compile_cache_model_card.md

Compile Cache — MI300X

Generated: 2026-07-21T05:53:03Z

Purpose

Records the exact configuration under which the compile cache (inductor + Triton) was built. If unexpected AUTOTUNE blocks appear on a later run, compare the current environment against these values to pinpoint the cache-key mismatch.

Hardware

• GPU: NVIDIA A100 80GB PCIe
• GPU count (world_size): 1 (1)
• Architecture: gfx942 (AMD MI300X)

Software Versions

Package	Version
Torch	2.7.1+cu126
HIP / ROCm	N/A
Triton	3.3.1
ck4inductor	N/A
Python	3.11.15

Compile Configuration

Setting	Value
Compile backend	inductor
Compile mode	N/A
GEMM backends	ATEN
TORCHINDUCTOR_CK_DIR	N/A
TORCHINDUCTOR_CACHE_DIR	/workspace/inductor_cache
TRITON_CACHE_DIR	/workspace/triton_cache
TRITON_MAX_BLOCK_SIZE	N/A
TORCHINDUCTOR_COMPILE_THREADS	1

Model

• Architecture: EnhancedDSRModel (~1.743B total params)
• Trainable params: fp32 (~337M — new alignment modules)
• Frozen backbone: bf16 (~1.406B — RWKV-7 1.5B World)
• Sequence length: 1024 tokens

Cache Statistics

Cache	Files
inductor_cache	615
triton_cache	6170
tar.zst size	91.6 MB

Autotune Cache Key

The inductor autotune winner cache key is a hash of the full candidate set including which backends are enabled. Adding/removing from TORCHINDUCTOR_MAX_AUTOTUNE_GEMM_BACKENDS (currently ATEN) will invalidate all cached winners and force a full AUTOTUNE re-run.

Recompilation Trigger Checklist

If AUTOTUNE blocks appear on a pod that should have cache hits, verify:

1. TORCHINDUCTOR_MAX_AUTOTUNE_GEMM_BACKENDS == ATEN
2. TORCHINDUCTOR_CK_DIR points to valid CK root: N/A
3. library/src symlink inside ck4inductor package exists (needed for gemm_universal kernel enumeration)
4. TORCHINDUCTOR_CACHE_DIR == /workspace/inductor_cache (with extracted cache files present)
5. TRITON_CACHE_DIR == /workspace/triton_cache (must NOT be ~/.triton — that is wiped on restart)
6. Torch version matches: 2.7.1+cu126 — different minor/patch versions can invalidate kernel hashes
7. Triton version matches: 3.3.1