fix(qwen): PR #35936 overlay — sidecar pattern to resolve Genesis RO-mount conflict

v0.5.0 mounted vLLM PR #35936 overlay files as :ro bind-mounts at vLLM's site-packages paths. On Genesis-loaded composes, this blocked Genesis P64 (qwen3coder MTP streaming early-return), P68 (auto force tool_choice= required), and P69 (long-context tool-format reminder) — all three write hooks into chat_completion/serving.py at vllm-import time. Genesis reported "Errno 30: Read-only file system" and warned "partial state risk; container should be torn down." Reported by @ygafarov on club-3090#120: https://github.com/noonghunna/club-3090/issues/120#issuecomment-4443236686 This commit converts all 23 PR #35936-mounted composes to a sidecar pattern: 1. Patched files are bind-mounted at side paths under /etc/club3090/ (pr35936-chat-completion-serving.py, pr35936-engine-serving.py). 2. install.sh is bind-mounted at /etc/club3090/install-pr35936.sh. 3. Each compose entrypoint invokes `bash /etc/club3090/install-pr35936.sh` BEFORE Genesis runs (apply_all in Genesis-loaded composes, or vllm serve invocation in Genesis-less ones). 4. install.sh copies the files into vLLM's site-packages via `cp`, landing them in the container's RW layer. Result: Genesis can write its hooks freely on top of our content. No RO-mount errors. Host patches dir stays canonical. Three entrypoint variants handled: - A) Existing apply_all invocation (9 composes — Genesis-loaded): insert install.sh before `python3 -m vllm._genesis.patches.apply_all` - B) Simple `exec vllm serve` (13 composes — non-Genesis): prepend install.sh before exec - C) No entrypoint at all (1 compose — qwopus): add new entrypoint that runs install.sh then exec's vllm serve Validated 2026-05-13 on single/long-text.yml (Genesis-loaded): [club3090/pr35936] chat_completion/serving.py installed from /etc/club3090/... [INFO:genesis.apply_all] applied: P64 qwen3coder MTP streaming early-return fix — P64 applied: 2 files modified, 0 idempotent [INFO:genesis.apply_all] applied: P68/P69 long-context tool-call adherence — Hook injected into create_chat_completion Zero `Read-only file system` errors anywhere in boot log. Pattern is forward-compatible: if a future PR #35936 patch update touches chat_completion/serving.py (e.g. streaming hunks), the sidecar layout already supports layering our content under Genesis writes — no further plumbing changes needed. Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
docs: clarify MODEL_DIR — second drive / HF cache / Windows-WSL
2026-05-13 17:01:33 +00:00 · 2026-05-13 14:32:10 +00:00 · 2026-05-12 22:11:24 +00:00 · 2026-05-12 21:23:33 +00:00
30 changed files with 407 additions and 54 deletions
@@ -16,6 +16,44 @@ history; SemVer takes over from `v0.3.0` onward.

 ---

+## v0.5.0 — 2026-05-12
+
+
+### ✨ Features
+
+- feat(qwen): ship froggeric chat-template fixes as default-on ([84498d4](https://github.com/noonghunna/club-3090/commit/84498d47aaf7a2fdb7c0203d53bb64414a64b6c1))
+- feat(vllm): add PR #35936 required-tool fallback overlay ([28b16b5](https://github.com/noonghunna/club-3090/commit/28b16b5dc9d602a8e1c4b8d5496aa82cbca7f95d))
+- feat(qwen-tq3): add CLUB3090_TQ_K1_SKIP_MTP layer-filter for PR #40914 K+1 dispatch ([6b2a7d5](https://github.com/noonghunna/club-3090/commit/6b2a7d553b6164bb4854e80fc0acdf8dcec18a87))
+
+
+### 🎯 New models + serving paths
+
+- compose(tq3-mtp-genesis): pin to Genesis v7.72.2 known-good vLLM nightly ([570fa71](https://github.com/noonghunna/club-3090/commit/570fa71240a12ab642693c0570851611e933f8c4))
+
+
+### 📊 Benchmarks + cross-rig data
+
+- bench(matrix): @ygafarov first heterogeneous Ampere + Blackwell eGPU dual ([1770931](https://github.com/noonghunna/club-3090/commit/1770931729a354bad319c8b58bdee143fe6ebce2))
+
+
+### 📝 Documentation
+
+- docs(dtype-matrix): more polish — RDNA naming, FP8 maturity caveats, AMD detection ([62b3b45](https://github.com/noonghunna/club-3090/commit/62b3b455a9ea5146cbf5576febd0c8b25a8c0fa1))
+- docs(dtype-matrix): polish nuances + add Intel and AMD vendor sections ([3d4548c](https://github.com/noonghunna/club-3090/commit/3d4548c50422da07c16fcba2a59d6f42f268355b))
+- docs(dtype-matrix): per-arch hardware accelerator matrix for compose optimization ([9c6d3cf](https://github.com/noonghunna/club-3090/commit/9c6d3cfba1de9d9079d6eafc9ff68c352cac7197))
+- docs(faq): add 'INT8 PTH doesn't scale at concurrency — is that a bug?' ([df53287](https://github.com/noonghunna/club-3090/commit/df53287b1c26ec83be8a30eec24baf2bddc993eb))
+- docs(tq3-mtp): writeup + charts for the Genesis-backed TQ3+MTP path ([c2b1c93](https://github.com/noonghunna/club-3090/commit/c2b1c93872f84fa9afa3bbe41360dc42be28c066))
+- docs(qwen-tq3): close round-4 — #40914 not shippable, route to nomtp + Genesis ([9fba037](https://github.com/noonghunna/club-3090/commit/9fba03788e30151f9bf8c85f279260696954d094))
+- docs(qwen-tq3): re-tombstone tq3-mtp.yml after round-3 MTP-skip validation ([063d3e9](https://github.com/noonghunna/club-3090/commit/063d3e943ce8da9cc69bd30c68b87426aca6202e))
+
+
+### 🧹 Maintenance
+
+- refactor(qwen): rename int8-tq3 → tq3-* family + add no-MTP + Genesis variants ([6182922](https://github.com/noonghunna/club-3090/commit/6182922225dfeee1c28084d1ff917bfd25539520))
+
+
+
+[Pin: `git checkout v0.5.0`] · [Full diff](https://github.com/noonghunna/club-3090/compare/v0.4.0...v0.5.0)
 ## v0.4.0 — 2026-05-11


@@ -67,6 +67,9 @@ git clone https://github.com/noonghunna/club-3090.git
 cd club-3090

 # 2. Download + SHA-verify the model (~20 GB; clones Genesis patches too)
+#    (asks you where to put model weights — pick in-repo default, ~/models, or
+#     a custom path on a different drive. To skip the prompt:
+#     `export MODEL_DIR=/mnt/your-drive/models` before running. See FAQ + .env.example.)
 bash scripts/setup.sh qwen3.6-27b

 # 3. Pick a config + boot it (interactive wizard — asks engine / cards / workload)
@@ -130,9 +130,26 @@ If your numbers on the same compose look different from ours by >15%, the most l

 ## Setup

-### `bash scripts/setup.sh qwen3.6-27b` is downloading 20+ GB. Where does it go?
+### `bash scripts/setup.sh qwen3.6-27b` is downloading 20+ GB. Where does it go? / Can I put models on a different drive?

-`<repo>/models-cache/` by default. Override with `MODEL_DIR=/path/to/your/scratch bash scripts/setup.sh qwen3.6-27b`. See [`.env.example`](../.env.example) for all env vars.
+Yes. The knob is `MODEL_DIR`, with **four ways** to set it (priority order):
+
+1. **`MODEL_DIR` env var in your shell** — takes precedence over everything:
+   ```bash
+   export MODEL_DIR=/mnt/your-second-drive/models
+   bash scripts/setup.sh qwen3.6-27b
+   ```
+2. **`.env` file at repo root** — picked up automatically on every script run. See [`.env.example`](../.env.example).
+3. **Interactive prompt** — `bash scripts/setup.sh qwen3.6-27b` with nothing set offers three choices: in-repo default, `~/models`, or custom path. After you pick custom, it asks "Save `MODEL_DIR=/your/path` to `.env` so we skip this next time?" — say `Y` and it persists for every subsequent `launch.sh` / `switch.sh` / `bench.sh` call.
+4. **Silent fallback** — `<repo>/models-cache/`. Functional but pollutes the git tree; not recommended.
+
+Every script that touches model paths reads from the same `MODEL_DIR`. The compose YAMLs' volume mount is `${MODEL_DIR:-...}:/root/.cache/huggingface` — once set, every container reads + writes there.
+
+**HF env-var integration** — we don't directly respect `HF_HOME` / `HF_HUB_CACHE` because we mount a host directory INTO the container's `/root/.cache/huggingface`, not the host's HF cache. The internal layout inside `MODEL_DIR` matches HF's repo-cache convention (`<MODEL_DIR>/<repo-subdir>/`), so models downloaded by `setup.sh` are byte-compatible with anything that reads HF's local cache layout. Two clean workarounds if you already have an HF cache you want to reuse:
+- Set `MODEL_DIR=$HF_HOME/hub`
+- Or symlink between them
+
+**On Windows / WSL2** — same mechanism. Docker Desktop handles path translation. Use Windows paths (`D:\models`) from PowerShell or WSL paths (`/mnt/d/models`) from WSL. If you flip between Linux and Windows on the same rig, point `MODEL_DIR` at a drive both OSes can see — the model files themselves are OS-agnostic.

