[FSDP][Collectives] skipping allgather when world size is 1 #160135

anshul-si · 2025-08-07T20:20:10Z

Summary: In its current state, FSDP collectives uses cuda synchronizations and communication ops regardless of what the world size is. However, now that replicate will use FSDP, there will be instances where group size = 1 and these synchronizations and ops will be used needlessly. I have updated fsdp_params group to skip the foreach_all_gather and foreach_all_gather_copy_out APIs when world_size ‎ = 1. I have created a test that uses CommDebugMode to verify that the all gather comm has been removed. I also edited an affected test which used 1-way FSDP by verifying and changing its assert statements for CommDebugMode. Below, I have included the link to the profile trace verifying these two APIs were skipped and two test commands.

https://interncache-all.fbcdn.net/manifold/perfetto-artifacts/tree/ui/index.html#!/?url=https://interncache-all.fbcdn.net/manifold/perfetto_internal_traces/tree/shared_trace/anshulsi_f846ac3b-9467-4060-8e36-8cc3bc4449c3_devgpu263.prn2.facebook.com_652183.1753822140871934814.pt.trace.json

Stack from ghstack (oldest at bottom):

cc @H-Huang @awgu @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @pragupta

[ghstack-poisoned]

pytorch-bot · 2025-08-07T20:20:14Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/160135

📄 Preview Python docs built from this PR
📄 Preview C++ docs built from this PR
❓ Need help or want to give feedback on the CI? Visit the bot commands wiki or our office hours

Note: Links to docs will display an error until the docs builds have been completed.

✅ You can merge normally! (2 Unrelated Failures)

As of commit 5516a61 with merge base c8205cb ():

FLAKY - The following jobs failed but were likely due to flakiness present on trunk:

pull / linux-jammy-py3.9-clang12 / test (crossref, 2, 2, linux.2xlarge) (gh) (similar failure)
cpp/test_jit 1/1 failed!
pull / linux-jammy-py3.9-clang12 / test (default, 2, 5, linux.4xlarge) (gh) (similar failure)
cpp/test_jit 1/1 failed!

This comment was automatically generated by Dr. CI and updates every 15 minutes.

ghstack-source-id: 21e30b1 Pull Request resolved: #160135

cc H-Huang awgu wanchaol fegin fduwjj wz337 wconstab d4l3k pragupta [ghstack-poisoned]

weifengpy · 2025-08-12T02:38:33Z

torch/distributed/fsdp/_fully_shard/_fsdp_param_group.py

+                fsdp_param.init_all_gather_outputs(
+                    [all_gather_input.numel()],
+                    [all_gather_input.dtype],
+                    1,


nit: use world_size for better readability

weifengpy · 2025-08-12T02:43:18Z

attach a trace showing all-gather are indeed skipped and there is only copy_ overhead?

**Summary:** In its current state, FSDP collectives uses cuda synchronizations and communication ops regardless of what the world size is. However, now that replicate will use FSDP, there will be instances where group size = 1 and these synchronizations and ops will be used needlessly. I have updated fsdp_params group to skip the foreach_all_gather and foreach_all_gather_copy_out APIs when world_size ‎ = 1. I have created a test that uses CommDebugMode to verify that the all gather comm has been removed. I also edited an affected test which used 1-way FSDP by verifying and changing its assert statements for CommDebugMode. Below, I have included the link to the profile trace verifying these two APIs were skipped and two test commands. https://interncache-all.fbcdn.net/manifold/perfetto-artifacts/tree/ui/index.html#!/?url=https://interncache-all.fbcdn.net/manifold/perfetto_internal_traces/tree/shared_trace/anshulsi_f846ac3b-9467-4060-8e36-8cc3bc4449c3_devgpu263.prn2.facebook.com_652183.1753822140871934814.pt.trace.json cc H-Huang awgu wanchaol fegin fduwjj wz337 wconstab d4l3k pragupta [ghstack-poisoned]

anshul-si · 2025-08-12T18:29:40Z

@pytorchbot merge

pytorchmergebot · 2025-08-12T18:31:32Z

Merge started

Your change will be merged once all checks pass (ETA 0-4 Hours).

Learn more about merging in the wiki.

