Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# commitmoe-qwen35-fp8-layers-v2
|
| 2 |
+
|
| 3 |
+
Per-layer columnar streams for Qwen3.5-35B-A3B-FP8 expert-routing prediction.
|
| 4 |
+
|
| 5 |
+
Rebuilt from the raw collector shards so that **all 40 layers** are available in
|
| 6 |
+
the columnar form, not only the L6 lead-2 slice.
|
| 7 |
+
|
| 8 |
+
- `layers_v1/chunk_10k/` — corpus prompts 0-9,999, rebuilt from `raw` shards
|
| 9 |
+
(source list taken from `index_10k.pt`, 3,126 shards; note the collector's
|
| 10 |
+
`manifest.jsonl` is stale at 938 shards and must not be used as the source list)
|
| 11 |
+
|
| 12 |
+
Layout per stream: `{hidden,logits}/layer=NN/part_*.safetensors`, plus `labels/`.
|
| 13 |
+
Columns carry their own `prompt_idx` / `tok`, so joins key on values, never on
|
| 14 |
+
row position.
|
| 15 |
+
|
| 16 |
+
Existing repos are untouched; this is an additive republish.
|