 ### How do I keep my install up-to-date?

@@ -69,7 +69,8 @@ Run `bash scripts/maintenance/list-image-pins.sh` for a live snapshot.
 |---|---|---|---|
 | [#35936](https://github.com/vllm-project/vllm/pull/35936) — `tool_choice="required"` falls back to configured tool parser | 🟡 Open / **local overlay active** | Qwen3-Coder with `--tool-call-parser qwen3_coder` emits XML-style tool calls. On pinned nightly `1acd67a79`, non-streaming `tool_choice="required"` validates JSON only, bypasses the configured parser, and returns `tool_calls=[]`. MLS-Bench hits this when `thinking.enabled=false`. | Vendored overlay: [`models/qwen3.6-27b/vllm/patches/vllm-pr35936-required-fallback/README.md`](../models/qwen3.6-27b/vllm/patches/vllm-pr35936-required-fallback/README.md). Drop when #35936 or equivalent lands in our pinned image. |
 | [#40361](https://github.com/vllm-project/vllm/pull/40361) — Marlin pad-sub-tile-n | 🟡 Open, mergeable, **stale 13d** (last update 2026-04-20) | All 4 dual-card composes + `dual-nvlink.yml` + `dual-nvlink-turbo.yml` mount the patched files vendored in-repo at `models/qwen3.6-27b/vllm/patches/vllm-marlin-pad/`. Drops out as a setup dependency when this merges + propagates. | Vendored mount: see [`models/qwen3.6-27b/vllm/patches/vllm-marlin-pad/README.md`](../models/qwen3.6-27b/vllm/patches/vllm-marlin-pad/README.md). Queued for rebase + ping next week (see "Active follow-ups" table above). |
-| [#40807](https://github.com/vllm-project/vllm/issues/40807) — `.tolist()` cudagraph crash on continuation-prefill | ⚫ Local workaround | Single-card TQ3 + spec-decode + chunked-prefill blocked without it. We ship a file-edit patch. | `patch_tolist_cudagraph.py` runs in `setup.sh`. Drop when upstream fixes the sync. |
+| [#40807](https://github.com/vllm-project/vllm/issues/40807) — `.tolist()` cudagraph crash on continuation-prefill | ✅ **Retired locally** (2026-05-05 Genesis v7.72.2 bump) — Genesis ships [P78 `TOLIST_CAPTURE_GUARD`](../models/qwen3.6-27b/vllm/compose/dual/tq3-mtp-genesis.yml) as the equivalent fix. Currently disabled (`=0`) on `tq3-mtp-genesis.yml` after rebench-full leg 6 (2026-05-11) passed clean with it off — apparent root cause is now covered by Genesis PN34 (workspace-lock relax) + post-#41434 attention rework. Non-Genesis composes on `1acd67a79` pin run without any guard for this bug; unvalidated at long-context TurboQuant chunked-prefill (worth testing per cferra's vllm#41403 validation pass — see vllm#40798 row below). | None active. Drop the Genesis env var permanently if a future v7.73.x rebench leaves it OFF without regression. |
+| [#40798](https://github.com/vllm-project/vllm/pull/40798) + [#42215](https://github.com/vllm-project/vllm/pull/42215) — share decode scratch workspace pre-CUDA-graph + decode-kernel warmup | 🟡 Open, validated cross-rig | Pair closes the `AssertionError: Workspace is locked but allocation requires NMB` crash that fires at `turboquant_attn.py:_continuation_prefill` for ≥48K-token chunked-prefill with TurboQuant KV. Independently validated on 2× 3090 sm_86 by cferra (vllm#41403 [comment](https://github.com/vllm-project/vllm/issues/41403#issuecomment-4435164709), 2026-05-12). | Genesis [PN34 `WORKSPACE_LOCK_RELAX`](../models/qwen3.6-27b/vllm/compose/dual/tq3-mtp-genesis.yml) addresses the same symptom via a different mechanism (relax-lock vs reserve-before-capture). On non-Genesis composes (`1acd67a79` pin) we currently have no guard — re-validate against this PR pair once they propagate to a nightly we pin to, then A/B PN34 vs upstream. |
 | [#40849](https://github.com/vllm-project/vllm/pull/40849) — MTP draft online-quant propagation | 🟡 Open / Genesis backport active | Closes Cliff 1 on FP8+MTP path (`tools-text.yml`). | Genesis PN8 backport: `GENESIS_ENABLE_PN8_MTP_DRAFT_ONLINE_QUANT=1`. |
 | [#40914](https://github.com/vllm-project/vllm/pull/40914) — Sandermage K+1 verify routing | 🟡 Open, ❌ negative on our Qwen3.6-27B stack | **Reframed 2026-05-11:** the synthetic `seq_lens` K+1 route is not the P67-equivalent we need here. Local rebase on post-#41434 nightly made MTP acceptance look perfect (AL=4.0 / ~100%) but produced `!`-flood needle corruption plus tool/multi-turn timeouts. Dropping it improved verify-stress from 3/7 to 5/7, but TQ3/TQ4/k8v4 + MTP still fail long-context needles. | Do not ship Genesis-free TQ+MTP on #40914 alone. Use `dual/tq3-nomtp.yml` without Genesis, or `dual/tq3-mtp-genesis.yml` with Genesis P67/P67b. |
 | [#40334](https://github.com/vllm-project/vllm/pull/40334) — DFlash `combine_hidden_states` dtype mismatch | 🟡 Open | All `dual-dflash*.yml` need `--dtype bfloat16` flag to work around. | Composes set `--dtype bfloat16`. Drop when this lands. |
@@ -44,8 +44,14 @@ services:
      - ../../patches/vllm-marlin-pad/marlin.py:/usr/local/lib/python3.12/dist-packages/vllm/model_executor/kernels/linear/mixed_precision/marlin.py:ro
      - ../../patches/vllm-marlin-pad/MPLinearKernel.py:/usr/local/lib/python3.12/dist-packages/vllm/model_executor/kernels/linear/mixed_precision/MPLinearKernel.py:ro
      # vLLM PR #35936 required-tool fallback (drop when upstream lands).
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/chat_completion/serving.py:ro
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/engine/serving.py:ro
+      # vLLM PR #35936 required-tool fallback — sidecar pattern (drop when upstream lands).
+      # Bind-mounted at side paths so install.sh can copy into vLLM's site-packages
+      # BEFORE Genesis runs — avoids the RO-mount conflict with Genesis P64/P68/P69
+      # which write to chat_completion/serving.py at vllm-import time. See
+      # patches/vllm-pr35936-required-fallback/install.sh + README.md.
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/etc/club3090/pr35936-chat-completion-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/etc/club3090/pr35936-engine-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/install.sh:/etc/club3090/install-pr35936.sh:ro
      # froggeric/Qwen-Fixed-Chat-Templates qwen3.6 — fixes 7 default-template
      # bugs (empty <think></think> spam, </thinking> hallucination, unclosed
      # think before tool call, no-user-query crash, developer role, etc.).
@@ -74,6 +80,8 @@ services:
      - bash
      - -c
      - |
+        # Install PR #35936 overlay before vllm imports (drop when upstream lands).
+        bash /etc/club3090/install-pr35936.sh
        exec vllm serve ${VLLM_ENFORCE_EAGER:+--enforce-eager} "$@"
      - --
    command:
@@ -73,8 +73,14 @@ services:
      - ../../patches/vllm-marlin-pad/marlin.py:/usr/local/lib/python3.12/dist-packages/vllm/model_executor/kernels/linear/mixed_precision/marlin.py:ro
      - ../../patches/vllm-marlin-pad/MPLinearKernel.py:/usr/local/lib/python3.12/dist-packages/vllm/model_executor/kernels/linear/mixed_precision/MPLinearKernel.py:ro
      # vLLM PR #35936 required-tool fallback (drop when upstream lands).
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/chat_completion/serving.py:ro
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/engine/serving.py:ro
+      # vLLM PR #35936 required-tool fallback — sidecar pattern (drop when upstream lands).
+      # Bind-mounted at side paths so install.sh can copy into vLLM's site-packages
+      # BEFORE Genesis runs — avoids the RO-mount conflict with Genesis P64/P68/P69
+      # which write to chat_completion/serving.py at vllm-import time. See
+      # patches/vllm-pr35936-required-fallback/install.sh + README.md.
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/etc/club3090/pr35936-chat-completion-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/etc/club3090/pr35936-engine-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/install.sh:/etc/club3090/install-pr35936.sh:ro
    environment:
      - HUGGING_FACE_HUB_TOKEN=${HF_TOKEN:-}
      - VLLM_WORKER_MULTIPROC_METHOD=spawn
@@ -103,6 +109,8 @@ services:
      - |
        # VLLM_ENFORCE_EAGER=1 in .env disables CUDA graphs — use on
        # hardware where graph capture causes OOM or instability (e.g. WSL2).
+        # Install PR #35936 overlay before vllm imports (drop when upstream lands).
+        bash /etc/club3090/install-pr35936.sh
        exec vllm serve ${VLLM_ENFORCE_EAGER:+--enforce-eager} "$@"
      - --
    command:
@@ -62,8 +62,14 @@ services:
      - ../../patches/vllm-marlin-pad/marlin.py:/usr/local/lib/python3.12/dist-packages/vllm/model_executor/kernels/linear/mixed_precision/marlin.py:ro
      - ../../patches/vllm-marlin-pad/MPLinearKernel.py:/usr/local/lib/python3.12/dist-packages/vllm/model_executor/kernels/linear/mixed_precision/MPLinearKernel.py:ro
      # vLLM PR #35936 required-tool fallback (drop when upstream lands).
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/chat_completion/serving.py:ro
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/engine/serving.py:ro
+      # vLLM PR #35936 required-tool fallback — sidecar pattern (drop when upstream lands).
+      # Bind-mounted at side paths so install.sh can copy into vLLM's site-packages
+      # BEFORE Genesis runs — avoids the RO-mount conflict with Genesis P64/P68/P69
+      # which write to chat_completion/serving.py at vllm-import time. See
+      # patches/vllm-pr35936-required-fallback/install.sh + README.md.
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/etc/club3090/pr35936-chat-completion-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/etc/club3090/pr35936-engine-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/install.sh:/etc/club3090/install-pr35936.sh:ro
      # froggeric/Qwen-Fixed-Chat-Templates qwen3.6 — fixes 7 default-template
      # bugs (empty <think></think> spam, </thinking> hallucination, unclosed
      # think before tool call, no-user-query crash, developer role, etc.).
