feat: vllm yay

2026-05-01 10:38:43 -04:00
parent 75eba8703f
commit 74ff71803b
5 changed files with 220 additions and 57 deletions
--- a/modules/nixos/services/llama-swap/config.nix
+++ b/modules/nixos/services/llama-swap/config.nix
@@ -358,60 +358,21 @@ in
      };
    };

-    "vllm-qwen3.5-27b-thinking" = {
-      name = "vLLM Qwen3.5 (27B) - Thinking";
-      macros.ctx = "196608";
-      proxy = "http://127.0.0.1:\${PORT}";
-      cmd = ''
-        ${pkgs.docker}/bin/docker run --rm --device=nvidia.com/gpu=all \
-          --name ''${MODEL_ID} \
-          -e PYTORCH_ALLOC_CONF=expandable_segments:True \
-          -v /mnt/ssd/vLLM:/root/.cache/huggingface \
-          -p ''${PORT}:8000 \
-          --ipc=host vllm/vllm-openai:latest \
-          --served-model-name ''${MODEL_ID} \
-          --model cyankiwi/Qwen3.5-27B-AWQ-4bit \
-          --max-model-len 24576 \
-          --kv-cache-dtype auto \
-          --max-num-seqs 4 \
-          --max-num-batched-tokens 4096 \
-          --enable-chunked-prefill \
-          --gpu-memory-utilization 0.95 \
-          --language-model-only \
-          --speculative-config '{"method":"mtp","num_speculative_tokens":3}' \
-          --enable-prefix-caching \
-          --enforce-eager \
-          --block-size 32 \
-          --swap-space 4 \
-          --tensor-parallel-size 1 \
-          --reasoning-parser qwen3 \
-          --enable-auto-tool-choice \
-          --default-chat-template-kwargs '{"enable_thinking": true}' \
-          --tool-call-parser qwen3_coder
-      '';
-      cmdStop = "docker stop \${MODEL_ID}";
-
-      metadata = {
-        type = [
-          "text-generation"
-          "coding"
-        ];
-      };
-    };
-
    # https://github.com/Lorbus/qwen36-27b-single-3090
-    # Model: Lorbus/Qwen3.6-27B-int4-AutoRound (auto_round int4)
-    # Genesis v7.14+ patches for MTP streaming + tool adherence
-    # Text-only (no vision) to maximize KV budget for ~100k context
-    "vllm-qwen3.6-27b-thinking" = {
-      name = "vLLM Qwen3.6 (27B) - Thinking";
-      macros.ctx = "75000";
+    # Long-text variant - 185K context, text-only (no vision)
+    # TurboQuant 3-bit KV + MTP n=3 + PN12/P104 cliff-closure stack
+    "vllm-qwen3.6-27b-long-text" = {
+      name = "vLLM Qwen3.6 (27B) - Long Text";
+      macros.ctx = "185000";
      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_tolist_cudagraph.py;
            exec vllm serve
            --served-model-name ''${MODEL_ID}
@@ -420,10 +381,10 @@ in
            --dtype float16
            --tensor-parallel-size 1
            --max-model-len ''${ctx}
-            --gpu-memory-utilization 0.97
+            --gpu-memory-utilization 0.975
            --max-num-seqs 1
-            --max-num-batched-tokens 2048
-            --kv-cache-dtype fp8_e5m2
+            --max-num-batched-tokens 4128
+            --kv-cache-dtype turboquant_3bit_nc
            --language-model-only
            --trust-remote-code
            --reasoning-parser qwen3
@@ -431,6 +392,7 @@ in
            --tool-call-parser qwen3_coder
            --enable-prefix-caching
            --enable-chunked-prefill
+            --no-scheduler-reserve-full-isl
            --speculative-config '{\"method\":\"mtp\",\"num_speculative_tokens\":3}'
            --host 0.0.0.0
            --port 8000
@@ -455,12 +417,110 @@ in
            -e CUDA_DEVICE_ORDER=PCI_BUS_ID \
            -e VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 \
            -e VLLM_MARLIN_USE_ATOMIC_ADD=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_P68_AUTO_FORCE_TOOL=1 \
