Updated example scripts and README

README.md

@@ -13,7 +13,7 @@ By [Apoorv Khandelwal](https://apoorvkh.com) and [Peter Curtin](https://github.c
 
 ---
 
-**`torchrunx`** is a *functional* utility for distributing PyTorch code across devices. This is a [more convenient, robust, and featureful](#torchrunx-uniquely-offers) alternative to CLI-based launchers, like `torchrun`, `accelerate launch`, and `deepspeed`.
+**`torchrunx`** is a *functional* utility for distributing PyTorch code across devices. This is a [more convenient, robust, and featureful](https://torchrun.xyz/features.html) alternative to CLI-based launchers, like `torchrun`, `accelerate launch`, and `deepspeed`.
 
 It enables complex workflows within a single script and has useful features even if only using 1 GPU.
 
@@ -29,20 +29,13 @@ Requires: Linux. If using multiple machines: SSH & shared filesystem.
 
 Suppose we have some distributed training function (which needs to run on every GPU):
 
-```python
-def distributed_training(model: nn.Module, num_steps: int) -> nn.Module: ...
-```
-
-<details>
-<summary><b>Implementation of <code>distributed_training</code> (click to expand)</b></summary>
-
 ```python
 from __future__ import annotations
 import os
 import torch
 import torch.nn as nn
 
-def distributed_training(num_steps: int = 10) -> nn.Module | None:
+def distributed_training(output_dir: str, num_steps: int = 10) -> str | None:
     rank = int(os.environ['RANK'])
     local_rank = int(os.environ['LOCAL_RANK'])
 
@@ -62,10 +55,13 @@ def distributed_training(num_steps: int = 10) -> nn.Module | None:
         optimizer.step()
 
     if rank == 0:
-        return model.cpu()
-```
+        os.makedirs(output_dir, exist_ok=True)
+        checkpoint_path = os.path.join(output_dir, "model.pt")
+        torch.save(model, checkpoint_path)
+        return checkpoint_path
 
-</details>
+    return None
+```
 
 We can distribute and run this function (e.g. on 2 machines x 2 GPUs) using **`torchrunx`**!
 
@@ -82,18 +78,20 @@ launcher = torchrunx.Launcher(
 
 results = launcher.run(
     distributed_training,
-    num_steps = 10
+    output_dir = "outputs",
+    num_steps = 10,
 )
 ```
 
 Once completed, you can retrieve the results and process them as you wish.
 
 ```python
-trained_model: nn.Module = results.rank(0)
-                     # or: results.index(hostname="localhost", local_rank=0)
+checkpoint_path: str = results.rank(0)
+                 # or: results.index(hostname="localhost", local_rank=0)
 
 # and continue your script
-torch.save(trained_model.state_dict(), "outputs/model.pth")
+model = torch.load(checkpoint_path, weights_only=False)
+model.eval()
 ```
 
 **See more examples where we fine-tune LLMs using:**
@@ -102,43 +100,4 @@ torch.save(trained_model.state_dict(), "outputs/model.pth")
   - [PyTorch Lightning](https://torchrun.xyz/examples/lightning.html)
   - [Accelerate](https://torchrun.xyz/examples/accelerate.html)
 