@@ -97,6 +103,8 @@ services:
      - |
        # VLLM_ENFORCE_EAGER=1 in .env disables CUDA graphs — use on
        # hardware where graph capture causes OOM or instability (e.g. WSL2).
+        # Install PR #35936 overlay before vllm imports (drop when upstream lands).
+        bash /etc/club3090/install-pr35936.sh
        exec vllm serve ${VLLM_ENFORCE_EAGER:+--enforce-eager} "$@"
      - --
    command:
@@ -86,8 +86,14 @@ services:
      - ../../patches/vllm-marlin-pad/marlin.py:/usr/local/lib/python3.12/dist-packages/vllm/model_executor/kernels/linear/mixed_precision/marlin.py:ro
      - ../../patches/vllm-marlin-pad/MPLinearKernel.py:/usr/local/lib/python3.12/dist-packages/vllm/model_executor/kernels/linear/mixed_precision/MPLinearKernel.py:ro
      # vLLM PR #35936 required-tool fallback (drop when upstream lands).
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/chat_completion/serving.py:ro
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/engine/serving.py:ro
+      # vLLM PR #35936 required-tool fallback — sidecar pattern (drop when upstream lands).
+      # Bind-mounted at side paths so install.sh can copy into vLLM's site-packages
+      # BEFORE Genesis runs — avoids the RO-mount conflict with Genesis P64/P68/P69
+      # which write to chat_completion/serving.py at vllm-import time. See
+      # patches/vllm-pr35936-required-fallback/install.sh + README.md.
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/etc/club3090/pr35936-chat-completion-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/etc/club3090/pr35936-engine-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/install.sh:/etc/club3090/install-pr35936.sh:ro
      # froggeric/Qwen-Fixed-Chat-Templates qwen3.6 — fixes 7 default-template
      # bugs (empty <think></think> spam, </thinking> hallucination, unclosed
      # think before tool call, no-user-query crash, developer role, etc.).
@@ -121,6 +127,8 @@ services:
      - |
        # VLLM_ENFORCE_EAGER=1 in .env disables CUDA graphs — use on
        # hardware where graph capture causes OOM or instability (e.g. WSL2).
+        # Install PR #35936 overlay before vllm imports (drop when upstream lands).
+        bash /etc/club3090/install-pr35936.sh
        exec vllm serve ${VLLM_ENFORCE_EAGER:+--enforce-eager} "$@"
      - --
    command:
@@ -74,8 +74,14 @@ services:
      - ../../patches/vllm-marlin-pad/marlin.py:/usr/local/lib/python3.12/dist-packages/vllm/model_executor/kernels/linear/mixed_precision/marlin.py:ro
      - ../../patches/vllm-marlin-pad/MPLinearKernel.py:/usr/local/lib/python3.12/dist-packages/vllm/model_executor/kernels/linear/mixed_precision/MPLinearKernel.py:ro
      # vLLM PR #35936 required-tool fallback (drop when upstream lands).
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/chat_completion/serving.py:ro
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/engine/serving.py:ro
+      # vLLM PR #35936 required-tool fallback — sidecar pattern (drop when upstream lands).
+      # Bind-mounted at side paths so install.sh can copy into vLLM's site-packages
+      # BEFORE Genesis runs — avoids the RO-mount conflict with Genesis P64/P68/P69
+      # which write to chat_completion/serving.py at vllm-import time. See
+      # patches/vllm-pr35936-required-fallback/install.sh + README.md.
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/etc/club3090/pr35936-chat-completion-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/etc/club3090/pr35936-engine-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/install.sh:/etc/club3090/install-pr35936.sh:ro
      # froggeric/Qwen-Fixed-Chat-Templates qwen3.6 — fixes 7 default-template
      # bugs (empty <think></think> spam, </thinking> hallucination, unclosed
      # think before tool call, no-user-query crash, developer role, etc.).
@@ -112,6 +118,8 @@ services:
        # hardware where Cliff 2 GDN activation spikes occur at runtime
        # (~50-65K active context tokens). Costs ~20-30% TPS in exchange for
        # stability. See docs/HARDWARE.md "Note for WSL2 / Windows users".
+        # Install PR #35936 overlay before vllm imports (drop when upstream lands).
+        bash /etc/club3090/install-pr35936.sh
        exec vllm serve ${VLLM_ENFORCE_EAGER:+--enforce-eager} "$@"
      - --
    command:
@@ -44,8 +44,14 @@ services:
      - ../../patches/vllm-marlin-pad/marlin.py:/usr/local/lib/python3.12/dist-packages/vllm/model_executor/kernels/linear/mixed_precision/marlin.py:ro
      - ../../patches/vllm-marlin-pad/MPLinearKernel.py:/usr/local/lib/python3.12/dist-packages/vllm/model_executor/kernels/linear/mixed_precision/MPLinearKernel.py:ro
      # vLLM PR #35936 required-tool fallback (drop when upstream lands).
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/chat_completion/serving.py:ro
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/engine/serving.py:ro
+      # vLLM PR #35936 required-tool fallback — sidecar pattern (drop when upstream lands).
+      # Bind-mounted at side paths so install.sh can copy into vLLM's site-packages
+      # BEFORE Genesis runs — avoids the RO-mount conflict with Genesis P64/P68/P69
+      # which write to chat_completion/serving.py at vllm-import time. See
+      # patches/vllm-pr35936-required-fallback/install.sh + README.md.
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/etc/club3090/pr35936-chat-completion-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/etc/club3090/pr35936-engine-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/install.sh:/etc/club3090/install-pr35936.sh:ro
      # froggeric/Qwen-Fixed-Chat-Templates qwen3.6 — fixes 7 default-template
      # bugs (empty <think></think> spam, </thinking> hallucination, unclosed
      # think before tool call, no-user-query crash, developer role, etc.).
@@ -74,6 +80,8 @@ services:
      - bash
      - -c
      - |
+        # Install PR #35936 overlay before vllm imports (drop when upstream lands).
+        bash /etc/club3090/install-pr35936.sh
        exec vllm serve ${VLLM_ENFORCE_EAGER:+--enforce-eager} "$@"
      - --
    command:
@@ -94,8 +94,14 @@ services:
      - ../../patches/vllm-marlin-pad/marlin.py:/usr/local/lib/python3.12/dist-packages/vllm/model_executor/kernels/linear/mixed_precision/marlin.py:ro
      - ../../patches/vllm-marlin-pad/MPLinearKernel.py:/usr/local/lib/python3.12/dist-packages/vllm/model_executor/kernels/linear/mixed_precision/MPLinearKernel.py:ro
      # vLLM PR #35936 required-tool fallback (drop when upstream lands).
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/chat_completion/serving.py:ro
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/engine/serving.py:ro
+      # vLLM PR #35936 required-tool fallback — sidecar pattern (drop when upstream lands).
+      # Bind-mounted at side paths so install.sh can copy into vLLM's site-packages
+      # BEFORE Genesis runs — avoids the RO-mount conflict with Genesis P64/P68/P69
+      # which write to chat_completion/serving.py at vllm-import time. See
+      # patches/vllm-pr35936-required-fallback/install.sh + README.md.
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/etc/club3090/pr35936-chat-completion-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/etc/club3090/pr35936-engine-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/install.sh:/etc/club3090/install-pr35936.sh:ro
      # froggeric/Qwen-Fixed-Chat-Templates qwen3.6 — fixes 7 default-template
      # bugs (empty <think></think> spam, </thinking> hallucination, unclosed
      # think before tool call, no-user-query crash, developer role, etc.).
@@ -131,6 +137,8 @@ services:
      - |
        # VLLM_ENFORCE_EAGER=1 in .env disables CUDA graphs — use on
        # hardware where graph capture causes OOM or instability (e.g. WSL2).
+        # Install PR #35936 overlay before vllm imports (drop when upstream lands).
+        bash /etc/club3090/install-pr35936.sh
        exec vllm serve ${VLLM_ENFORCE_EAGER:+--enforce-eager} "$@"
      - --
    command:
@@ -90,8 +90,14 @@ services:
      - ../../patches/vllm-marlin-pad/marlin.py:/usr/local/lib/python3.12/dist-packages/vllm/model_executor/kernels/linear/mixed_precision/marlin.py:ro
      - ../../patches/vllm-marlin-pad/MPLinearKernel.py:/usr/local/lib/python3.12/dist-packages/vllm/model_executor/kernels/linear/mixed_precision/MPLinearKernel.py:ro
      # vLLM PR #35936 required-tool fallback (drop when upstream lands).
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/chat_completion/serving.py:ro
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/engine/serving.py:ro
+      # vLLM PR #35936 required-tool fallback — sidecar pattern (drop when upstream lands).
+      # Bind-mounted at side paths so install.sh can copy into vLLM's site-packages
+      # BEFORE Genesis runs — avoids the RO-mount conflict with Genesis P64/P68/P69
+      # which write to chat_completion/serving.py at vllm-import time. See
+      # patches/vllm-pr35936-required-fallback/install.sh + README.md.
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/etc/club3090/pr35936-chat-completion-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/etc/club3090/pr35936-engine-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/install.sh:/etc/club3090/install-pr35936.sh:ro
      # froggeric/Qwen-Fixed-Chat-Templates qwen3.6 — fixes 7 default-template
      # bugs (empty <think></think> spam, </thinking> hallucination, unclosed
      # think before tool call, no-user-query crash, developer role, etc.).
@@ -127,6 +133,8 @@ services:
      - |
        # VLLM_ENFORCE_EAGER=1 in .env disables CUDA graphs — use on
        # hardware where graph capture causes OOM or instability (e.g. WSL2).
+        # Install PR #35936 overlay before vllm imports (drop when upstream lands).
