Introduce cuda_p2p based fused_all_gather_matmul and fused_matmul_reduce_scatter #126634
Conversation
This file contains 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
…uce_scatter [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/126634
Note: Links to docs will display an error until the docs builds have been completed. ❗ 1 Active SEVsThere are 1 currently active SEVs. If your PR is affected, please view them below: ✅ You can merge normally! (3 Unrelated Failures)As of commit 29e6b1f with merge base ff65b18 ( FLAKY - The following jobs failed but were likely due to flakiness present on trunk:
UNSTABLE - The following job failed but was likely due to flakiness present on trunk and has been marked as unstable:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
pytorch-bot
bot
added
the
oncall: distributed
…_matmul_reduce_scatter" ## Context See context [here](#122163). ## This PR Introduces `cuda_p2p` based `fused_all_gather_matmul` and `fused_matmul_reduce_scatter` dispatcher ops which performs micro-pipelining TP for `all-gather -> matmul` and `matmul -> reduce-scatter` respectively. Fusion vs. decomposition - in principle, the micro-pipelining is achieved via decomposition. However, in practice, today Inductor can't deal with the decomposed patterns well. So instead performing decomposition in Inductor, we fuse the patterns to be decomposed and dispatch them to corresponding operators that handle decomposition + micropipelining. cc mrshenli pritamdamania87 zhaojuanmao satgera gqchen aazzolini osalpekar jiayisuse H-Huang kwen2501 awgu penguinwu fegin XilunWu wanchaol fduwjj wz337 tianyu-l wconstab yf225 chauhang d4l3k [ghstack-poisoned]
Chillee
approved these changes
May 21, 2024
wanchaol
approved these changes
May 21, 2024
|
|
||
|
|
||
| @contextmanager | ||
| def test_with_non_cuda_p2p_group(): |
There was a problem hiding this comment.
nit: these test utils should move to the torch.testing package instead?
| ag_shape = list(A_shard.shape) | ||
| ag_shape[gather_dim] *= group_size | ||
| ag_out = A_shard.new_empty(ag_shape) | ||
| return ag_out, [ag_out @ B for B in Bs] |
There was a problem hiding this comment.
for meta formulas, wondering if this matmul would actually incur computation or just call the matmul meta kernel(i guess it's the later one?)
There was a problem hiding this comment.
Yeah we are calling the meta kernels to deduce device, shape, and strides.
…_matmul_reduce_scatter" ## Context See context [here](#122163). ## This PR Introduces `cuda_p2p` based `fused_all_gather_matmul` and `fused_matmul_reduce_scatter` dispatcher ops which performs micro-pipelining TP for `all-gather -> matmul` and `matmul -> reduce-scatter` respectively. Fusion vs. decomposition - in principle, the micro-pipelining is achieved via decomposition. However, in practice, today Inductor can't deal with the decomposed patterns well. So instead performing decomposition in Inductor, we fuse the patterns to be decomposed and dispatch them to corresponding operators that handle decomposition + micropipelining. cc mrshenli pritamdamania87 zhaojuanmao satgera gqchen aazzolini osalpekar jiayisuse H-Huang kwen2501 awgu penguinwu fegin XilunWu wanchaol fduwjj wz337 tianyu-l wconstab yf225 chauhang d4l3k [ghstack-poisoned]
…_matmul_reduce_scatter" ## Context See context [here](#122163). ## This PR Introduces `cuda_p2p` based `fused_all_gather_matmul` and `fused_matmul_reduce_scatter` dispatcher ops which performs micro-pipelining TP for `all-gather -> matmul` and `matmul -> reduce-scatter` respectively. Fusion vs. decomposition - in principle, the micro-pipelining is achieved via decomposition. However, in practice, today Inductor can't deal with the decomposed patterns well. So instead performing decomposition in Inductor, we fuse the patterns to be decomposed and dispatch them to corresponding operators that handle decomposition + micropipelining. cc mrshenli pritamdamania87 zhaojuanmao satgera gqchen aazzolini osalpekar jiayisuse H-Huang kwen2501 awgu penguinwu fegin XilunWu wanchaol fduwjj wz337 tianyu-l wconstab yf225 chauhang d4l3k [ghstack-poisoned]
