feat(llama-swap): sync vLLM configs from club-3090, add evan API key

Sync all three vLLM model configs from club-3090 master (ae4846f).
Update to Genesis v7.65 full PROD env set with new patches.
Update docker image to nightly-7a1eb8ac. Add torch_compile and
triton cache dirs. Add agent setup guide (AGENTS.md).

Add 'evan' API key to llama-swap sops secrets.

modules/nixos/services/llama-swap/AGENTS.md

@@ -0,0 +1,41 @@
+# llama-swap Module — Agent Guide
+
+## Syncing vLLM Configs from club-3090
+
+The three vLLM model configs in `config.nix` (`vllm-qwen3.6-27b-long-text`, `vllm-qwen3.6-27b-long-vision`, `vllm-qwen3.6-27b-tools-text`) are derived from the club-3090 repo's Docker Compose files. Each config block has a `Synced from:` comment with the commit hash it was last aligned to.
+
+### Source Files
+
+The upstream compose files live at https://github.com/noonghunna/club-3090 under `models/qwen3.6-27b/vllm/compose/`:
+
+| config.nix model ID               | Compose file                        |
+|------------------------------------|-------------------------------------|
+| `vllm-qwen3.6-27b-long-text`      | `docker-compose.long-text.yml`      |
+| `vllm-qwen3.6-27b-long-vision`    | `docker-compose.long-vision.yml`    |
+| `vllm-qwen3.6-27b-tools-text`     | `docker-compose.tools-text.yml`     |
+
+### Sync Process
+
+1. **Fetch the latest compose files** from https://github.com/noonghunna/club-3090 (master branch) and note the HEAD commit hash.
+2. **Diff each compose file** against the current config.nix block. The mapping is:
+   - Compose `command:` args → Nix `vllmCmd` string (the `exec vllm serve ...` block)
+   - Compose `environment:` → Nix docker `-e` flags
+   - Compose `volumes:` → Nix docker `-v` flags
+   - Compose `image:` → Nix docker image tag at the end of the `docker run` command
+   - Compose `entrypoint:` → Nix `vllmCmd` preamble (the `set -e; pip install ...; python3 ...` lines before `exec vllm serve`)
+3. **Apply changes** to `config.nix`. Key things to watch:
+   - `--max-model-len` and `--gpu-memory-utilization` — these change across versions
+   - Genesis env vars — the full set grows frequently; add new ones, remove deprecated ones
+   - Sidecar patches — old patches get absorbed into Genesis; drop them from entrypoint + volume mounts
+   - Docker image tag — update when the compose files move to a new nightly
+4. **Keep `patch_timings_07351e088.py`** — this is our own patch, not from club-3090. Always retain it in the entrypoint and volume mounts.
+5. **Update the `Synced from:` comment** on each config block with the new commit hash and date.
+6. **Update `setup-qwen36-vllm.sh`** if the upstream `patches/` directory changed (new patches added, old ones removed). The setup script downloads sidecar patches and creates cache directories.
+7. **Verify syntax**: `nix-instantiate --parse config.nix`
+
+### Structural Notes
+
+- `config.nix` uses Nix string interpolation. Newlines in `vllmCmd` are flattened to spaces via `builtins.replaceStrings` before passing to `docker run -c`.
+- We pin `CUDA_VISIBLE_DEVICES=0` and `CUDA_DEVICE_ORDER=PCI_BUS_ID` (not in compose files) because the host has multiple GPUs and llama-swap's concurrency matrix manages GPU assignment.
+- Volume mounts use `/mnt/ssd/vLLM/` paths (Models, Patches, Cache) — these match what `setup-qwen36-vllm.sh` creates.
+- The `patches/` subdirectory in this module contains our custom timings patch and its source `.patch` file — unrelated to club-3090's `patches/` dir.

modules/nixos/services/llama-swap/config.nix

@@ -97,19 +97,19 @@ in
       };
     };
 
