use index_put only in kv cache update to reduce number of operators #3786
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/pytorch/executorch/3786
Note: Links to docs will display an error until the docs builds have been completed. ❌ 2 New FailuresAs of commit 6574788 with merge base 0333390 ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D57949659 |
…ytorch#3786) Summary: Pull Request resolved: pytorch#3786 The decomposition from ``` class IndexPut(torch.nn.Module): def __init__(self): super().__init__() def forward(self, x, input_pos, value): x[:, :, input_pos] = value return x ``` is ``` opcode name target args kwargs ------------- --------------- -------------------------- ----------------------------------------------- -------- placeholder x x () {} placeholder input_pos input_pos () {} placeholder value value () {} call_function slice_1 aten.slice.Tensor (x, 0, 0, 9223372036854775807) {} call_function slice_2 aten.slice.Tensor (slice_1, 1, 0, 9223372036854775807) {} call_function index_put aten.index_put.default (slice_2, [None, None, input_pos], value) {} call_function slice_3 aten.slice.Tensor (x, 0, 0, 9223372036854775807) {} call_function slice_scatter aten.slice_scatter.default (slice_3, index_put, 1, 0, 9223372036854775807) {} call_function slice_scatter_1 aten.slice_scatter.default (x, slice_scatter, 0, 0, 9223372036854775807) {} output output output ((slice_scatter_1, slice_scatter_1),) {} ``` however `x[:, :, input_pos] = value` really is just updating the content inside `x` with value, essentially just `index_put` By replacing `x[:, :, input_pos] = value` with `torch.ops.aten.index_put_(x, [None, None, input_pos], value)`, we reduce the number of operators from 6 to 1. ``` class IndexPut(torch.nn.Module): def __init__(self): super().__init__() def forward(self, x, indices, values): torch.ops.aten.index_put_(x, [None, None, input_pos], value) return x ``` decomposition is ``` opcode name target args kwargs ------------- --------- ---------------------- ----------------------------------- -------- placeholder x x () {} placeholder input_pos input_pos () {} placeholder value value () {} call_function index_put aten.index_put.default (x, [None, None, input_pos], value) {} output output output ((index_put, index_put),) {} ``` A more proper way to address this in long term is via pattern matching to replace the patterns with the simplified pattern Perf: For stories, before the diff ``` I 00:00:03.437290 executorch:runner.cpp:419] Prompt Tokens: 9 Generated Tokens: 118 I 00:00:03.437295 executorch:runner.cpp:425] Model Load Time: 0.763000 (seconds) I 00:00:03.437301 executorch:runner.cpp:435] Total inference time: 2.661000 (seconds) Rate: 44.344231 (tokens/second) I 00:00:03.437305 executorch:runner.cpp:443] Prompt evaluation: 0.185000 (seconds) Rate: 48.648649 (tokens/second) I 00:00:03.437309 executorch:runner.cpp:454] Generated 118 tokens: 2.476000 (seconds) Rate: 47.657512 (tokens/second) I 00:00:03.437313 executorch:runner.cpp:462] Time to first generated token: 0.206000 (seconds) I 00:00:03.437315 executorch:runner.cpp:469] Sampling time over 127 tokens: 0.042000 (seconds) ``` After the diff ``` I 00:00:03.195257 executorch:runner.cpp:419] Prompt Tokens: 9 Generated Tokens: 118 I 00:00:03.195295 executorch:runner.cpp:425] Model Load Time: 0.683000 (seconds) I 00:00:03.195314 executorch:runner.cpp:435] Total inference time: 2.502000 (seconds) Rate: 47.162270 (tokens/second) I 00:00:03.195319 executorch:runner.cpp:443] Prompt evaluation: 0.175000 (seconds) Rate: 51.428571 (tokens/second) I 00:00:03.195323 executorch:runner.cpp:454] Generated 118 tokens: 2.327000 (seconds) Rate: 50.709067 (tokens/second) I 00:00:03.195327 executorch:runner.cpp:462] Time to first generated token: 0.195000 (seconds) I 00:00:03.195330 executorch:runner.cpp:469] Sampling time over 127 tokens: 0.049000 (seconds) ``` Differential Revision: D57949659
|
This pull request was exported from Phabricator. Differential Revision: D57949659 |
cccclai
force-pushed
the
export-D57949659
branch
from
d9aed5a to
6574788
Compare
YifanShenSZ
reviewed
Jul 18, 2024
| ) | ||
| v_out = torch.ops.aten.index_put_( | ||
| self.v_cache, [None, None, input_pos], v_val | ||
| ) |
There was a problem hiding this comment.
Is the 2nd v_out redundant?
larryliu0820
approved these changes
Jul 18, 2024
|
@cccclai Do you have any performance numbers for before and after comparison? |
about 5% IIRC. |
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.
Summary:
The decomposition from
is
however
x[:, :, input_pos] = valuereally is just updating the content insidexwith value, essentially justindex_putBy replacing
x[:, :, input_pos] = valuewithtorch.ops.aten.index_put_(x, [None, None, input_pos], value), we reduce the number of operators from 6 to 1.decomposition is
A more proper way to address this in long term is via pattern matching to replace the patterns with the simplified pattern
Differential Revision: D57949659