compile cache model card update

phase02_cache/compile_cache_model_card_only_decoder_phase02.md

@@ -1,6 +1,6 @@
 # Compile Cache — MI300X
 
-Generated: 2026-07-12T17:09:31Z
+Generated: 2026-07-13T14:40:00Z
 
 ## Purpose
 Records the exact configuration under which the compile cache (inductor + Triton) was built.
@@ -8,29 +8,29 @@ If unexpected AUTOTUNE blocks appear on a later run, compare the current environ
 against these values to pinpoint the cache-key mismatch.
 
 ## Hardware
-- GPU: AMD Instinct MI300X VF
+- GPU: NVIDIA A100 80GB PCIe
 - GPU count (world_size): 1 (1)
 - Architecture: gfx942 (AMD MI300X)
 
 ## Software Versions
 | Package | Version |
 |---------|---------|
-| Torch   | 2.11.0+rocm7.2 |
-| HIP / ROCm | 7.2.26015 |
-| Triton  | 3.7.0 |
+| Torch   | 2.11.0+cu128 |
+| HIP / ROCm | N/A |
+| Triton  | 3.6.0 |
 | ck4inductor | N/A |
-| Python  | 3.11.15 |
+| Python  | 3.11.10 |
 
 ## Compile Configuration
 | Setting | Value |
 |---------|-------|
 | Compile backend | inductor |
 | Compile mode | max-autotune-no-cudagraphs |
-| GEMM backends | ATEN |
+| GEMM backends | ATEN,TRITON |
 | TORCHINDUCTOR\_CK\_DIR | `N/A` |
-| TORCHINDUCTOR\_CACHE\_DIR | `/dev/shm/inductor_cache` |
-| TRITON\_CACHE\_DIR | `/dev/shm/triton_cache` |
-| TRITON\_MAX\_BLOCK\_SIZE | 32 |
+| TORCHINDUCTOR\_CACHE\_DIR | `/tmp/ml_cache/inductor_cache` |
+| TRITON\_CACHE\_DIR | `/tmp/ml_cache/triton_cache` |
+| TRITON\_MAX\_BLOCK\_SIZE | N/A |
 | TORCHINDUCTOR\_COMPILE\_THREADS | 1 |
 
 ## Model
@@ -42,22 +42,22 @@ against these values to pinpoint the cache-key mismatch.
 ## Cache Statistics
 | Cache | Files |
 |-------|-------|
-| inductor\_cache | 1028 |
-| triton\_cache | 36863 |
-| tar.zst size | 3813.3 MB |
+| inductor\_cache | 1027 |
+| triton\_cache | 26763 |
+| tar.zst size | 902.2 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
+`TORCHINDUCTOR_MAX_AUTOTUNE_GEMM_BACKENDS` (currently `ATEN,TRITON`) 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`
+1. `TORCHINDUCTOR_MAX_AUTOTUNE_GEMM_BACKENDS` == `ATEN,TRITON`
 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` == `/dev/shm/inductor_cache` (with extracted cache files present)
-5. `TRITON_CACHE_DIR` == `/dev/shm/triton_cache` (must NOT be `~/.triton` — that is wiped on restart)
-6. Torch version matches: `2.11.0+rocm7.2` — different minor/patch versions can invalidate kernel hashes
-7. Triton version matches: `3.7.0`
+4. `TORCHINDUCTOR_CACHE_DIR` == `/tmp/ml_cache/inductor_cache` (with extracted cache files present)
+5. `TRITON_CACHE_DIR` == `/tmp/ml_cache/triton_cache` (must NOT be `~/.triton` — that is wiped on restart)
+6. Torch version matches: `2.11.0+cu128` — different minor/patch versions can invalidate kernel hashes
+7. Triton version matches: `3.6.0`