[compile] Enable sequence parallelism matching w/o custom ops enabled #27126
+410
−432
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
Signed-off-by: angelayi <yiangela7@gmail.com>
angelayi
force-pushed
the
sp_custom_op
branch
from
c1efc65 to
ed10d76
Compare
angelayi
requested review from
ProExpertProg,
WoosukKwon,
hmellor,
houseroad,
mgoin,
robertgshaw2-redhat,
simon-mo,
tlrmchlsmth,
yewentao256,
youkaichao and
zou3519
as code owners
| """Base helper for RMSNorm and RMSNorm + Quantization functionalization.""" | ||
| def get_first_out_wrapper(fn): | ||
| @functools.wraps(fn) | ||
| def wrapper(*args): |
There was a problem hiding this comment.
Does this work? I thought that during tracing the pattern matching tracer will think that args is a single parameter
There was a problem hiding this comment.
yes! updated the test to assert the number of all_reduce/all_gather ops in the graph!
ProExpertProg
added this to the
vllm==v0.12.0/torch==2.9.0 compilation improvements milestone
Signed-off-by: angelayi <yiangela7@gmail.com> Co-authored-by: Luka Govedič <lgovedic@redhat.com>
angelayi
force-pushed
the
sp_custom_op
branch
from
ed10d76 to
5d66118
Compare
ProExpertProg
approved these changes
Oct 20, 2025
Sure! |
Co-authored-by: Luka Govedič <lgovedic@redhat.com> Signed-off-by: angelayi <yiangela7@gmail.com>
ProExpertProg
added
the
ready
|
@angelayi I have below error if not specify |
|
@angelayi It seems odd to me that enabling AsyncTP results in higher latency for Llama-70B. From our earlier benchmark, we observed about a 10% reduce in average latency for prefill stage with AsyncTP enabled for the same model on 4XH200. |
ProExpertProg
approved these changes
Oct 28, 2025
Comment on lines
+148
to
+152
| if inductor_graph_partition and "fp4" in model_name.lower(): | ||
| pytest.skip( | ||
| "Known bug for fp4 fusion & inductor partition: " | ||
| "https://github.com/vllm-project/vllm/issues/26988" | ||
| ) |
There was a problem hiding this comment.
#26988 was fixed
Suggested change
| if inductor_graph_partition and "fp4" in model_name.lower(): | |
| pytest.skip( | |
| "Known bug for fp4 fusion & inductor partition: " | |
| "https://github.com/vllm-project/vllm/issues/26988" | |
| ) |
Comment on lines
+236
to
+240
| if inductor_graph_partition and "fp4" in model_name.lower(): | ||
| pytest.skip( | ||
| "Known bug for fp4 fusion & inductor partition: " | ||
| "https://github.com/vllm-project/vllm/issues/26988" | ||
| ) |
There was a problem hiding this comment.
Suggested change
| if inductor_graph_partition and "fp4" in model_name.lower(): | |
| pytest.skip( | |
| "Known bug for fp4 fusion & inductor partition: " | |
| "https://github.com/vllm-project/vllm/issues/26988" | |
| ) |
Comment on lines
+333
to
+337
| if inductor_graph_partition and "fp4" in model_name.lower(): | ||
| pytest.skip( | ||
| "Known bug for fp4 fusion & inductor partition: " | ||
| "https://github.com/vllm-project/vllm/issues/26988" | ||
| ) |
There was a problem hiding this comment.
Suggested change
| if inductor_graph_partition and "fp4" in model_name.lower(): | |
| pytest.skip( | |
| "Known bug for fp4 fusion & inductor partition: " | |
| "https://github.com/vllm-project/vllm/issues/26988" | |
| ) |
This was referenced Oct 29, 2025
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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.
Purpose
Based on #24604, modified sequence-parallelism pass to do custom op matching w/o needing to enable the custom op
Test Plan
pytest -sv tests/compile/test_sequence_parallelism.pyPerformance numbers
I did some benchmarking with the command on H100 w/o flashinfer
while varying
"pass_config": {"enable_async_tp": true, "enable_sequence_parallelism": true}vs."pass_config": {"enable_async_tp": false, "enable_sequence_parallelism": false}"custom_ops":["+quant_fp8", "+rms_norm"]vs."custom_ops":[]