[llama3-8B] add flex_attention model flavor #1884
Draft
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
[ghstack-poisoned]
XilunWu
requested review from
fegin,
tianyu-l,
wconstab and
wwwjn
as code owners
This was referenced Oct 15, 2025
meta-cla
bot
added
the
CLA Signed
**Summary** Enable llama3-8B model to use `flex_attention`. **Test** `CONFIG_FILE=./torchtitan/models/llama3/train_configs/llama3_8b.toml NGPU=4 LOG_RANK=0,1,2,3 ./run_train.sh --model.flavor "8B_flex_attn" --activation_checkpoint.mode "full" --parallelism.context_parallel_degree 2` ``` [rank3]:[titan] 2025-10-15 10:37:01,390 - root - INFO - step: 1 loss: 12.2465 grad_norm: 4.5253 memory: 33.21GiB(34.95%) tps: 86 tflops: 4.98 mfu: 0.50% [rank0]:[titan] 2025-10-15 10:37:01,391 - root - INFO - step: 1 loss: 12.2465 grad_norm: 4.5253 memory: 33.21GiB(34.95%) tps: 87 tflops: 5.02 mfu: 0.51% [rank1]:[titan] 2025-10-15 10:37:01,391 - root - INFO - step: 1 loss: 12.2465 grad_norm: 4.5253 memory: 33.21GiB(34.95%) tps: 87 tflops: 5.03 mfu: 0.51% [rank2]:[titan] 2025-10-15 10:37:01,391 - root - INFO - step: 1 loss: 12.2465 grad_norm: 4.5253 memory: 33.21GiB(34.95%) tps: 86 tflops: 4.99 mfu: 0.50% [rank2]:[titan] 2025-10-15 10:37:07,957 - root - INFO - step: 10 loss: 10.2391 grad_norm: 16.1413 memory: 44.84GiB(47.19%) tps: 5,615 tflops: 325.20 mfu: 32.88% [rank3]:[titan] 2025-10-15 10:37:07,957 - root - INFO - step: 10 loss: 10.2391 grad_norm: 16.1413 memory: 44.84GiB(47.19%) tps: 5,615 tflops: 325.20 mfu: 32.88% [rank0]:[titan] 2025-10-15 10:37:07,957 - root - INFO - step: 10 loss: 10.2391 grad_norm: 16.1413 memory: 44.84GiB(47.19%) tps: 5,616 tflops: 325.22 mfu: 32.88% [rank1]:[titan] 2025-10-15 10:37:07,956 - root - INFO - step: 10 loss: 10.2391 grad_norm: 16.1413 memory: 44.84GiB(47.19%) tps: 5,615 tflops: 325.20 mfu: 32.88% [rank3]:[titan] 2025-10-15 10:37:15,308 - root - INFO - step: 20 loss: 8.5913 grad_norm: 7.1942 memory: 44.84GiB(47.19%) tps: 5,573 tflops: 322.73 mfu: 32.63% [rank0]:[titan] 2025-10-15 10:37:15,308 - root - INFO - step: 20 loss: 8.5913 grad_norm: 7.1942 memory: 44.84GiB(47.19%) tps: 5,573 tflops: 322.74 mfu: 32.63% [rank1]:[titan] 2025-10-15 10:37:15,308 - root - INFO - step: 20 loss: 8.5913 grad_norm: 7.1942 memory: 44.84GiB(47.19%) tps: 5,573 tflops: 322.73 mfu: 32.63% [rank2]:[titan] 2025-10-15 10:37:15,308 - root - INFO - step: 20 loss: 8.5913 grad_norm: 7.1942 memory: 44.84GiB(47.19%) tps: 5,573 tflops: 322.73 mfu: 32.63% [rank3]:[titan] 2025-10-15 10:37:22,689 - root - INFO - step: 30 loss: 7.7257 grad_norm: 2.7261 memory: 44.84GiB(47.19%) tps: 5,550 tflops: 321.44 mfu: 32.50% [rank0]:[titan] 2025-10-15 10:37:22,689 - root - INFO - step: 30 loss: 7.7257 grad_norm: 2.7261 memory: 44.84GiB(47.19%) tps: 5,550 tflops: 321.44 mfu: 32.50% [rank1]:[titan] 2025-10-15 10:37:22,689 - root - INFO - step: 30 loss: 7.7257 grad_norm: 2.7261 memory: 44.84GiB(47.19%) tps: 5,550 tflops: 321.44 mfu: 32.50% [rank2]:[titan] 2025-10-15 10:37:22,689 - root - INFO - step: 30 loss: 7.7257 grad_norm: 2.7261 memory: 44.84GiB(47.19%) tps: 5,550 tflops: 321.44 mfu: 32.50% [rank3]:[titan] 2025-10-15 10:37:30,028 - root - INFO - step: 40 loss: 7.4543 grad_norm: 3.6042 memory: 44.84GiB(47.19%) tps: 5,582 tflops: 323.26 mfu: 32.69% [rank0]:[titan] 2025-10-15 10:37:30,028 - root - INFO - step: 40 loss: 7.4543 grad_norm: 3.6042 memory: 44.84GiB(47.19%) tps: 5,582 tflops: 323.27 mfu: 32.69% [rank1]:[titan] 2025-10-15 10:37:30,028 - root - INFO - step: 40 loss: 7.4543 grad_norm: 3.6042 memory: 44.84GiB(47.19%) tps: 5,582 tflops: 323.25 mfu: 32.69% ``` [ghstack-poisoned]
This was referenced Oct 16, 2025
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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.
Stack from ghstack (oldest at bottom):
Summary
Enable llama3-8B model to use
flex_attention.Test
CONFIG_FILE=./torchtitan/models/llama3/train_configs/llama3_8b.toml NGPU=4 LOG_RANK=0,1,2,3 ./run_train.sh --model.flavor "8B_flex_attn" --activation_checkpoint.mode "full" --parallelism.context_parallel_degree 2