+        bash /etc/club3090/install-pr35936.sh
        exec vllm serve ${VLLM_ENFORCE_EAGER:+--enforce-eager} "$@"
      - --
    command:
@@ -87,8 +87,14 @@ services:
      - ../../patches/genesis/vllm/_genesis:/usr/local/lib/python3.12/dist-packages/vllm/_genesis:ro
      - ../../patches/local/qwen3coder_tool_parser_deferred_commit.py:/patches/qwen3coder_tool_parser_deferred_commit.py:ro
      # vLLM PR #35936 required-tool fallback (drop when upstream lands).
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/chat_completion/serving.py:ro
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/engine/serving.py:ro
+      # vLLM PR #35936 required-tool fallback — sidecar pattern (drop when upstream lands).
+      # Bind-mounted at side paths so install.sh can copy into vLLM's site-packages
+      # BEFORE Genesis runs — avoids the RO-mount conflict with Genesis P64/P68/P69
+      # which write to chat_completion/serving.py at vllm-import time. See
+      # patches/vllm-pr35936-required-fallback/install.sh + README.md.
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/etc/club3090/pr35936-chat-completion-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/etc/club3090/pr35936-engine-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/install.sh:/etc/club3090/install-pr35936.sh:ro
      # froggeric/Qwen-Fixed-Chat-Templates qwen3.6 — fixes 7 default-template
      # bugs (empty <think></think> spam, </thinking> hallucination, unclosed
      # think before tool call, no-user-query crash, developer role, etc.).
@@ -222,6 +228,9 @@ services:
          echo "         bash scripts/setup.sh qwen3.6-27b" >&2
          exit 1
        fi
+        # Install PR #35936 overlay BEFORE Genesis runs so Genesis can write hooks
+        # to chat_completion/serving.py without hitting RO-mount errors.
+        bash /etc/club3090/install-pr35936.sh
        python3 -m vllm._genesis.patches.apply_all
        # Tool-parser deferred-commit fix for qwen3coder SSE-silence bug (issue #72).
        # Drops out when vllm-project/vllm lands the upstream fix.
@@ -80,8 +80,14 @@ services:
      - ../../patches/vllm-marlin-pad/marlin.py:/usr/local/lib/python3.12/dist-packages/vllm/model_executor/kernels/linear/mixed_precision/marlin.py:ro
      - ../../patches/vllm-marlin-pad/MPLinearKernel.py:/usr/local/lib/python3.12/dist-packages/vllm/model_executor/kernels/linear/mixed_precision/MPLinearKernel.py:ro
      # vLLM PR #35936 required-tool fallback (drop when upstream lands).
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/chat_completion/serving.py:ro
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/engine/serving.py:ro
+      # vLLM PR #35936 required-tool fallback — sidecar pattern (drop when upstream lands).
+      # Bind-mounted at side paths so install.sh can copy into vLLM's site-packages
+      # BEFORE Genesis runs — avoids the RO-mount conflict with Genesis P64/P68/P69
+      # which write to chat_completion/serving.py at vllm-import time. See
+      # patches/vllm-pr35936-required-fallback/install.sh + README.md.
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/etc/club3090/pr35936-chat-completion-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/etc/club3090/pr35936-engine-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/install.sh:/etc/club3090/install-pr35936.sh:ro
      # froggeric/Qwen-Fixed-Chat-Templates qwen3.6 — fixes 7 default-template
      # bugs (empty <think></think> spam, </thinking> hallucination, unclosed
      # think before tool call, no-user-query crash, developer role, etc.).
@@ -119,6 +125,8 @@ services:
        # hardware where Cliff 2 GDN activation spikes occur at runtime
        # (~50-65K active context tokens). Costs ~20-30% TPS in exchange for
        # stability. See docs/HARDWARE.md "Note for WSL2 / Windows users".
+        # Install PR #35936 overlay before vllm imports (drop when upstream lands).
+        bash /etc/club3090/install-pr35936.sh
        exec vllm serve ${VLLM_ENFORCE_EAGER:+--enforce-eager} "$@"
      - --
    command:
@@ -49,8 +49,14 @@ services:
      - ../../patches/vllm-marlin-pad/marlin.py:/usr/local/lib/python3.12/dist-packages/vllm/model_executor/kernels/linear/mixed_precision/marlin.py:ro
      - ../../patches/vllm-marlin-pad/MPLinearKernel.py:/usr/local/lib/python3.12/dist-packages/vllm/model_executor/kernels/linear/mixed_precision/MPLinearKernel.py:ro
      # vLLM PR #35936 required-tool fallback (drop when upstream lands).
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/chat_completion/serving.py:ro
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/engine/serving.py:ro
+      # vLLM PR #35936 required-tool fallback — sidecar pattern (drop when upstream lands).
+      # Bind-mounted at side paths so install.sh can copy into vLLM's site-packages
+      # BEFORE Genesis runs — avoids the RO-mount conflict with Genesis P64/P68/P69
+      # which write to chat_completion/serving.py at vllm-import time. See
+      # patches/vllm-pr35936-required-fallback/install.sh + README.md.
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/etc/club3090/pr35936-chat-completion-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/etc/club3090/pr35936-engine-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/install.sh:/etc/club3090/install-pr35936.sh:ro
    environment:
      - HUGGING_FACE_HUB_TOKEN=${HF_TOKEN:-}
      - VLLM_WORKER_MULTIPROC_METHOD=spawn
@@ -69,6 +75,14 @@ services:
            - driver: nvidia
              count: all
              capabilities: [gpu]
+    entrypoint:
+      - bash
+      - -c
+      - |
+        # Install PR #35936 overlay before vllm imports (drop when upstream lands).
+        bash /etc/club3090/install-pr35936.sh
+        exec vllm serve ${VLLM_ENFORCE_EAGER:+--enforce-eager} "$@"
+      - --
    command:
      - --model
      - /root/.cache/huggingface/qwopus3.6-27b-int4-recipe-d-bf16mtp
@@ -194,7 +194,14 @@ services:
      # PN54 — GDN contiguous-call deduplication (Cliff 2b OOM mitigation).
      - GENESIS_ENABLE_PN54=${GENESIS_ENABLE_PN54:-0}
      # Explicit OFFs to match Sandermage's PROD env-var set verbatim:
-      # P78 (P78_TOLIST_CAPTURE_GUARD) — superseded by our patch_tolist_cudagraph.py
+      # P78 (P78_TOLIST_CAPTURE_GUARD) — vllm#40807 `.tolist()` cudagraph guard.
+      #   Was OFF here because we previously shipped our own patch_tolist_cudagraph.py
+      #   sidecar covering the same bug; that sidecar was retired 2026-05-05 (Genesis
+      #   v7.72.2 bump). We keep P78=0 because rebench-full leg 6 (2026-05-11) passed
+      #   clean with it off — 7/7 verify-stress, 0/100 silent-empty soak, 18/30 aider.
+      #   PN34 (workspace-lock relax) + post-#41434 attention rework appear to cover
+      #   the symptom on this Genesis-pinned nightly. Flip to 1 if a future bench
+      #   surfaces .tolist() cudagraph corruption.
      # P81 (FP8 block-scaled M<=8) — FP8-specific, no-op on our TQ3 path
      # P82 — biased on small-batch single-stream Lorbus INT4 + MTP K=3 (Sander PROD)
      - GENESIS_ENABLE_P78_TOLIST_CAPTURE_GUARD=0
@@ -91,8 +91,14 @@ services:
      # cudagraph path off the table entirely.
      - ../../patches/vllm-pr40798-rebased/v1/worker/gpu_model_runner.py:/usr/local/lib/python3.12/dist-packages/vllm/v1/worker/gpu_model_runner.py:ro
      # vLLM PR #35936 required-tool fallback (drop when upstream lands).
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/chat_completion/serving.py:ro
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/engine/serving.py:ro
+      # vLLM PR #35936 required-tool fallback — sidecar pattern (drop when upstream lands).
+      # Bind-mounted at side paths so install.sh can copy into vLLM's site-packages
+      # BEFORE Genesis runs — avoids the RO-mount conflict with Genesis P64/P68/P69
+      # which write to chat_completion/serving.py at vllm-import time. See
+      # patches/vllm-pr35936-required-fallback/install.sh + README.md.
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/etc/club3090/pr35936-chat-completion-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/etc/club3090/pr35936-engine-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/install.sh:/etc/club3090/install-pr35936.sh:ro
      # froggeric/Qwen-Fixed-Chat-Templates qwen3.6 — fixes 7 default-template
      # bugs (empty <think></think> spam, </thinking> hallucination, unclosed
      # think before tool call, no-user-query crash, developer role, etc.).
@@ -122,6 +128,8 @@ services:
      - bash
      - -c
      - |
+        # Install PR #35936 overlay before vllm imports (drop when upstream lands).
+        bash /etc/club3090/install-pr35936.sh
        exec vllm serve ${VLLM_ENFORCE_EAGER:+--enforce-eager} "$@"
      - --
    command:
@@ -75,8 +75,14 @@ services:
      # used; nightly already has the kernel + caller for the rest.
      - ../../patches/vllm-pr40798-rebased/v1/worker/gpu_model_runner.py:/usr/local/lib/python3.12/dist-packages/vllm/v1/worker/gpu_model_runner.py:ro
      # vLLM PR #35936 required-tool fallback (drop when upstream lands).
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/chat_completion/serving.py:ro
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/engine/serving.py:ro
+      # vLLM PR #35936 required-tool fallback — sidecar pattern (drop when upstream lands).
+      # Bind-mounted at side paths so install.sh can copy into vLLM's site-packages
+      # BEFORE Genesis runs — avoids the RO-mount conflict with Genesis P64/P68/P69
+      # which write to chat_completion/serving.py at vllm-import time. See
+      # patches/vllm-pr35936-required-fallback/install.sh + README.md.
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/etc/club3090/pr35936-chat-completion-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/etc/club3090/pr35936-engine-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/install.sh:/etc/club3090/install-pr35936.sh:ro
      # froggeric/Qwen-Fixed-Chat-Templates qwen3.6 — fixes 7 default-template
      # bugs (empty <think></think> spam, </thinking> hallucination, unclosed
      # think before tool call, no-user-query crash, developer role, etc.).
@@ -105,6 +111,8 @@ services:
      - bash
      - -c
      - |
+        # Install PR #35936 overlay before vllm imports (drop when upstream lands).
+        bash /etc/club3090/install-pr35936.sh
        exec vllm serve ${VLLM_ENFORCE_EAGER:+--enforce-eager} "$@"
      - --
    command:
@@ -71,8 +71,14 @@ services:
      - ../../patches/genesis/vllm/_genesis:/usr/local/lib/python3.12/dist-packages/vllm/_genesis:ro
      - ../../patches/local/qwen3coder_tool_parser_deferred_commit.py:/patches/qwen3coder_tool_parser_deferred_commit.py:ro
      # vLLM PR #35936 required-tool fallback (drop when upstream lands).
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/chat_completion/serving.py:ro
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/engine/serving.py:ro
+      # vLLM PR #35936 required-tool fallback — sidecar pattern (drop when upstream lands).
+      # Bind-mounted at side paths so install.sh can copy into vLLM's site-packages
+      # BEFORE Genesis runs — avoids the RO-mount conflict with Genesis P64/P68/P69
+      # which write to chat_completion/serving.py at vllm-import time. See
+      # patches/vllm-pr35936-required-fallback/install.sh + README.md.
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/etc/club3090/pr35936-chat-completion-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/etc/club3090/pr35936-engine-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/install.sh:/etc/club3090/install-pr35936.sh:ro
      # froggeric/Qwen-Fixed-Chat-Templates qwen3.6 — fixes 7 default-template
      # bugs (empty <think></think> spam, </thinking> hallucination, unclosed
      # think before tool call, no-user-query crash, developer role, etc.).
@@ -204,6 +210,9 @@ services:
          echo "         bash scripts/setup.sh qwen3.6-27b" >&2
          exit 1
        fi
+        # Install PR #35936 overlay BEFORE Genesis runs so Genesis can write hooks
+        # to chat_completion/serving.py without hitting RO-mount errors.
+        bash /etc/club3090/install-pr35936.sh
        python3 -m vllm._genesis.patches.apply_all
        # Tool-parser deferred-commit fix for qwen3coder SSE-silence bug (issue #72).
        # Drops out when vllm-project/vllm lands the upstream fix.
@@ -79,8 +79,14 @@ services:
      - ../../patches/vllm-marlin-pad/marlin.py:/usr/local/lib/python3.12/dist-packages/vllm/model_executor/kernels/linear/mixed_precision/marlin.py:ro
      - ../../patches/vllm-marlin-pad/MPLinearKernel.py:/usr/local/lib/python3.12/dist-packages/vllm/model_executor/kernels/linear/mixed_precision/MPLinearKernel.py:ro
      # vLLM PR #35936 required-tool fallback (drop when upstream lands).
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/chat_completion/serving.py:ro
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/engine/serving.py:ro
+      # vLLM PR #35936 required-tool fallback — sidecar pattern (drop when upstream lands).
+      # Bind-mounted at side paths so install.sh can copy into vLLM's site-packages
+      # BEFORE Genesis runs — avoids the RO-mount conflict with Genesis P64/P68/P69
+      # which write to chat_completion/serving.py at vllm-import time. See
+      # patches/vllm-pr35936-required-fallback/install.sh + README.md.
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/etc/club3090/pr35936-chat-completion-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/etc/club3090/pr35936-engine-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/install.sh:/etc/club3090/install-pr35936.sh:ro
      # froggeric/Qwen-Fixed-Chat-Templates qwen3.6 — fixes 7 default-template
      # bugs (empty <think></think> spam, </thinking> hallucination, unclosed
      # think before tool call, no-user-query crash, developer role, etc.).
@@ -114,6 +120,8 @@ services:
      - |
        # VLLM_ENFORCE_EAGER=1 in .env disables CUDA graphs — use on
        # hardware where graph capture causes OOM or instability (e.g. WSL2).
+        # Install PR #35936 overlay before vllm imports (drop when upstream lands).
+        bash /etc/club3090/install-pr35936.sh
        exec vllm serve ${VLLM_ENFORCE_EAGER:+--enforce-eager} "$@"
      - --
    command:
@@ -82,8 +82,14 @@ services:
      - ../../patches/vllm-marlin-pad/marlin.py:/usr/local/lib/python3.12/dist-packages/vllm/model_executor/kernels/linear/mixed_precision/marlin.py:ro
      - ../../patches/vllm-marlin-pad/MPLinearKernel.py:/usr/local/lib/python3.12/dist-packages/vllm/model_executor/kernels/linear/mixed_precision/MPLinearKernel.py:ro
      # vLLM PR #35936 required-tool fallback (drop when upstream lands).
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/chat_completion/serving.py:ro
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/engine/serving.py:ro
+      # vLLM PR #35936 required-tool fallback — sidecar pattern (drop when upstream lands).
+      # Bind-mounted at side paths so install.sh can copy into vLLM's site-packages
+      # BEFORE Genesis runs — avoids the RO-mount conflict with Genesis P64/P68/P69
+      # which write to chat_completion/serving.py at vllm-import time. See
+      # patches/vllm-pr35936-required-fallback/install.sh + README.md.
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/etc/club3090/pr35936-chat-completion-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/etc/club3090/pr35936-engine-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/install.sh:/etc/club3090/install-pr35936.sh:ro
      # froggeric/Qwen-Fixed-Chat-Templates qwen3.6 — fixes 7 default-template
      # bugs (empty <think></think> spam, </thinking> hallucination, unclosed
      # think before tool call, no-user-query crash, developer role, etc.).
@@ -120,6 +126,8 @@ services:
        # hardware where Cliff 2 GDN activation spikes occur at runtime
        # (~50-65K active context tokens). Costs ~20-30% TPS in exchange for
        # stability. See docs/HARDWARE.md "Note for WSL2 / Windows users".
+        # Install PR #35936 overlay before vllm imports (drop when upstream lands).
+        bash /etc/club3090/install-pr35936.sh
        exec vllm serve ${VLLM_ENFORCE_EAGER:+--enforce-eager} "$@"
      - --
    command:
@@ -148,8 +148,14 @@ services:
      # workspace_lock_disable — relaxes vllm#39226 strict assertion until
      # Sandermage's P98 marker fix lands. v0.20-only requirement.
      # vLLM PR #35936 required-tool fallback (drop when upstream lands).
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/chat_completion/serving.py:ro
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/engine/serving.py:ro
+      # vLLM PR #35936 required-tool fallback — sidecar pattern (drop when upstream lands).
+      # Bind-mounted at side paths so install.sh can copy into vLLM's site-packages
+      # BEFORE Genesis runs — avoids the RO-mount conflict with Genesis P64/P68/P69
+      # which write to chat_completion/serving.py at vllm-import time. See
+      # patches/vllm-pr35936-required-fallback/install.sh + README.md.
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/etc/club3090/pr35936-chat-completion-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/etc/club3090/pr35936-engine-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/install.sh:/etc/club3090/install-pr35936.sh:ro
      # froggeric/Qwen-Fixed-Chat-Templates qwen3.6 — fixes 7 default-template
      # bugs (empty <think></think> spam, </thinking> hallucination, unclosed
      # think before tool call, no-user-query crash, developer role, etc.).
@@ -275,6 +281,9 @@ services:
          echo "         bash scripts/setup.sh qwen3.6-27b" >&2
          exit 1
        fi
+        # Install PR #35936 overlay BEFORE Genesis runs so Genesis can write hooks
+        # to chat_completion/serving.py without hitting RO-mount errors.
+        bash /etc/club3090/install-pr35936.sh
        python3 -m vllm._genesis.patches.apply_all
        # Tool-parser deferred-commit fix for qwen3coder SSE-silence bug (issue #72).
        # Drops out when vllm-project/vllm lands the upstream fix.
@@ -119,8 +119,14 @@ services:
      # workspace_lock_disable — relaxes vllm#39226 strict assertion until
      # Sandermage's P98 marker fix lands. v0.20-only requirement.
      # vLLM PR #35936 required-tool fallback (drop when upstream lands).
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/chat_completion/serving.py:ro
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/engine/serving.py:ro
+      # vLLM PR #35936 required-tool fallback — sidecar pattern (drop when upstream lands).
+      # Bind-mounted at side paths so install.sh can copy into vLLM's site-packages
+      # BEFORE Genesis runs — avoids the RO-mount conflict with Genesis P64/P68/P69
+      # which write to chat_completion/serving.py at vllm-import time. See
+      # patches/vllm-pr35936-required-fallback/install.sh + README.md.
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/etc/club3090/pr35936-chat-completion-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/etc/club3090/pr35936-engine-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/install.sh:/etc/club3090/install-pr35936.sh:ro
      # froggeric/Qwen-Fixed-Chat-Templates qwen3.6 — fixes 7 default-template
      # bugs (empty <think></think> spam, </thinking> hallucination, unclosed
      # think before tool call, no-user-query crash, developer role, etc.).
@@ -206,6 +212,9 @@ services:
          echo "         bash scripts/setup.sh qwen3.6-27b" >&2
          exit 1
        fi
+        # Install PR #35936 overlay BEFORE Genesis runs so Genesis can write hooks
+        # to chat_completion/serving.py without hitting RO-mount errors.
+        bash /etc/club3090/install-pr35936.sh
        python3 -m vllm._genesis.patches.apply_all
        # Tool-parser deferred-commit fix for qwen3coder SSE-silence bug (issue #72).
        # Drops out when vllm-project/vllm lands the upstream fix.
@@ -136,8 +136,14 @@ services:
      # surface natively (PN12 native + PN25 + PN17 + P15B). See migration
      # commit history if you need to resurrect.
      # vLLM PR #35936 required-tool fallback (drop when upstream lands).
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/chat_completion/serving.py:ro
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/engine/serving.py:ro
+      # vLLM PR #35936 required-tool fallback — sidecar pattern (drop when upstream lands).
+      # Bind-mounted at side paths so install.sh can copy into vLLM's site-packages
+      # BEFORE Genesis runs — avoids the RO-mount conflict with Genesis P64/P68/P69
+      # which write to chat_completion/serving.py at vllm-import time. See
+      # patches/vllm-pr35936-required-fallback/install.sh + README.md.
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/etc/club3090/pr35936-chat-completion-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/etc/club3090/pr35936-engine-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/install.sh:/etc/club3090/install-pr35936.sh:ro
      # froggeric/Qwen-Fixed-Chat-Templates qwen3.6 — fixes 7 default-template
      # bugs (empty <think></think> spam, </thinking> hallucination, unclosed
      # think before tool call, no-user-query crash, developer role, etc.).
@@ -300,6 +306,9 @@ services:
          echo "         bash scripts/setup.sh qwen3.6-27b" >&2
          exit 1
        fi
+        # Install PR #35936 overlay BEFORE Genesis runs so Genesis can write hooks
+        # to chat_completion/serving.py without hitting RO-mount errors.
+        bash /etc/club3090/install-pr35936.sh
        python3 -m vllm._genesis.patches.apply_all
        # Tool-parser deferred-commit fix for qwen3coder SSE-silence bug (issue #72).
        # Drops out when vllm-project/vllm lands the upstream fix.
@@ -146,8 +146,14 @@ services:
      # surface natively (PN12 native + PN25 + PN17 + P15B). See migration
      # commit history if you need to resurrect.
      # vLLM PR #35936 required-tool fallback (drop when upstream lands).
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/chat_completion/serving.py:ro
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/engine/serving.py:ro
+      # vLLM PR #35936 required-tool fallback — sidecar pattern (drop when upstream lands).
+      # Bind-mounted at side paths so install.sh can copy into vLLM's site-packages
+      # BEFORE Genesis runs — avoids the RO-mount conflict with Genesis P64/P68/P69
+      # which write to chat_completion/serving.py at vllm-import time. See
+      # patches/vllm-pr35936-required-fallback/install.sh + README.md.
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/etc/club3090/pr35936-chat-completion-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/etc/club3090/pr35936-engine-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/install.sh:/etc/club3090/install-pr35936.sh:ro
      # froggeric/Qwen-Fixed-Chat-Templates qwen3.6 — fixes 7 default-template
      # bugs (empty <think></think> spam, </thinking> hallucination, unclosed
      # think before tool call, no-user-query crash, developer role, etc.).
@@ -317,6 +323,9 @@ services:
          echo "         bash scripts/setup.sh qwen3.6-27b" >&2
          exit 1
        fi
+        # Install PR #35936 overlay BEFORE Genesis runs so Genesis can write hooks
+        # to chat_completion/serving.py without hitting RO-mount errors.
+        bash /etc/club3090/install-pr35936.sh
        python3 -m vllm._genesis.patches.apply_all
        # Tool-parser deferred-commit fix for qwen3coder SSE-silence bug (issue #72).
        # Drops out when vllm-project/vllm lands the upstream fix.
@@ -114,8 +114,14 @@ services:
      # workspace_lock_disable — relaxes vllm#39226 strict assertion until
      # Sandermage's P98 marker fix lands. v0.20-only requirement.
      # vLLM PR #35936 required-tool fallback (drop when upstream lands).
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/chat_completion/serving.py:ro
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/engine/serving.py:ro
+      # vLLM PR #35936 required-tool fallback — sidecar pattern (drop when upstream lands).
+      # Bind-mounted at side paths so install.sh can copy into vLLM's site-packages
+      # BEFORE Genesis runs — avoids the RO-mount conflict with Genesis P64/P68/P69
+      # which write to chat_completion/serving.py at vllm-import time. See
+      # patches/vllm-pr35936-required-fallback/install.sh + README.md.
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/etc/club3090/pr35936-chat-completion-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/etc/club3090/pr35936-engine-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/install.sh:/etc/club3090/install-pr35936.sh:ro
      # froggeric/Qwen-Fixed-Chat-Templates qwen3.6 — fixes 7 default-template
      # bugs (empty <think></think> spam, </thinking> hallucination, unclosed
      # think before tool call, no-user-query crash, developer role, etc.).
@@ -241,6 +247,9 @@ services:
          echo "         bash scripts/setup.sh qwen3.6-27b" >&2
          exit 1
        fi
+        # Install PR #35936 overlay BEFORE Genesis runs so Genesis can write hooks
+        # to chat_completion/serving.py without hitting RO-mount errors.
+        bash /etc/club3090/install-pr35936.sh
        python3 -m vllm._genesis.patches.apply_all
        # Tool-parser deferred-commit fix for qwen3coder SSE-silence bug (issue #72).
        # Drops out when vllm-project/vllm lands the upstream fix.
@@ -46,8 +46,14 @@ services:
      - ../../cache/torch_compile:/root/.cache/vllm/torch_compile_cache
      - ../../cache/triton:/root/.triton/cache
      # vLLM PR #35936 required-tool fallback (drop when upstream lands).
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/chat_completion/serving.py:ro
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/engine/serving.py:ro
+      # vLLM PR #35936 required-tool fallback — sidecar pattern (drop when upstream lands).
+      # Bind-mounted at side paths so install.sh can copy into vLLM's site-packages
+      # BEFORE Genesis runs — avoids the RO-mount conflict with Genesis P64/P68/P69
+      # which write to chat_completion/serving.py at vllm-import time. See
+      # patches/vllm-pr35936-required-fallback/install.sh + README.md.
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/etc/club3090/pr35936-chat-completion-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/etc/club3090/pr35936-engine-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/install.sh:/etc/club3090/install-pr35936.sh:ro
      # froggeric/Qwen-Fixed-Chat-Templates qwen3.6 — fixes 7 default-template
      # bugs (empty <think></think> spam, </thinking> hallucination, unclosed
      # think before tool call, no-user-query crash, developer role, etc.).
@@ -82,6 +88,8 @@ services:
      - |
        # VLLM_ENFORCE_EAGER=1 in .env disables CUDA graphs — use on
        # hardware where graph capture causes OOM or instability (e.g. WSL2).
+        # Install PR #35936 overlay before vllm imports (drop when upstream lands).
+        bash /etc/club3090/install-pr35936.sh
        exec vllm serve ${VLLM_ENFORCE_EAGER:+--enforce-eager} "$@"
      - --
    command:
@@ -51,8 +51,14 @@ services:
      - ../../patches/genesis/vllm/_genesis:/usr/local/lib/python3.12/dist-packages/vllm/_genesis:ro
      - ../../patches/local/qwen3coder_tool_parser_deferred_commit.py:/patches/qwen3coder_tool_parser_deferred_commit.py:ro
      # vLLM PR #35936 required-tool fallback (drop when upstream lands).
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/chat_completion/serving.py:ro
-      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/engine/serving.py:ro
+      # vLLM PR #35936 required-tool fallback — sidecar pattern (drop when upstream lands).
+      # Bind-mounted at side paths so install.sh can copy into vLLM's site-packages
+      # BEFORE Genesis runs — avoids the RO-mount conflict with Genesis P64/P68/P69
+      # which write to chat_completion/serving.py at vllm-import time. See
+      # patches/vllm-pr35936-required-fallback/install.sh + README.md.
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/chat_completion/serving.py:/etc/club3090/pr35936-chat-completion-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/vllm/entrypoints/openai/engine/serving.py:/etc/club3090/pr35936-engine-serving.py:ro
+      - ../../patches/vllm-pr35936-required-fallback/install.sh:/etc/club3090/install-pr35936.sh:ro
      # froggeric/Qwen-Fixed-Chat-Templates qwen3.6 — fixes 7 default-template
      # bugs (empty <think></think> spam, </thinking> hallucination, unclosed
      # think before tool call, no-user-query crash, developer role, etc.).
@@ -130,6 +136,9 @@ services:
          echo "         bash scripts/setup.sh qwen3.6-27b" >&2
          exit 1
        fi
+        # Install PR #35936 overlay BEFORE Genesis runs so Genesis can write hooks
+        # to chat_completion/serving.py without hitting RO-mount errors.
+        bash /etc/club3090/install-pr35936.sh
        python3 -m vllm._genesis.patches.apply_all
        # Tool-parser deferred-commit fix for qwen3coder SSE-silence bug (issue #72).
        # Drops out when vllm-project/vllm lands the upstream fix.
@@ -49,10 +49,48 @@ This overlay therefore lands the PR's intent in **both** places in
   as before so XML-emitting paths keep working.

 The chat completion file is vendored unchanged from the pinned nightly as
-a matching drop-in mount target. The PR's streaming-side hunks target a
-pre-parser-manager code path that has been refactored on `nightly-1acd67a79`;
-non-streaming clients (MLS-Bench, our curl repro, most agent harnesses)
-exercise only the engine-side fix.
+a future-ready slot. The PR's streaming-side hunks target a pre-parser-manager
+code path that has been refactored on `nightly-1acd67a79`; non-streaming
+clients (MLS-Bench, our curl repro, most agent harnesses) exercise only the
+engine-side fix.
+
+## Installation: sidecar pattern (v0.5.1+)
+
+The overlay is **not** mounted directly at vLLM's site-packages paths. Instead:
+
+1. Both files are bind-mounted at side paths under `/etc/club3090/`:
+   - `pr35936-chat-completion-serving.py`
+   - `pr35936-engine-serving.py`
+2. `install.sh` is bind-mounted at `/etc/club3090/install-pr35936.sh`
+3. The compose entrypoint invokes `bash /etc/club3090/install-pr35936.sh`
+   **before** any other patch step (`python3 -m vllm._genesis.patches.apply_all`
+   in Genesis-loaded composes, or `exec vllm serve` in Genesis-less ones).
+4. `install.sh` copies our files into vLLM's site-packages with `cp`. The
+   destination becomes a writable file in the container's RW layer.
+
+### Why a sidecar instead of an RO bind-mount
+
+v0.5.0 originally mounted both files directly at vLLM's site-packages paths
+with `:ro`. That broke 8 Genesis-loaded composes because Genesis P64
+(qwen3coder MTP streaming early-return), P68 (auto force tool_choice=required),
+and P69 (long-context tool-format reminder) all write hooks into
+`chat_completion/serving.py` at vllm-import time. The RO mount blocked those
+writes with `Errno 30: Read-only file system`, and Genesis explicitly warned
+"partial state risk; container should be torn down." Reported by @ygafarov in
+[#120](https://github.com/noonghunna/club-3090/issues/120#issuecomment-4443236686);
+sidecar pattern shipped in v0.5.1.
+
+The sidecar resolves it cleanly:
+- Our patched files land in the container's RW layer (not bind-mounted RO)
+- Genesis can write its hooks on top freely
+- Host patches dir stays canonical (no Genesis hooks bleeding into our git tree)
+- Each container restart starts fresh; Genesis re-applies on a clean copy
+
+If a future patch to our overlay also touches `chat_completion/serving.py`
+(e.g. when PR #35936's streaming hunks land for a nightly we pin to), the
+sidecar layout supports it without further changes — our patched content
+sits on disk in the install.sh source path, gets installed before Genesis
+runs, Genesis layers its hooks on top.

