float8 rowwise training: add FSDP workaround #1629
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Summary: Adds the workaround from pytorch/pytorch#141881 to the torchao float8 rowwise recipe, to reduce memory usage when FSDP is on. Test Plan: tested in torchtitan, LLaMa 3 8B 8H100 training with rowwise peak memory decreased from 67GiB to 59GiB Reviewers: Subscribers: Tasks: Tags:
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/1629
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 066f889 with merge base 47f96f1 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
vkuzo
added
topic: performance
vkuzo
added a commit
to pytorch/torchtitan
that referenced
this pull request
Jan 27, 2025
Summary: This is an example of how to call float8 training with rowwise scaling from torchao. TODO: finalize API in torchao, and finalize how we want to expose it in torchtitan, and optimize performance. ``` // baseline (bf16 + compile) > with-proxy CONFIG_FILE="./train_configs/llama3_8b.toml" ./run_llama_train.sh --training.compile ... step: 20 loss: 8.4931 memory: 47.65GiB(50.16%) tps: 5,760 mfu: 33.73% // experiment (rowwise float8 + compile) > with-proxy CONFIG_FILE="./train_configs/llama3_8b.toml" ./run_llama_train.sh --float8.enable_float8_linear --training.compile ... // torchao main branch step: 40 loss: 7.3818 memory: 66.81GiB(70.33%) tps: 6,412 mfu: 37.55% // torchao with pytorch/ao#1629 step: 20 loss: 8.3823 memory: 58.55GiB(61.63%) tps: 6,424 mfu: 37.62% // for comparison, tensorwise float8 with float8 all-gather (on main branch) with-proxy CONFIG_FILE="./train_configs/llama3_8b.toml" ./run_llama_train.sh --float8.enable_float8_linear --training.compile --float8.enable_fsdp_float8_all_gather --float8.precompute_float8_dynamic_scale_for_fsdp ... step: 20 loss: 8.4258 memory: 47.32GiB(49.81%) tps: 7,186 mfu: 42.08% ``` Test Plan: Reviewers: Subscribers: Tasks: Tags:
facebook-github-bot
added
the
CLA Signed
danielvegamyhre
approved these changes
Jan 28, 2025
vkuzo
added a commit
to pytorch/torchtitan
that referenced
this pull request
Feb 7, 2025
Summary: This is an example of how to call float8 training with rowwise scaling from torchao. TODO: finalize API in torchao, and finalize how we want to expose it in torchtitan, and optimize performance. ``` // baseline (bf16 + compile) > with-proxy CONFIG_FILE="./train_configs/llama3_8b.toml" ./run_llama_train.sh --training.compile ... step: 20 loss: 8.4931 memory: 47.65GiB(50.16%) tps: 5,760 mfu: 33.73% // experiment (rowwise float8 + compile) > with-proxy CONFIG_FILE="./train_configs/llama3_8b.toml" ./run_llama_train.sh --float8.enable_float8_linear --training.compile ... // torchao main branch step: 40 loss: 7.3818 memory: 66.81GiB(70.33%) tps: 6,412 mfu: 37.55% // torchao with pytorch/ao#1629 step: 20 loss: 8.3823 memory: 58.55GiB(61.63%) tps: 6,424 mfu: 37.62% // for comparison, tensorwise float8 with float8 all-gather (on main branch) with-proxy CONFIG_FILE="./train_configs/llama3_8b.toml" ./run_llama_train.sh --float8.enable_float8_linear --training.compile --float8.enable_fsdp_float8_all_gather --float8.precompute_float8_dynamic_scale_for_fsdp ... step: 20 loss: 8.4258 memory: 47.32GiB(49.81%) tps: 7,186 mfu: 42.08% ``` Test Plan: Reviewers: Subscribers: Tasks: Tags:
vkuzo
added a commit
to pytorch/torchtitan
that referenced
this pull request
Feb 7, 2025
Summary: This is an example of how to call float8 training with rowwise scaling from torchao. TODO: finalize API in torchao, and finalize how we want to expose it in torchtitan, and optimize performance. ``` // baseline (bf16 + compile) > with-proxy CONFIG_FILE="./train_configs/llama3_8b.toml" ./run_llama_train.sh --training.compile ... step: 20 loss: 8.4931 memory: 47.65GiB(50.16%) tps: 5,760 mfu: 33.73% // experiment (rowwise float8 + compile) > with-proxy CONFIG_FILE="./train_configs/llama3_8b.toml" ./run_llama_train.sh --float8.enable_float8_linear --training.compile ... // torchao main branch step: 40 loss: 7.3818 memory: 66.81GiB(70.33%) tps: 6,412 mfu: 37.55% // torchao with pytorch/ao#1629 step: 20 loss: 8.3823 memory: 58.55GiB(61.63%) tps: 6,424 mfu: 37.62% // for comparison, tensorwise float8 with float8 all-gather (on main branch) with-proxy CONFIG_FILE="./train_configs/llama3_8b.toml" ./run_llama_train.sh --float8.enable_float8_linear --training.compile --float8.enable_fsdp_float8_all_gather --float8.precompute_float8_dynamic_scale_for_fsdp ... step: 20 loss: 8.4258 memory: 47.32GiB(49.81%) tps: 7,186 mfu: 42.08% ``` Test Plan: Reviewers: Subscribers: Tasks: Tags:
vkuzo
added a commit
to pytorch/torchtitan
that referenced
this pull request
Feb 16, 2025
Summary: This is an example of how to call float8 training with rowwise scaling from torchao. TODO: finalize API in torchao, and finalize how we want to expose it in torchtitan, and optimize performance. ``` // baseline (bf16 + compile) > with-proxy CONFIG_FILE="./train_configs/llama3_8b.toml" ./run_llama_train.sh --training.compile ... step: 20 loss: 8.4931 memory: 47.65GiB(50.16%) tps: 5,760 mfu: 33.73% // experiment (rowwise float8 + compile) > with-proxy CONFIG_FILE="./train_configs/llama3_8b.toml" ./run_llama_train.sh --float8.enable_float8_linear --training.compile ... // torchao main branch step: 40 loss: 7.3818 memory: 66.81GiB(70.33%) tps: 6,412 mfu: 37.55% // torchao with pytorch/ao#1629 step: 20 loss: 8.3823 memory: 58.55GiB(61.63%) tps: 6,424 mfu: 37.62% // for comparison, tensorwise float8 with float8 all-gather (on main branch) with-proxy CONFIG_FILE="./train_configs/llama3_8b.toml" ./run_llama_train.sh --float8.enable_float8_linear --training.compile --float8.enable_fsdp_float8_all_gather --float8.precompute_float8_dynamic_scale_for_fsdp ... step: 20 loss: 8.4258 memory: 47.32GiB(49.81%) tps: 7,186 mfu: 42.08% ``` Test Plan: Reviewers: Subscribers: Tasks: Tags:
vkuzo
added a commit
to pytorch/torchtitan
that referenced
this pull request
Feb 20, 2025
Summary: This is an example of how to call float8 training with rowwise scaling from torchao. TODO: finalize API in torchao, and finalize how we want to expose it in torchtitan, and optimize performance. ``` // baseline (bf16 + compile) > with-proxy CONFIG_FILE="./train_configs/llama3_8b.toml" ./run_llama_train.sh --training.compile ... step: 20 loss: 8.4931 memory: 47.65GiB(50.16%) tps: 5,760 mfu: 33.73% // experiment (rowwise float8 + compile) > with-proxy CONFIG_FILE="./train_configs/llama3_8b.toml" ./run_llama_train.sh --float8.enable_float8_linear --training.compile ... // torchao main branch step: 40 loss: 7.3818 memory: 66.81GiB(70.33%) tps: 6,412 mfu: 37.55% // torchao with pytorch/ao#1629 step: 20 loss: 8.3823 memory: 58.55GiB(61.63%) tps: 6,424 mfu: 37.62% // for comparison, tensorwise float8 with float8 all-gather (on main branch) with-proxy CONFIG_FILE="./train_configs/llama3_8b.toml" ./run_llama_train.sh --float8.enable_float8_linear --training.compile --float8.enable_fsdp_float8_all_gather --float8.precompute_float8_dynamic_scale_for_fsdp ... step: 20 loss: 8.4258 memory: 47.32GiB(49.81%) tps: 7,186 mfu: 42.08% ``` Test Plan: Reviewers: Subscribers: Tasks: Tags:
vkuzo
added a commit
to pytorch/torchtitan
that referenced
this pull request
Feb 26, 2025
Summary: This is an example of how to call float8 training with rowwise scaling from torchao. TODO: finalize API in torchao, and finalize how we want to expose it in torchtitan, and optimize performance. ``` // baseline (bf16 + compile) > with-proxy CONFIG_FILE="./train_configs/llama3_8b.toml" ./run_llama_train.sh --training.compile ... step: 20 loss: 8.4931 memory: 47.65GiB(50.16%) tps: 5,760 mfu: 33.73% // experiment (rowwise float8 + compile) > with-proxy CONFIG_FILE="./train_configs/llama3_8b.toml" ./run_llama_train.sh --float8.enable_float8_linear --training.compile ... // torchao main branch step: 40 loss: 7.3818 memory: 66.81GiB(70.33%) tps: 6,412 mfu: 37.55% // torchao with pytorch/ao#1629 step: 20 loss: 8.3823 memory: 58.55GiB(61.63%) tps: 6,424 mfu: 37.62% // for comparison, tensorwise float8 with float8 all-gather (on main branch) with-proxy CONFIG_FILE="./train_configs/llama3_8b.toml" ./run_llama_train.sh --float8.enable_float8_linear --training.compile --float8.enable_fsdp_float8_all_gather --float8.precompute_float8_dynamic_scale_for_fsdp ... step: 20 loss: 8.4258 memory: 47.32GiB(49.81%) tps: 7,186 mfu: 42.08% ``` Test Plan: Reviewers: Subscribers: Tasks: Tags:
vkuzo
added a commit
to pytorch/torchtitan
that referenced
this pull request
Feb 26, 2025
Summary: This is an example of how to call float8 training with rowwise scaling from torchao. TODO: finalize API in torchao, and finalize how we want to expose it in torchtitan, and optimize performance. ``` // baseline (bf16 + compile) > with-proxy CONFIG_FILE="./train_configs/llama3_8b.toml" ./run_llama_train.sh --training.compile ... step: 20 loss: 8.4931 memory: 47.65GiB(50.16%) tps: 5,760 mfu: 33.73% // experiment (rowwise float8 + compile) > with-proxy CONFIG_FILE="./train_configs/llama3_8b.toml" ./run_llama_train.sh --float8.enable_float8_linear --training.compile ... // torchao main branch step: 40 loss: 7.3818 memory: 66.81GiB(70.33%) tps: 6,412 mfu: 37.55% // torchao with pytorch/ao#1629 step: 20 loss: 8.3823 memory: 58.55GiB(61.63%) tps: 6,424 mfu: 37.62% // for comparison, tensorwise float8 with float8 all-gather (on main branch) with-proxy CONFIG_FILE="./train_configs/llama3_8b.toml" ./run_llama_train.sh --float8.enable_float8_linear --training.compile --float8.enable_fsdp_float8_all_gather --float8.precompute_float8_dynamic_scale_for_fsdp ... step: 20 loss: 8.4258 memory: 47.32GiB(49.81%) tps: 7,186 mfu: 42.08% ``` Test Plan: Reviewers: Subscribers: Tasks: Tags:
vkuzo
added a commit
to pytorch/torchtitan
that referenced
this pull request
Feb 27, 2025
Summary: This is an example of how to call float8 training with rowwise scaling from torchao. TODO: finalize API in torchao, and finalize how we want to expose it in torchtitan, and optimize performance. ``` // baseline (bf16 + compile) > with-proxy CONFIG_FILE="./train_configs/llama3_8b.toml" ./run_llama_train.sh --training.compile ... step: 20 loss: 8.4931 memory: 47.65GiB(50.16%) tps: 5,760 mfu: 33.73% // experiment (rowwise float8 + compile) > with-proxy CONFIG_FILE="./train_configs/llama3_8b.toml" ./run_llama_train.sh --float8.enable_float8_linear --training.compile ... // torchao main branch step: 40 loss: 7.3818 memory: 66.81GiB(70.33%) tps: 6,412 mfu: 37.55% // torchao with pytorch/ao#1629 step: 20 loss: 8.3823 memory: 58.55GiB(61.63%) tps: 6,424 mfu: 37.62% // for comparison, tensorwise float8 with float8 all-gather (on main branch) with-proxy CONFIG_FILE="./train_configs/llama3_8b.toml" ./run_llama_train.sh --float8.enable_float8_linear --training.compile --float8.enable_fsdp_float8_all_gather --float8.precompute_float8_dynamic_scale_for_fsdp ... step: 20 loss: 8.4258 memory: 47.32GiB(49.81%) tps: 7,186 mfu: 42.08% ``` Test Plan: Reviewers: Subscribers: Tasks: Tags:
vkuzo
added a commit
to pytorch/torchtitan
that referenced
this pull request
Feb 27, 2025
Summary: This is an example of how to call float8 training with rowwise scaling from torchao. TODO: finalize API in torchao, and finalize how we want to expose it in torchtitan, and optimize performance. ``` // baseline (bf16 + compile) > with-proxy CONFIG_FILE="./train_configs/llama3_8b.toml" ./run_llama_train.sh --training.compile ... step: 20 loss: 8.4931 memory: 47.65GiB(50.16%) tps: 5,760 mfu: 33.73% // experiment (rowwise float8 + compile) > with-proxy CONFIG_FILE="./train_configs/llama3_8b.toml" ./run_llama_train.sh --float8.enable_float8_linear --training.compile ... // torchao main branch step: 40 loss: 7.3818 memory: 66.81GiB(70.33%) tps: 6,412 mfu: 37.55% // torchao with pytorch/ao#1629 step: 20 loss: 8.3823 memory: 58.55GiB(61.63%) tps: 6,424 mfu: 37.62% // for comparison, tensorwise float8 with float8 all-gather (on main branch) with-proxy CONFIG_FILE="./train_configs/llama3_8b.toml" ./run_llama_train.sh --float8.enable_float8_linear --training.compile --float8.enable_fsdp_float8_all_gather --float8.precompute_float8_dynamic_scale_for_fsdp ... step: 20 loss: 8.4258 memory: 47.32GiB(49.81%) tps: 7,186 mfu: 42.08% ``` Test Plan: Reviewers: Subscribers: Tasks: Tags:
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Summary:
Adds the workaround from
pytorch/pytorch#141881 to the torchao float8 rowwise recipe, to reduce memory usage when FSDP is on.
Test Plan: tested in torchtitan, LLaMa 3 8B 8H100 training with rowwise peak memory decreased from 67GiB to 59GiB
Reviewers:
Subscribers:
Tasks:
Tags: