Upload code/1028_run/README.md with huggingface_hub

code/1028_run/README.md

@@ -125,26 +125,30 @@ Merges episode npz + videos, renumbers indices, writes `collect_meta.json`.
 
 ### Speed comparison
 
-| Setting | it/s | Time for 130 epochs |
-|---------|------|---------------------|
-| `1027_run`: batch=64, workers=4, FP32 | ~1.1 | ~10h |
-| `1028_run`: batch=256, workers=16, AMP (FP16) | ~3–4 (est.) | ~2.5–3h |
+| Setting | it/s | Batches/epoch | Time for 130 epochs |
+|---------|------|---------------|---------------------|
+| `1027_run`: batch=64, workers=4, FP32 | ~1.1 | 324 | ~10h |
+| `1028_run`: batch=256, workers=16, AMP (FP16) | ~2.5 | 39 | **~78 min** |
+
+**Measured speedup: 7.7× on training** (tested on battery_v1, L40s GPU).
 
 **Speedup sources:**
-1. **batch_size 64→256**: more GPU utilization per step (4×)
-2. **num_workers 4→16**: parallel zarr.zip loading (4×, capped by I/O)
-3. **AMP mixed precision**: FP16 forward/backward, ~1.5–2× on L40s tensor cores
+1. **batch_size 64→256**: 4× fewer batches per epoch
+2. **AMP mixed precision** (`+training.use_amp=True`): FP16 forward/backward via `torch.cuda.amp.GradScaler` + `autocast`, ~2× throughput on L40s tensor cores
+3. **num_workers 4→16**: parallel zarr.zip cache loading
 
-AMP is implemented in `train_equi_workspace.py` via `torch.cuda.amp.GradScaler` + `autocast`. Disable with `training.use_amp=False`.
+AMP is patched into `train_equi_workspace.py`. Disable with `+training.use_amp=False`.
 
 ### Collection speedup
 
 | Setting | Total collect time (50 demos) |
 |---------|-------------------------------|
-| `1027_run`: 2x parallel (GPU 0+2) | ~5–8h (with GPU crash risk) |
-| `1028_run`: 4x parallel (GPU 0/1/2/3) | ~2.5–4h (est.) |
+| `1027_run`: 2x parallel (GPU 0+2) | ~5–8h |
+| `1028_run`: 4x parallel (GPU 0/1/2/3) | **~2.5h** |
+
+**Measured: battery_v1 50 demos in 2h 26min** (4x parallel, 15+15+15+20 shards).
 
-~300s per demo × 50 demos / 4 GPUs = ~62 min minimum. With Isaac Sim init overhead ~4h total.
+~270s per demo × 50 demos / 4 GPUs = ~56 min minimum. With Isaac Sim init overhead (~3 min/proc × 4 staggered) total ~2.5h.
 
 ---