Questions? Feedback? Please reach out to the PyTorch DevX Team

Advanced Debugging

Check the merge workflow status
here

weifengpy · 2025-08-12T20:41:57Z

test/distributed/_composable/test_composability/test_2d_composability.py


            fwd_comm_counts = fwd_comm_mode.get_comm_counts()
-            self.assertEqual(len(fwd_comm_counts), 2)
+            self.assertEqual(len(fwd_comm_counts), 1)


I missed this: I missed this unit test. what's that fwd_comm_counts, bwd_comm_counts changed?

This is a test case that uses 1-way FSDP. Since I skip over foreach_gather and foreach_gather_copyout when world_size is 1, the all_gather count drops to 0, which is also why the number of different entries in fwd_comm_counts dict becomes 1. Same thing happens in the backward pass.

got you. thanks for clarifying

…orch#160136) **Summary:** In its current state, FSDP collectives uses cuda synchronizations and communication ops regardless of what the world size is. However, now that replicate will use FSDP, there will be instances where group size = 1 and these synchronizations and ops will be used needlessly. I have updated fsdp_collectives to skip reduce_scatter in the foreach_reduce API when world_size ‎ = 1. I have created edited a test that uses CommDebugMode to verify that the reduce_scatter has been removed. I also edited an affected test which used 1-way FSDP by verifying and changing its assert statements for CommDebugMode. I have also added a test command. **Test Cases** 1. pytest test/distributed/_composable/fsdp/test_fully_shard_training.py -k test_train_parity_single_worldsize1 2. pytest test/distributed/_composable/test_composability/test_2d_composability.py -k test_tp_with_fsdp_offloading Pull Request resolved: pytorch#160136 Approved by: https://github.com/weifengpy ghstack dependencies: pytorch#160135

…vice movements (pytorch#160147) **Summary:** In order to ensure that replicate acts as intended (a specialized version of hsdp) we need to make sure that it can pass the same tests that fully_shard can for training. To this end, I have added three test cases, one to test input device movement and the other two to test parameter registration during the forward and backward pass of a model. **Test Cases** 1. pytest test/distributed/_composable/test_replicate_training.py -k test_root_move_forward_input_to_device 2. pytest test/distributed/_composable/test_replicate_training.py -k TestReplicateRegisteredParams Pull Request resolved: pytorch#160147 Approved by: https://github.com/weifengpy ghstack dependencies: pytorch#160135, pytorch#160136

…orch#160136) **Summary:** In its current state, FSDP collectives uses cuda synchronizations and communication ops regardless of what the world size is. However, now that replicate will use FSDP, there will be instances where group size = 1 and these synchronizations and ops will be used needlessly. I have updated fsdp_collectives to skip reduce_scatter in the foreach_reduce API when world_size ‎ = 1. I have created edited a test that uses CommDebugMode to verify that the reduce_scatter has been removed. I also edited an affected test which used 1-way FSDP by verifying and changing its assert statements for CommDebugMode. I have also added a test command. **Test Cases** 1. pytest test/distributed/_composable/fsdp/test_fully_shard_training.py -k test_train_parity_single_worldsize1 2. pytest test/distributed/_composable/test_composability/test_2d_composability.py -k test_tp_with_fsdp_offloading Pull Request resolved: pytorch#160136 Approved by: https://github.com/weifengpy ghstack dependencies: pytorch#160135

…vice movements (pytorch#160147) **Summary:** In order to ensure that replicate acts as intended (a specialized version of hsdp) we need to make sure that it can pass the same tests that fully_shard can for training. To this end, I have added three test cases, one to test input device movement and the other two to test parameter registration during the forward and backward pass of a model. **Test Cases** 1. pytest test/distributed/_composable/test_replicate_training.py -k test_root_move_forward_input_to_device 2. pytest test/distributed/_composable/test_replicate_training.py -k TestReplicateRegisteredParams Pull Request resolved: pytorch#160147 Approved by: https://github.com/weifengpy ghstack dependencies: pytorch#160135, pytorch#160136

