[Inductor] auto-chunker #136702
Open
[Inductor] auto-chunker #136702
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]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/136702
Note: Links to docs will display an error until the docs builds have been completed. ❌ 2 New Failures, 1 Unrelated FailureAs of commit 4fb4e55 with merge base e8de914 ( NEW FAILURES - The following jobs have failed:
FLAKY - The following job failed but was likely due to flakiness present on trunk:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
This was referenced Oct 19, 2024
cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
This was referenced Oct 23, 2024
cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
Contributor
|
@shunting314 @jansel should we try to get this in? what would it take? i can take a pass reviewing it |
Contributor
Author
|
@v0i0 thanks for offering review. That would be great! I just need find bandwidth to resolve comments, do more tests and get this in. |
Contributor
|
Yeah, I think we should get this landed. Re-request review when you want me to take another look. |
v0i0
reviewed
Dec 12, 2025
Contributor
v0i0
left a comment
v0i0
left a comment
There was a problem hiding this comment.
this is amazing! what does the generated code look like? why is this a bunch of invoke_subgraph rather than vmap w/ chunk_size or hop.scan?
| aten.neg.default, | ||
| ] | ||
| ) | ||
| def propagate_general_copy_metadata(out_node, ignore_broadcast=False): |
Contributor
There was a problem hiding this comment.
while i admittedly have no idea how to do it, this seems like a special case of propagating along vmap/batching rules. if they exist somewhere, you could probably cut this down a lot.
Contributor
Author
For linear+cross entropy , the fx graph before chunking is : https://gist.github.com/shunting314/231e4e5c15923f3d16e069768628daaf , the fx graph after chunking is: https://gist.github.com/shunting314/2583db2ea0ea36bdad4ed1a5732bbc49 . |
Contributor
Author
Thanks for mentioning the similarity with vmap. yea, maybe some functionality can be borrowed/shared. I'm not sure. But for auto-chunker, the different chunks have to be handled serially so we reuse intermediate buffers. hop.scan was not available when I started prototyping auto-chunker. |
The AutoChunker defines the following chunking metadata and propagate them thru the subgraphs that we chunk 1. scale_by: The AutoChunker is only enabled if there is a single scalar tangent. To decouple the dependency on tangent for the bwd subgraph so we can compute them in fwd, we pretend the tangent is 1 first and record it in 'scale_by' . This metadata get propagated and when we cancel the chunking effect in the end of bwd subgraph, we apply the scaling. 2. chunk_dim: record which dimension of the tensor get chunked 3. need_sum: if true, the original Tensor is the sum (rather than concat) of each chunked tensors. One important implementation detail is, we need put chunked subgraph in a HOP (use invoke_subgraph here). Otherwise Inductor fuse across these subgraphs and results in no peak memory saving. This is still a prototype since I assume all chunked input for the chunking subgraph are chunked at the same dimension. But this can be not true. As discussed offline with Jason and Horace, a principled way to resolve this is to propagate the chunking metadata in the backward direction. Here are some early benchmarking result on GPT2. - 64 chunks: - final 19 iters avg: 242.550ms - peak memory consumption: 12603 MiB - 32 chunks: - final 19 iters avg: 206.180ms - peak memory consumption: 12880 MiB - 16 chunks - final 19 iters avg: 196.997ms - peak memory consumption: 13267 MiB - 8 chunks - final 19 iters avg: 194.924ms - peak memory consumption: 14049 MiB With 64 chunks, our peak memory is smaller than llm.c's 13.4GB. I also tried the AutoChunker on PT2 OSS benchmarks to verify the numerical. By default our accuracy test picks a very small batch size. This makes AutoChunker get skipped. I force batch_size to be 16 for BertForMaskedLM to trigger the AutoChunker and verified the numerical correctness. FYI jansel Chillee eellison . Will send an update when this is fully ready for review after I resolve the hacky things mentioned above. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 amjames chauhang aakhundov ColinPeppler desertfire ngimel [ghstack-poisoned]
Contributor
Amazing! Would it be possible to run this against a bunch of training code (e.g. torchtitan w/ num_gpus=1 & torchbench) to see if anything is missing / get some more memory savings data? |
Contributor
Author
Sure. That's actually the tests I plan to do and share results. torchtitan/torchtune are both good choices. I'll share more results Our dashboard (HF/torchbench/TIMM) may be stretchy to test AutoChunker since for the small workload, AutoChunker most likely will be skipped. |
