Ship thinking-off runtime default

SHA256SUMS

@@ -1,4 +1,4 @@
-a0df2b53e4cd23e4ccb7b20ed44773f0ee6b721cdcb1f0f07282151c4a350195  README.md
+3b9154b6872131ef0c75778b06fb45217a4a47ae19f75feb922af6c9b62d1ced  README.md
 87e8ab7ed85be38e270830b157061a99af89fb85f51eb12c411e6dcecf5d88e8  RELEASE_MANIFEST.json
 7c3d8064baf55779579b48c58fa083e0bda2a639c857b4bde56bab4008d28002  docs/btl-3-compact-gguf-exporter.md
 d1f83f8a4652b7d07417067e964af864aaebf559bb9b0c76a54a0c357c3ffe86  docs/launch-btl3-compact.md
@@ -25,8 +25,8 @@ ef85e198bda0c18a42650ea1a576bffdc3f8749d2f8dcec739b9be34efeebb2d  licenses/LICEN
 87cfc2262dbf499daff47db1339293198fba61431f623fd7226004d6714d58c8  licenses/THIRD_PARTY_NOTICES.md
 2ddf9527620a17a2a6739d184a7096c45712092e6589128792ec6254e94dc30c  model/BTL-3-Compact-AVQ2.gguf
 94f29bbed6a22c35b992c5c6ebf0e7c92f13b836b90f36f461c9cf2f0f1d010d  runtimes/preview/BTL-3-Compact-linux-arm64-cuda/LICENSE.llama.cpp
-043658e4900b04a80fb9d4247c4fe8b04f5adc3706325996dc176b202f9baa24  runtimes/preview/BTL-3-Compact-linux-arm64-cuda/bin/btl3-server
-328388c97746e3781f0fe8520ec73aa4bb8ad55978c2a62385964fff33e711d0  runtimes/preview/BTL-3-Compact-linux-arm64-cuda/bundle-manifest.json
+b6e370b019f80cb605ac78846156233a816aa61d423ff5a7e923fcd7e8670abc  runtimes/preview/BTL-3-Compact-linux-arm64-cuda/bin/btl3-server
+0982584bae74781ee188a0a0a626f40ef333d4492c69f0699f6eb4468ae7aff2  runtimes/preview/BTL-3-Compact-linux-arm64-cuda/bundle-manifest.json
 7761d260387b95fa13b7ec101453abca45ef972d3e6ccc8bcf54251e1ac0470b  runtimes/preview/BTL-3-Compact-linux-arm64-cuda/lib/libcublas.so.13
 3839e12c3b69ed8eb732e5d5f8e749403a582a9d706f779017c785152d9e62f9  runtimes/preview/BTL-3-Compact-linux-arm64-cuda/lib/libcublasLt.so.13
 7bdba2b5b08cbdc85203c41cc94598adedb1bcfea7cb574ca693ac73599e4e63  runtimes/preview/BTL-3-Compact-linux-arm64-cuda/lib/libcudart.so.13
@@ -49,8 +49,8 @@ b67f9d10eaa85a500180ddecf6ecdf0e9c100b0bdf4c10e32b0bf657ca50e494  runtimes/previ
 7d5450cb2d142651b8afa315b5f238efc805dad827d91ba367d8516bc9d49e7a  runtimes/supported/BTL-3-Compact-macos-arm64/LICENSE.OpenSSL
 94f29bbed6a22c35b992c5c6ebf0e7c92f13b836b90f36f461c9cf2f0f1d010d  runtimes/supported/BTL-3-Compact-macos-arm64/LICENSE.llama.cpp
 de4350f194d6f393bd3ea1065b024d7f92fea0a8f2e435f6557dc0c6696b3a74  runtimes/supported/BTL-3-Compact-macos-arm64/bin/btl3-ollama-bridge
-f14fe48ab3d5f0662f7ffe524cd6db8bc4d9bb2d892e54147c1e290cd2532d75  runtimes/supported/BTL-3-Compact-macos-arm64/bin/btl3-server
-3ac935decfa3c67faaa210dee3b07417611cc1abcf13c5fcf5850fac3cbd399e  runtimes/supported/BTL-3-Compact-macos-arm64/bundle-manifest.json
+e4dc8efdbeadfc74c27516115b808b3589da1b8b27d0f330498dacd9b1f3d8a9  runtimes/supported/BTL-3-Compact-macos-arm64/bin/btl3-server
+e192aee70a8473f595bb106d0f7b5604fac6c389c08ed00d55f51673fcca3b27  runtimes/supported/BTL-3-Compact-macos-arm64/bundle-manifest.json
 1f6ea0bde7b3d4ea462b047e12905c8b554b111d2808deb33fc3add67b4d0895  runtimes/supported/BTL-3-Compact-macos-arm64/lib/libcrypto.3.dylib
 a5edff95c1a5f9c6fd3072ea65cb8b75e44fb92bbb1b35901ec43fe62cbc6eff  runtimes/supported/BTL-3-Compact-macos-arm64/lib/libggml-base.0.13.1.dylib
 14c8604076154609c9767a936953f7c2e419099a509e129629467329b5d21aad  runtimes/supported/BTL-3-Compact-macos-arm64/lib/libggml-blas.0.13.1.dylib

