[simplefsdp auto-bucketing] ir node runtime estimation #157572
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
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/157572
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New Failure, 2 Unrelated FailuresAs of commit 4a0c2e8 with merge base 05c19d1 ( NEW FAILURE - The following job has failed:
BROKEN TRUNK - The following jobs failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This PR needs a
|
bdhirsh
reviewed
Jul 7, 2025
| process_group = c10d.distributed_c10d._get_default_group() | ||
|
|
||
| tensor_input = create_real_tensor(inputs.data.get_size(), inputs.data.get_dtype(), inputs.data.get_device()) | ||
| tensor_output = create_real_tensor(kernel.layout.size, kernel.layout.dtype, kernel.layout.device) |
There was a problem hiding this comment.
one downside to allocating real tensors at compile time to do the comms estimation is that if we are already close to peak memory at the time that we do the estimation, we risk compile causing an OOM. IIRC we've seen similar problems before with inductor, since it allocates tensors to do autotuning.
Mostly just calling this out as a potential downside. Although maybe this is still the best option, given that your work showed this is noticeably more accurate than other forms of comms estimation? Runtime PGO is another option but it is a much bigger lift to set up. Or just brainstorming: maybe we can also avoid these allocations being problematic by either offloading other tensors to CPU, or e.g. doing the runtime estimation with small tensor shapes, and having a way to accurately estimate the runtime for larger shapes.
There was a problem hiding this comment.
These tensors are deleted immediately in L126 with del tensor_input, tensor_output. Do you think there might be potential memory overhead even after they are deleted? 👀
wconstab
reviewed
Jul 7, 2025
| end_evt.record() | ||
| end_evt.synchronize() | ||
|
|
||
| torch.cuda.synchronize() |
There was a problem hiding this comment.
nit: doing cuda sync / event sync should not be necessary to get timings from the event. I think i'd suggest just removing them.
There was a problem hiding this comment.
We will need end_evt.synchronize() to ensure the end_event has finished before calculating the runtime in start_evt.elapsed_time(end_evt). Otherwise, it will throw error. I think torch.cuda.synchronize() is not necessary here tho.
This was referenced Jul 11, 2025
Git-Hub-Chris
pushed a commit
to Git-Hub-Chris/PyTorch
that referenced
this pull request
Jul 15, 2025
ghstack-source-id: 40c2647 Pull-Request: pytorch/pytorch#157572
Computation Estimation: - BaseSchedulerNode/FusedSchedulerNode: Get the generated triton code and use [do_bench](https://github.com/pytorch/pytorch/blob/85111cd165f108ffabb4a90083d59d7a867ebd9f/torch/_inductor/codegen/triton.py#L4234) mode to benchmark the runtime. - ExternKernelSchedulerNode: Realize faketensor into realtensor, and log the cuda event time. Communication Estimation: - NCCL estimation: estimate communication using NCCL's [ncclGroupSimulateEnd API](https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/api/group.html?highlight=ncclsiminfo_t) from this [PR](#149343). - Profiling-based estimation: Realize faketensor into realtensor, and log the cuda event time for performing the communication. Note: By default, I use Profiling-based estimation to estimate communication. The reason is it gives better estimation performance in multi-node settings. At the end, it aggregates the estimated runtime from all ranks, and take the median runtime as the final estimated results. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov [ghstack-poisoned]
ruisizhang123
added a commit
that referenced
this pull request
Aug 11, 2025
…cket helper function" This pr is based on Diff D67292294 from yf225. Major changes are: - Change the function structure to be compatible with auto-bucketing - Group bucketed nodes & dependencies with GroupedSchedulerNodes for easier reordering. Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #160282 * #158609 * #158321 * #158098 * __->__ #158097 * #157572 cc H-Huang awgu wanchaol fegin fduwjj wz337 wconstab d4l3k voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov [ghstack-poisoned]
ruisizhang123
added a commit
that referenced
this pull request
Aug 11, 2025
…tion" This pr is based on Diff D67292294 from yf225. Major changes are: - Change the function structure to be compatible with auto-bucketing - Group bucketed nodes & dependencies with GroupedSchedulerNodes for easier reordering. Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #160282 * #158609 * #158321 * #158098 * __->__ #158097 * #157572 cc H-Huang awgu wanchaol fegin fduwjj wz337 wconstab d4l3k voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov [ghstack-poisoned]
ruisizhang123
added a commit
that referenced
this pull request
Aug 11, 2025