-    # https://github.com/noonghunna/club-3090/tree/v0.20-experimental/models/qwen3.6-27b/vllm
-    # Long-text variant - v0.20 experimental single-3090 profile, text-only (no vision)
-    # TurboQuant 3-bit KV + MTP n=3 + workspace-lock sidecar.
+    # https://github.com/noonghunna/club-3090/tree/master/models/qwen3.6-27b/vllm
+    # Synced from: club-3090 ae4846f (2026-05-02) — docker-compose.long-text.yml
+    # Long-text variant - 180K context, text-only (no vision)
+    # TurboQuant 3-bit KV + MTP n=3 + Genesis v7.65 full PROD env set
     "vllm-qwen3.6-27b-long-text" = {
       name = "vLLM Qwen3.6 (27B) - Long Text";
-      macros.ctx = "214000";
+      macros.ctx = "180000";
       proxy = "http://127.0.0.1:\${PORT}";
       cmd =
         let
           vllmCmd = ''
             set -e; pip install xxhash pandas scipy -q;
             python3 -m vllm._genesis.patches.apply_all;
-            python3 /patches/patch_pn12_compile_safe_custom_op.py;
             python3 /patches/patch_workspace_lock_disable.py;
             python3 /patches/patch_tolist_cudagraph.py;
             python3 /patches/patch_timings_07351e088.py;
@@ -120,7 +120,7 @@ in
             --dtype float16
             --tensor-parallel-size 1
             --max-model-len ''${ctx}
-            --gpu-memory-utilization 0.985
+            --gpu-memory-utilization 0.95
             --max-num-seqs 1
             --max-num-batched-tokens 4128
             --kv-cache-dtype turboquant_3bit_nc
@@ -156,20 +156,64 @@ in
             -e CUDA_DEVICE_ORDER=PCI_BUS_ID \
             -e VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 \
             -e VLLM_MARLIN_USE_ATOMIC_ADD=1 \
+            -e TRITON_CACHE_DIR=/root/.triton/cache \
+            -e GENESIS_ENABLE_P4=1 \
+            -e GENESIS_ENABLE_P58_ASYNC_PLACEHOLDER_FIX=1 \
+            -e GENESIS_ENABLE_P60_GDN_NGRAM_FIX=1 \
+            -e GENESIS_ENABLE_P60B_TRITON_KERNEL=1 \
+            -e GENESIS_ENABLE_P61_QWEN3_MULTI_TOOL=1 \
+            -e GENESIS_ENABLE_P61B_STREAMING_OVERLAP=1 \
+            -e GENESIS_ENABLE_P62_STRUCT_OUT_SPEC_TIMING=1 \
+            -e GENESIS_ENABLE_P64_QWEN3CODER_MTP_STREAMING=1 \
             -e GENESIS_ENABLE_P65_TURBOQUANT_SPEC_CG_DOWNGRADE=1 \
             -e GENESIS_ENABLE_P66_CUDAGRAPH_SIZE_FILTER=1 \
-            -e GENESIS_ENABLE_P64_QWEN3CODER_MTP_STREAMING=1 \
+            -e GENESIS_ENABLE_P67_TQ_MULTI_QUERY_KERNEL=1 \
+            -e GENESIS_ENABLE_P68_AUTO_FORCE_TOOL=1 \
+            -e GENESIS_ENABLE_P69_LONG_CTX_TOOL_REMINDER=1 \
+            -e GENESIS_P68_P69_LONG_CTX_THRESHOLD_CHARS=50000 \
+            -e GENESIS_ENABLE_P72_PROFILE_RUN_CAP=1 \
+            -e GENESIS_PROFILE_RUN_CAP_M=4128 \
+            -e GENESIS_ENABLE_P74_CHUNK_CLAMP=1 \
+            -e GENESIS_ENABLE_P83=1 \
+            -e GENESIS_ENABLE_P85=1 \
+            -e GENESIS_ENABLE_P87=1 \
+            -e GENESIS_ENABLE_P91=1 \
+            -e GENESIS_ENABLE_P94=1 \
+            -e GENESIS_ENABLE_P98=1 \
+            -e GENESIS_ENABLE_P99=1 \
+            -e GENESIS_ENABLE_P100=1 \
             -e GENESIS_ENABLE_P101=1 \
             -e GENESIS_ENABLE_P103=1 \
+            -e GENESIS_ENABLE_PN8_MTP_DRAFT_ONLINE_QUANT=1 \
+            -e GENESIS_ENABLE_PN9_INDEPENDENT_DRAFTER_ATTN=1 \
+            -e GENESIS_ENABLE_PN11_GDN_AB_CONTIGUOUS=1 \
             -e GENESIS_ENABLE_PN12_FFN_INTERMEDIATE_POOL=1 \
             -e GENESIS_ENABLE_PN13_CUDA_GRAPH_LAMBDA_ARITY=1 \
+            -e GENESIS_ENABLE_PN14_TQ_DECODE_OOB_CLAMP=1 \
             -e GENESIS_ENABLE_PN17_FA2_LSE_CLAMP=1 \
+            -e GENESIS_ENABLE_PN25_SILU_INDUCTOR_SAFE=1 \
+            -e GENESIS_ENABLE_PN30_DS_LAYOUT_SPEC_DECODE=1 \
+            -e GENESIS_PREALLOC_TOKEN_BUDGET=4128 \
+            -e GENESIS_BUFFER_MODE=shared \
+            -e GENESIS_ENABLE_P78_TOLIST_CAPTURE_GUARD=0 \
+            -e GENESIS_ENABLE_P81_FP8_BLOCK_SCALED_M_LE_8=0 \
+            -e GENESIS_ENABLE_P82=0 \
+            -e GENESIS_P82_THRESHOLD_SINGLE=0.3 \
             -e GENESIS_ENABLE_PN19_SCOPED_MAX_SPLIT=1 \
-            -e GENESIS_ENABLE_P98=1 \
+            -e GENESIS_ENABLE_PN22_LOCAL_ARGMAX_TP=1 \
+            -e GENESIS_ENABLE_PN26_SPARSE_V=1 \
+            -e GENESIS_PN26_SPARSE_V_BLOCK_KV=8 \
+            -e GENESIS_PN26_SPARSE_V_NUM_WARPS=4 \
+            -e GENESIS_PN26_SPARSE_V_THRESHOLD=0.01 \
+            -e GENESIS_ENABLE_P38B_COMPILE_SAFE=1 \
+            -e GENESIS_ENABLE_P15B_FA_VARLEN_CLAMP=1 \
+            -e VLLM_SSM_CONV_STATE_LAYOUT=DS \
+            -e VLLM_USE_FUSED_MOE_GROUPED_TOPK=1 \
             -v /mnt/ssd/vLLM/Models:/root/.cache/huggingface \
+            -v /mnt/ssd/vLLM/Cache/torch_compile:/root/.cache/vllm/torch_compile_cache \
+            -v /mnt/ssd/vLLM/Cache/triton:/root/.triton/cache \
             -v /mnt/ssd/vLLM/Patches/genesis/vllm/_genesis:/usr/local/lib/python3.12/dist-packages/vllm/_genesis:ro \
             -v /mnt/ssd/vLLM/Patches/patch_tolist_cudagraph.py:/patches/patch_tolist_cudagraph.py:ro \
-            -v /mnt/ssd/vLLM/Patches/patch_pn12_compile_safe_custom_op.py:/patches/patch_pn12_compile_safe_custom_op.py:ro \
             -v /mnt/ssd/vLLM/Patches/patch_workspace_lock_disable.py:/patches/patch_workspace_lock_disable.py:ro \
             -v /mnt/ssd/vLLM/Patches/patch_timings_07351e088.py:/patches/patch_timings_07351e088.py:ro \
             -p ''${PORT}:8000 \
@@ -188,20 +232,19 @@ in
     };
 