The AutoChunker defines the following chunking metadata and propagate them thru the subgraphs that we chunk 1. scale_by: The AutoChunker is only enabled if there is a single scalar tangent. To decouple the dependency on tangent for the bwd subgraph so we can compute them in fwd, we pretend the tangent is 1 first and record it in 'scale_by' . This metadata get propagated and when we cancel the chunking effect in the end of bwd subgraph, we apply the scaling. 2. chunk_dim: record which dimension of the tensor get chunked 3. need_sum: if true, the original Tensor is the sum (rather than concat) of each chunked tensors. One important implementation detail is, we need put chunked subgraph in a HOP (use invoke_subgraph here). Otherwise Inductor fuse across these subgraphs and results in no peak memory saving. This is still a prototype since I assume all chunked input for the chunking subgraph are chunked at the same dimension. But this can be not true. As discussed offline with Jason and Horace, a principled way to resolve this is to propagate the chunking metadata in the backward direction. Here are some early benchmarking result on GPT2. - 64 chunks: - final 19 iters avg: 242.550ms - peak memory consumption: 12603 MiB - 32 chunks: - final 19 iters avg: 206.180ms - peak memory consumption: 12880 MiB - 16 chunks - final 19 iters avg: 196.997ms - peak memory consumption: 13267 MiB - 8 chunks - final 19 iters avg: 194.924ms - peak memory consumption: 14049 MiB With 64 chunks, our peak memory is smaller than llm.c's 13.4GB. I also tried the AutoChunker on PT2 OSS benchmarks to verify the numerical. By default our accuracy test picks a very small batch size. This makes AutoChunker get skipped. I force batch_size to be 16 for BertForMaskedLM to trigger the AutoChunker and verified the numerical correctness. FYI jansel Chillee eellison . Will send an update when this is fully ready for review after I resolve the hacky things mentioned above. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 amjames chauhang aakhundov ColinPeppler desertfire ngimel [ghstack-poisoned]
The AutoChunker defines the following chunking metadata and propagate them thru the subgraphs that we chunk 1. scale_by: The AutoChunker is only enabled if there is a single scalar tangent. To decouple the dependency on tangent for the bwd subgraph so we can compute them in fwd, we pretend the tangent is 1 first and record it in 'scale_by' . This metadata get propagated and when we cancel the chunking effect in the end of bwd subgraph, we apply the scaling. 2. chunk_dim: record which dimension of the tensor get chunked 3. need_sum: if true, the original Tensor is the sum (rather than concat) of each chunked tensors. One important implementation detail is, we need put chunked subgraph in a HOP (use invoke_subgraph here). Otherwise Inductor fuse across these subgraphs and results in no peak memory saving. This is still a prototype since I assume all chunked input for the chunking subgraph are chunked at the same dimension. But this can be not true. As discussed offline with Jason and Horace, a principled way to resolve this is to propagate the chunking metadata in the backward direction. Here are some early benchmarking result on GPT2. - 64 chunks: - final 19 iters avg: 242.550ms - peak memory consumption: 12603 MiB - 32 chunks: - final 19 iters avg: 206.180ms - peak memory consumption: 12880 MiB - 16 chunks - final 19 iters avg: 196.997ms - peak memory consumption: 13267 MiB - 8 chunks - final 19 iters avg: 194.924ms - peak memory consumption: 14049 MiB With 64 chunks, our peak memory is smaller than llm.c's 13.4GB. I also tried the AutoChunker on PT2 OSS benchmarks to verify the numerical. By default our accuracy test picks a very small batch size. This makes AutoChunker get skipped. I force batch_size to be 16 for BertForMaskedLM to trigger the AutoChunker and verified the numerical correctness. FYI jansel Chillee eellison . Will send an update when this is fully ready for review after I resolve the hacky things mentioned above. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 amjames chauhang aakhundov ColinPeppler desertfire ngimel [ghstack-poisoned]
The AutoChunker defines the following chunking metadata and propagate them thru the subgraphs that we chunk 1. scale_by: The AutoChunker is only enabled if there is a single scalar tangent. To decouple the dependency on tangent for the bwd subgraph so we can compute them in fwd, we pretend the tangent is 1 first and record it in 'scale_by' . This metadata get propagated and when we cancel the chunking effect in the end of bwd subgraph, we apply the scaling. 