…_matmul_reduce_scatter" ## Context See context [here](#122163). ## This PR Introduces `cuda_p2p` based `fused_all_gather_matmul` and `fused_matmul_reduce_scatter` dispatcher ops which performs micro-pipelining TP for `all-gather -> matmul` and `matmul -> reduce-scatter` respectively. Fusion vs. decomposition - in principle, the micro-pipelining is achieved via decomposition. However, in practice, today Inductor can't deal with the decomposed patterns well. So instead performing decomposition in Inductor, we fuse the patterns to be decomposed and dispatch them to corresponding operators that handle decomposition + micropipelining. cc mrshenli pritamdamania87 zhaojuanmao satgera gqchen aazzolini osalpekar jiayisuse H-Huang kwen2501 awgu penguinwu fegin XilunWu wanchaol fduwjj wz337 tianyu-l wconstab yf225 chauhang d4l3k [ghstack-poisoned]
yifuwang
added a commit
that referenced
this pull request
May 21, 2024
…uce_scatter ghstack-source-id: cfada01c278b4ed552914d073147c77aa29e6a04 Pull Request resolved: #126634
…_matmul_reduce_scatter" ## Context See context [here](#122163). ## This PR Introduces `cuda_p2p` based `fused_all_gather_matmul` and `fused_matmul_reduce_scatter` dispatcher ops which performs micro-pipelining TP for `all-gather -> matmul` and `matmul -> reduce-scatter` respectively. Fusion vs. decomposition - in principle, the micro-pipelining is achieved via decomposition. However, in practice, today Inductor can't deal with the decomposed patterns well. So instead performing decomposition in Inductor, we fuse the patterns to be decomposed and dispatch them to corresponding operators that handle decomposition + micropipelining. cc mrshenli pritamdamania87 zhaojuanmao satgera gqchen aazzolini osalpekar jiayisuse H-Huang kwen2501 awgu penguinwu fegin XilunWu wanchaol fduwjj wz337 tianyu-l wconstab yf225 chauhang d4l3k [ghstack-poisoned]
…_matmul_reduce_scatter" ## Context See context [here](#122163). ## This PR Introduces `cuda_p2p` based `fused_all_gather_matmul` and `fused_matmul_reduce_scatter` dispatcher ops which performs micro-pipelining TP for `all-gather -> matmul` and `matmul -> reduce-scatter` respectively. Fusion vs. decomposition - in principle, the micro-pipelining is achieved via decomposition. However, in practice, today Inductor can't deal with the decomposed patterns well. So instead performing decomposition in Inductor, we fuse the patterns to be decomposed and dispatch them to corresponding operators that handle decomposition + micropipelining. cc mrshenli pritamdamania87 zhaojuanmao satgera gqchen aazzolini osalpekar jiayisuse H-Huang kwen2501 awgu penguinwu fegin XilunWu wanchaol fduwjj wz337 tianyu-l wconstab yf225 chauhang d4l3k [ghstack-poisoned]
…_matmul_reduce_scatter" [ghstack-poisoned]
…_matmul_reduce_scatter" [ghstack-poisoned]
|
@pytorchbot merge |
Merge startedYour 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 |
bigfootjon
pushed a commit
that referenced
this pull request
Jun 5, 2024
…uce_scatter (#126634) Pull Request resolved: #126634 Approved by: https://github.com/Chillee, https://github.com/wanchaol (cherry picked from commit 1071437)
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):
cc @mrshenli @pritamdamania87 @zhaojuanmao @satgera @gqchen @aazzolini @osalpekar @jiayisuse @H-Huang @kwen2501 @awgu @penguinwu @fegin @XilunWu @wanchaol @fduwjj @wz337 @tianyu-l @wconstab @yf225 @chauhang @d4l3k