     # https://github.com/noonghunna/club-3090/tree/master/models/qwen3.6-27b/vllm
-    # Long-vision variant - 140K context with vision tower active
-    # TurboQuant 3-bit KV + MTP n=3 + PN12/P104 cliff-closure stack
+    # Synced from: club-3090 ae4846f (2026-05-02) — docker-compose.long-vision.yml
+    # Long-vision variant - 145K context with vision tower active
+    # TurboQuant 3-bit KV + MTP n=3 + Genesis v7.65 full PROD env set
     "vllm-qwen3.6-27b-long-vision" = {
       name = "vLLM Qwen3.6 (27B) - Long Vision";
-      macros.ctx = "140000";
+      macros.ctx = "145000";
       proxy = "http://127.0.0.1:\${PORT}";
       cmd =
         let
           vllmCmd = ''
             set -e; pip install xxhash pandas scipy -q;
             python3 -m vllm._genesis.patches.apply_all;
-            python3 /patches/patch_pn12_ffn_pool_anchor.py;
-            python3 /patches/patch_pn12_compile_safe_custom_op.py;
-            python3 /patches/patch_fa_max_seqlen_clamp.py;
+            python3 /patches/patch_workspace_lock_disable.py;
             python3 /patches/patch_tolist_cudagraph.py;
             python3 /patches/patch_timings_07351e088.py;
             exec vllm serve
@@ -235,7 +278,7 @@ in
             -e VLLM_WORKER_MULTIPROC_METHOD=spawn \
             -e NCCL_CUMEM_ENABLE=0 \
             -e NCCL_P2P_DISABLE=1 \
-            -e VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=1 \
+            -e VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=0 \
             -e VLLM_NO_USAGE_STATS=1 \
             -e PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True,max_split_size_mb:512 \
             -e VLLM_FLOAT32_MATMUL_PRECISION=high \
@@ -246,25 +289,69 @@ in
             -e CUDA_DEVICE_ORDER=PCI_BUS_ID \
             -e VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 \
             -e VLLM_MARLIN_USE_ATOMIC_ADD=1 \
+            -e TRITON_CACHE_DIR=/root/.triton/cache \
+            -e GENESIS_ENABLE_P4=1 \
+            -e GENESIS_ENABLE_P58_ASYNC_PLACEHOLDER_FIX=1 \
+            -e GENESIS_ENABLE_P60_GDN_NGRAM_FIX=1 \
+            -e GENESIS_ENABLE_P60B_TRITON_KERNEL=1 \
+            -e GENESIS_ENABLE_P61_QWEN3_MULTI_TOOL=1 \
+            -e GENESIS_ENABLE_P61B_STREAMING_OVERLAP=1 \
+            -e GENESIS_ENABLE_P62_STRUCT_OUT_SPEC_TIMING=1 \
+            -e GENESIS_ENABLE_P64_QWEN3CODER_MTP_STREAMING=1 \
             -e GENESIS_ENABLE_P65_TURBOQUANT_SPEC_CG_DOWNGRADE=1 \
             -e GENESIS_ENABLE_P66_CUDAGRAPH_SIZE_FILTER=1 \
-            -e GENESIS_ENABLE_P64_QWEN3CODER_MTP_STREAMING=1 \
+            -e GENESIS_ENABLE_P67_TQ_MULTI_QUERY_KERNEL=1 \
+            -e GENESIS_ENABLE_P68_AUTO_FORCE_TOOL=1 \
+            -e GENESIS_ENABLE_P69_LONG_CTX_TOOL_REMINDER=1 \
+            -e GENESIS_P68_P69_LONG_CTX_THRESHOLD_CHARS=50000 \
+            -e GENESIS_ENABLE_P72_PROFILE_RUN_CAP=1 \
+            -e GENESIS_PROFILE_RUN_CAP_M=4128 \
+            -e GENESIS_ENABLE_P74_CHUNK_CLAMP=1 \
+            -e GENESIS_ENABLE_P83=1 \
+            -e GENESIS_ENABLE_P85=1 \
+            -e GENESIS_ENABLE_P87=1 \
