Repository navigation
[Fix] dp-attention: Align prefill graph buckets and decode padding across DP ranks - #46138
Conversation
This comment has been minimized.
This comment has been minimized.
|
/rerun-test test/registered/unit/layers/test_dp_attention.py test/registered/unit/layers/attention/test_dsa_backend_dp_padding.py test/registered/unit/model_executor/runner/test_prefill_cuda_graph_padding.py test/registered/unit/model_executor/test_prefill_cuda_graph_runner.py test/registered/unit/model_executor/test_prefill_cuda_graph_runner_helpers.py |
|
🚀 |
Co-authored-by: Chunan Zeng <zcnrex@gmail.com>
Co-authored-by: Chunan Zeng <zcnrex@gmail.com> Co-authored-by: Denji-kk <51020465+Denji-kk@users.noreply.github.com>
9343589 to
d0ec8c6
Compare
7e9662d to
193e54b
Compare
|
🚀 |
|
/rerun-test test/registered/unit/layers/test_dp_attention.py test/registered/unit/layers/attention/test_dsa_backend_dp_padding.py test/registered/unit/model_executor/runner/test_prefill_cuda_graph_padding.py test/registered/unit/model_executor/test_prefill_cuda_graph_runner.py test/registered/unit/model_executor/test_prefill_cuda_graph_runner_helpers.py |
|
🚀 |
|
/rerun-test test/registered/unit/layers/test_dp_attention.py test/registered/unit/layers/attention/test_dsa_backend_dp_padding.py test/registered/unit/model_executor/runner/test_prefill_cuda_graph_padding.py test/registered/unit/model_executor/test_prefill_cuda_graph_runner.py test/registered/unit/model_executor/test_prefill_cuda_graph_runner_helpers.py |
|
🚀 |
|
/rerun-test test/registered/unit/layers/test_dp_attention.py test/registered/unit/layers/attention/test_dsa_backend_dp_padding.py test/registered/unit/test_runtime_context.py test/registered/unit/model_executor/runner/test_prefill_cuda_graph_padding.py test/registered/unit/model_executor/test_prefill_cuda_graph_runner.py test/registered/unit/model_executor/test_prefill_cuda_graph_runner_helpers.py test/registered/unit/model_executor/runner/test_prefill_shared_read_done.py |
|
🚀 |
|
Preview deployment for your docs. Learn more about Mintlify Previews.
💡 Tip: Enable Automations to automatically generate PRs for you. |
# Conflicts: # python/sglang/srt/layers/layer_boundary/fusions/custom_all_reduce_fusion.py # test/registered/unit/layer_boundary/test_custom_all_reduce_fusion.py
|
/run-full-ci |
Summary
Two DP-attention padding changes:
Changes to existing code
forward_batch_info:global_num_tokensis rounded up to the smallest captured prefill bucket on every rank.DpPaddingMode: decode selects SUM_LEN under the condition above, both eagerly and inget_default_mode_in_cuda_graph, which now receives the captured per-rank token counts.AttentionBackend.skips_prefill_metadata_on_idle_dp_rank(defaultFalse): when a backend sets it,PrefillCudaGraphRunnerskipsinit_forward_metadataand the read snapshot on a rank with no real tokens. No existing backend sets it, so their replay is unchanged.The prefill change affects every model that uses DP attention with breakable prefill graphs; the decode padding change affects DP-attention decode without idle ranks.
Verification
Cut byte-for-byte from #45925 (MiniMax-M3.1-Flash support) at 10e655a.
CI States
Latest PR Test (Base): ✅ Run #40160842644
Latest PR Test (Extra): ✅ Run #40160842552
Latest PR Test (AMD ROCm 10): ⏳ Run #40160842664