repack: compile cache model card
Browse files- compile_cache_model_card.md +68 -0
compile_cache_model_card.md
ADDED
|
@@ -0,0 +1,68 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Compile Cache — MI300X
|
| 2 |
+
|
| 3 |
+
Generated: 2026-04-12T04:49:32Z (by repack_and_upload_compile_cache.sh)
|
| 4 |
+
|
| 5 |
+
## Purpose
|
| 6 |
+
Records the exact configuration under which the compile cache (inductor + Triton) was built.
|
| 7 |
+
If unexpected AUTOTUNE blocks appear on a later run, compare the current environment
|
| 8 |
+
against these values to pinpoint the cache-key mismatch.
|
| 9 |
+
|
| 10 |
+
## Hardware
|
| 11 |
+
- GPU: AMD Radeon Graphics
|
| 12 |
+
- GPU count: 1
|
| 13 |
+
- Architecture: gfx942 (AMD MI300X)
|
| 14 |
+
|
| 15 |
+
## Software Versions
|
| 16 |
+
| Package | Version |
|
| 17 |
+
|---------|---------|
|
| 18 |
+
| Torch | 2.6.0+rocm6.2.4 |
|
| 19 |
+
| HIP / ROCm | 6.2.41134-65d174c3e |
|
| 20 |
+
| Triton | 3.1.0 |
|
| 21 |
+
| ck4inductor | 0.0.11 |
|
| 22 |
+
| Python | 3.10.13 |
|
| 23 |
+
|
| 24 |
+
## Compile Configuration
|
| 25 |
+
| Setting | Value |
|
| 26 |
+
|---------|-------|
|
| 27 |
+
| Compile backend | inductor |
|
| 28 |
+
| Compile mode | max-autotune-no-cudagraphs |
|
| 29 |
+
| GEMM backends | ATEN,TRITON,CK |
|
| 30 |
+
| TORCHINDUCTOR\_CK\_DIR | `N/A` |
|
| 31 |
+
| TORCHINDUCTOR\_CACHE\_DIR | `/workspace/inductor_cache` |
|
| 32 |
+
| TRITON\_CACHE\_DIR | `/workspace/triton_cache` |
|
| 33 |
+
| TRITON\_MAX\_BLOCK\_SIZE | 64 |
|
| 34 |
+
| TORCHINDUCTOR\_COMPILE\_THREADS | 4 |
|
| 35 |
+
|
| 36 |
+
## Model
|
| 37 |
+
- Architecture: EnhancedDSRModel (~1.743B total params)
|
| 38 |
+
- Trainable params: fp32 (~337M — new alignment modules)
|
| 39 |
+
- Frozen backbone: bf16 (~1.406B — RWKV-7 1.5B World)
|
| 40 |
+
- Sequence length: 1024 tokens
|
| 41 |
+
|
| 42 |
+
## Cache Statistics (at time of pack)
|
| 43 |
+
| Cache | Files |
|
| 44 |
+
|-------|-------|
|
| 45 |
+
| inductor\_cache | 7654 |
|
| 46 |
+
| triton\_cache | 59993 |
|
| 47 |
+
| tar.gz size | 443 MB |
|
| 48 |
+
|
| 49 |
+
## HuggingFace
|
| 50 |
+
- Repo: `RaninduH/torchCompiledModels`
|
| 51 |
+
- File: `my_compile_cache.tar.gz`
|
| 52 |
+
- Model card: `compile_cache_model_card.md`
|
| 53 |
+
|
| 54 |
+
## Autotune Cache Key
|
| 55 |
+
The inductor autotune winner cache key is a **hash of the full candidate set**
|
| 56 |
+
including which backends are enabled. Adding/removing from
|
| 57 |
+
`TORCHINDUCTOR_MAX_AUTOTUNE_GEMM_BACKENDS` (currently `ATEN,TRITON,CK`) will
|
| 58 |
+
invalidate all cached winners and force a full AUTOTUNE re-run.
|
| 59 |
+
|
| 60 |
+
## Recompilation Trigger Checklist
|
| 61 |
+
If AUTOTUNE blocks appear on a pod that should have cache hits, verify:
|
| 62 |
+
1. `TORCHINDUCTOR_MAX_AUTOTUNE_GEMM_BACKENDS` == `ATEN,TRITON,CK`
|
| 63 |
+
2. `TORCHINDUCTOR_CK_DIR` points to valid CK root: `N/A`
|
| 64 |
+
3. `library/src` symlink inside ck4inductor package exists (needed for gemm_universal kernel enumeration)
|
| 65 |
+
4. `TORCHINDUCTOR_CACHE_DIR` == `/workspace/inductor_cache` (with extracted cache files present)
|
| 66 |
+
5. `TRITON_CACHE_DIR` == `/workspace/triton_cache` (must NOT be `~/.triton` — that is wiped on restart)
|
| 67 |
+
6. Torch version matches: `2.6.0+rocm6.2.4` — different minor/patch versions can invalidate kernel hashes
|
| 68 |
+
7. Triton version matches: `3.1.0`
|