You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Copy file name to clipboardExpand all lines: benchmarks/float8/training/README.md
+3-3
Original file line number
Diff line number
Diff line change
@@ -1,17 +1,17 @@
1
1
# Float8 training benchmarking
2
2
3
-
The `float8_training_benchmark.sh` script in this directory can be used to launch a Llama3 8b training run with [torchtitan](https://github.com/pytorch/torchtitan) training run, and parse the logs to calculate the median tokens/sec and peak memory usage for you.
3
+
The `torchtitan_benchmark.sh` script in this directory can be used to launch a Llama3 8b training run with [torchtitan](https://github.com/pytorch/torchtitan) training run, and parse the logs to calculate the median tokens/sec and peak memory usage for you.
Training parameters can be configured via environment variables.
10
10
11
11
- Required:
12
12
-`TORCHTITAN_ROOT`: Root directory of torchtitan in your local filesystem
13
13
- Optional:
14
-
-`FLOAT8_RECIPE_WITH_BEST_SETTINGS`: "rowwise" or "tensorwise". Applies float8 training with the specified scaling recipe, as well as additional training configs which are optimal for that scaling recipe. See `float8_training_benchmark.sh` for more details.
14
+
-`FLOAT8_RECIPE_WITH_BEST_SETTINGS`: "rowwise" or "tensorwise". Applies float8 training with the specified scaling recipe, as well as additional training configs which are optimal for that scaling recipe. See `torchtitan_benchmark.sh` for more details.
15
15
-`BATCH_SIZE`: Defaults to 1.
16
16
-`STEPS`: Defaults to 100.
17
17
-`EXTRA_ARGS`: Extra arguments to pass to torchtitan training script. See [torchtitan](https://github.com/pytorch/torchtitan) docs for the full list of options.
0 commit comments