-            -e GENESIS_ENABLE_P69_LONG_CTX_TOOL_REMINDER=1 \
+            -e GENESIS_ENABLE_P101=1 \
+            -e GENESIS_ENABLE_P103=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_PN17_FA2_LSE_CLAMP=1 \
+            -e GENESIS_ENABLE_P37=1 \
            -v /mnt/ssd/vLLM/Models:/root/.cache/huggingface \
            -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 \
+            -p ''${PORT}:8000 \
+            --entrypoint /bin/bash \
+            vllm/vllm-openai:nightly-07351e0883470724dd5a7e9730ed10e01fc99d08 \
+            -c "${vllmCmdFlat}"
+        '';
+      cmdStop = "docker stop \${MODEL_ID}";
+
+      metadata = {
+        type = [
+          "text-generation"
+          "coding"
+        ];
+      };
+    };
+
+    # https://github.com/Lorbus/qwen36-27b-single-3090
+    # Long-vision variant - 140K context with vision tower active
+    # TurboQuant 3-bit KV + MTP n=3 + PN12/P104 cliff-closure stack
+    "vllm-qwen3.6-27b-long-vision" = {
+      name = "vLLM Qwen3.6 (27B) - Long Vision";
+      macros.ctx = "140000";
+      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_tolist_cudagraph.py;
+            exec vllm serve
+            --served-model-name ''${MODEL_ID}
+            --model /root/.cache/huggingface/qwen3.6-27b-autoround-int4
+            --quantization auto_round
+            --dtype float16
+            --tensor-parallel-size 1
+            --max-model-len ''${ctx}
+            --gpu-memory-utilization 0.95
+            --max-num-seqs 1
+            --max-num-batched-tokens 4128
+            --kv-cache-dtype turboquant_3bit_nc
+            --trust-remote-code
+            --reasoning-parser qwen3
+            --enable-auto-tool-choice
+            --tool-call-parser qwen3_coder
+            --enable-prefix-caching
+            --enable-chunked-prefill
+            --no-scheduler-reserve-full-isl
+            --speculative-config '{\"method\":\"mtp\",\"num_speculative_tokens\":3}'
+            --host 0.0.0.0
+            --port 8000
+          '';
+          vllmCmdFlat = builtins.replaceStrings [ "\n" ] [ " " ] vllmCmd;
+        in
+        ''
+          ${pkgs.docker}/bin/docker run --rm --device=nvidia.com/gpu=all \
+            --name ''${MODEL_ID} \
+            --ipc=host \
+            -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_NO_USAGE_STATS=1 \
+            -e PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True,max_split_size_mb:512 \
+            -e VLLM_FLOAT32_MATMUL_PRECISION=high \
+            -e VLLM_USE_FLASHINFER_SAMPLER=1 \
+            -e OMP_NUM_THREADS=1 \
+            -e CUDA_DEVICE_MAX_CONNECTIONS=8 \
+            -e CUDA_VISIBLE_DEVICES=0 \
+            -e CUDA_DEVICE_ORDER=PCI_BUS_ID \
+            -e VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 \
+            -e VLLM_MARLIN_USE_ATOMIC_ADD=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_P101=1 \
+            -e GENESIS_ENABLE_P103=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_PN17_FA2_LSE_CLAMP=1 \
+            -v /mnt/ssd/vLLM/Models:/root/.cache/huggingface \
+            -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 \
            -p ''${PORT}:8000 \
            --entrypoint /bin/bash \
            vllm/vllm-openai:nightly-07351e0883470724dd5a7e9730ed10e01fc99d08 \
--- a/modules/nixos/services/llama-swap/setup-qwen36-vllm.sh
+++ b/modules/nixos/services/llama-swap/setup-qwen36-vllm.sh
@@ -0,0 +1,103 @@
+#!/usr/bin/env bash
+# Setup script for vLLM Qwen3.6-27B on a single 3090.