+            -e GENESIS_ENABLE_P91=1 \
+            -e GENESIS_ENABLE_P94=1 \
+            -e GENESIS_ENABLE_P98=1 \
+            -e GENESIS_ENABLE_P99=1 \
+            -e GENESIS_ENABLE_P100=1 \
             -e GENESIS_ENABLE_P101=1 \
             -e GENESIS_ENABLE_P103=1 \
+            -e GENESIS_ENABLE_PN8_MTP_DRAFT_ONLINE_QUANT=1 \
+            -e GENESIS_ENABLE_PN9_INDEPENDENT_DRAFTER_ATTN=1 \
+            -e GENESIS_ENABLE_PN11_GDN_AB_CONTIGUOUS=1 \
             -e GENESIS_ENABLE_PN12_FFN_INTERMEDIATE_POOL=1 \
             -e GENESIS_ENABLE_PN13_CUDA_GRAPH_LAMBDA_ARITY=1 \
-            -e GENESIS_ENABLE_FA_MAX_SEQLEN_CLAMP=1 \
+            -e GENESIS_ENABLE_PN14_TQ_DECODE_OOB_CLAMP=1 \
             -e GENESIS_ENABLE_PN17_FA2_LSE_CLAMP=1 \
+            -e GENESIS_ENABLE_PN19_SCOPED_MAX_SPLIT=1 \
+            -e GENESIS_ENABLE_PN22_LOCAL_ARGMAX_TP=1 \
+            -e GENESIS_ENABLE_PN26_SPARSE_V=1 \
+            -e GENESIS_PN26_SPARSE_V_BLOCK_KV=8 \
+            -e GENESIS_PN26_SPARSE_V_NUM_WARPS=4 \
+            -e GENESIS_PN26_SPARSE_V_THRESHOLD=0.01 \
+            -e GENESIS_ENABLE_P38B_COMPILE_SAFE=1 \
+            -e GENESIS_ENABLE_P15B_FA_VARLEN_CLAMP=1 \
+            -e GENESIS_ENABLE_PN25_SILU_INDUCTOR_SAFE=1 \
+            -e GENESIS_ENABLE_PN30_DS_LAYOUT_SPEC_DECODE=1 \
+            -e GENESIS_PREALLOC_TOKEN_BUDGET=4128 \
+            -e GENESIS_BUFFER_MODE=shared \
+            -e GENESIS_ENABLE_P78_TOLIST_CAPTURE_GUARD=0 \
+            -e GENESIS_ENABLE_P81_FP8_BLOCK_SCALED_M_LE_8=0 \
+            -e GENESIS_ENABLE_P82=0 \
+            -e GENESIS_P82_THRESHOLD_SINGLE=0.3 \
+            -e VLLM_SSM_CONV_STATE_LAYOUT=DS \
+            -e VLLM_USE_FUSED_MOE_GROUPED_TOPK=1 \
             -v /mnt/ssd/vLLM/Models:/root/.cache/huggingface \
+            -v /mnt/ssd/vLLM/Cache/torch_compile:/root/.cache/vllm/torch_compile_cache \
+            -v /mnt/ssd/vLLM/Cache/triton:/root/.triton/cache \
             -v /mnt/ssd/vLLM/Patches/genesis/vllm/_genesis:/usr/local/lib/python3.12/dist-packages/vllm/_genesis:ro \
             -v /mnt/ssd/vLLM/Patches/patch_tolist_cudagraph.py:/patches/patch_tolist_cudagraph.py:ro \
-            -v /mnt/ssd/vLLM/Patches/patch_pn12_ffn_pool_anchor.py:/patches/patch_pn12_ffn_pool_anchor.py:ro \
-            -v /mnt/ssd/vLLM/Patches/patch_pn12_compile_safe_custom_op.py:/patches/patch_pn12_compile_safe_custom_op.py:ro \
-            -v /mnt/ssd/vLLM/Patches/patch_fa_max_seqlen_clamp.py:/patches/patch_fa_max_seqlen_clamp.py:ro \
+            -v /mnt/ssd/vLLM/Patches/patch_workspace_lock_disable.py:/patches/patch_workspace_lock_disable.py:ro \
             -v /mnt/ssd/vLLM/Patches/patch_timings_07351e088.py:/patches/patch_timings_07351e088.py:ro \
             -p ''${PORT}:8000 \
             --entrypoint /bin/bash \
-            vllm/vllm-openai:nightly-07351e0883470724dd5a7e9730ed10e01fc99d08 \
+            vllm/vllm-openai:nightly-7a1eb8ac2ec4ea69338c51dc7afd4b15010abfa8 \
             -c "${vllmCmdFlat}"
         '';
       cmdStop = "${pkgs.docker}/bin/docker stop \${MODEL_ID}";
@@ -279,9 +366,9 @@ in
     };
 
