Skip to content

ProcessGroupBabyNCCL: support multiple streams and use event on start #91

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Merged
merged 1 commit into from
Jan 31, 2025
Merged

ProcessGroupBabyNCCL: support multiple streams and use event on start #91

merged 1 commit into from
Jan 31, 2025

Conversation

d4l3k
Copy link
Member

@d4l3k d4l3k commented Jan 30, 2025 •
edited
Loading

This overhauls how we handle ProcessGroupBabyNCCL cuda events and streams. It also enables the PG test suite for BabyNCCL similar to #89 which enabled it for BabyGloo.

Key changes:

  • BabyNCCL now is stream aware and will allocate one CUDA stream per stream_id in the parent process. This avoids issues when multiple streams are using NCCL and avoids any unintentional synchronization by using a single stream in the child process.
  • BabyNCCL uses a CUDA event to synchronize the start of the NCCL operation. Previously BabyNCCL would immediately start running the NCCL operation even if the launching stream hadn't completed yet.
  • Added _pg_tests for BabyNCCL
  • Unified the codepath and got rid of _BabyNCCLWork/WORK_CLASS
  • Detect when the subprocess has crashed and error out early
  • Propagate device_id to the child process

Test plan:

pytest torchft/process_group_test.py

test integration with torchtitan #82 with patch P1722192431

@d4l3k d4l3k requested review from H-Huang and fegin January 30, 2025 19:26
@facebook-github-bot facebook-github-bot added the CLA Signed This label is managed by the Meta Open Source bot. label Jan 30, 2025
@d4l3k d4l3k force-pushed the d4l3k/baby_nccl_streams branch 3 times, most recently from 037f346 to 7d48c52 Compare January 30, 2025 23:56
@d4l3k d4l3k requested a review from kwen2501 January 30, 2025 23:56
@d4l3k d4l3k force-pushed the d4l3k/baby_nccl_streams branch from 7d48c52 to 0f8b44a Compare January 30, 2025 23:57
fegin
fegin reviewed Jan 31, 2025
View reviewed changes
torchft/process_group.py
if stream_key not in streams:
streams[stream_key] = torch.cuda.Stream(
device=stream_device
)
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Are we going to have zombie stream if there are multiple failures? Will this cause memory leakage?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

My understanding is streams are specific to the cuda context/process so this will be cleaned up just fine when it gets killed

@d4l3k d4l3k force-pushed the d4l3k/baby_nccl_streams branch from 0f8b44a to 30296e9 Compare January 31, 2025 00:33
@d4l3k d4l3k requested a review from fegin January 31, 2025 00:33
@d4l3k d4l3k merged commit 2b23017 into main Jan 31, 2025
6 checks passed
@d4l3k d4l3k deleted the d4l3k/baby_nccl_streams branch January 31, 2025 01:23
@fegin fegin mentioned this pull request Jan 31, 2025
Closed
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 3, 2025
@fegin
Update base for Update on "[WIP][RFC] TorchFT integration"
d95dff9 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new group joins. 

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN socketProgress: Connection closed by remote peer devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by modifying the command. 

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment line 42 in `torchtitan/utils.py`


**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1 NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2 --experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3 NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2 --experimental.enable_torchft --experimental.ft_replica_group_id=1
```



[ghstack-poisoned]
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 3, 2025
@fegin
Update on "[WIP][RFC] TorchFT integration"
dfcb08d 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new group joins. 

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN socketProgress: Connection closed by remote peer devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by modifying the command. 

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment line 42 in `torchtitan/utils.py`


**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1 NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2 --experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3 NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2 --experimental.enable_torchft --experimental.ft_replica_group_id=1
```