runtimes/preview/BTL-3-Compact-linux-arm64-cuda/bin/btl3-server

@@ -0,0 +1,104 @@
+#!/bin/sh
+set -eu
+
+bundle_root=$(CDPATH= cd -- "$(dirname -- "$0")/.." && pwd)
+server="$bundle_root/libexec/llama-server"
+model_name="BTL-3-Compact-AVQ2.gguf"
+
+find_model() {
+    if [ -n "${BTL3_MODEL:-}" ]; then
+        printf '%s\n' "$BTL3_MODEL"
+        return
+    fi
+    for candidate in \
+        "$bundle_root/model/$model_name" \
+        "$bundle_root/../../artifacts/release/$model_name" \
+        "$PWD/$model_name"
+    do
+        if [ -f "$candidate" ]; then
+            printf '%s\n' "$candidate"
+            return
+        fi
+    done
+    return 1
+}
+
+detect_gpu_memory() {
+    if [ -n "${BTL3_GPU_MEMORY_MIB:-}" ]; then
+        printf '%s\n' "$BTL3_GPU_MEMORY_MIB"
+        return
+    fi
+    if command -v nvidia-smi >/dev/null 2>&1; then
+        value=$(nvidia-smi --query-gpu=memory.total \
+            --format=csv,noheader,nounits 2>/dev/null | head -1 | tr -dc '0-9')
+        if [ -n "$value" ]; then
+            printf '%s\n' "$value"
+            return
+        fi
+    fi
+    case "$(uname -m)" in
+        aarch64|arm64)
+            awk '/MemTotal:/ { printf "%d\n", $2 / 1024 }' /proc/meminfo
+            return
+            ;;
+    esac
+    return 1
+}
+
+choose_context() {
+    memory=$1
+    if [ "$memory" -ge 96000 ]; then printf '%s\n' 131072
+    elif [ "$memory" -ge 48000 ]; then printf '%s\n' 98304
+    elif [ "$memory" -ge 28000 ]; then printf '%s\n' 65536
+    elif [ "$memory" -ge 20000 ]; then printf '%s\n' 32768
+    else printf '%s\n' 16384
+    fi
+}
+
+model=$(find_model || true)
+if [ -z "$model" ] || [ ! -f "$model" ]; then
+    echo "BTL-3 model not found. Set BTL3_MODEL to $model_name." >&2
+    exit 2
+fi
+if [ ! -x "$server" ]; then
+    echo "Packaged llama-server is missing or not executable: $server" >&2
+    exit 2
+fi
+memory_mib=$(detect_gpu_memory || true)
+if [ -z "$memory_mib" ]; then
+    echo "NVIDIA GPU memory could not be detected; set BTL3_GPU_MEMORY_MIB." >&2
+    exit 2
+fi
+
+host=${BTL3_HOST:-127.0.0.1}
+port=${BTL3_PORT:-8080}
+ctx=${BTL3_CTX_SIZE:-$(choose_context "$memory_mib")}
+parallel=${BTL3_PARALLEL:-1}
+aliases=${BTL3_MODEL_ALIASES:-BTL-3}
+gpu_layers=${BTL3_GPU_LAYERS:-99}
+case "${BTL3_ENABLE_THINKING:-false}" in
+    1|true|TRUE) chat_template_kwargs='{"enable_thinking":true}' ;;
+    *) chat_template_kwargs='{"enable_thinking":false}' ;;
+esac
+export LD_LIBRARY_PATH="$bundle_root/lib${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"
+export GGML_BACKEND_PATH="$bundle_root/lib/libggml-cuda.so"
+case "$(uname -m)" in
+    aarch64|arm64) export GGML_CUDA_ENABLE_UNIFIED_MEMORY=1 ;;
+esac
+
+set -- \
+    --model "$model" --alias "$aliases" --host "$host" --port "$port" \
+    --ctx-size "$ctx" --parallel "$parallel" --n-gpu-layers "$gpu_layers" \
+    --jinja --reasoning auto --reasoning-format deepseek --cont-batching \
+    --chat-template-kwargs "$chat_template_kwargs" \
+    --cache-ram 0 --no-warmup --no-ui "$@"
+if [ -n "${BTL3_API_KEY:-}" ]; then
+    set -- "$@" --api-key "$BTL3_API_KEY"
+fi
+if [ "${BTL3_PRINT_COMMAND:-0}" = "1" ]; then
+    printf 'executable=%s\nmodel=%s\nhost=%s\nport=%s\nctx_size=%s\ngpu_memory_mib=%s\nthinking_enabled=%s\n' \
+        "$server" "$model" "$host" "$port" "$ctx" "$memory_mib" \
+        "$( [ "$chat_template_kwargs" = '{"enable_thinking":true}' ] && printf true || printf false )"
+    exit 0
+fi
+exec "$server" "$@"