… with greedy algorithm" ## Greedy Algorithm Design ### Runtime estimation estimation is done in [__->__ #157572]. #### Communication estimation: - Optimization: While realize communication data into realtensor gives us good estimation of a node, it could incur significant overhead in auto-bucketing. - Calibration-based estimation: We sample 20 samples evenly in a range of [min(fwd_ag_tensor_list), 0.3* sum(fwd_ag_tensor_list)], and benchmark its communication time. This also applies to backward reduce scatter tensors. - When a new IR comes, it will search the saved communication dictionary for the closest tensor size, and use the closest tensor runtime as its predicted runtime. #### Cache estimated runtime Here, we add a cache to save pre-estimated results. Leverage a comm cache dictionary & comp cache dictionary to save pre-estimated inductor ir. If it meets a new IR node with the same key, it will skip the estimation for the new IR node. - For comm, the key is communication type, input tensor size & output tensor size. - For comp, the key is (i) generated triton code for BaseSchedulerNode/FusedSchedulerNode; (ii) Extern Kernel args ### Greedy Algorithm Implementation Core idea: the greedy algorithm decide if the node will be bucketed with the previous one based on several criteria below. The reordering will by itself reorder with the previous computation. bucketing is done in [__->__ #158097] reordering is done in [__->__ #158098] #### FWD Pass: - (i) the bucketed AG communication could be overlapped by the previous computation; - (ii) the bucketed AG copy-in/out memory doesn’t exceed peak memory; - (iii) bucketed AG communication size doesn’t exceed 0.3* sum(fwd_ag_tensor_list), such that the estimated AG communication time is always in the calibration bound. #### BWD Pass: - (i) the bucketed AG + RS communication could be overlapped by the previous computation; - (ii) the bucketed AG+RS copy-in/out memory doesn’t exceed peak memory; - (iii) RS always have future compute to overlap it, such that its final exposed communication is small; - (iv) bucketed AG/RS communication size doesn’t exceed 0.3* sum(fwd_ag_tensor_list)/0.3* sum(bwd_rs_tensor_list), such that the estimated AG/RS communication time is always in the calibration bound. cc H-Huang awgu wanchaol fegin fduwjj wz337 wconstab d4l3k pragupta voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben [ghstack-poisoned]
ruisizhang123
added a commit
that referenced
this pull request
Aug 11, 2025
…orithm" ## Greedy Algorithm Design ### Runtime estimation estimation is done in [__->__ #157572]. #### Communication estimation: - Optimization: While realize communication data into realtensor gives us good estimation of a node, it could incur significant overhead in auto-bucketing. - Calibration-based estimation: We sample 20 samples evenly in a range of [min(fwd_ag_tensor_list), 0.3* sum(fwd_ag_tensor_list)], and benchmark its communication time. This also applies to backward reduce scatter tensors. - When a new IR comes, it will search the saved communication dictionary for the closest tensor size, and use the closest tensor runtime as its predicted runtime. #### Cache estimated runtime Here, we add a cache to save pre-estimated results. Leverage a comm cache dictionary & comp cache dictionary to save pre-estimated inductor ir. If it meets a new IR node with the same key, it will skip the estimation for the new IR node. - For comm, the key is communication type, input tensor size & output tensor size. - For comp, the key is (i) generated triton code for BaseSchedulerNode/FusedSchedulerNode; (ii) Extern Kernel args ### Greedy Algorithm Implementation Core idea: the greedy algorithm decide if the node will be bucketed with the previous one based on several criteria below. The reordering will by itself reorder with the previous computation. bucketing is done in [__->__ #158097] reordering is done in [__->__ #158098] #### FWD Pass: - (i) the bucketed AG communication could be overlapped by the previous computation; - (ii) the bucketed AG copy-in/out memory doesn’t exceed peak memory; - (iii) bucketed AG communication size doesn’t exceed 0.3* sum(fwd_ag_tensor_list), such that the estimated AG communication time is always in the calibration bound. #### BWD Pass: - (i) the bucketed AG + RS communication could be overlapped by the previous computation; - (ii) the bucketed AG+RS copy-in/out memory doesn’t exceed peak memory; - (iii) RS always have future compute to overlap it, such that its final exposed communication is small; - (iv) bucketed AG/RS communication size doesn’t exceed 0.3* sum(fwd_ag_tensor_list)/0.3* sum(bwd_rs_tensor_list), such that the estimated AG/RS communication time is always in the calibration bound. cc H-Huang awgu wanchaol fegin fduwjj wz337 wconstab d4l3k pragupta voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben [ghstack-poisoned]
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.
Computation Estimation:
Communication Estimation:
Note: By default, I use Profiling-based estimation to estimate communication. The reason is it gives better estimation performance in multi-node settings.
At the end, it aggregates the estimated runtime from all ranks, and take the median runtime as the final estimated results.
Stack from ghstack (oldest at bottom):
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @chenyang78 @kadeng @muchulee8 @amjames @chauhang @aakhundov