[ghstack-poisoned]
@fegin fegin mentioned this pull request Feb 11, 2025
Merged
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 11, 2025
@fegin
Integrate TorchFT
7cf91ff 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: bf6f0c5
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 12, 2025
@fegin
Integrate TorchFT
1b1a41d 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: 40d4964
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 12, 2025
@fegin
Integrate TorchFT
073b226 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: 4e04a58
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 12, 2025
@fegin
Integrate TorchFT
330744f 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: bf6f0c5
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 12, 2025
@fegin
Integrate TorchFT
eba0a53 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: 3690680
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 12, 2025
@fegin
Integrate TorchFT
b5e09d0 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: 088581c
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 12, 2025
@fegin
Integrate TorchFT
33213a3 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: 7f44395
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 12, 2025
@fegin
Integrate TorchFT
3b2c908 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: a5168a6
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 12, 2025
@fegin
Integrate TorchFT
8c61e5c 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: 82766ce
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 12, 2025
@fegin
Integrate TorchFT
11ace68 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: c90068b
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 12, 2025
@fegin
Integrate TorchFT
68d468e 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: 19016cc
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 21, 2025
@fegin
Integrate TorchFT
5409a57 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: 208678e
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 21, 2025
@fegin
Integrate TorchFT
57f6b85 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: fb41c66
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 21, 2025
@fegin
Integrate TorchFT
fff7a77 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: c898dc9
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 24, 2025
@fegin
Integrate TorchFT
58510b5 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: ec7fd5c
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 25, 2025
@fegin
Integrate TorchFT
67645cd 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: 9fba357
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 25, 2025
@fegin
Integrate TorchFT
e6e2309 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: e8ce81d
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 25, 2025
@fegin
Integrate TorchFT
28f8c50 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: 440da0f
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 25, 2025
@fegin
Integrate TorchFT
db6a75c 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: 1760ebf
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 25, 2025
@fegin
Integrate TorchFT
b806414 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: 5763454
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 25, 2025
@fegin
Integrate TorchFT
7c70988 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: 5f9a731
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 26, 2025
@fegin
Integrate TorchFT
d45965f 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: 8fa713a
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 26, 2025
@fegin
Integrate TorchFT
3b2669e 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: eeb3a32
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 26, 2025
@fegin
Integrate TorchFT
94d34b9 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: 2faf8ce
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 27, 2025
@fegin
Integrate TorchFT
fef1f07 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: d85feb5
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 27, 2025
@fegin
Integrate TorchFT
3514de3 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: 34c0f96
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 27, 2025
@fegin
Integrate TorchFT
88fea27 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: 64a7afa
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 27, 2025
@fegin
Integrate TorchFT
ff3864f 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: 45c6bee
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 27, 2025
@fegin
Integrate TorchFT
cfe89b2 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: cb2d40f
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 27, 2025
@fegin
Integrate TorchFT
5e1e8f3 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: 376f937
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 27, 2025
@fegin
Integrate TorchFT
f3300c3 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: 5a8ed07
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 27, 2025
@fegin
Integrate TorchFT
ee770e3 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: b428002
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 28, 2025
@fegin
Integrate TorchFT
307ec5b 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: 0a01915
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 28, 2025
@fegin
Integrate TorchFT
3aa61e5 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: f07ae76
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 28, 2025
@fegin
Integrate TorchFT
6be50f7 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: 1264c99
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 28, 2025
@fegin
Integrate TorchFT
c523e50 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: 2b401c8
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 28, 2025
@fegin
Integrate TorchFT
9add0ec 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: 4217630
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 28, 2025
@fegin
Integrate TorchFT
11248c6 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: 76ba987
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 28, 2025
@fegin
Integrate TorchFT
9cf0e84 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: 3e70806
Pull Request resolved: #834
fegin added a commit to pytorch/torchtitan that referenced this pull request Feb 28, 2025
@fegin
Integrate TorchFT
0f5bafa 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: 3e70806
Pull Request resolved: #834
MaxiBoether pushed a commit to eth-easl/torchtitan-mixtera that referenced this pull request Apr 17, 2025
@fegin @MaxiBoether
Integrate TorchFT
f78ec62 
**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: meta-pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: meta-pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with meta-pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with meta-pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: 3e70806
Pull Request resolved: pytorch#834
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@fegin fegin fegin approved these changes

@H-Huang H-Huang Awaiting requested review from H-Huang

@kwen2501 kwen2501 Awaiting requested review from kwen2501

Assignees

No one assigned

Labels

CLA Signed This label is managed by the Meta Open Source bot.

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

3 participants

@d4l3k @fegin @facebook-github-bot