[NVBUG-6448152][perf] TEST ONLY; DO NOT REVIEW: exclude active transfers from scheduler capacity

[chienchunhung]

TEST ONLY; DO NOT REVIEW

This is a one-factor performance diagnostic for the GB300 PP4 disaggregated workload.

Hypothesis

The CuTe-off discriminator completed 512/512 at 814.02 output tok/s, so automatic PP CuTe BF16 selection does not explain the current approximately 2x regression. The next bounded current-only candidate is the transfer-in-progress scheduler-accounting change, which makes generation-side transfers consume logical scheduler request capacity.

Diagnostic change

• When TRTLLM_NVBUG_6448152_EXCLUDE_DISAGG_GEN_TRANSFER_FROM_CAPACITY=1, exclude only DISAGG_GENERATION_TRANS_IN_PROGRESS from the logical max-request count in guaranteed-no-evict and max-utilization scheduling.
• Preserve self-KV, cross-KV, and PEFT reservation, transfer-budget admission, polling, sequence slots, transfer traffic, and asynchronous consensus.
• Preserve default and static-batch behavior.
• Emit a bounded NVBUG6448152_SCHED event=effective_config marker.
• Enable the diagnostic only in the exact GB300 perf-sanity YAML. CuTe kernel auto-selection remains enabled.
• Keep the image, workload, request count, scheduler policy, timeouts, and coordinator unchanged.

Verification

• One DCO-signed diagnostic commit directly on the exact official asynchronous-coordinator head.
• Focused coverage added for default accounting, the opt-in bypass in both scheduler policies, and preserved PEFT resource gating.
• Existing transfer-budget admission coverage is unchanged.
• Clang-format, YAML validation, codespell, and diff checks passed.
• The focused CUDA C++ test binary cannot run on the local macOS/ARM host; the exact GB300 result is pending.

Interpretation

• A valid complete result at or above 1402 output tok/s would implicate logical scheduler accounting for active transfers.
• A valid result near 800 output tok/s would rule down this factor and move the next discriminator to the runtime/container stack.

Do not review or merge this diagnostic.

[chienchunhung]

/bot run --stage-list "GB300-12_GPUs-3_Nodes-PyTorch-Disagg-PerfSanity-CTX1-NODE1-GPU4-GEN1-NODE2-GPU8-Post-Merge-1"

[tensorrt-cicd]

PR_Github #60984 [ run ] triggered by Bot. Commit: add1341 Link to invocation

[tensorrt-cicd]

PR_Github #60984 [ run ] completed with state FAILURE. Commit: add1341
/LLM/main/L0_MergeRequest_PR pipeline #49244 (Partly Tested) completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

• Please check the failed tests and fix your PR
• If you cannot view the failures, ask the CI triggerer to share details
• Once fixed, request an NVIDIA team member to trigger CI again

CI Agent Failure Analysis

Link to invocation