feat(pd): MLA cache layout support + PD RDMA perf/stability#308
Open
xiaguan wants to merge 39 commits into
Open
feat(pd): MLA cache layout support + PD RDMA perf/stability#308xiaguan wants to merge 39 commits into
xiaguan wants to merge 39 commits into
Conversation
This dev/test host's default CUDA runtime is 13.3, but the default build feature is cudarc/cuda-12080, whose runtime bindings reference cudaEventElapsedTime_v2 — a symbol absent from the CUDA 13 libcudart. The release test hook therefore dlopen-failed on every CUDA test. Build and run the pre-commit release tests with --features cuda-13 so the bindings match the installed runtime. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
These experiment/benchmark/debug logs belong in the standalone docs repo (project knowledge), not in the pegaflow PR surface. Relocated to /data/pegadev/docs/projects/. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
Adds PD (Prefill/Decode disaggregation) connector support for the MLA cache layout, plus the surrounding PD RDMA performance/stability work and a save-only PegaFlow connector mode.
master...HEAD: 39 files, +3384 / −1216. Branch is 0 commits behind master, so it merges cleanly.Main changes
PD connector — MLA cache layout (
python/pegaflow/pd_connector/)layout.py/metadata.py/kv_params.py: support MLA (single latent KV) cache layout in addition to the existing FlashAttention HND layoutprefill_worker.py/decode_worker.py/rdma.py: PD RDMA push path reworked — parallelized decode RDMA waits, reduced dispatch/handshake/wait-registration overhead, stabilized producer release, improved push throughputscheduler.py/proxy.py/worker.py: plumbing for the aboveConnector (
python/pegaflow/connector/)feat(connector): save-only PegaFlow mode (feat(connector): add save-only pegaflow mode #300)Server / transfer (Rust)
fix(server): isolate CUDA tensor registry on a dedicated thread (fix(server): isolate CUDA tensor registry on a dedicated thread #301)pegaflow-serverregistry/http_server/service changes +http_cleanup_hang_repro.rsregression testrefactor(metrics): sharedbuild_bucketshelper for histogram buckets (refactor(metrics): use build_buckets helper for histogram buckets #298)pegaflow-transferverbs_domain tweaksPyO3 bindings (
python/src/lib.rs,pegaflow.pyi) updated for the new connector surface.CI / dev infra
fix(ci): installlibibverbs-devin release wheel build (fix(ci): install libibverbs-dev in release wheel build #303)ci: run the pre-commitcargo test --releasehook with--features cuda-13, matching the CUDA 13 runtime on the dev/test hosts (the defaultcuda-12080bindings referencecudaEventElapsedTime_v2, absent in CUDA 13 libcudart)Tests
test_pd_connector_layout.py,test_combine_hashes.pyexpansion,test_pd_connector_flow.py(renamed fromtest_pd_connector.py),test_vllm_save_only_e2e.py, sharedpd_connector_test_utils.pyTest status
Rust workspace tests pass under
--features cuda-13(via the pre-commit hook). The vLLM correctness E2E and PD integration gates should be run on the GPU machine before merge perCLAUDE.md.