2. chunk_dim: record which dimension of the tensor get chunked 3. need_sum: if true, the original Tensor is the sum (rather than concat) of each chunked tensors. One important implementation detail is, we need put chunked subgraph in a HOP (use invoke_subgraph here). Otherwise Inductor fuse across these subgraphs and results in no peak memory saving. This is still a prototype since I assume all chunked input for the chunking subgraph are chunked at the same dimension. But this can be not true. As discussed offline with Jason and Horace, a principled way to resolve this is to propagate the chunking metadata in the backward direction. Here are some early benchmarking result on GPT2. - 64 chunks: - final 19 iters avg: 242.550ms - peak memory consumption: 12603 MiB - 32 chunks: - final 19 iters avg: 206.180ms - peak memory consumption: 12880 MiB - 16 chunks - final 19 iters avg: 196.997ms - peak memory consumption: 13267 MiB - 8 chunks - final 19 iters avg: 194.924ms - peak memory consumption: 14049 MiB With 64 chunks, our peak memory is smaller than llm.c's 13.4GB. I also tried the AutoChunker on PT2 OSS benchmarks to verify the numerical. By default our accuracy test picks a very small batch size. This makes AutoChunker get skipped. I force batch_size to be 16 for BertForMaskedLM to trigger the AutoChunker and verified the numerical correctness. FYI jansel Chillee eellison . Will send an update when this is fully ready for review after I resolve the hacky things mentioned above. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 amjames chauhang aakhundov ColinPeppler desertfire ngimel [ghstack-poisoned]
The AutoChunker defines the following chunking metadata and propagate them thru the subgraphs that we chunk 1. scale_by: The AutoChunker is only enabled if there is a single scalar tangent. To decouple the dependency on tangent for the bwd subgraph so we can compute them in fwd, we pretend the tangent is 1 first and record it in 'scale_by' . This metadata get propagated and when we cancel the chunking effect in the end of bwd subgraph, we apply the scaling. 2. chunk_dim: record which dimension of the tensor get chunked 3. need_sum: if true, the original Tensor is the sum (rather than concat) of each chunked tensors. One important implementation detail is, we need put chunked subgraph in a HOP (use invoke_subgraph here). Otherwise Inductor fuse across these subgraphs and results in no peak memory saving. This is still a prototype since I assume all chunked input for the chunking subgraph are chunked at the same dimension. But this can be not true. As discussed offline with Jason and Horace, a principled way to resolve this is to propagate the chunking metadata in the backward direction. Here are some early benchmarking result on GPT2. - 64 chunks: - final 19 iters avg: 242.550ms - peak memory consumption: 12603 MiB - 32 chunks: - final 19 iters avg: 206.180ms - peak memory consumption: 12880 MiB - 16 chunks - final 19 iters avg: 196.997ms - peak memory consumption: 13267 MiB - 8 chunks - final 19 iters avg: 194.924ms - peak memory consumption: 14049 MiB With 64 chunks, our peak memory is smaller than llm.c's 13.4GB. I also tried the AutoChunker on PT2 OSS benchmarks to verify the numerical. By default our accuracy test picks a very small batch size. This makes AutoChunker get skipped. I force batch_size to be 16 for BertForMaskedLM to trigger the AutoChunker and verified the numerical correctness. FYI jansel Chillee eellison . Will send an update when this is fully ready for review after I resolve the hacky things mentioned above. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 amjames chauhang aakhundov ColinPeppler desertfire ngimel [ghstack-poisoned]
The AutoChunker defines the following chunking metadata and propagate them thru the subgraphs that we chunk 1. scale_by: The AutoChunker is only enabled if there is a single scalar tangent. To decouple the dependency on tangent for the bwd subgraph so we can compute them in fwd, we pretend the tangent is 1 first and record it in 'scale_by' . This metadata get propagated and when we cancel the chunking effect in the end of bwd subgraph, we apply the scaling. 2. chunk_dim: record which dimension of the tensor get chunked 3. need_sum: if true, the original Tensor is the sum (rather than concat) of each chunked tensors. One important implementation detail is, we need put chunked subgraph in a HOP (use invoke_subgraph here). Otherwise Inductor fuse across these subgraphs and results in no peak memory saving. This is still a prototype since I assume all chunked input for the chunking subgraph are chunked at the same dimension. But this can be not true. As discussed offline with Jason and Horace, a principled way to resolve this is to propagate the chunking metadata in the backward direction. Here are some early benchmarking result on GPT2. - 64 chunks: - final 19 iters avg: 242.550ms - peak memory consumption: 12603 MiB - 32 chunks: - final 19 iters avg: 206.180ms - peak memory consumption: 12880 MiB - 16 chunks - final 19 iters avg: 196.997ms - peak memory consumption: 13267 MiB - 8 chunks - final 19 iters avg: 194.924ms - peak memory consumption: 14049 MiB With 64 chunks, our peak memory is smaller than llm.c's 13.4GB. I also tried the AutoChunker on PT2 OSS benchmarks to verify the numerical. By default our accuracy test picks a very small batch size. This makes AutoChunker get skipped. I force batch_size to be 16 for BertForMaskedLM to trigger the AutoChunker and verified the numerical correctness. FYI jansel Chillee eellison . Will send an update when this is fully ready for review after I resolve the hacky things mentioned above. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 amjames chauhang aakhundov ColinPeppler desertfire ngimel [ghstack-poisoned]