…ation with torchtitan" **Summary:** During this experiment to integrate the new replicate function into torchtitan, I used pytorch/pytorch#162021, which has not been landed. However, since this is more about making replicate more efficient rather than changing replicate's core code, pytorch/pytorch#160135, which has landed, should be fine. pytorch/pytorch#160133 is the last time replicate_with_fsdp.py and its replicate api were touched. In order to enable the new replicate, which uses a 2D device mesh (since it is a specialized version of HSDP), I changed the parallelism code to include dp_shard dim = 1 only if dp_replicate > 1, and created device mesh that I pass down in apply_ddp. **Test Case** 1. CONFIG_FILE="./torchtitan/models/llama3/train_configs/debug_model.toml" ./run_train.sh Expected output of this experiment should be something like: [rank0]:[titan] 2025-09-15 17:38:26,676 - root - INFO - Starting job: Llama 3 debug training [rank0]:[titan] 2025-09-15 17:38:29,094 - root - WARNING - ENV[TORCH_NCCL_ASYNC_ERROR_HANDLING] = 1 will be overridden to 3 based on job config **[rank0]:[titan] 2025-09-15 17:38:29,097 - root - INFO - Building 2-D device mesh with ['dp_replicate', 'dp_shard'], [8, 1]** [rank0]:[titan] 2025-09-15 17:38:29,104 - root - INFO - [GC] Initial GC collection 0.00 seconds [rank0]:NCCL version 2.27.5+cuda12.6 [rank0]:[titan] 2025-09-15 17:38:35,439 - root - INFO - Loading tokenizer from tokenizer.json [rank0]:[titan] 2025-09-15 17:38:35,441 - root - INFO - Preparing c4_test dataset from tests/assets/c4_test [rank0]:[titan] 2025-09-15 17:38:35,894 - root - INFO - Building llama3 debugmodel with TransformerModelArgs(_enforced='This field is used to enforce all fields have defaults.', dim=256, n_layers=6, n_heads=16, n_kv_heads=None, vocab_size=2000, multiple_of=256, ffn_dim_multiplier=None, norm_eps=1e-05, rope_theta=500000, max_seq_len=2048, depth_init=True, use_flex_attn=False, attn_mask_type='causal', eos_id=0) [rank0]:[titan] 2025-09-15 17:38:35,931 - root - INFO - CUDA capacity: NVIDIA H100 with 94.99GiB memory [rank0]:[titan] 2025-09-15 17:38:35,950 - root - INFO - Model llama3 debugmodel size: 6,139,136 total parameters [rank0]:[titan] 2025-09-15 17:38:35,951 - root - INFO - Applied selective activation checkpointing to the model **[rank0]:[titan] 2025-09-15 17:38:35,972 - root - INFO - Applied DDP to the model** [rank0]:[titan] 2025-09-15 17:38:36,153 - root - INFO - Peak FLOPS used for computing MFU: 9.890e+14 [rank0]:[titan] 2025-09-15 17:38:36,153 - root - INFO - CUDA memory usage for model: 0.04GiB(0.04%) [rank0]:[titan] 2025-09-15 17:38:36,154 - root - WARNING - model.safetensors.index.json not found at hf_assets_path: ./tests/assets/tokenizer/model.safetensors.index.json. Defaulting to saving a single safetensors file if checkpoint is saved in HF format [rank0]:[titan] 2025-09-15 17:38:36,154 - root - INFO - Mixed precision training is handled by AMP [rank0]:[titan] 2025-09-15 17:38:36,154 - root - INFO - Trainer is initialized with local batch size 8, global batch size 64, gradient accumulation steps 1, sequence length 2048, total steps 10 (warmup 2) [ghstack-poisoned]

…160135) **Summary:** In its current state, FSDP collectives uses cuda synchronizations and communication ops regardless of what the world size is. However, now that replicate will use FSDP, there will be instances where group size = 1 and these synchronizations and ops will be used needlessly. I have updated fsdp_params group to skip the foreach_all_gather and foreach_all_gather_copy_out APIs when world_size ‎ = 1. I have created a test that uses CommDebugMode to verify that the all gather comm has been removed. I also edited an affected test which used 1-way FSDP by verifying and changing its assert statements for CommDebugMode. Below, I have included the link to the profile trace verifying these two APIs were skipped and two test commands. https://interncache-all.fbcdn.net/manifold/perfetto-artifacts/tree/ui/index.html#!/?url=https://interncache-all.fbcdn.net/manifold/perfetto_internal_traces/tree/shared_trace/anshulsi_f846ac3b-9467-4060-8e36-8cc3bc4449c3_devgpu263.prn2.facebook.com_652183.1753822140871934814.pt.trace.json Pull Request resolved: pytorch#160135 Approved by: https://github.com/weifengpy