     # https://github.com/noonghunna/club-3090/tree/master/models/qwen3.6-27b/vllm
+    # Synced from: club-3090 ae4846f (2026-05-02) — docker-compose.tools-text.yml
     # Tools-text variant - 75K context, text-only (no vision)
-    # fp8_e5m2 KV + MTP n=3. This is the repo's validated long-context
-    # tool-calling profile and should be more stable than TurboQuant 128K.
+    # fp8_e5m2 KV + MTP n=3. IDE agents (Cline, Cursor, OpenCode, etc.)
     "vllm-qwen3.6-27b-tools-text" = {
       name = "vLLM Qwen3.6 (27B) - Tools Text";
       macros.ctx = "75000";
@@ -324,7 +411,7 @@ in
             -e VLLM_WORKER_MULTIPROC_METHOD=spawn \
             -e NCCL_CUMEM_ENABLE=0 \
             -e NCCL_P2P_DISABLE=1 \
-            -e VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=1 \
+            -e VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=0 \
             -e VLLM_NO_USAGE_STATS=1 \
             -e PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True,max_split_size_mb:512 \
             -e VLLM_FLOAT32_MATMUL_PRECISION=high \
@@ -335,15 +422,31 @@ in
             -e CUDA_DEVICE_ORDER=PCI_BUS_ID \
             -e VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 \
             -e VLLM_MARLIN_USE_ATOMIC_ADD=1 \
+            -e TRITON_CACHE_DIR=/root/.triton/cache \
+            -e GENESIS_ENABLE_P58_ASYNC_PLACEHOLDER_FIX=1 \
             -e GENESIS_ENABLE_P64_QWEN3CODER_MTP_STREAMING=1 \
+            -e GENESIS_ENABLE_P66_CUDAGRAPH_SIZE_FILTER=1 \
+            -e GENESIS_ENABLE_P68_AUTO_FORCE_TOOL=1 \
+            -e GENESIS_ENABLE_P69_LONG_CTX_TOOL_REMINDER=1 \
+            -e GENESIS_P68_P69_LONG_CTX_THRESHOLD_CHARS=50000 \
+            -e GENESIS_ENABLE_P72_PROFILE_RUN_CAP=1 \
+            -e GENESIS_PROFILE_RUN_CAP_M=4128 \
+            -e GENESIS_ENABLE_P74_CHUNK_CLAMP=1 \
+            -e GENESIS_ENABLE_P94=1 \
             -e GENESIS_ENABLE_PN8_MTP_DRAFT_ONLINE_QUANT=1 \
+            -e GENESIS_ENABLE_PN13_CUDA_GRAPH_LAMBDA_ARITY=1 \
+            -e GENESIS_ENABLE_PN14_TQ_DECODE_OOB_CLAMP=1 \
+            -e GENESIS_ENABLE_PN17_FA2_LSE_CLAMP=1 \
+            -e GENESIS_ENABLE_PN19_SCOPED_MAX_SPLIT=1 \
             -v /mnt/ssd/vLLM/Models:/root/.cache/huggingface \
+            -v /mnt/ssd/vLLM/Cache/torch_compile:/root/.cache/vllm/torch_compile_cache \
+            -v /mnt/ssd/vLLM/Cache/triton:/root/.triton/cache \
             -v /mnt/ssd/vLLM/Patches/genesis/vllm/_genesis:/usr/local/lib/python3.12/dist-packages/vllm/_genesis:ro \
             -v /mnt/ssd/vLLM/Patches/patch_tolist_cudagraph.py:/patches/patch_tolist_cudagraph.py:ro \
             -v /mnt/ssd/vLLM/Patches/patch_timings_07351e088.py:/patches/patch_timings_07351e088.py:ro \
             -p ''${PORT}:8000 \
             --entrypoint /bin/bash \
-            vllm/vllm-openai:nightly-07351e0883470724dd5a7e9730ed10e01fc99d08 \
+            vllm/vllm-openai:nightly-7a1eb8ac2ec4ea69338c51dc7afd4b15010abfa8 \
             -c "${vllmCmdFlat}"
         '';
       cmdStop = "${pkgs.docker}/bin/docker stop \${MODEL_ID}";

