[WIP] bench: add GLM-5.2 PD AgentX workflow

[llying-001]

Description

Summary

This WIP PR adds a reproducible GLM-5.2 1P1D benchmark and validation workflow for two MI355X nodes.

Planned benchmark coverage includes:

• Pinned xiaobochen SGLang/AITER optimization layer
• Two-stage image build and cross-node digest verification
• Node inventory and selection
• Diagnostic RDMA/Mooncake preflight
• Correctness, long-context, tool-call, and GSM8K validation
• KV-aware AgentX C8/3600s measurement
• Analysis and evidence reporting

Supporting changes already included

• Avoid ephemeral-port reuse for delayed KV-event publishers.
• Add explicit Mooncake producer-write diagnostics while retaining READ as the public default.
• Fix engine-ready timeout and RDMA mount propagation in the GLM-5.2 example.
• Isolate the fork-specific rejection-sampling patch behind an explicit build opt-in.
• Support the fork’s /sglang source layout.

WIP status

The supporting shared changes have been pushed. bench/glm5p2_1p1d/ is still local and has not yet been committed to this PR.

Validation

• Focused unit tests: 49 passed
• Complete tests/unit/tools: 31 passed
• Ruff, formatting, shell syntax, and DCO checks passed

Remaining validation

The main libionic1-only image still needs to be rebuilt and rerun for production Mooncake WRITE on the target cluster.

[Copilot]

🟡 Changes recommended

The new ready-timeout unit test injects stub sglang modules into sys.modules without restoring them, which can leak into other tests and cause session-level flakiness.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

This PR adds supporting infrastructure for a reproducible GLM-5.2 PD benchmark workflow by tightening engine startup/port-allocation contracts, expanding Mooncake KV preflight diagnostics to cover producer-push (WRITE), and scoping a fork-only SGLang patch behind an explicit build opt-in.

Changes:

• Add Mooncake preflight read/write operation selection and WRITE-side verification, with accompanying unit tests and documentation.
• Make delayed-bind publishers avoid ephemeral-port reuse by consistently using free_tcp_port_block(1) and by hardening free_tcp_port_block’s contract.
• Introduce a fork-only SGLang rejection-sampling patch applied only when explicitly enabled at image build time, and document its scope/status.

File summaries

File	Description
tests/unit/tools/test_preflight_kv.py	Adds unit coverage for Mooncake preflight opcode selection, WRITE dispatch, and target-side verification.
tests/unit/engine/test_ready_timeout.py	Adds tests asserting the INFERA_ENGINE_READY_TIMEOUT contract across SGLang/vLLM wrappers.
tests/unit/common/test_net_ports.py	Updates port-allocation tests for non-ephemeral single-port blocks and invalid count handling.
manual/reference/environment.md	Documents INFERA_PREFLIGHT_MOONCAKE_OPCODE and INFERA_ENGINE_READY_TIMEOUT.
infera/tools/preflight/network/mooncakeperf.py	Adds WRITE-mode (producer-push) Mooncake probe support plus verification/reporting updates.
infera/engine/vllm/main.py	Uses free_tcp_port_block(1) for KV-events port selection to avoid ephemeral reuse during long startup.
infera/engine/sglang/worker.py	Uses free_tcp_port_block(dp_size) even for DP=1 to avoid ephemeral reuse during long startup.
infera/engine/atom/main.py	Uses free_tcp_port_block(1) for ATOM KV-events port selection to avoid ephemeral reuse during long startup.
infera/common/net.py	Clarifies free_tcp_port vs free_tcp_port_block guidance and rejects non-positive block sizes.
examples/sglang_1p1d_glm5.2/engine/up.sh	Passes through HOST_RDMA_MOUNT into the container environment.
examples/sglang_1p1d_glm5.2/engine/leg.sh	Renames the example startup timeout env var to INFERA_ENGINE_READY_TIMEOUT.
deploy/docker/patches/sglang_disagg/README.md	Updates patch documentation and describes the fork-only rejection-sampling patch rationale.
deploy/docker/patches/sglang_disagg_fork/patch_pd_disable_implicit_rocm_rejection_sampling.py	New fork-specific patch script disabling implicit ROCm rejection sampling for PD decode.
deploy/docker/patch.upstream.status.md	Documents the fork-only patch scope and retirement criteria.
deploy/docker/Dockerfile.sglang	Derives SGLANG_DIR from the installed package location and applies the fork-only patch behind a build arg.

Review details

• Files reviewed: 15/15 changed files
• Comments generated: 2
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.