The AutoChunker defines the following chunking metadata and propagate them thru the subgraphs that we chunk 1. scale_by: The AutoChunker is only enabled if there is a single scalar tangent. To decouple the dependency on tangent for the bwd subgraph so we can compute them in fwd, we pretend the tangent is 1 first and record it in 'scale_by' . This metadata get propagated and when we cancel the chunking effect in the end of bwd subgraph, we apply the scaling. 2. chunk_dim: record which dimension of the tensor get chunked 3. need_sum: if true, the original Tensor is the sum (rather than concat) of each chunked tensors. One important implementation detail is, we need put chunked subgraph in a HOP (use invoke_subgraph here). Otherwise Inductor fuse across these subgraphs and results in no peak memory saving. This is still a prototype since I assume all chunked input for the chunking subgraph are chunked at the same dimension. But this can be not true. As discussed offline with Jason and Horace, a principled way to resolve this is to propagate the chunking metadata in the backward direction. Here are some early benchmarking result on GPT2. - 64 chunks: - final 19 iters avg: 242.550ms - peak memory consumption: 12603 MiB - 32 chunks: - final 19 iters avg: 206.180ms - peak memory consumption: 12880 MiB - 16 chunks - final 19 iters avg: 196.997ms - peak memory consumption: 13267 MiB - 8 chunks - final 19 iters avg: 194.924ms - peak memory consumption: 14049 MiB With 64 chunks, our peak memory is smaller than llm.c's 13.4GB. I also tried the AutoChunker on PT2 OSS benchmarks to verify the numerical. By default our accuracy test picks a very small batch size. This makes AutoChunker get skipped. I force batch_size to be 16 for BertForMaskedLM to trigger the AutoChunker and verified the numerical correctness. FYI jansel Chillee eellison . Will send an update when this is fully ready for review after I resolve the hacky things mentioned above. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 amjames chauhang aakhundov ColinPeppler desertfire ngimel [ghstack-poisoned]
The AutoChunker defines the following chunking metadata and propagate them thru the subgraphs that we chunk 1. scale_by: The AutoChunker is only enabled if there is a single scalar tangent. To decouple the dependency on tangent for the bwd subgraph so we can compute them in fwd, we pretend the tangent is 1 first and record it in 'scale_by' . This metadata get propagated and when we cancel the chunking effect in the end of bwd subgraph, we apply the scaling. 2. chunk_dim: record which dimension of the tensor get chunked 3. need_sum: if true, the original Tensor is the sum (rather than concat) of each chunked tensors. One important implementation detail is, we need put chunked subgraph in a HOP (use invoke_subgraph here). Otherwise Inductor fuse across these subgraphs and results in no peak memory saving. This is still a prototype since I assume all chunked input for the chunking subgraph are chunked at the same dimension. But this can be not true. As discussed offline with Jason and Horace, a principled way to resolve this is to propagate the chunking metadata in the backward direction. Here are some early benchmarking result on GPT2. - 64 chunks: - final 19 iters avg: 242.550ms - peak memory consumption: 12603 MiB - 32 chunks: - final 19 iters avg: 206.180ms - peak memory consumption: 12880 MiB - 16 chunks - final 19 iters avg: 196.997ms - peak memory consumption: 13267 MiB - 8 chunks - final 19 iters avg: 194.924ms - peak memory consumption: 14049 MiB With 64 chunks, our peak memory is smaller than llm.c's 13.4GB. I also tried the AutoChunker on PT2 OSS benchmarks to verify the numerical. By default our accuracy test picks a very small batch size. This makes AutoChunker get skipped. I force batch_size to be 16 for BertForMaskedLM to trigger the AutoChunker and verified the numerical correctness. FYI jansel Chillee eellison . Will send an update when this is fully ready for review after I resolve the hacky things mentioned above. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 amjames chauhang aakhundov ColinPeppler desertfire ngimel [ghstack-poisoned]