-**Refer to our [API](https://torchrun.xyz/api.html) and [Usage](https://torchrun.xyz/usage/general.html) for many more capabilities!**
-
----
-
-## `torchrunx` uniquely offers
-
-1. **An automatic launcher that "just works" for everyone** 🚀
-
-> `torchrunx` is an SSH-based, pure-Python library that is universally easy to install.<br>
-> No system-specific dependencies and orchestration for *automatic* multi-node distribution.
-
-2. **Conventional CLI commands** 🖥️
-
-> Run familiar commands, like `python my_script.py ...`, and customize arguments as you wish.
-> 
-> Other launchers override `python` in a cumbersome way: e.g. `torchrun --nproc_per_node=2 --nnodes=2 --node_rank=0 --master_addr=100.43.331.111 --master_port=1234 my_script.py ...`.
-
-3. **Support for more complex workflows in a single script** 🎛️
-
-> Your workflow may have steps that are complex (e.g. pre-train, fine-tune, test) or may different parallelizations (e.g. training on 8 GPUs, testing on 1 GPU). In these cases, CLI-based launchers require each step to live in its own script. Our library treats these steps in a modular way, so they can cleanly fit together in a single script!
->
-> 
-> We clean memory leaks as we go, so previous steps won't crash or adversely affect future steps.
-
-4. **Better handling of system failures. No more zombies!** 🧟
-
-> With `torchrun`, your "work" is inherently coupled to your main Python process. If the system kills one of your workers (e.g. due to RAM OOM or segmentation faults), there is no way to fail gracefully in Python. Your processes might hang for 10 minutes (the NCCL timeout) or become perpetual zombies.
->
-> 
-> `torchrunx` decouples "launcher" and "worker" processes. If the system kills a worker, our launcher immediately raises a `WorkerFailure` exception, which users can handle as they wish. We always clean up all nodes, so no more zombies!
-
-5. **Bonus features** 🎁
-
-> - Return objects from distributed functions.
-> - [Automatic detection of SLURM environments.](https://torchrun.xyz/usage/slurm.html)
-> - Start multi-node training from Python notebooks!
-> - Our library is fully typed!
-> - Custom, fine-grained handling of [logging](https://torchrun.xyz/usage/logging.html), [environment variables](https://torchrun.xyz/usage/general.html#environment-variables), and [exception propagation](https://torchrun.xyz/usage/general.html#exceptions). We have nice defaults too: no more interleaved logs and irrelevant exceptions!
-
-**On our [roadmap](https://github.com/apoorvkh/torchrunx/issues?q=is%3Aopen+is%3Aissue+label%3Aenhancement): higher-order parallelism, support for debuggers, and more!**
+**Refer to our [API](https://torchrun.xyz/api.html), [Features](https://torchrun.xyz/features.html), and [Usage](https://torchrun.xyz/usage/general.html) for many more capabilities!**

docs/source/artifacts/accelerate_help.txt

@@ -22,12 +22,15 @@ usage: accelerate_train.py [-h] [OPTIONS]
 │     ``"/etc/ssh/ssh_config"``. (default: None)                             │
 │ --launcher.backend {None,nccl,gloo,mpi,ucc}                                │
 │     `Backend                                                               │
-│     <https://pytorch.org/docs/stable/distributed.html#torch.distributed.B… │
+│     <https://pytorch.org/docs/stable/distributed.html#torch.distributed.Ba │
+│     ckend>`_                                                               │
 │             for worker process group. By default, NCCL (GPU backend).      │
 │             Use GLOO for CPU backend. ``None`` for no process group.       │
 │     (default: nccl)                                                        │
-│ --launcher.timeout INT                                                     │
+│ --launcher.worker-timeout INT                                              │
 │     Worker process group timeout (seconds). (default: 600)                 │
+│ --launcher.agent-timeout INT                                               │
+│     Agent communication timeout (seconds). (default: 180)                  │
 │ --launcher.copy-env-vars [STR [STR ...]]                                   │
 │     Environment variables to copy from the launcher process to workers.    │
 │     Supports Unix pattern matching syntax. (default: PATH LD_LIBRARY       │

docs/source/artifacts/argparse_cli_help.txt

@@ -1,7 +1,8 @@
 usage: -c [-h] [--hostnames HOSTNAMES [HOSTNAMES ...]]
           [--workers-per-host WORKERS_PER_HOST [WORKERS_PER_HOST ...]]
           [--ssh-config-file SSH_CONFIG_FILE]
-          [--backend {nccl,gloo,mpi,ucc,None}] [--timeout TIMEOUT]
+          [--backend {nccl,gloo,mpi,ucc,None}]
+          [--worker-timeout WORKER_TIMEOUT] [--agent-timeout AGENT_TIMEOUT]
           [--copy-env-vars COPY_ENV_VARS [COPY_ENV_VARS ...]]
           [--extra-env-vars [EXTRA_ENV_VARS ...]] [--env-file ENV_FILE]
 
