RaninduH commited on
Commit
554c358
·
1 Parent(s): e7c5b9f

compile cache model card update

Browse files
Files changed (1) hide show
  1. compile_cache_model_card.md +10 -10
compile_cache_model_card.md CHANGED
@@ -1,6 +1,6 @@
1
- # Compile Cache — MI300X
2
 
3
- Generated: 2026-05-01T20:24:06Z
4
 
5
  ## Purpose
6
  Records the exact configuration under which the compile cache (inductor + Triton) was built.
@@ -17,7 +17,7 @@ against these values to pinpoint the cache-key mismatch.
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 | 3.10.12 |
23
 
@@ -35,15 +35,15 @@ against these values to pinpoint the cache-key mismatch.
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
43
  | Cache | Files |
44
  |-------|-------|
45
- | inductor\_cache | 997 |
46
- | triton\_cache | 7227 |
47
  | tar.gz size | 110.3 MB |
48
 
49
  ## Autotune Cache Key
@@ -58,6 +58,6 @@ If AUTOTUNE blocks appear on a pod that should have cache hits, verify:
58
  2. `TORCHINDUCTOR_CK_DIR` points to valid CK root: `/home/hotaisle/.local/lib/python3.10/site-packages/ck4inductor`
59
  3. `library/src` symlink inside ck4inductor package exists (needed for gemm_universal kernel enumeration)
60
  4. `TORCHINDUCTOR_CACHE_DIR` == `/tmp/ml_cache/inductor_cache` (with extracted cache files present)
61
- 5. `TRITON_CACHE_DIR` == `/tmp/ml_cache/triton_cache` (must NOT be `~/.triton` — that is wiped on restart)
62
- 6. Torch version matches: `2.6.0+rocm6.2.4` — different minor/patch versions can invalidate kernel hashes
63
- 7. Triton version matches: `3.1.0`
 
1
+ # Compile Cache MI300X
2
 
3
+ Generated: 2026-05-07T17:26:34Z
4
 
5
  ## Purpose
6
  Records the exact configuration under which the compile cache (inductor + Triton) was built.
 
17
  |---------|---------|
18
  | Torch | 2.6.0+rocm6.2.4 |
19
  | HIP / ROCm | 6.2.41134-65d174c3e |
20
+ | Triton | 3.2.0 |
21
  | ck4inductor | 0.0.11 |
22
  | Python | 3.10.12 |
23
 
 
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
43
  | Cache | Files |
44
  |-------|-------|
45
+ | inductor\_cache | 1019 |
46
+ | triton\_cache | 8202 |
47
  | tar.gz size | 110.3 MB |
48
 
49
  ## Autotune Cache Key
 
58
  2. `TORCHINDUCTOR_CK_DIR` points to valid CK root: `/home/hotaisle/.local/lib/python3.10/site-packages/ck4inductor`
59
  3. `library/src` symlink inside ck4inductor package exists (needed for gemm_universal kernel enumeration)
60
  4. `TORCHINDUCTOR_CACHE_DIR` == `/tmp/ml_cache/inductor_cache` (with extracted cache files present)
61
+ 5. `TRITON_CACHE_DIR` == `/tmp/ml_cache/triton_cache` (must NOT be `~/.triton` that is wiped on restart)
62
+ 6. Torch version matches: `2.6.0+rocm6.2.4` different minor/patch versions can invalidate kernel hashes
63
+ 7. Triton version matches: `3.2.0`