…orch#160136) **Summary:** In its current state, FSDP collectives uses cuda synchronizations and communication ops regardless of what the world size is. However, now that replicate will use FSDP, there will be instances where group size = 1 and these synchronizations and ops will be used needlessly. I have updated fsdp_collectives to skip reduce_scatter in the foreach_reduce API when world_size ‎ = 1. I have created edited a test that uses CommDebugMode to verify that the reduce_scatter has been removed. I also edited an affected test which used 1-way FSDP by verifying and changing its assert statements for CommDebugMode. I have also added a test command. **Test Cases** 1. pytest test/distributed/_composable/fsdp/test_fully_shard_training.py -k test_train_parity_single_worldsize1 2. pytest test/distributed/_composable/test_composability/test_2d_composability.py -k test_tp_with_fsdp_offloading Pull Request resolved: pytorch#160136 Approved by: https://github.com/weifengpy ghstack dependencies: pytorch#160135

…vice movements (pytorch#160147) **Summary:** In order to ensure that replicate acts as intended (a specialized version of hsdp) we need to make sure that it can pass the same tests that fully_shard can for training. To this end, I have added three test cases, one to test input device movement and the other two to test parameter registration during the forward and backward pass of a model. **Test Cases** 1. pytest test/distributed/_composable/test_replicate_training.py -k test_root_move_forward_input_to_device 2. pytest test/distributed/_composable/test_replicate_training.py -k TestReplicateRegisteredParams Pull Request resolved: pytorch#160147 Approved by: https://github.com/weifengpy ghstack dependencies: pytorch#160135, pytorch#160136

…ation with torchtitan" **Summary:** During this experiment to integrate the new replicate function into torchtitan, I used pytorch/pytorch#162021, which has not been landed. However, since this is more about making replicate more efficient rather than changing replicate's core code, pytorch/pytorch#160135, which has landed, should be fine. pytorch/pytorch#160133 is the last time replicate_with_fsdp.py and its replicate api were touched. In order to enable the new replicate, which uses a 2D device mesh (since it is a specialized version of HSDP), I changed the parallelism code to include dp_shard dim = 1 only if dp_replicate > 1, and created device mesh that I pass down in apply_ddp. **Test Case** 1. CONFIG_FILE="./torchtitan/models/llama3/train_configs/debug_model.toml" ./run_train.sh Expected output of this experiment should be something like: [rank0]:[titan] 2025-09-15 17:38:26,676 - root - INFO - Starting job: Llama 3 debug training [rank0]:[titan] 2025-09-15 17:38:29,094 - root - WARNING - ENV[TORCH_NCCL_ASYNC_ERROR_HANDLING] = 1 will be overridden to 3 based on job config **[rank0]:[titan] 2025-09-15 17:38:29,097 - root - INFO - Building 2-D device mesh with ['dp_replicate', 'dp_shard'], [8, 1]** [rank0]:[titan] 2025-09-15 17:38:29,104 - root - INFO - [GC] Initial GC collection 0.00 seconds [rank0]:NCCL version 2.27.5+cuda12.6 [rank0]:[titan] 2025-09-15 17:38:35,439 - root - INFO - Loading tokenizer from tokenizer.json [rank0]:[titan] 2025-09-15 17:38:35,441 - root - INFO - Preparing c4_test dataset from tests/assets/c4_test [rank0]:[titan] 2025-09-15 17:38:35,894 - root - INFO - Building llama3 debugmodel with TransformerModelArgs(_enforced='This field is used to enforce all fields have defaults.', dim=256, n_layers=6, n_heads=16, n_kv_heads=None, vocab_size=2000, multiple_of=256, ffn_dim_multiplier=None, norm_eps=1e-05, rope_theta=500000, max_seq_len=2048, depth_init=True, use_flex_attn=False, attn_mask_type='causal', eos_id=0) [rank0]:[titan] 2025-09-15 17:38:35,931 - root - INFO - CUDA capacity: NVIDIA H100 with 94.99GiB memory [rank0]:[titan] 2025-09-15 17:38:35,950 - root - INFO - Model llama3 debugmodel size: 6,139,136 total parameters [rank0]:[titan] 2025-09-15 17:38:35,951 - root - INFO - Applied selective activation checkpointing to the model **[rank0]:[titan] 2025-09-15 17:38:35,972 - root - INFO - Applied DDP to the model** [rank0]:[titan] 2025-09-15 17:38:36,153 - root - INFO - Peak FLOPS used for computing MFU: 9.890e+14 [rank0]:[titan] 2025-09-15 17:38:36,153 - root - INFO - CUDA memory usage for model: 0.04GiB(0.04%) [rank0]:[titan] 2025-09-15 17:38:36,154 - root - WARNING - model.safetensors.index.json not found at hf_assets_path: ./tests/assets/tokenizer/model.safetensors.index.json. Defaulting to saving a single safetensors file if checkpoint is saved in HF format [rank0]:[titan] 2025-09-15 17:38:36,154 - root - INFO - Mixed precision training is handled by AMP [rank0]:[titan] 2025-09-15 17:38:36,154 - root - INFO - Trainer is initialized with local batch size 8, global batch size 64, gradient accumulation steps 1, sequence length 2048, total steps 10 (warmup 2) [ghstack-poisoned]