runtimes/preview/BTL-3-Compact-linux-arm64-cuda/bundle-manifest.json

@@ -0,0 +1,101 @@
+{
+  "schema_version": 1,
+  "bundle": "BTL-3 Compact Linux arm64 CUDA (DGX Spark)",
+  "target": "linux-arm64",
+  "cuda": {
+    "toolkit": "13.0.2",
+    "architectures": "121-real"
+  },
+  "status": "cross-compiled; NVIDIA runtime conformance pending",
+  "external_model": {
+    "filename": "BTL-3-Compact-AVQ2.gguf",
+    "bytes": 8392369600,
+    "sha256": "2ddf9527620a17a2a6739d184a7096c45712092e6589128792ec6254e94dc30c"
+  },
+  "files": {
+    "LICENSE.llama.cpp": {
+      "bytes": 1078,
+      "sha256": "94f29bbed6a22c35b992c5c6ebf0e7c92f13b836b90f36f461c9cf2f0f1d010d"
+    },
+    "bin/btl3-server": {
+      "bytes": 3297,
+      "sha256": "b6e370b019f80cb605ac78846156233a816aa61d423ff5a7e923fcd7e8670abc"
+    },
+    "lib/libcublas.so.13": {
+      "bytes": 61918408,
+      "sha256": "7761d260387b95fa13b7ec101453abca45ef972d3e6ccc8bcf54251e1ac0470b"
+    },
+    "lib/libcublasLt.so.13": {
+      "bytes": 629929568,
+      "sha256": "3839e12c3b69ed8eb732e5d5f8e749403a582a9d706f779017c785152d9e62f9"
+    },
+    "lib/libcudart.so.13": {
+      "bytes": 707752,
+      "sha256": "7bdba2b5b08cbdc85203c41cc94598adedb1bcfea7cb574ca693ac73599e4e63"
+    },
+    "lib/libggml-base.so.0": {
+      "bytes": 872704,
+      "sha256": "b18971c1a35490194c8aa1666a57494e6324064155017add217f0b20ade657ba"
+    },
+    "lib/libggml-base.so.0.13.1": {
+      "bytes": 872704,
+      "sha256": "b18971c1a35490194c8aa1666a57494e6324064155017add217f0b20ade657ba"
+    },
+    "lib/libggml-cpu.so": {
+      "bytes": 1000112,
+      "sha256": "b61975da167b511286b5df2d6cc384d3ba9eb78d559927d784352d420a4d59db"
+    },
+    "lib/libggml-cuda.so": {
+      "bytes": 43037040,
+      "sha256": "9edb3b2a88fe59fe6d33d9e6a70f8bede4df68b3acfc5e00a63c53e34163985c"
+    },
+    "lib/libggml.so.0": {
+      "bytes": 77824,
+      "sha256": "3ab1d08c26865f5ff7ed753818cd84bc1fa9438e633eae7e3a53b633794973f5"
+    },
+    "lib/libggml.so.0.13.1": {
+      "bytes": 77824,
+      "sha256": "3ab1d08c26865f5ff7ed753818cd84bc1fa9438e633eae7e3a53b633794973f5"
+    },
+    "lib/libllama-cli-impl.so": {
+      "bytes": 1576888,
+      "sha256": "cde7a5d92f37a641b8962fd2a4a516edfadb19a53b426ab392d10995053df7c0"
+    },
+    "lib/libllama-common.so.0": {
+      "bytes": 5074576,
+      "sha256": "a154d3ee1da598eb55ff0ecfbdcd6898fa449a1dbac7c2afcf22263eac7abae6"
+    },
+    "lib/libllama-common.so.0.0.0": {
+      "bytes": 5074576,
+      "sha256": "a154d3ee1da598eb55ff0ecfbdcd6898fa449a1dbac7c2afcf22263eac7abae6"
+    },
+    "lib/libllama-server-impl.so": {
+      "bytes": 3194872,
+      "sha256": "44c74d0f766c9f0fa1bf0482eabf12927b9a5615e18c348c8cd19e3c2c952f25"
+    },
+    "lib/libllama.so.0": {
+      "bytes": 3630056,
+      "sha256": "78addbbc0eb03da6208eab469e8e17a989e47f3e3c99674aa3203fd658694021"
+    },
+    "lib/libllama.so.0.0.0": {
+      "bytes": 3630056,
+      "sha256": "78addbbc0eb03da6208eab469e8e17a989e47f3e3c99674aa3203fd658694021"
+    },
+    "lib/libmtmd.so.0": {
+      "bytes": 1316176,
+      "sha256": "60a791c271d779f4f11321a13dd380424ca20909f530b035ea3d568ab2b87178"
+    },
+    "lib/libmtmd.so.0.0.0": {
+      "bytes": 1316176,
+      "sha256": "60a791c271d779f4f11321a13dd380424ca20909f530b035ea3d568ab2b87178"
+    },
+    "libexec/llama-cli": {
+      "bytes": 72448,
+      "sha256": "5e1c37e59fb2cd37acd8df0fe87a5e1f7a74a496af4ae278012757d84a1b1939"
+    },
+    "libexec/llama-server": {
+      "bytes": 72456,
+      "sha256": "b67f9d10eaa85a500180ddecf6ecdf0e9c100b0bdf4c10e32b0bf657ca50e494"
+    }
+  }
+}