The AutoChunker defines the following chunking metadata and propagate them thru the subgraphs that we chunk 1. scale_by: The AutoChunker is only enabled if there is a single scalar tangent. To decouple the dependency on tangent for the bwd subgraph so we can compute them in fwd, we pretend the tangent is 1 first and record it in 'scale_by' . This metadata get propagated and when we cancel the chunking effect in the end of bwd subgraph, we apply the scaling. 2. chunk_dim: record which dimension of the tensor get chunked 3. need_sum: if true, the original Tensor is the sum (rather than concat) of each chunked tensors. One important implementation detail is, we need put chunked subgraph in a HOP (use invoke_subgraph here). Otherwise Inductor fuse across these subgraphs and results in no peak memory saving. This is still a prototype since I assume all chunked input for the chunking subgraph are chunked at the same dimension. But this can be not true. As discussed offline with Jason and Horace, a principled way to resolve this is to propagate the chunking metadata in the backward direction. Here are some early benchmarking result on GPT2. - 64 chunks: - final 19 iters avg: 242.550ms - peak memory consumption: 12603 MiB - 32 chunks: - final 19 iters avg: 206.180ms - peak memory consumption: 12880 MiB - 16 chunks - final 19 iters avg: 196.997ms - peak memory consumption: 13267 MiB - 8 chunks - final 19 iters avg: 194.924ms - peak memory consumption: 14049 MiB With 64 chunks, our peak memory is smaller than llm.c's 13.4GB. I also tried the AutoChunker on PT2 OSS benchmarks to verify the numerical. By default our accuracy test picks a very small batch size. This makes AutoChunker get skipped. I force batch_size to be 16 for BertForMaskedLM to trigger the AutoChunker and verified the numerical correctness. FYI jansel Chillee eellison . Will send an update when this is fully ready for review after I resolve the hacky things mentioned above. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 amjames chauhang aakhundov ColinPeppler desertfire ngimel [ghstack-poisoned]
The AutoChunker defines the following chunking metadata and propagate them thru the subgraphs that we chunk 1. scale_by: The AutoChunker is only enabled if there is a single scalar tangent. To decouple the dependency on tangent for the bwd subgraph so we can compute them in fwd, we pretend the tangent is 1 first and record it in 'scale_by' . This metadata get propagated and when we cancel the chunking effect in the end of bwd subgraph, we apply the scaling. 2. chunk_dim: record which dimension of the tensor get chunked 3. need_sum: if true, the original Tensor is the sum (rather than concat) of each chunked tensors. One important implementation detail is, we need put chunked subgraph in a HOP (use invoke_subgraph here). Otherwise Inductor fuse across these subgraphs and results in no peak memory saving. This is still a prototype since I assume all chunked input for the chunking subgraph are chunked at the same dimension. But this can be not true. As discussed offline with Jason and Horace, a principled way to resolve this is to propagate the chunking metadata in the backward direction. Here are some early benchmarking result on GPT2. - 64 chunks: - final 19 iters avg: 242.550ms - peak memory consumption: 12603 MiB - 32 chunks: - final 19 iters avg: 206.180ms - peak memory consumption: 12880 MiB - 16 chunks - final 19 iters avg: 196.997ms - peak memory consumption: 13267 MiB - 8 chunks - final 19 iters avg: 194.924ms - peak memory consumption: 14049 MiB With 64 chunks, our peak memory is smaller than llm.c's 13.4GB. I also tried the AutoChunker on PT2 OSS benchmarks to verify the numerical. By default our accuracy test picks a very small batch size. This makes AutoChunker get skipped. I force batch_size to be 16 for BertForMaskedLM to trigger the AutoChunker and verified the numerical correctness. FYI jansel Chillee eellison . Will send an update when this is fully ready for review after I resolve the hacky things mentioned above. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 amjames chauhang aakhundov ColinPeppler desertfire ngimel [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.
Stack from ghstack (oldest at bottom):
The AutoChunker defines the following chunking metadata and propagate them thru the subgraphs that we chunk
One important implementation detail is, we need put chunked subgraph in a HOP (use invoke_subgraph here). Otherwise Inductor fuse across these subgraphs and results in no peak memory saving.
This is still a prototype since I assume all chunked input for the chunking subgraph are chunked at the same dimension. But this can be not true. As discussed offline with Jason and Horace, a principled way to resolve this is to propagate the chunking metadata in the backward direction.
Here are some early benchmarking result on GPT2.
With 64 chunks, our peak memory is smaller than llm.c's 13.4GB.
I also tried the AutoChunker on PT2 OSS benchmarks to verify the numerical. By default our accuracy test picks a very small batch size. This makes AutoChunker get skipped. I force batch_size to be 16 for BertForMaskedLM to trigger the AutoChunker and verified the numerical correctness.
FYI @jansel @Chillee @eellison . Will send an update when this is fully ready for review after I resolve the hacky things mentioned above.
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @amjames @chauhang @aakhundov @ColinPeppler @desertfire @ngimel