…replicate integration with torchtitan" **Summary:** During this experiment to integrate the new replicate function into torchtitan, I used pytorch/pytorch#162021, which has not been landed. However, since this is more about making replicate more efficient rather than changing replicate's core code, pytorch/pytorch#160135, which has landed, should be fine. pytorch/pytorch#160133 is the last time replicate_with_fsdp.py and its replicate api were touched. In order to enable the new replicate, which uses a 2D device mesh (since it is a specialized version of HSDP), I changed the parallelism code to include dp_shard dim = 1 only if dp_replicate > 1, and created device mesh that I pass down in apply_ddp. Below is a link comparing the loss curves for Llama3.1-8B models: one configured with dimension sharding (2) and tensor parallelism (4), and the other with dimension replication (2) and sharding (4). <img width="1266" height="483" alt="image" src="https://github.com/user-attachments/assets/40198bc5-5e3f-486b-be56-12111e010e0c" /> https://fburl.com/mlhub/btkos8ok **Test Case** 1. CONFIG_FILE="./torchtitan/models/llama3/train_configs/debug_model.toml" ./run_train.sh Expected output of this experiment should be something like: [rank0]:[titan] 2025-09-15 17:38:26,676 - root - INFO - Starting job: Llama 3 debug training [rank0]:[titan] 2025-09-15 17:38:29,094 - root - WARNING - ENV[TORCH_NCCL_ASYNC_ERROR_HANDLING] = 1 will be overridden to 3 based on job config **[rank0]:[titan] 2025-09-15 17:38:29,097 - root - INFO - Building 2-D device mesh with ['dp_replicate', 'dp_shard'], [8, 1]** [rank0]:[titan] 2025-09-15 17:38:29,104 - root - INFO - [GC] Initial GC collection 0.00 seconds [rank0]:NCCL version 2.27.5+cuda12.6 [rank0]:[titan] 2025-09-15 17:38:35,439 - root - INFO - Loading tokenizer from tokenizer.json [rank0]:[titan] 2025-09-15 17:38:35,441 - root - INFO - Preparing c4_test dataset from tests/assets/c4_test [rank0]:[titan] 2025-09-15 17:38:35,894 - root - INFO - Building llama3 debugmodel with TransformerModelArgs(_enforced='This field is used to enforce all fields have defaults.', dim=256, n_layers=6, n_heads=16, n_kv_heads=None, vocab_size=2000, multiple_of=256, ffn_dim_multiplier=None, norm_eps=1e-05, rope_theta=500000, max_seq_len=2048, depth_init=True, use_flex_attn=False, attn_mask_type='causal', eos_id=0) [rank0]:[titan] 2025-09-15 17:38:35,931 - root - INFO - CUDA capacity: NVIDIA H100 with 94.99GiB memory [rank0]:[titan] 2025-09-15 17:38:35,950 - root - INFO - Model llama3 debugmodel size: 6,139,136 total parameters [rank0]:[titan] 2025-09-15 17:38:35,951 - root - INFO - Applied selective activation checkpointing to the model **[rank0]:[titan] 2025-09-15 17:38:35,972 - root - INFO - Applied DDP to the model** [rank0]:[titan] 2025-09-15 17:38:36,153 - root - INFO - Peak FLOPS used for computing MFU: 9.890e+14 [rank0]:[titan] 2025-09-15 17:38:36,153 - root - INFO - CUDA memory usage for model: 0.04GiB(0.04%) [rank0]:[titan] 2025-09-15 17:38:36,154 - root - WARNING - model.safetensors.index.json not found at hf_assets_path: ./tests/assets/tokenizer/model.safetensors.index.json. Defaulting to saving a single safetensors file if checkpoint is saved in HF format [rank0]:[titan] 2025-09-15 17:38:36,154 - root - INFO - Mixed precision training is handled by AMP [rank0]:[titan] 2025-09-15 17:38:36,154 - root - INFO - Trainer is initialized with local batch size 8, global batch size 64, gradient accumulation steps 1, sequence length 2048, total steps 10 (warmup 2) [ghstack-poisoned]