+#
+# Downloads the model, clones Genesis patches, and fetches all sidecar
+# patches into place under /mnt/ssd/vLLM/.
+#
+# Idempotent — safe to re-run; skips steps already completed.
+#
+# Prerequisites: git (with git-lfs), docker
+
+set -euo pipefail
+
+MODEL_DIR="/mnt/ssd/vLLM/Models"
+MODEL_SUBDIR="qwen3.6-27b-autoround-int4"
+PATCHES_DIR="/mnt/ssd/vLLM/Patches"
+GENESIS_DIR="${PATCHES_DIR}/genesis"
+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"
+
+# Base URL for sidecar patches (club-3090 repo, master branch)
+PATCH_BASE_URL="https://raw.githubusercontent.com/noonghunna/club-3090/master/models/qwen3.6-27b/vllm/patches"
+
+# ---------- Preflight Checks ----------
+for cmd in git git-lfs; do
+  if ! command -v "$cmd" &>/dev/null; then
+    echo "ERROR: '$cmd' not found in PATH." >&2
+    exit 1
+  fi
+done
+
+# ---------- Create Directories ----------
+echo "Creating directories..."
+mkdir -p "${MODEL_DIR}" "${PATCHES_DIR}"
+
+# ---------- Download Model ----------
+if [ -d "${MODEL_DIR}/${MODEL_SUBDIR}/.git" ]; then
+  echo "Model already cloned at ${MODEL_DIR}/${MODEL_SUBDIR}, skipping."
+else
+  echo "Cloning Lorbus/Qwen3.6-27B-int4-AutoRound (with LFS)..."
+  git clone https://huggingface.co/Lorbus/Qwen3.6-27B-int4-AutoRound \
+    "${MODEL_DIR}/${MODEL_SUBDIR}"
+  echo "Model cloned."
+fi
+
+# ---------- Clone 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."
+else
+  echo "Cloning Genesis patches..."
+  git clone https://github.com/Sandermage/genesis-vllm-patches "${GENESIS_DIR}"
+  echo "Genesis patches cloned."
+fi
+
+# ---------- Download Sidecar Patches ----------
+# Fetched from club-3090 repo so this script is self-contained.
+download_patch() {
+  local dest="$1"
+  local filename
+  filename="$(basename "$dest")"
+  if [ -f "${dest}" ]; then
+    echo "Patch ${filename} already present, skipping."
+  else
+    echo "Downloading ${filename}..."
+    python3 -c "
+import urllib.request, sys
+url = '${PATCH_BASE_URL}/' + sys.argv[2]
+try:
+    urllib.request.urlretrieve(url, sys.argv[1])
+    print('Downloaded from GitHub.')
+except Exception as e:
+    print(f'Download failed: {e}', file=sys.stderr)
+    sys.exit(1)
+" "${dest}" "${filename}"
+    echo "Patch ${filename} written."
+  fi
+}
+
+download_patch "${TOLIST_PATCH}"
+download_patch "${PN12_FFN_PATCH}"
+download_patch "${PN12_COMPILE_PATCH}"
+download_patch "${FA_CLAMP_PATCH}"
+
+# ---------- Summary ----------
+echo ""
+echo "=== Setup Complete ==="
+echo "  Model:   ${MODEL_DIR}/${MODEL_SUBDIR}"
+echo "  Genesis: ${GENESIS_DIR}"
+echo "  Patch:   ${TOLIST_PATCH}"
+echo ""
+echo "Expected layout:"
+echo "  /mnt/ssd/vLLM/"
+echo "  ├── Models/"
+echo "  │   └── qwen3.6-27b-autoround-int4/          (model weights)"
+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)"