modules/nixos/services/llama-swap/default.nix

@@ -88,6 +88,11 @@ in
 
     # Create Config
     sops = {
+      secrets = {
+        "llama_swap_api_keys/evan" = {
+          sopsFile = lib.snowfall.fs.get-file "secrets/common/llama-swap.yaml";
+        };
+      };
       secrets = {
         "llama_swap_api_keys/pi" = {
           sopsFile = lib.snowfall.fs.get-file "secrets/common/llama-swap.yaml";
@@ -101,6 +106,7 @@ in
           recursiveUpdate cfg.config {
             apiKeys = [
               config.sops.placeholder."llama_swap_api_keys/pi"
+              config.sops.placeholder."llama_swap_api_keys/evan"
             ];
           }
         );

modules/nixos/services/llama-swap/setup-qwen36-vllm.sh

@@ -1,33 +1,38 @@
 #!/usr/bin/env bash
 # Setup script for vLLM Qwen3.6-27B on a single 3090.
 #
-# Downloads the model, clones Genesis patches, and fetches all sidecar
+# Downloads the model, clones Genesis patches (pinned), applies setup-time
+# source patches to the Genesis tree, and fetches all boot-time sidecar
 # patches into place under /mnt/ssd/vLLM/.
 #
 # Idempotent — safe to re-run; skips steps already completed.
 #
-# Prerequisites: git (with git-lfs), docker
+# Prerequisites: git (with git-lfs), docker, python3
 
 set -euo pipefail
 
 MODEL_DIR="/mnt/ssd/vLLM/Models"
 MODEL_SUBDIR="qwen3.6-27b-autoround-int4"
 PATCHES_DIR="/mnt/ssd/vLLM/Patches"
+CACHE_DIR="/mnt/ssd/vLLM/Cache"
 GENESIS_DIR="${PATCHES_DIR}/genesis"
+
+# Pin Genesis to the validated commit (bump requires re-testing all composes)
+GENESIS_PIN="${GENESIS_PIN:-fc89395}"
+
 TOLIST_PATCH="${PATCHES_DIR}/patch_tolist_cudagraph.py"
-PN12_FFN_PATCH="${PATCHES_DIR}/patch_pn12_ffn_pool_anchor.py"
-PN12_COMPILE_PATCH="${PATCHES_DIR}/patch_pn12_compile_safe_custom_op.py"
-FA_CLAMP_PATCH="${PATCHES_DIR}/patch_fa_max_seqlen_clamp.py"
 WORKSPACE_LOCK_PATCH="${PATCHES_DIR}/patch_workspace_lock_disable.py"
+PN25_REGISTER_PATCH="${PATCHES_DIR}/patch_pn25_genesis_register_fix.py"
+PN30_DST_PATCH="${PATCHES_DIR}/patch_pn30_dst_shaped_temp_fix.py"
+PR40798_PATCH="${PATCHES_DIR}/patch_pr40798_workspace.py"
 TIMINGS_PATCH="${PATCHES_DIR}/patch_timings_07351e088.py"
 TIMINGS_PATCH_URL="${TIMINGS_PATCH_URL:-https://gitea.va.reichard.io/evan/nix/raw/branch/master/modules/nixos/services/llama-swap/patches/patch_timings_07351e088.py}"
 
