_{Comment @coderabbitai help to get the list of available commands and usage tips.}

@lowsfer could you review this?

/bot run --disable-fail-fast

PR_Github #55425 [ run ] triggered by Bot. Commit: 75e8beb Link to invocation

PR_Github #55425 [ run ] completed with state SUCCESS. Commit: 75e8beb
/LLM/main/L0_MergeRequest_PR pipeline #44363 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

Thanks for triggering CI. I cannot access the internal failed test details from the L0 report, but the public GitHub wrapper shows 33703 passed, 10 failed, 12923 skipped.

I pushed ffd097c4f to narrow the concurrent-decode constraint so it is bounded by the explicit KV cache token budget when kv_cache_config.max_tokens is set:

• keep the original max_batch_size floor for unconstrained/free-memory sizing
• use min(max_batch_size, max_tokens // tokens_per_block) when max_tokens is explicit
• keep at least one decode slot, matching the existing StorageManager baseline floor

This should avoid forcing impossible min-slot constraints in small-budget test configurations while preserving the intended fix for the high-concurrency windowed-pool deadlock case. I also added a focused unit check for the bounded case.

Local checks I could run here:

• PYTHONPYCACHEPREFIX=/private/tmp/trtllm-pycache /usr/bin/python3 -m py_compile tensorrt_llm/_torch/pyexecutor/kv_cache_manager_v2.py tests/unittest/_torch/executor/test_per_layer_head_dim.py
• git diff --check

I could not run the full target pytest locally because this checkout does not have the full TensorRT-LLM test dependency environment installed (transformers is missing after importing the test conftest). Could you please rerun L0 when convenient? If the same internal failures remain, please share the failed test names/log snippets and I can tighten the fix further.

Small correction: I amended the fix with DCO sign-off and force-pushed it as 062bc375. DCO is passing again on the current PR head.

I rebased the branch onto current upstream/main and resolved the conflict in kv_cache_manager_v2.py by preserving both upstream enable_stats wiring and this PR's concurrent-decode KV-cache slot constraint / typical_step fallback.

Pushed new head: 207015d98.

Local checks:

• PYTHONPYCACHEPREFIX=/tmp/tensorrt-pycache /usr/bin/python3 -m py_compile tensorrt_llm/_torch/pyexecutor/kv_cache_manager_v2.py
• git diff --check upstream/main...HEAD

The public GitHub checks currently show DCO passing; NVIDIA internal L0 still needs to be re-triggered/shared by an NVIDIA maintainer if it remains blocked.