[Klaud Cold] Add KLAUD_DEBUG.md operational playbook + AGENTS.md reference

AGENTS.md

@@ -2,6 +2,8 @@
 
 Guidance for AI agents working with InferenceX.
 
+> **Before debugging a failing Klaud-Cold / claude/* image-bump PR, read [`KLAUD_DEBUG.md`](KLAUD_DEBUG.md).** It captures recurring failure modes (tokenizer crash, vLLM CUDA-graph OOM, B300 sglang regressions, cluster docker/perms/disk issues), the exact workarounds, and gh-CLI gotchas — most cron-PR failures are already cataloged there.
+
 ## Project Overview
 
 InferenceX is an open-source automated benchmarking system that tracks LLM inference performance across hardware (NVIDIA B200/H100/H200/GB200, AMD MI300X/MI325X/MI355X) and software stacks (vLLM, SGLang, TensorRT-LLM, ATOM). Results published to https://inferencex.com/.

KLAUD_DEBUG.md

@@ -0,0 +1,230 @@
+# KLAUD_DEBUG.md — Operational Knowledge for Recipe-Bump PRs
+
+A running playbook of failures the Klaud-Cold image-bump cron has hit, the diagnoses, and the fixes/workarounds applied. **Read this first** before debugging a new failing claude/* PR — most failure modes here recur.
+
+When you fix something not yet listed, add it here so the next session doesn't re-learn it.
+
+---
+
+## 1. PR setup-stage failures
+
+### 1.1 `perf-changelog.yaml`: deletion-not-allowed
+**Symptom:** the `setup` job fails before any sweep runs with
+```
+ValueError: Deletions are not allowed in /home/runner/work/InferenceX/InferenceX/perf-changelog.yaml.
+Only additions to the changelog are permitted. Found deleted line: ...
+```
+**Root cause:** Cron-PR branches go stale; when main merges new changelog entries, the PR's local snapshot of `perf-changelog.yaml` no longer covers them, so the validator sees the missing lines as deletions. A naive rebase can also strip trailing whitespace from unrelated entries — same effect (e.g. `pr-link: ...1311  ` → `pr-link: ...1311`).
+
+**Fix (canonical):**
+```bash
+# In the PR's worktree, after `git merge origin/main` conflicts on perf-changelog.yaml:
+git checkout origin/main -- perf-changelog.yaml          # take main's bytes verbatim
+cat >> perf-changelog.yaml <<EOF                          # then append THIS PR's entry at tail
+
+- config-keys:
+    - <recipe-key>
+  description:
+    - "<one-line summary>"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/<N>
+EOF
+python3 -c "import yaml; yaml.safe_load(open('perf-changelog.yaml'))"
+```
+
+Do **not** try a 3-way merge of `perf-changelog.yaml` — whitespace edits will silently re-trigger the deletion check.
+
+---
+
+## 2. Bench-client tokenizer crash (sglang v0.5.12 images)
+
+**Symptom:** the sglang server loads the model fine; the bench client crashes before sending any request with
+```
+AttributeError: LlamaTokenizer has no attribute all_special_tokens_extended
+File "/opt/venv/.../vllm/transformers_utils/tokenizer.py:101" in get_cached_tokenizer
+```
+
+**Root cause:** `utils/bench_serving/benchmark_serving.py:48-51` prefers vLLM's `get_tokenizer` (which calls `get_cached_tokenizer` → probes `.all_special_tokens_extended`). The newer `transformers` library bundled inside sglang v0.5.12 images no longer exposes that attribute on `LlamaTokenizer`.
+
+**Fix (in-PR workaround):** swap import preference so the local `backend_request_func.get_tokenizer` (HF AutoTokenizer-based, no `get_cached_tokenizer` probe) wins:
+```python
+try:
+    from backend_request_func import get_tokenizer
+except ImportError:
+    from vllm.transformers_utils.tokenizer import get_tokenizer
+```
+Affects PRs touching sglang v0.5.12 images on any AMD platform. Apply per-PR (each PR's branch needs its own swap commit) until a global fix lands.
+
+---
+
+## 3. vLLM v0.21.x / v0.20.x: GPU OOM at model-load
+
+**Symptom:** vLLM workers OOM during weight loading or right after warmup:
+- `HSA_STATUS_ERROR_OUT_OF_RESOURCES: Available Free mem : 0 MB` (AMD)
+- `torch.OutOfMemoryError: CUDA out of memory. ... GPU N has X GiB of which Y MiB is free` (NVIDIA)
+- vLLM may also log `_check_enough_kv_cache_memory` failing with **negative** available bytes (e.g. `-25.24 GiB`).
+
+**Root cause:** v0.21.0 (and v0.20.2+) enabled an aggressive CUDA-graph memory profiler that pre-reserves a large chunk of VRAM up front (~30% on B200), shrinking effective `--gpu-memory-utilization` well below what the flag says. Old SHA-pinned custom images had a smaller footprint, so the recipe's existing `0.95` setting now starves the KV cache.
+
+**Fix:** in `benchmarks/single_node/<recipe>.sh`, either:
+1. **Lower `--gpu-memory-utilization`** (`0.95 → 0.90`, sometimes 0.85). Matches the H100/H200/B200 NVIDIA pattern. Smallest blast radius.
+2. **Disable the profiler entirely** for cases where lowering isn't enough: `export VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=0` before `vllm serve`. Matches `benchmarks/single_node/agentic/kimik2.5_fp4_b200.sh:65`.
+
+Seen on: #1395 (kimik2.5-fp4-b200-vllm — needed env var), #1403 (gptoss-fp4-mi300x-vllm — needed 0.90), #1461 (dsv4-fp8-h200-vllm — needed 0.90).
+
+---
+
+## 4. Custom DSV4 image → generic v0.5.12 OOMs
+
+**Symptom:** DSV4 recipes work on their SHA-pinned `lmsysorg/sglang:deepseek-v4-hopper@sha256:...` (or `deepseek-v4-b300`, `deepseek-v4-blackwell`) custom builds, but OOM on weights load when bumped to the generic `v0.5.12-cu130` release tag. Example: DSV4-Pro FP8+MTP weights consume ~125.43 GB / 141 GB per H200, leaving `-4.05 GB` for KV cache.
+
+**Root cause:** The custom DSV4 images use a different weight layout / EAGLE draft handling that fits in less memory than the generic release. The release tag isn't a drop-in replacement.
+
+**Fix:** keep DSV4 recipes pinned to their custom SHA-pinned image until upstream sglang gains the same DSV4-specific weight handling. Bumping to the generic tag is currently NOT viable.
+
+Seen on: #1460 (dsv4-fp8-h200-sglang+mtp).
+
+---
+
+## 5. Upstream sglang v0.5.12 B300 regressions
+
+Two distinct upstream regressions on NVIDIA B300 (Blackwell, `sm_120`) shipped in `lmsysorg/sglang:v0.5.12-cu130`:
+
+### 5a. DeepGemm TMA-descriptor crash (GLM-5-FP8)
+**Symptom:** CUDA graph capture aborts with `CUDA_ERROR_ILLEGAL_ADDRESS (700)` at `/deepgemm/csrc/.../runtime_utils.hpp:143` on the **first batch size** for **every TP rank**. Server never serves a prompt.
+
+**Workarounds (any one):**
+1. `--fp8-gemm-runner-backend cutlass` to bypass DeepGemm via CUTLASS.
+2. `export SGL_ENABLE_JIT_DEEPGEMM=0` before `python -m sglang.launch_server` to skip JIT DeepGemm.
+3. Pin recipe to `lmsysorg/sglang:v0.5.11-cu130`.
+
+Filed upstream: sgl-project/sglang#25551. Seen on #1421.
+
+### 5b. trtllm GEMM bug at bs=128 + MTP / EAGLE (GLM-5-NVFP4)
+**Symptom:** EAGLE draft CUDA graph capture crashes immediately at the largest batch size with `RuntimeError ... trtllm_batched_gemm_runner.cu:276 ... numBatches=256, GemmMNK 128x1024x6144`. The target model captures fine; only the draft model crashes.
+
+**Workarounds:**
+1. Cap `--cuda-graph-max-bs` and `--max-running-requests` to 64 in the launch script to avoid the bs=128 trigger.
+2. Comment out the MTP/EAGLE scenarios on B300 in the recipe.
+3. Pin to v0.5.11-cu130.
+
+Seen on #1420.
+
+### 5c. flash_attn SM-arch assertion (qwen3.5-bf16)
+**Symptom:** All 4 TP workers AssertionError on first forward pass:
+```
+File "/opt/venv/.../sglang/srt/layers/attention/flashattention_backend.py:..."
+  assert sm_100 <= arch <= sm_110f
+```
+B300 is `sm_120`, outside the asserted range. Server never becomes healthy; warmup times out at 600s.
+
+**Fix:** Needs sglang image with flash_attn supporting `sm_120` — no local workaround. Pin to v0.5.11-cu130 in the meantime.
+
+Seen on #1422.
+
+---
+
+## 6. Cluster infrastructure (AMD MI355X / MI300X / MI325X)
+
+### 6.1 `mia1-p01-g09 / g19 / g37` (amd-tw-mi355) — persistently drained
+- **g09**: `pyxis is broken`
+- **g19**: `Kill task failed (JobId=N StepId=N)`
+- **g37**: `permission issues with GHA runner workflows : Not responding` (down since Mar 2026)
+
+If a sweep job lands on any of these, it'll never start. Nothing to do at the recipe level — these stay drained until ops fixes them.
+
+### 6.2 `mia1-p01-g11 / g12 / g31` — docker socket perms
+**Symptom:** mi355x jobs fail with `permission denied while trying to connect to the docker API at unix:///var/run/docker.sock` during the `docker stop $(docker ps -a -q)` cleanup step, cascading into SLURM job expiration.
+**Fix:** ops needs to fix docker group / socket perms on these nodes. Recipe-level workaround: none.
+
+### 6.3 `chi-mi300x-049` — `/nvme_home` disk-full
+**Symptom:** pyxis container extraction fails with `No space left on device` writing to `/nvme_home/gharunner/.local/share/enroot/pyxis_*/opt/rocm-*/...`. The `/nvme_home` partition is hosted under `/` on this node and has been chronically near-full.
+
+**Fix already landed:** `runners/launch_mi300x-amds.sh` now pins salloc to only known-good mi300x nodes (`chi-mi300x-[034-036,054,057-058]`) — see PR #1462. `chi-mi300x-049` is held in `State=DOWN` by a watchdog on the controller (`/home/gharunner/_audit/drain_049_watchdog.sh`) that re-applies the drain every 10s if SLURM auto-clears it (which it does on dynamic-norm nodes).
+
+### 6.4 `chi-mi325x-pod1-017` — orphaned port-8888 process
+**Symptom:** sglang server bind fails with `[Errno 98] Address already in use` on port 8888. Held by an MLPerf accuracy run started outside SLURM.
+**Fix:** SSH to controller, find the holder via `ss -tlnp | grep :8888`, `kill` the PID. If recurring, file with the team running MLPerf experiments.
+
+### 6.5 Cluster controller layout
+- **amd-vultr-mi300**: SLURM controller for 7 mi300x nodes (3 down, see 6.3).
+- **amd-vultr-mi325**: SLURM controller for 6 mi325x nodes.
+- **amd-tw-mi355**: jumpbox → ssh further to compute nodes (`mia1-p01-gNN`). 12 nodes (3 drained, see 6.1).
+- `/home` is NFS-mounted across clusters from `chi-mi325x-pod1-001:/nfs/homes`, **root-writable**.
+- `/tmp` and `/nvme_home` are per-node local; HF cache lives at node-local `/raid/hf-hub-cache/` (2.7T per mi300x node).
+- Use `srun -w <FQDN>` (with the **full FQDN**, not the short hostname) from the controller to run admin commands on a compute node.
+
+### 6.6 Drain watchdog pattern
+SLURM auto-clears `State=DRAIN` on `DYNAMIC_NORM` nodes when they re-register. To keep a node out of the pool sticky-style, use `State=DOWN` AND start a watchdog:
+```bash
+# on the controller, as root
+nohup bash -c '
+  while true; do
+    s=$(scontrol show node <FQDN> 2>/dev/null | grep -oE "State=[A-Z+_]+")
+    if ! echo "$s" | grep -qE "DOWN|DRAIN"; then
+      scontrol update NodeName=<FQDN> State=DOWN Reason="watchdog" >/dev/null 2>&1
+    fi
+    sleep 10
+  done
+' > /home/gharunner/_audit/drain_<node>_watchdog.log 2>&1 &
+```
+Doesn't survive controller reboots — for permanent removal a SLURM admin should edit `slurm.conf`.
+
+---
+
+## 7. Docker image tag gotchas
+
+**Don't invent a "release" tag pattern from a date-suffixed nightly.** `lmsysorg/sglang-rocm:v0.5.12-rocm720-mi35x` does **not** exist — only the dated `v0.5.12-rocm720-mi35x-20260517` does. All MI355X `sglang-rocm:rocm720` tags follow the dated-nightly pattern.
+
+Before bumping an image, verify the target tag exists:
+```bash
+curl -sI "https://hub.docker.com/v2/repositories/lmsysorg/sglang-rocm/tags/v0.5.12-rocm720-mi35x"
+# 200 → exists; 404 → doesn't
+```
+
+Or check whether any other recipe on main uses the proposed tag — if zero uses, suspect.
+
+---
+
+## 8. CI: rerun mechanics
+
+- `gh run rerun <id> --failed` only works when the workflow run is **completed** with `conclusion=failure`. If the run is still `queued`/`in_progress`, the call returns "cannot be rerun".
+- To abandon an in-flight run and start fresh, push an **empty commit** to the PR branch:
+  ```bash
+  git commit --allow-empty -m "Re-trigger sweep"
+  git push
+  ```
+  The old run will be auto-cancelled by `workflow/cancel-sweep-on-merge` (provided the head SHA changed).
+- For a `cancelled` run (not `failure`), use `gh run rerun <id>` without `--failed` to re-run everything.
+
+---
+
+## 9. gh CLI gotchas
+
+- **`gh pr edit` silently aborts** on a Projects-classic deprecation GraphQL error. Title/body updates won't apply. Use `gh api -X PATCH "repos/<org>/<repo>/pulls/<N>" -f title="..." -F body=@file.md` instead.
+- Same issue for adding labels — use `gh api -X POST "repos/<org>/<repo>/issues/<N>/labels" -f "labels[]=<name>"`.
+- `gh pr view ... --jq .headRefName` output can have a trailing `\r`. Strip it: `gh pr view <N> --json headRefName --jq .headRefName | tr -d '\r\n'`. Otherwise shell concatenation produces `branchunners/launch_mi300x-amds.sh`-style corruption.
+- `gh pr list --json statusCheckRollup` **truncates** each PR's rollup — never trust it for per-check filters. Re-query each PR individually with `gh pr view <N> --json statusCheckRollup`.
+- `gh` and the GitHub Actions API: `conclusion` is `""` (empty string, not `null`) for in-flight checks, so `jq`'s `// .status` fallback doesn't trigger. Use:
+  ```jq
+  def state: if (.conclusion // "") != "" then .conclusion else .status end;
+  ```
+
+---
+
+## 10. PR conventions for this repo
+
+- Image-bump / new-recipe PRs I open on behalf of the user (or that the user creates) get the **`[Klaud Cold]`** title prefix.
+- Add the `full-sweep-enabled` label so a full sweep actually runs (`gh api -X POST ... labels[]=full-sweep-enabled`). Without it, the sweep is mostly SKIPPED.
+- After any code change that shifts a PR's scope (drops a recipe, changes an image tag), **update the PR title AND body in the same step** and **verify** with `gh pr view <N> --json title,body` — `gh pr edit` silently fails (see §9).
+- `utils/merge_with_reuse.sh <N>` is the merge entrypoint; it handles the `perf-changelog.yaml` auto-append.
+
+---
+
+## 11. Useful slash commands (defined in `.claude/commands/`)
+
+- `/find-mergeable-claude-prs` — lists `claude/*` PRs whose full sweep finished all-green.
+- `/list-claude-pr-status` — lists READY/RUNNING (and optionally FAILED) state per `claude/*` PR.
+- `/fix-klaud-cron-prs` — diagnoses failing `claude/*` PRs by reading their failed job logs.
+- `/merge-prs <N> [<N>...]` — sequential merge via `utils/merge_with_reuse.sh`.
+
+Each command file is self-contained; read them to understand the exact jq filters they use.