Refactor HT mode

[Binyang Li (Binyang2014)]

No description provided.

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[Copilot]

Pull request overview

This PR refactors the EP high-throughput (HT) backend into a single direct-dispatch/direct-combine path (moving implementation under src/ext/ep/high-throughput/ and simplifying the API/config surface), and adds an opt-in low-latency (LL) TOKEN_MAJOR padding initialization mode intended to support fixed-capacity kernels without a CPU sync.

Changes:

• Refactor HT to a direct receive-pool design (new high-throughput/{dispatch,combine,layout,runtime}.cu, simplified config, updated bindings and Python backend).
• Add token_major_init_padding plumbing for LL TOKEN_MAJOR dispatch (C++/CUDA + Python + tests/bench flags/docs).
• Improve Unix socket server lifecycle management (explicit shutdown() + destructor cleanup).

Reviewed changes

Copilot reviewed 30 out of 30 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
test/python/ep/test_low_latency_multirank.py	Adds CLI flag and validation for compact TOKEN_MAJOR offsets + optional padding initialization checks.
test/python/ep/test_intranode_multirank.py	Updates HT test to match new direct path and validates returned combined weights.
test/python/ep/run_ep_bench.py	Plumbs --token-major-init-padding through benchmark command construction.
test/python/ep/ep_bench_ll.py	Adds --token-major-init-padding and passes it into LL communicator config.
src/ext/ep/README.md	Updates LL TOKEN_MAJOR description (offsets + padding) and rewrites HT description for direct-only path.
src/ext/ep/moe_runtime.hpp	Extends LL MoERuntime ctor with initializeTokenMajorPadding state.
src/ext/ep/moe_runtime.cc	Wires initializeTokenMajorPadding into LL dispatch workload; keeps combine disabled for it.
src/ext/ep/low_latency/dispatch.cu	Implements token-major offsets and optional padding metadata initialization; tightens dispatch argument asserts.
src/ext/ep/low_latency/combine.cu	Switches token-major addressing to rank offsets; tightens combine argument asserts.
src/ext/ep/include/launch.cuh	Updates rank-switch macro signature; sets cooperative launch attributes via SETUP_LAUNCH_CONFIG.
src/ext/ep/include/api.cuh	Renames HT namespace/API to high_throughput::*; updates docs for token-major layoutRange semantics and adds workload flag.
src/ext/ep/ht/layout.cu	Deleted legacy HT ring layout implementation (replaced by high-throughput/layout.cu).
src/ext/ep/ht/intranode_kernel.cu	Deleted legacy HT ring + optional-path kernels (replaced by direct-only kernels).
src/ext/ep/ht/config.hpp	Deleted legacy HT ring-tuning config.
src/ext/ep/ht/buffer.cuh	Deleted legacy HT ring buffer view helpers.
src/ext/ep/ht_runtime.hpp	Updates HT runtime interface/state to match direct-only backend and new config type.
src/ext/ep/ht_runtime.cc	Refactors HT runtime to direct-only dispatch/combine and adds collective “direct ready” gating.
src/ext/ep/high-throughput/runtime.cu	Implements HT barrier launcher under new namespace.
src/ext/ep/high-throughput/layout.cu	New HT routing layout kernel under high_throughput.
src/ext/ep/high-throughput/dispatch.cu	New cooperative direct dispatch kernel that writes into peer recv pools + drains metadata locally.
src/ext/ep/high-throughput/config.cuh	New simplified HT config and receive-pool sizing helpers.
src/ext/ep/high-throughput/combine.cu	New cooperative TMA gather/reduce combine kernel for HT.
src/ext/ep/CMakeLists.txt	Switches HT sources/includes from ht/ to high-throughput/.
src/ext/ep/bindings.cpp	Updates nanobind exports for new LL ctor arg and new HT config/runtime signatures.
src/core/unix_socket.cc	Adds shutdown() and destructor cleanup; adjusts stop() behavior/logging.
src/core/include/unix_socket.hpp	Declares ~UnixSocketServer() and new private shutdown() helper.
python/mscclpp/ep/types.py	Adds token_major_init_padding and LL token-major rank_offsets; trims HT combine context fields.
python/mscclpp/ep/README.md	Documents compact token-major offsets + padding option; updates HT combine-handle description.
python/mscclpp/ep/low_latency.py	Plumbs token-major padding option; uses offsets tensor; improves layout validation/errors.
python/mscclpp/ep/high_throughput.py	Updates HT backend to direct receive-pool only; removes ring-era tensors from dispatch/combine API.

Comments suppressed due to low confidence (1)

src/core/unix_socket.cc:153

• shutdown() is now called from both stop() and the destructor, but it is not fully idempotent: it always calls ::close(listenUnixSockFd_) (which may be -1) and unlinks listenUnixSockPath_ without clearing it. Making shutdown() guard close/unlink and clearing the path avoids spurious errors on repeated shutdown calls and makes cleanup safer if stop() is called explicitly before destruction.

void UnixSocketServer::shutdown() {
  *abortFlag_ = 1;
  if (mainThread_.joinable()) {
    mainThread_.join();
  }
  ::close(listenUnixSockFd_);
  listenUnixSockFd_ = -1;