…ation with torchtitan" **Summary:** During this experiment to integrate the new replicate function into torchtitan, I used pytorch/pytorch#162021, which has not been landed. However, since this is more about making replicate more efficient rather than changing replicate's core code, pytorch/pytorch#160135, which has landed, should be fine. pytorch/pytorch#160133 is the last time replicate_with_fsdp.py and its replicate api were touched. In order to enable the new replicate, which uses a 2D device mesh (since it is a specialized version of HSDP), I changed the parallelism code to include dp_shard dim = 1 only if dp_replicate > 1, and created device mesh that I pass down in apply_ddp. Below is a link comparing the loss curves for Llama3.1-8B models: one configured with dimension sharding (2) and tensor parallelism (4), and the other with dimension replication (2) and sharding (4). <img width="1266" height="483" alt="image" src="https://github.com/user-attachments/assets/40198bc5-5e3f-486b-be56-12111e010e0c" /> https://fburl.com/mlhub/btkos8ok **Test Case** 1. CONFIG_FILE="./torchtitan/models/llama3/train_configs/debug_model.toml" ./run_train.sh Expected output of this experiment should be something like: [rank0]:[titan] 2025-09-15 17:38:26,676 - root - INFO - Starting job: Llama 3 debug training [rank0]:[titan] 2025-09-15 17:38:29,094 - root - WARNING - ENV[TORCH_NCCL_ASYNC_ERROR_HANDLING] = 1 will be overridden to 3 based on job config **[rank0]:[titan] 2025-09-15 17:38:29,097 - root - INFO - Building 2-D device mesh with ['dp_replicate', 'dp_shard'], [8, 1]** [rank0]:[titan] 2025-09-15 17:38:29,104 - root - INFO - [GC] Initial GC collection 0.00 seconds [rank0]:NCCL version 2.27.5+cuda12.6 [rank0]:[titan] 2025-09-15 17:38:35,439 - root - INFO - Loading tokenizer from tokenizer.json [rank0]:[titan] 2025-09-15 17:38:35,441 - root - INFO - Preparing c4_test dataset from tests/assets/c4_test [rank0]:[titan] 2025-09-15 17:38:35,894 - root - INFO - Building llama3 debugmodel with TransformerModelArgs(_enforced='This field is used to enforce all fields have defaults.', dim=256, n_layers=6, n_heads=16, n_kv_heads=None, vocab_size=2000, multiple_of=256, ffn_dim_multiplier=None, norm_eps=1e-05, rope_theta=500000, max_seq_len=2048, depth_init=True, use_flex_attn=False, attn_mask_type='causal', eos_id=0) [rank0]:[titan] 2025-09-15 17:38:35,931 - root - INFO - CUDA capacity: NVIDIA H100 with 94.99GiB memory [rank0]:[titan] 2025-09-15 17:38:35,950 - root - INFO - Model llama3 debugmodel size: 6,139,136 total parameters [rank0]:[titan] 2025-09-15 17:38:35,951 - root - INFO - Applied selective activation checkpointing to the model **[rank0]:[titan] 2025-09-15 17:38:35,972 - root - INFO - Applied DDP to the model** [rank0]:[titan] 2025-09-15 17:38:36,153 - root - INFO - Peak FLOPS used for computing MFU: 9.890e+14 [rank0]:[titan] 2025-09-15 17:38:36,153 - root - INFO - CUDA memory usage for model: 0.04GiB(0.04%) [rank0]:[titan] 2025-09-15 17:38:36,154 - root - WARNING - model.safetensors.index.json not found at hf_assets_path: ./tests/assets/tokenizer/model.safetensors.index.json. Defaulting to saving a single safetensors file if checkpoint is saved in HF format [rank0]:[titan] 2025-09-15 17:38:36,154 - root - INFO - Mixed precision training is handled by AMP [rank0]:[titan] 2025-09-15 17:38:36,154 - root - INFO - Trainer is initialized with local batch size 8, global batch size 64, gradient accumulation steps 1, sequence length 2048, total steps 10 (warmup 2) [ghstack-poisoned]