Standardize card: usage-first
Browse files
CARD.md
CHANGED
|
@@ -60,7 +60,7 @@ boundaries are arbitrary; any partition composes exactly.
|
|
| 60 |
| `det_sum(x, dim)` | permutation-invariant reduction |
|
| 61 |
| `DetLinear` | linear module, deterministic forward and backward |
|
| 62 |
|
| 63 |
-
##
|
| 64 |
|
| 65 |
- Sharding and grouping invariance: `det_gemm` over full K equals the
|
| 66 |
`det_combine`/`det_finalize` composition of 2-, 4-, 8-, and 16-way
|
|
@@ -77,7 +77,7 @@ boundaries are arbitrary; any partition composes exactly.
|
|
| 77 |
- Determinism: repeated runs, and a training run whose forward is computed
|
| 78 |
by sharded composition, are bitwise identical.
|
| 79 |
|
| 80 |
-
## Scope
|
| 81 |
|
| 82 |
The kernel accumulates on fp64 scalar cores, not tensor cores; throughput
|
| 83 |
is measured and reported below, and the intended regimes are debugging
|
|
|
|
| 60 |
| `det_sum(x, dim)` | permutation-invariant reduction |
|
| 61 |
| `DetLinear` | linear module, deterministic forward and backward |
|
| 62 |
|
| 63 |
+
## Correctness
|
| 64 |
|
| 65 |
- Sharding and grouping invariance: `det_gemm` over full K equals the
|
| 66 |
`det_combine`/`det_finalize` composition of 2-, 4-, 8-, and 16-way
|
|
|
|
| 77 |
- Determinism: repeated runs, and a training run whose forward is computed
|
| 78 |
by sharded composition, are bitwise identical.
|
| 79 |
|
| 80 |
+
## Scope
|
| 81 |
|
| 82 |
The kernel accumulates on fp64 scalar cores, not tensor cores; throughput
|
| 83 |
is measured and reported below, and the intended regimes are debugging
|
README.md
CHANGED
|
@@ -60,7 +60,7 @@ boundaries are arbitrary; any partition composes exactly.
|
|
| 60 |
| `det_sum(x, dim)` | permutation-invariant reduction |
|
| 61 |
| `DetLinear` | linear module, deterministic forward and backward |
|
| 62 |
|
| 63 |
-
##
|
| 64 |
|
| 65 |
- Sharding and grouping invariance: `det_gemm` over full K equals the
|
| 66 |
`det_combine`/`det_finalize` composition of 2-, 4-, 8-, and 16-way
|
|
@@ -77,7 +77,7 @@ boundaries are arbitrary; any partition composes exactly.
|
|
| 77 |
- Determinism: repeated runs, and a training run whose forward is computed
|
| 78 |
by sharded composition, are bitwise identical.
|
| 79 |
|
| 80 |
-
## Scope
|
| 81 |
|
| 82 |
The kernel accumulates on fp64 scalar cores, not tensor cores; throughput
|
| 83 |
is measured and reported below, and the intended regimes are debugging
|
|
|
|
| 60 |
| `det_sum(x, dim)` | permutation-invariant reduction |
|
| 61 |
| `DetLinear` | linear module, deterministic forward and backward |
|
| 62 |
|
| 63 |
+
## Correctness
|
| 64 |
|
| 65 |
- Sharding and grouping invariance: `det_gemm` over full K equals the
|
| 66 |
`det_combine`/`det_finalize` composition of 2-, 4-, 8-, and 16-way
|
|
|
|
| 77 |
- Determinism: repeated runs, and a training run whose forward is computed
|
| 78 |
by sharded composition, are bitwise identical.
|
| 79 |
|
| 80 |
+
## Scope
|
| 81 |
|
| 82 |
The kernel accumulates on fp64 scalar cores, not tensor cores; throughput
|
| 83 |
is measured and reported below, and the intended regimes are debugging
|