 ## Validation

@@ -68,6 +106,12 @@ End-to-end checked 2026-05-12:
  agent completes loop with non-zero steps (was "No action returned after
  3 attempts" pre-overlay).

+Sidecar pattern validated 2026-05-13 on `single/long-text.yml` (Genesis-loaded):
+- `install.sh` runs before Genesis: "chat_completion/serving.py installed from /etc/club3090/..."
+- Genesis P64 succeeds: "P64 applied: 2 files modified, 0 idempotent" (was failing with `Read-only file system` pre-v0.5.1)
+- Genesis P68/P69 succeed: "Hook injected into create_chat_completion" (was failing pre-v0.5.1)
+- Zero `Read-only file system` errors anywhere in the boot log.
+
 ## Drop Trigger

 Remove this overlay when vLLM PR #35936, or an equivalent fix, is merged
@@ -0,0 +1,45 @@
+#!/usr/bin/env bash
+# Install PR #35936 required-tool fallback overlay into vLLM's site-packages.
+#
+# WHY a sidecar instead of an RO bind-mount:
+# Genesis P64 (qwen3coder MTP streaming early-return), P68 (auto-force
+# tool_choice=required), and P69 (long-context tool-format reminder) all
+# write hooks INTO chat_completion/serving.py at vllm-import time. An RO
+# bind-mount blocks those writes with `Errno 30: Read-only file system`,
+# and Genesis explicitly warns "partial state risk; container should be
+# torn down."
+#
+# Sidecar copies our files into the container's RW layer BEFORE Genesis
+# runs, so Genesis can write its hooks freely on top. Same pattern we
+# used for `patch_tolist_cudagraph.py` before Genesis absorbed it.
+#
+# Today chat_completion/serving.py is byte-identical to the upstream
+# nightly (PR #35936's streaming hunks don't apply on our pin), so
+# Genesis writes its hooks onto a clean copy. Slot stays future-ready
+# for when streaming hunks land — our changes will sit underneath
+# Genesis writes.
+#
+# Idempotent: cp overwrites unconditionally (per-boot fresh copy is
+# correct behaviour — Genesis re-applies on each fresh container).
+
+set -euo pipefail
+
+SRC_CC="${CLUB3090_PR35936_CHAT_COMPLETION_SRC:-/etc/club3090/pr35936-chat-completion-serving.py}"
+DST_CC="/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/chat_completion/serving.py"
+
+SRC_ENGINE="${CLUB3090_PR35936_ENGINE_SRC:-/etc/club3090/pr35936-engine-serving.py}"
+DST_ENGINE="/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/engine/serving.py"
+
+if [ -r "$SRC_CC" ]; then
+  cp "$SRC_CC" "$DST_CC"
+  echo "[club3090/pr35936] chat_completion/serving.py installed from $SRC_CC" >&2
+else
+  echo "[club3090/pr35936] WARN: $SRC_CC not found; chat_completion/serving.py left untouched" >&2
+fi
+
+if [ -r "$SRC_ENGINE" ]; then
+  cp "$SRC_ENGINE" "$DST_ENGINE"
+  echo "[club3090/pr35936] engine/serving.py installed from $SRC_ENGINE" >&2
+else
+  echo "[club3090/pr35936] WARN: $SRC_ENGINE not found; engine/serving.py left untouched (PR #35936 fix INACTIVE)" >&2
+fi
Author	SHA1	Message	Date
noonghunnaandClaude Opus 4.7	6617e1e090	fix(qwen): PR #35936 overlay — sidecar pattern to resolve Genesis RO-mount conflict Release / release (push) Failing after 52s Details v0.5.0 mounted vLLM PR #35936 overlay files as :ro bind-mounts at vLLM's site-packages paths. On Genesis-loaded composes, this blocked Genesis P64 (qwen3coder MTP streaming early-return), P68 (auto force tool_choice= required), and P69 (long-context tool-format reminder) — all three write hooks into chat_completion/serving.py at vllm-import time. Genesis reported "Errno 30: Read-only file system" and warned "partial state risk; container should be torn down." Reported by @ygafarov on club-3090#120: https://github.com/noonghunna/club-3090/issues/120#issuecomment-4443236686 This commit converts all 23 PR #35936-mounted composes to a sidecar pattern: 1. Patched files are bind-mounted at side paths under /etc/club3090/ (pr35936-chat-completion-serving.py, pr35936-engine-serving.py). 2. install.sh is bind-mounted at /etc/club3090/install-pr35936.sh. 3. Each compose entrypoint invokes `bash /etc/club3090/install-pr35936.sh` BEFORE Genesis runs (apply_all in Genesis-loaded composes, or vllm serve invocation in Genesis-less ones). 4. install.sh copies the files into vLLM's site-packages via `cp`, landing them in the container's RW layer. Result: Genesis can write its hooks freely on top of our content. No RO-mount errors. Host patches dir stays canonical. Three entrypoint variants handled: - A) Existing apply_all invocation (9 composes — Genesis-loaded): insert install.sh before `python3 -m vllm._genesis.patches.apply_all` - B) Simple `exec vllm serve` (13 composes — non-Genesis): prepend install.sh before exec - C) No entrypoint at all (1 compose — qwopus): add new entrypoint that runs install.sh then exec's vllm serve Validated 2026-05-13 on single/long-text.yml (Genesis-loaded): [club3090/pr35936] chat_completion/serving.py installed from /etc/club3090/... [INFO:genesis.apply_all] applied: P64 qwen3coder MTP streaming early-return fix — P64 applied: 2 files modified, 0 idempotent [INFO:genesis.apply_all] applied: P68/P69 long-context tool-call adherence — Hook injected into create_chat_completion Zero `Read-only file system` errors anywhere in boot log. Pattern is forward-compatible: if a future PR #35936 patch update touches chat_completion/serving.py (e.g. streaming hunks), the sidecar layout already supports layering our content under Genesis writes — no further plumbing changes needed. Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>	2026-05-13 17:01:33 +00:00
noonghunnaandClaude Opus 4.7	1678ca0c8b	docs: clarify MODEL_DIR — second drive / HF cache / Windows-WSL @tomByrer asked on disc #125 about pointing the scripts at a model dir on a second drive. The support has always been there (MODEL_DIR env var with 4-tier resolution: shell export, .env file, interactive prompt, silent fallback) but the docs were too terse to discover without running setup.sh interactively. Two changes: - docs/FAQ.md: expand the "where does it go?" entry from a one-liner to a full explanation covering all 4 resolution paths, HF env-var integration (we mount host dir INTO container cache, so HF_HOME isn't directly respected but layout is byte-compatible), and Windows/WSL2 path handling. - README.md: add a 3-line parenthetical to step 2 of the quickstart pointing at MODEL_DIR. Users who copy-paste the quickstart now see the option without having to dig. Both surfaces now point at the same canonical answer. Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>	2026-05-13 14:32:10 +00:00
noonghunnaandClaude Opus 4.7	14ffe45667	docs(upstream): correct stale vllm#40807 row + add #40798/#42215 row The UPSTREAM.md row for vllm#40807 (.tolist cudagraph crash) and the P78 comment in tq3-mtp-genesis.yml both claimed we ship a local patch_tolist_cudagraph.py sidecar. That sidecar was retired on 2026-05-05 when we bumped Genesis to v7.72.2 (which provides P78 as the equivalent guard) — but the docs never caught up. Two stale lines fixed: - docs/UPSTREAM.md vllm#40807 row: was "⚫ Local workaround", now "✅ Retired locally" with the actual current state (P78=0 on tq3-mtp-genesis.yml because rebench-full leg 6 passed clean without it; non-Genesis composes on 1acd67a79 pin run without any guard, unvalidated at long-context TurboQuant chunked-prefill). - tq3-mtp-genesis.yml line 197 comment: was "superseded by our patch_tolist_cudagraph.py" (the file we no longer ship), now explains the actual reason P78=0 (rebench evidence) plus the flip-to-1 trigger. Also adds a new UPSTREAM.md row for vllm#40798 + vllm#42215 — the upstream pair that closes the workspace-locked assertion at turboquant_attn.py:_continuation_prefill for ≥48K-token chunked- prefill with TurboQuant KV. Independently validated on 2× 3090 sm_86 by @cferra (vllm#41403 comment, 2026-05-12). On our stack, Genesis PN34 addresses the same symptom via a different mechanism (relax-lock vs reserve-before-capture); when the upstream pair propagates to a nightly we pin to, we'd A/B PN34 vs the pair. Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>	2026-05-12 22:11:24 +00:00
github-actions[bot]	722f998ff3	chore(changelog): regenerate for v0.5.0 [skip ci]	2026-05-12 21:23:33 +00:00