@@ -21,7 +22,10 @@ torchrunx:
   --backend {nccl,gloo,mpi,ucc,None}
                         For worker process group. Default: 'nccl'. Use 'gloo'
                         for CPU. 'None' to disable.
-  --timeout TIMEOUT     Worker process group timeout in seconds. Default: 600.
+  --worker-timeout WORKER_TIMEOUT
+                        Worker process group timeout in seconds. Default: 600.
+  --agent-timeout AGENT_TIMEOUT
+                        Agent communication timeout in seconds. Default: 180.
   --copy-env-vars COPY_ENV_VARS [COPY_ENV_VARS ...]
                         Environment variables to copy to workers. Supports
                         Unix pattern matching.

docs/source/artifacts/deepspeed_help.txt

@@ -1,5 +1,5 @@
-[2025-02-23 16:02:38,914] [WARNING] [real_accelerator.py:181:get_accelerator] Setting accelerator to CPU. If you have GPU or other accelerator, we were unable to detect it.
-[2025-02-23 16:02:38,930] [INFO] [real_accelerator.py:222:get_accelerator] Setting ds_accelerator to cpu (auto detect)
+[2025-06-25 15:33:02,489] [INFO] [real_accelerator.py:222:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+Warning: The cache directory for DeepSpeed Triton autotune, /users/akhand10/.triton/autotune, appears to be on an NFS system. While this is generally acceptable, if you experience slowdowns or hanging when DeepSpeed exits, it is recommended to set the TRITON_CACHE_DIR environment variable to a non-NFS path.
 usage: deepspeed_train.py [-h] [OPTIONS]
 
 ╭─ options ──────────────────────────────────────────────────────────────────╮
@@ -42,8 +42,10 @@ usage: deepspeed_train.py [-h] [OPTIONS]
 │             for worker process group. By default, NCCL (GPU backend).      │
 │             Use GLOO for CPU backend. ``None`` for no process group.       │
 │     (default: nccl)                                                        │
-│ --launcher.timeout INT                                                     │
+│ --launcher.worker-timeout INT                                              │
 │     Worker process group timeout (seconds). (default: 600)                 │
+│ --launcher.agent-timeout INT                                               │
+│     Agent communication timeout (seconds). (default: 180)                  │
 │ --launcher.copy-env-vars [STR [STR ...]]                                   │
 │     Environment variables to copy from the launcher process to workers.    │
 │     Supports Unix pattern matching syntax. (default: PATH LD_LIBRARY       │

docs/source/artifacts/lightning_help.txt

@@ -18,12 +18,15 @@ usage: lightning_train.py [-h] [OPTIONS]
 │     ``"/etc/ssh/ssh_config"``. (default: None)                             │
 │ --launcher.backend {None,nccl,gloo,mpi,ucc}                                │
 │     `Backend                                                               │
-│     <https://pytorch.org/docs/stable/distributed.html#torch.distributed.B… │
+│     <https://pytorch.org/docs/stable/distributed.html#torch.distributed.Ba │
+│     ckend>`_                                                               │
 │             for worker process group. By default, NCCL (GPU backend).      │
 │             Use GLOO for CPU backend. ``None`` for no process group.       │
 │     (default: nccl)                                                        │
-│ --launcher.timeout INT                                                     │
+│ --launcher.worker-timeout INT                                              │
 │     Worker process group timeout (seconds). (default: 600)                 │
+│ --launcher.agent-timeout INT                                               │
+│     Agent communication timeout (seconds). (default: 180)                  │
 │ --launcher.copy-env-vars [STR [STR ...]]                                   │
 │     Environment variables to copy from the launcher process to workers.    │
 │     Supports Unix pattern matching syntax. (default: PATH LD_LIBRARY       │