/opt/nim/start_server.sh

nvs:1000

/opt/nim

/bin/bash -lc set -euo pipefail; triplet="$(uname -m)-linux-gnu"; rs() { LD_LIBRARY_PATH="/rp/usr/lib/${triplet}:/rp/lib/${triplet}" /rp/usr/bin/rsync "$@"; }; for entry in /stage4/* /stage4/.[!.]*; do [ -e "$entry" ] ||
  [ -L "$entry" ] ||
  continue; name="$(basename "$entry")"; case "$name" in proc|sys|dev|tmp|run|mnt|media) continue;; esac; if [ -d "$entry" ] &&
  [ ! -L "$entry" ]; then rs -aH --numeric-ids --delete --exclude=/hosts --exclude=/resolv.conf --exclude=/hostname "$entry/" "/$name/"; else rs -aH --numeric-ids "$entry" "/$name"; fi; done; python3 -c "import nim_stack, sglang; from nim_stack.manifest.io import _generate_symlinks; print('squash sanity: imports ok')"; test -f /sgl-workspace/sglang/python/sglang/srt/models/glm5_next.py; grep -q 'glm53_always_think' /sgl-workspace/sglang/python/sglang/srt/parser/template_detection.py; grep -q 'data_ptr() % 32' /sgl-workspace/sglang/python/sglang/srt/models/qwen3_5.py; test -f /usr/local/lib/python3.12/dist-packages/nim_stack/engines/sglang/patches/external_model_package.py; grep -q 'ensure_external_model_package' /usr/local/lib/python3.12/dist-packages/nim_stack/engines/sglang/lifecycle.py; grep -q 'parse_sglang_args' /usr/local/lib/python3.12/dist-packages/nim_stack/engines/sglang/manifest/evaluator.py; grep -q 'reasoning_parser = S.INTERNAL("auto")' /usr/local/lib/python3.12/dist-packages/nim_stack/engines/sglang/transformers/apply_runtime_parameters.py; grep -q '_stop_str_window_start' /sgl-workspace/sglang/python/sglang/srt/managers/schedule_batch.py; grep -q 'class PoolCapacityError(ValueError)' /sgl-workspace/sglang/python/sglang/srt/model_executor/pool_configurator.py; grep -q 'def _measure_context' /sgl-workspace/sglang/python/sglang/srt/entrypoints/cpu_memory_profile.py; grep -q 'Cannot cancel a response with status' /sgl-workspace/sglang/python/sglang/srt/entrypoints/openai/serving_responses.py; grep -q 'default_max_tokens = None' /sgl-workspace/sglang/python/sglang/srt/entrypoints/openai/serving_responses.py; grep -q 'NIM_ENABLE_RESPONSE_STORE' /usr/local/lib/python3.12/dist-packages/nim_stack/compiler/ir/schema.py; grep -q 'def _token_id_upper_bound' /sgl-workspace/sglang/python/sglang/srt/entrypoints/openai/serving_tokenize.py; grep -q 'class VideoStreamUndecodableError(ValueError)' /sgl-workspace/sglang/python/sglang/srt/utils/video_decoder.py; test "$(grep -c 'MULTIMODAL_DEFAULT_MAX_OUTPUT_TOKENS' /sgl-workspace/sglang/python/sglang/srt/entrypoints/openai/serving_responses.py || true)" = 0; grep -q 'selectedModelProfileId' /usr/local/lib/python3.12/dist-packages/nim_stack/nimlib_impl/metadata.py; grep -q 'reasoning_off_prefill' /sgl-workspace/sglang/python/sglang/srt/parser/reasoning_parser.py; grep -q 'media_url_max_file_size_mb = S.INTERNAL(2048)' /usr/local/lib/python3.12/dist-packages/nim_stack/engines/sglang/transformers/apply_runtime_parameters.py; grep -q 'HEALTH_LOCATIONS_ISOLATED' /opt/nim/scripts/generate_nginx_config.sh; grep -q '_compressed_tensors_precision' /usr/local/lib/python3.12/dist-packages/nim_stack/nimlib_impl/model_info_derivation.py; grep -q 'EnvProfileSelector.select(artifact.manifest)' /usr/local/lib/python3.12/dist-packages/nim_stack/manifest/selector.py; grep -q 'media_url_timeout_secs=bind("REQUEST_TIMEOUT")' /usr/local/lib/python3.12/dist-packages/nim_stack/engines/sglang/dialect.py; grep -q 'media_url_timeout_secs = S.INTERNAL(120)' /usr/local/lib/python3.12/dist-packages/nim_stack/engines/sglang/transformers/apply_runtime_parameters.py; grep -q 'Checkpoint precision is unavailable' /usr/local/lib/python3.12/dist-packages/nim_stack/manifest/model_metadata.py; grep -q 'RECIPE_REGISTRY = RecipeRegistry(' /usr/local/lib/python3.12/dist-packages/nim_stack/engines/sglang/recipe_registry.py; grep -q '_MAX_NUM_SEQS_PER_GPU' /usr/local/lib/python3.12/dist-packages/nim_stack/engines/sglang/recipe_registry.py; grep -q 'architecture_aliases=' /usr/local/lib/python3.12/dist-packages/nim_stack/engines/sglang/recipe_registry.py; python3 /recipe_probe/probe.py; test -f /usr/local/lib/python3.12/dist-packages/sglang_meta_common/models/hf_layout.py; python3 -c "import importlib.metadata as md; names = {d.metadata['Name'].lower() for d in md.distributions()}; vendor = {n for n in names if n.startswith('sglang-met' 'a')}; assert vendor == {'sglang-meta-common', 'sglang-meta'}, sorted(vendor)"; test "${SGLANG_EXTERNAL_MODEL_PACKAGE:-}" = sglang_meta_common.models; test "${SGLANG_EXTERNAL_MM_PROCESSOR_PACKAGE:-}" = sglang_meta_common.processors; grep -q 'class ApplySGLangServingRecipe' /usr/local/lib/python3.12/dist-packages/nim_stack/engines/sglang/transformers/apply_serving_recipe.py; grep -q 'moe = config.get("moe")' /usr/local/lib/python3.12/dist-packages/nim_stack/compiler/analyses/model_config.py; grep -q 'NIM_MAX_REQUEST_BODY_MB' /opt/nim/scripts/nginx_env_vars.sh; grep -q 'client_max_body_size .*NIM_MAX_REQUEST_BODY_MB' /opt/nim/etc/nginx.conf.template; grep -q 'def get_lm_head_scheme' /sgl-workspace/sglang/python/sglang/srt/layers/quantization/compressed_tensors/compressed_tensors.py; grep -q '_abort_on_client_disconnect' /sgl-workspace/sglang/python/sglang/srt/entrypoints/openai/serving_responses.py; grep -q 'allowed_methods=frozenset({"GET"})' /sgl-workspace/sglang/python/sglang/srt/utils/common.py; grep -q '_to_audio_url_part' /sgl-workspace/sglang/python/sglang/srt/entrypoints/openai/protocol.py; grep -q 'delivery/107' /sgl-workspace/sglang/python/sglang/srt/multimodal/processors/glm4v.py; grep -q 'not dirname.startswith(".")' /usr/local/lib/python3.12/dist-packages/nim_stack/manifest/io.py; grep -q 'def _serves_in_place' /usr/local/lib/python3.12/dist-packages/nim_stack/nimlib_impl/model_download.py; test "$(grep -c '_materialize_dynamic_modules(workspace_path)' /usr/local/lib/python3.12/dist-packages/nim_stack/nimlib_impl/model_download.py)" = 2; grep -q 'class StartupInputError' /usr/local/lib/python3.12/dist-packages/nim_stack/manifest/preflight.py; grep -q 'This is a model-free NIM' /usr/local/lib/python3.12/dist-packages/nim_stack/manifest/preflight.py; grep -q 'This is a model-free NIM. Provide a model' /opt/nim/start_server.sh; test -z "$(find /sgl-workspace/sglang/python /usr/local/lib/python3.12/dist-packages /opt/nim -type f \( -name '*.orig' -o -name '*.rej' \) -print -quit)"; test "$(python3 -c 'import json; print(json.load(open("/usr/share/nodejs/@fastify/busboy/package.json"))["version"])')" = 3.2.1; test -f /usr/share/nodejs/@fastify/busboy/index.js; test ! -e /usr/share/nodejs/@fastify/busboy/lib/main.js; if [ "$(uname -m)" = x86_64 ]; then grep -q -a 'go1.26.9' /usr/local/lib/python3.12/dist-packages/mooncake/libetcd_wrapper.so; if grep -q -a 'go1.26.6' /usr/local/lib/python3.12/dist-packages/mooncake/libetcd_wrapper.so; then echo 'squash: the base go1.26.6 mooncake wrapper is still in place (030 did not take)' >&2; exit 1; fi; test ! -e /usr/local/lib/python3.12/dist-packages/mooncake_transfer_engine_cuda13.libs/libetcd_wrapper-3ad7b349.so; else test -z "$(find /usr/local/lib/python3.12/dist-packages -name 'libetcd_wrapper*.so' -print -quit)"; fi; test ! -e /opt/nim/patches; test -z "$(find /opt/nim/vendor -maxdepth 1 -name 'sglang_met[a]*' -print -quit 2>/dev/null)"; test -z "$(find /opt/nim/runtime-cache -not -user 1000 -print -quit)"; test -x /opt/nim/start_server.sh; test ! -e /usr/bin/rsync

SGLANG_CACHE_DIR=/opt/nim/runtime-cache/sglang TORCHINDUCTOR_CACHE_DIR=/opt/nim/runtime-cache/torchinductor TRITON_CACHE_DIR=/opt/nim/runtime-cache/triton HF_HOME= SGLANG_EXTERNAL_MODEL_PACKAGE=sglang_meta_common.models SGLANG_EXTERNAL_MM_PROCESSOR_PACKAGE=sglang_meta_common.processors FLASHINFER_NO_DOWNLOAD=1

root

[/bin/bash -lc]

—

/opt/nim/start_server.sh

nvs:1000