging”. Use the FlameCore kernels or open a blocker.
  • Keep runs tight. Use focused tests (cargo test -p training --features cuda,bf16_u16 -- --nocapture name_of_case). If a loop exceeds 10 minutes, you’re running the wrong suite.
  • No surprise memory spikes. Chunk parity dumps, enable tiling helpers, and watch device_fp32_alloc_bytes_during_infer—any non-zero value is a regression that must be fixed before merging.

Safe Workflow

  • Compose existing FlameCore ops; only request new kernels after proving the gap. Land changes in FlameCore → mirror → trainer.
  • Keep SDXL/SD3 worklists (docs/FLAME_SD35_*) current. Every STRICT change or arena tweak should be recorded before you close the PR.
  • Feature-gate diagnostics (cfg!(feature = "strict_bf16")). No ad-hoc env guards.

Validation Checklist

  • cargo fmt, cargo clippy --all-targets -- -D warnings, cargo test --all-targets.
  • cargo test --manifest-path crates/models/Cargo.toml --features cuda,bf16_u16 -- --nocapture sdxl_dtype_guard with STRICT_BF16=1.
  • Run scripts/check_sdxl_dtype_guard.sh (or equivalent) and log VRAM/time deltas; >0.5 % regression is a blocker.
  • Record allocator metrics (ARENA_TEMP_SOFTCAP_MB, TILE_VRAM_FRACTION, PARITY_MAX_DUMP_MB) in the test log whenever you tweak them so the next agent can reproduce your settings.

SDPA BF16 Tiling Plan (in progress)

  1. Remove FP32 fallback – eliminate the SDPA forward_f32 path under STRICT_BF16 so oversize workspace requests error out instead of upcasting.
  2. Runtime tiler – add a VRAM-aware planner (tile order Q → H → Dv) that allocates a single ring-buffer workspace and retries with smaller tiles on OOM.
  3. Integration & tests – wire the planner into attention, expose counters/tripwires, and add CI contract tests (tight/loose budgets, strict BF16 parity, dtype checks).

Escalation

If you must temporarily disable STRICT enforcement to unblock yourself, stop and escalate in docs/FLAME_SD35_BLOCKER_PLAN.md. We do not accept “temporary” BF16 relaxations—document the reproduction and wait for approval instead of merging hacks.