KEMBAR78
[AOTI] Fix a special case compile time data type codegen for sym int variables by YUNQIUGUO · Pull Request #138106 · pytorch/pytorch · GitHub
Skip to content

[AOTI] Fix a special case compile time data type codegen for sym int variables #138106

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Closed
wants to merge 1 commit into from
Closed

[AOTI] Fix a special case compile time data type codegen for sym int variables #138106

wants to merge 1 commit into from

Conversation

@YUNQIUGUO
Copy link
Contributor

@YUNQIUGUO YUNQIUGUO commented Oct 16, 2024 •
edited by pytorch-bot bot
Loading

Summary:
This change unblocks the CFR AOTI lowering runtime error.

TL;DR:

In this model, one triton kernel expects a scalar input dtype as i64, but getting an i32. The reason is "auto" can infer a smaller data type if the variable it passed in e.g. is i32. thus cause CUDA IMA.
Original problematic kernel: triton_poi_fused_add_ge_logical_and_logical_or_lt_46_grid_100.

This diff manually cast it to i64 for all symbolic arguments in compile time for i64 triton kernel inputs, instead of use auto var_x = {arg} in cpp wrapper code.

Test Plan:
Verified in FLB locally:

PYTORCH_NO_CUDA_MEMORY_CACHING=1 AOT_INDUCTOR_DEBUG_INTERMEDIATE_VALUE_PRINTER=3 TORCH_LOGS="output_code" TORCHINDUCTOR_MAX_AUTOTUNE=1 TORCH_SHOW_CPP_STACKTRACES=1 CUDA_LAUNCH_BLOCKING=1 ~/fbsource/buck-out/v2/gen/fbcode/98e643f8bb44fe9d/hpc/new/models/feed/benchmark/__feed_lower_benchmark__/feed_lower_benchmark.par --skip-eager --skip-flop-estimation --lower-backend="AOT_INDUCTOR" --sync-mode=0 --precision bf16 --output-precision bf16  --lower-presets="ifr_cint;disable_new_lowering_weights;disable_dper_passes:passes=fuse_parallel_linear_no_weight_change" --remove-unexpected-type-cast=False --load="manifold://ads_storage_fblearner/tree/user/facebook/fblearner/predictor/924293663/0/gpu_lowering/input.merge"```

Differential Revision: D64490039




cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang @aakhundov

@pytorch-bot
Copy link

pytorch-bot bot commented Oct 16, 2024 •
edited
Loading

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/138106

Note: Links to docs will display an error until the docs builds have been completed.

✅ You can merge normally! (1 Unrelated Failure)

As of commit 2267787 with merge base 620039c (image):

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.

@facebook-github-bot
Copy link
Contributor

This pull request was exported from Phabricator. Differential Revision: D64490039

@facebook-github-bot
Copy link
Contributor

This pull request was exported from Phabricator. Differential Revision: D64490039

@YUNQIUGUO YUNQIUGUO added the topic: not user facing topic category label Oct 16, 2024
@facebook-github-bot
Copy link
Contributor

This pull request was exported from Phabricator. Differential Revision: D64490039

YUNQIUGUO added a commit to YUNQIUGUO/pytorch that referenced this pull request Oct 16, 2024
@YUNQIUGUO @facebook-github-bot
[AOTI] Fix a special case compile time data type codegen for sym int …
68df8f6 
…variables (pytorch#138106)

Summary:

This change unblocks the CFR AOTI lowering runtime error.

TL;DR:

In this model, one triton kernel expects a scalar input dtype as i64, but getting an i32. The reason is "auto"  can infer a smaller data type if the variable get passed in e.g. is i32. thus cause CUDA IMA.
 Original problematic kernel: `triton_poi_fused_add_ge_logical_and_logical_or_lt_46_grid_100`. and third input `auto var_402 = u0`.

This diff explicitly specifies it to i64 for all symbolic arguments in compile time for i64 triton kernel inputs, instead of use `auto var_x = {arg}` in cpp wrapper code.