-# Base URLs for sidecar patches (club-3090 repo)
+# Base URL for sidecar patches (club-3090 repo)
 PATCH_BASE_URL="https://raw.githubusercontent.com/noonghunna/club-3090/master/models/qwen3.6-27b/vllm/patches"
-PATCH_EXPERIMENTAL_BASE_URL="https://raw.githubusercontent.com/noonghunna/club-3090/v0.20-experimental/models/qwen3.6-27b/vllm/patches"
 
 # ---------- Preflight Checks ----------
-for cmd in git git-lfs curl; do
+for cmd in git git-lfs curl python3; do
   if ! command -v "$cmd" &>/dev/null; then
     echo "ERROR: '$cmd' not found in PATH." >&2
     exit 1
@@ -36,7 +41,7 @@ done
 
 # ---------- Create Directories ----------
 echo "Creating directories..."
-mkdir -p "${MODEL_DIR}" "${PATCHES_DIR}"
+mkdir -p "${MODEL_DIR}" "${PATCHES_DIR}" "${CACHE_DIR}/torch_compile" "${CACHE_DIR}/triton"
 
 # ---------- Download Model ----------
 if [ -d "${MODEL_DIR}/${MODEL_SUBDIR}/.git" ]; then
@@ -48,16 +53,24 @@ else
   echo "Model cloned."
 fi
 
-# ---------- Clone Genesis Patches ----------
+# ---------- Clone / Pin Genesis Patches ----------
 if [ -d "${GENESIS_DIR}/.git" ]; then
-  echo "Genesis patches already cloned at ${GENESIS_DIR}, pulling latest..."
-  git -C "${GENESIS_DIR}" pull --ff-only || echo "Pull failed (non-fatal), using existing."
+  echo "Genesis already cloned — fetching + checking out ${GENESIS_PIN} ..."
+  (cd "${GENESIS_DIR}" && git fetch origin && git checkout "${GENESIS_PIN}" 2>&1 | tail -3)
 else
-  echo "Cloning Genesis patches..."
+  echo "Cloning Genesis patches at ${GENESIS_PIN} ..."
   git clone https://github.com/Sandermage/genesis-vllm-patches "${GENESIS_DIR}"
-  echo "Genesis patches cloned."
+  (cd "${GENESIS_DIR}" && git checkout "${GENESIS_PIN}")
 fi
 
+# Sanity Check — v7.14+ layout
+if [[ ! -d "${GENESIS_DIR}/vllm/_genesis" ]]; then
+  echo "ERROR: genesis tree at ${GENESIS_PIN} missing vllm/_genesis package." >&2
+  echo "       Re-run with GENESIS_PIN=<other-ref> to try a different version." >&2
+  exit 1
+fi
+echo "Genesis pinned to ${GENESIS_PIN} ($(cd "${GENESIS_DIR}" && git rev-parse --short HEAD))"
+
 # ---------- Download Sidecar Patches ----------
 # Fetched from club-3090 repo so this script is self-contained.
 download_patch() {
@@ -74,21 +87,47 @@ download_patch() {
 }
 
 download_patch "${TOLIST_PATCH}"
-download_patch "${PN12_FFN_PATCH}"
-download_patch "${PN12_COMPILE_PATCH}"
-download_patch "${FA_CLAMP_PATCH}"
+download_patch "${WORKSPACE_LOCK_PATCH}"
+download_patch "${PN25_REGISTER_PATCH}"
+download_patch "${PN30_DST_PATCH}"
+download_patch "${PR40798_PATCH}"
 