runtimes/supported/BTL-3-Compact-macos-arm64/bin/btl3-server

@@ -43,6 +43,10 @@ ctx=${BTL3_CTX_SIZE:-32768}
 parallel=${BTL3_PARALLEL:-1}
 gpu_layers=${BTL3_GPU_LAYERS:-99}
 aliases=${BTL3_MODEL_ALIASES:-BTL-3}
+case "${BTL3_ENABLE_THINKING:-false}" in
+    1|true|TRUE) chat_template_kwargs='{"enable_thinking":true}' ;;
+    *) chat_template_kwargs='{"enable_thinking":false}' ;;
+esac
 
 set -- \
     --model "$model" \
@@ -55,6 +59,7 @@ set -- \
     --jinja \
     --reasoning auto \
     --reasoning-format deepseek \
+    --chat-template-kwargs "$chat_template_kwargs" \
     --cont-batching \
     --cache-ram 0 \
     --no-mmap \
@@ -66,8 +71,9 @@ if [ -n "${BTL3_API_KEY:-}" ]; then
     set -- "$@" --api-key "$BTL3_API_KEY"
 fi
 if [ "${BTL3_PRINT_COMMAND:-0}" = "1" ]; then
-    printf 'executable=%s\nmodel=%s\nhost=%s\nport=%s\nctx_size=%s\n' \
-        "$server" "$model" "$host" "$port" "$ctx"
+    printf 'executable=%s\nmodel=%s\nhost=%s\nport=%s\nctx_size=%s\nthinking_enabled=%s\n' \
+        "$server" "$model" "$host" "$port" "$ctx" \
+        "$( [ "$chat_template_kwargs" = '{"enable_thinking":true}' ] && printf true || printf false )"
     exit 0
 fi
 exec "$server" "$@"

runtimes/supported/BTL-3-Compact-macos-arm64/bundle-manifest.json

@@ -25,8 +25,8 @@
       "sha256": "de4350f194d6f393bd3ea1065b024d7f92fea0a8f2e435f6557dc0c6696b3a74"
     },
     "bin/btl3-server": {
-      "bytes": 1780,
-      "sha256": "f14fe48ab3d5f0662f7ffe524cd6db8bc4d9bb2d892e54147c1e290cd2532d75"
+      "bytes": 2132,
+      "sha256": "e4dc8efdbeadfc74c27516115b808b3589da1b8b27d0f330498dacd9b1f3d8a9"
     },
     "lib/libcrypto.3.dylib": {
       "bytes": 4845712,