Test Plan:
Verified in FLB locally:

```
PYTORCH_NO_CUDA_MEMORY_CACHING=1 AOT_INDUCTOR_DEBUG_INTERMEDIATE_VALUE_PRINTER=3 TORCH_LOGS="output_code" TORCHINDUCTOR_MAX_AUTOTUNE=1 TORCH_SHOW_CPP_STACKTRACES=1 CUDA_LAUNCH_BLOCKING=1 ~/fbsource/buck-out/v2/gen/fbcode/98e643f8bb44fe9d/hpc/new/models/feed/benchmark/__feed_lower_benchmark__/feed_lower_benchmark.par --skip-eager --skip-flop-estimation --lower-backend="AOT_INDUCTOR" --sync-mode=0 --precision bf16 --output-precision bf16  --lower-presets="ifr_cint;disable_new_lowering_weights;disable_dper_passes:passes=fuse_parallel_linear_no_weight_change" --remove-unexpected-type-cast=False --load="manifold://ads_storage_fblearner/tree/user/facebook/fblearner/predictor/924293663/0/gpu_lowering/input.merge"```

Differential Revision: D64490039
@facebook-github-bot
Copy link
Contributor

This pull request was exported from Phabricator. Differential Revision: D64490039

@facebook-github-bot
Copy link
Contributor

This pull request was exported from Phabricator. Differential Revision: D64490039

YUNQIUGUO added a commit to YUNQIUGUO/pytorch that referenced this pull request Oct 17, 2024
@YUNQIUGUO @facebook-github-bot
[AOTI] Fix a special case compile time data type codegen for sym int …
98dacff 
…variables (pytorch#138106)

Summary:

This change unblocks the CFR AOTI lowering runtime error.

TL;DR:

In this model, one triton kernel expects a scalar input dtype as i64, but getting an i32. The reason is "auto"  can infer a smaller data type if the variable get passed in e.g. is i32. thus cause CUDA IMA.
 Original problematic kernel: `triton_poi_fused_add_ge_logical_and_logical_or_lt_46_grid_100`. and third input `auto var_402 = u0`.

This diff explicitly specifies it to i64 for all symbolic arguments in compile time for i64 triton kernel inputs, instead of use `auto var_x = {arg}` in cpp wrapper code.

Test Plan:
Verified in FLB locally:

```
PYTORCH_NO_CUDA_MEMORY_CACHING=1 AOT_INDUCTOR_DEBUG_INTERMEDIATE_VALUE_PRINTER=3 TORCH_LOGS="output_code" TORCHINDUCTOR_MAX_AUTOTUNE=1 TORCH_SHOW_CPP_STACKTRACES=1 CUDA_LAUNCH_BLOCKING=1 ~/fbsource/buck-out/v2/gen/fbcode/98e643f8bb44fe9d/hpc/new/models/feed/benchmark/__feed_lower_benchmark__/feed_lower_benchmark.par --skip-eager --skip-flop-estimation --lower-backend="AOT_INDUCTOR" --sync-mode=0 --precision bf16 --output-precision bf16  --lower-presets="ifr_cint;disable_new_lowering_weights;disable_dper_passes:passes=fuse_parallel_linear_no_weight_change" --remove-unexpected-type-cast=False --load="manifold://ads_storage_fblearner/tree/user/facebook/fblearner/predictor/924293663/0/gpu_lowering/input.merge"```

Differential Revision: D64490039
@facebook-github-bot
Copy link
Contributor

This pull request was exported from Phabricator. Differential Revision: D64490039

YUNQIUGUO added a commit to YUNQIUGUO/pytorch that referenced this pull request Oct 17, 2024
@YUNQIUGUO @facebook-github-bot
[AOTI] Fix a special case compile time data type codegen for sym int …
4be7f82 
…variables (pytorch#138106)

Summary:

This change unblocks the CFR AOTI lowering runtime error.

TL;DR:

In this model, one triton kernel expects a scalar input dtype as i64, but getting an i32. The reason is "auto"  can infer a smaller data type if the variable get passed in e.g. is i32. thus cause CUDA IMA.
 Original problematic kernel: `triton_poi_fused_add_ge_logical_and_logical_or_lt_46_grid_100`. and third input `auto var_402 = u0`.

This diff explicitly specifies it to i64 for all symbolic arguments in compile time for i64 triton kernel inputs, instead of use `auto var_x = {arg}` in cpp wrapper code.

Test Plan:
Verified in FLB locally:

```
PYTORCH_NO_CUDA_MEMORY_CACHING=1 AOT_INDUCTOR_DEBUG_INTERMEDIATE_VALUE_PRINTER=3 TORCH_LOGS="output_code" TORCHINDUCTOR_MAX_AUTOTUNE=1 TORCH_SHOW_CPP_STACKTRACES=1 CUDA_LAUNCH_BLOCKING=1 ~/fbsource/buck-out/v2/gen/fbcode/98e643f8bb44fe9d/hpc/new/models/feed/benchmark/__feed_lower_benchmark__/feed_lower_benchmark.par --skip-eager --skip-flop-estimation --lower-backend="AOT_INDUCTOR" --sync-mode=0 --precision bf16 --output-precision bf16  --lower-presets="ifr_cint;disable_new_lowering_weights;disable_dper_passes:passes=fuse_parallel_linear_no_weight_change" --remove-unexpected-type-cast=False --load="manifold://ads_storage_fblearner/tree/user/facebook/fblearner/predictor/924293663/0/gpu_lowering/input.merge"```

Differential Revision: D64490039
@facebook-github-bot
Copy link
Contributor

This pull request was exported from Phabricator. Differential Revision: D64490039

@facebook-github-bot
Copy link
Contributor

This pull request was exported from Phabricator. Differential Revision: D64490039

@facebook-github-bot
Copy link
Contributor

This pull request was exported from Phabricator. Differential Revision: D64490039

@facebook-github-bot
Copy link
Contributor

This pull request was exported from Phabricator. Differential Revision: D64490039

YUNQIUGUO added a commit to YUNQIUGUO/pytorch that referenced this pull request Oct 17, 2024
@YUNQIUGUO @facebook-github-bot
[AOTI] Fix a special case compile time data type codegen for sym int …
2d258c8 
…variables (pytorch#138106)

Summary:

This change unblocks the CFR AOTI lowering runtime error.

TL;DR:

In this model, one triton kernel expects a scalar input dtype as i64, but getting an i32. The reason is "auto"  can infer a smaller data type if the variable get passed in e.g. is i32. thus cause CUDA IMA.
 Original problematic kernel: `triton_poi_fused_add_ge_logical_and_logical_or_lt_46_grid_100`. and third input `auto var_402 = u0`.

This diff explicitly specifies it to i64 for all symbolic arguments in compile time for i64 triton kernel inputs, instead of use `auto var_x = {arg}` in cpp wrapper code.

Test Plan:
Verified in FLB locally:

```
PYTORCH_NO_CUDA_MEMORY_CACHING=1 AOT_INDUCTOR_DEBUG_INTERMEDIATE_VALUE_PRINTER=3 TORCH_LOGS="output_code" TORCHINDUCTOR_MAX_AUTOTUNE=1 TORCH_SHOW_CPP_STACKTRACES=1 CUDA_LAUNCH_BLOCKING=1 ~/fbsource/buck-out/v2/gen/fbcode/98e643f8bb44fe9d/hpc/new/models/feed/benchmark/__feed_lower_benchmark__/feed_lower_benchmark.par --skip-eager --skip-flop-estimation --lower-backend="AOT_INDUCTOR" --sync-mode=0 --precision bf16 --output-precision bf16  --lower-presets="ifr_cint;disable_new_lowering_weights;disable_dper_passes:passes=fuse_parallel_linear_no_weight_change" --remove-unexpected-type-cast=False --load="manifold://ads_storage_fblearner/tree/user/facebook/fblearner/predictor/924293663/0/gpu_lowering/input.merge"```

Differential Revision: D64490039
@facebook-github-bot
Copy link
Contributor

This pull request was exported from Phabricator. Differential Revision: D64490039

@YUNQIUGUO @facebook-github-bot
[AOTI] Fix a special case compile time data type codegen for sym int …
2267787 
…variables (pytorch#138106)

Summary:

This change unblocks the CFR AOTI lowering runtime error.

TL;DR:

In this model, one triton kernel expects a scalar input dtype as i64, but getting an i32. The reason is "auto"  can infer a smaller data type if the variable get passed in e.g. is i32. thus cause CUDA IMA.
 Original problematic kernel: `triton_poi_fused_add_ge_logical_and_logical_or_lt_46_grid_100`. and third input `auto var_402 = u0`.

This diff explicitly specifies it to i64 for all symbolic arguments in compile time for i64 triton kernel inputs, instead of use `auto var_x = {arg}` in cpp wrapper code.

Test Plan:
Verified in FLB locally:

```
PYTORCH_NO_CUDA_MEMORY_CACHING=1 AOT_INDUCTOR_DEBUG_INTERMEDIATE_VALUE_PRINTER=3 TORCH_LOGS="output_code" TORCHINDUCTOR_MAX_AUTOTUNE=1 TORCH_SHOW_CPP_STACKTRACES=1 CUDA_LAUNCH_BLOCKING=1 ~/fbsource/buck-out/v2/gen/fbcode/98e643f8bb44fe9d/hpc/new/models/feed/benchmark/__feed_lower_benchmark__/feed_lower_benchmark.par --skip-eager --skip-flop-estimation --lower-backend="AOT_INDUCTOR" --sync-mode=0 --precision bf16 --output-precision bf16  --lower-presets="ifr_cint;disable_new_lowering_weights;disable_dper_passes:passes=fuse_parallel_linear_no_weight_change" --remove-unexpected-type-cast=False --load="manifold://ads_storage_fblearner/tree/user/facebook/fblearner/predictor/924293663/0/gpu_lowering/input.merge"```

Differential Revision: D64490039
@facebook-github-bot
Copy link
Contributor

This pull request was exported from Phabricator. Differential Revision: D64490039

@pytorch-bot pytorch-bot bot added the ciflow/trunk Trigger trunk jobs on your pull request label Oct 19, 2024
@facebook-github-bot
Copy link
Contributor

@pytorchbot merge

(Initiating merge automatically since Phabricator Diff has merged)

@pytorchmergebot
Copy link
Collaborator

Merge started

Your 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

Advanced Debugging
Check the merge workflow status
here

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@ColinPeppler ColinPeppler ColinPeppler approved these changes

Assignees

No one assigned

Labels

ciflow/inductor ciflow/trunk Trigger trunk jobs on your pull request fb-exported Merged module: inductor topic: not user facing topic category

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

4 participants

@YUNQIUGUO @facebook-github-bot @pytorchmergebot @ColinPeppler