-# ---------- Download v0.20 Workspace Patch ----------
-if [ -f "${WORKSPACE_LOCK_PATCH}" ]; then
-  echo "Patch $(basename "${WORKSPACE_LOCK_PATCH}") already present, skipping."
-else
-  echo "Downloading $(basename "${WORKSPACE_LOCK_PATCH}") from v0.20-experimental..."
-  curl -fsSL \
-    "${PATCH_EXPERIMENTAL_BASE_URL}/$(basename "${WORKSPACE_LOCK_PATCH}")" \
-    -o "${WORKSPACE_LOCK_PATCH}"
-  echo "Patch $(basename "${WORKSPACE_LOCK_PATCH}") written."
+# ---------- Apply Setup-Time Genesis Source Patches ----------
+# These modify the Genesis checkout in-place. The Genesis tree is mounted
+# read-only into the container, so these MUST run at setup time, not boot.
+
+# The setup-time patches use hardcoded relative paths like
+# "models/qwen3.6-27b/vllm/patches/genesis/vllm/_genesis/..."
+# rooted at the club-3090 repo root. Our Genesis clone lives at
+# ${PATCHES_DIR}/genesis, so we create a temporary symlink tree
+# so the relative paths resolve correctly.
+PATCH_WORKDIR="$(mktemp -d)"
+mkdir -p "${PATCH_WORKDIR}/models/qwen3.6-27b/vllm/patches"
+ln -sfn "${GENESIS_DIR}" "${PATCH_WORKDIR}/models/qwen3.6-27b/vllm/patches/genesis"
+
+# PN25 Worker-Spawn Registration Fix
+# Registers the PN25 opaque op at module import time instead of inside
+# the compiled forward path. Required for TP=1 spawned workers.
+if [ -f "${PN25_REGISTER_PATCH}" ]; then
+  echo "Applying PN25 genesis register fix to Genesis tree..."
+  (cd "${PATCH_WORKDIR}" && python3 "${PN25_REGISTER_PATCH}") || {
+    echo "WARN: PN25 register fix did not apply cleanly. PN25 may not work in workers." >&2
+  }
 fi
 
+# PN30 DS Conv-State Dst-Shaped Temp Fix
+# Corrects DS layout corruption in PN30's speculative decode path by
+# building a destination-shaped temp instead of compacting source tail.
+if [ -f "${PN30_DST_PATCH}" ]; then
+  echo "Applying PN30 dst-shaped temp fix to Genesis tree..."
+  (cd "${PATCH_WORKDIR}" && python3 "${PN30_DST_PATCH}") || {
+    echo "WARN: PN30 dst-shaped temp fix did not apply cleanly. DS layout may not work." >&2
+  }
+fi
+
+# Clean Up Symlink Workdir
+rm -rf "${PATCH_WORKDIR}"
+
 # ---------- Download Timing Patch ----------
 tmp_timings_patch="$(mktemp)"
 trap 'rm -f "${tmp_timings_patch}"' EXIT
@@ -108,21 +147,22 @@ fi
 echo ""
 echo "=== Setup Complete ==="
 echo "  Model:   ${MODEL_DIR}/${MODEL_SUBDIR}"
-echo "  Genesis: ${GENESIS_DIR}"
-echo "  Patch:   ${TOLIST_PATCH}"
-echo "  Workspace: ${WORKSPACE_LOCK_PATCH}"
-echo "  Timings:   ${TIMINGS_PATCH}"
+echo "  Genesis: ${GENESIS_DIR} (pinned: ${GENESIS_PIN})"
+echo "  Cache:   ${CACHE_DIR}/{torch_compile,triton}"
 echo ""
 echo "Expected layout:"
 echo "  /mnt/ssd/vLLM/"
 echo "  ├── Models/"
 echo "  │   └── qwen3.6-27b-autoround-int4/          (model weights)"
+echo "  ├── Cache/"
+echo "  │   ├── torch_compile/                        (torch.compile cache)"
+echo "  │   └── triton/                               (Triton kernel cache)"
 echo "  └── Patches/"
-echo "      ├── genesis/                               (Genesis v7.14+ repo)"
-echo "      │   └── vllm/_genesis/                     (mounted into container)"
-echo "      ├── patch_tolist_cudagraph.py              (cudagraph capture fix)"
-echo "      ├── patch_pn12_ffn_pool_anchor.py          (PN12 FFN pool anchor fix)"
-echo "      ├── patch_pn12_compile_safe_custom_op.py   (PN12 compile-safe custom op)"
-echo "      ├── patch_fa_max_seqlen_clamp.py           (FA softmax_lse clamp — P104)"
-echo "      ├── patch_workspace_lock_disable.py        (v0.20 WorkspaceManager lock workaround)"
-echo "      └── patch_timings_07351e088.py             (llama.cpp-compatible timings)"
+echo "      ├── genesis/                               (Genesis v7.65 @ ${GENESIS_PIN})"
+echo "      │   └── vllm/_genesis/                     (mounted into container, PN25+PN30 applied)"
+echo "      ├── patch_tolist_cudagraph.py              (boot-time: cudagraph capture fix)"
+echo "      ├── patch_workspace_lock_disable.py        (boot-time: v0.20 WorkspaceManager lock workaround)"
+echo "      ├── patch_pn25_genesis_register_fix.py     (setup-time: applied to Genesis tree)"
+echo "      ├── patch_pn30_dst_shaped_temp_fix.py      (setup-time: applied to Genesis tree)"
+echo "      ├── patch_pr40798_workspace.py             (PR40798 workspace fix)"
+echo "      └── patch_timings_07351e088.py             (boot-time: llama.cpp-compatible timings)"