nemo-evaluator-launcher 0.1.19__py3-none-any.whl → 0.1.56__py3-none-any.whl

nemo_evaluator_launcher/common/container_metadata/utils.py

@@ -0,0 +1,63 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+"""Utility functions for container metadata processing."""
+
+
+def parse_container_image(container_image: str) -> tuple[str, str, str, str]:
+    """Parse a container image string into registry type, registry URL, repository, and tag.
+
+    Args:
+        container_image: Container image string (e.g., "nvcr.io/nvidia/eval-factory/simple-evals:25.10")
+
+    Returns:
+        Tuple of (registry_type, registry_url, repository, tag)
+    """
+    # Split tag from image
+    if ":" in container_image:
+        image_part, tag = container_image.rsplit(":", 1)
+    else:
+        image_part = container_image
+        tag = "latest"
+
+    # Parse registry and repository
+    parts = image_part.split("/")
+    if len(parts) < 2:
+        raise ValueError(f"Invalid container image format: {container_image}")
+
+    # Check if first part is a registry (contains '.' or is 'localhost')
+    if "." in parts[0] or parts[0] == "localhost":
+        registry_host = parts[0]
+        # Determine registry type
+        if "gitlab" in registry_host.lower():
+            registry_type = "gitlab"
+        elif "nvcr.io" in registry_host:
+            registry_type = "nvcr"
+        else:
+            registry_type = "nvcr"  # Default to nvcr for other registries
+
+        # Check if registry has a port
+        if ":" in registry_host:
+            registry_url = registry_host
+        else:
+            registry_url = registry_host
+        repository = "/".join(parts[1:])
+    else:
+        # Default registry (Docker Hub)
+        registry_type = "nvcr"
+        registry_url = "registry-1.docker.io"
+        repository = image_part
+
+    return registry_type, registry_url, repository, tag

nemo_evaluator_launcher/common/helpers.py

@@ -14,7 +14,6 @@
 # limitations under the License.
 #
 import base64
-import copy
 import datetime
 from dataclasses import dataclass
 from typing import Optional
@@ -22,6 +21,7 @@ from typing import Optional
 import yaml
 from omegaconf import DictConfig, OmegaConf
 
+from nemo_evaluator_launcher.cli.version import get_versions
 from nemo_evaluator_launcher.common.logging_utils import logger
 
 
@@ -35,35 +35,97 @@ class CmdAndReadableComment:
     # A debuggale readable comment that can be passed along for accompanying
     # the actual command
     debug: str
+    # Whether the content might be potentially unsafe. This is a flag useful for
+    # downstream callers who want to raise exceptions e.g. when a script was
+    # saved that would execute this command.
+    is_potentially_unsafe: bool = False
 
 
-def _yaml_to_echo_command(
-    yaml_str: str, filename: str = "config_ef.yaml"
-) -> CmdAndReadableComment:
-    """Create a safe (see below) echo command saving a yaml to file.
+def _str_to_echo_command(str_to_save: str, filename: str) -> CmdAndReadableComment:
+    """Create a safe (see below) echo command saving a string to file.
 
     Safety in this context means the ability to pass such echo command through the
     `bash -c '...'` boundaries for example.
 
     Naturally, enconding with base64 creates debuggability issues. For that, the second
-    output of the function is the yaml string with bash comment signs prepended.
+    output of the function is the string with bash comment signs prepended.
     """
-    yaml_str_b64 = base64.b64encode(yaml_str.encode("utf-8")).decode("utf-8")
+    str_to_save_b64 = base64.b64encode(str_to_save.encode("utf-8")).decode("utf-8")
     debug_str = "\n".join(
-        [f"# Contents of {filename}"] + ["# " + s for s in yaml_str.splitlines()]
+        [f"# Contents of {filename}"] + ["# " + s for s in str_to_save.splitlines()]
     )
     return CmdAndReadableComment(
-        cmd=f'echo "{yaml_str_b64}" | base64 -d > {filename}', debug=debug_str
+        cmd=f'echo "{str_to_save_b64}" | base64 -d > {filename}', debug=debug_str
     )
 
 
+def _set_nested_optionally_overriding(
+    d: dict, keys: list[str], val: object, *, override_if_exists: bool = False
+):
+    """Sets d[...keys....] = value, creating keys all the way"""
+    temp = d
+    for key in keys[:-1]:
+        temp = temp.setdefault(key, {})
+    if override_if_exists or keys[-1] not in temp:
+        temp[keys[-1]] = val
+
+
+_MIGRATION_MESSAGE = """
+`overrides` field is no longer supported. Use `nemo_evaluator_config` field instead, e.g.:
+
+1. If you are using overrides in your yaml config, replace:
+
+```yaml
+evaluation:
+  overrides:
+    config.params.temperature: 0.6
+    config.params.top_p: 0.95
+```
+
+with:
+
+```yaml
+evaluation:
+  nemo_evaluator_config:
+    config:
+      params:
+        temperature: 0.6
+        top_p: 0.95
+```
+
+2. If you are using overrides in your cli command, replace:
+
+```bash
+nemo-evaluator-launcher run --config my_config.yaml \\
+    -o evaluation.overrides.config.params.temperature=0.6 \\
+    -o evaluation.overrides.config.params.top_p=0.95
+```
+
+with:
+
+```bash
+nemo-evaluator-launcher run --config my_config.yaml \\
+    -o evaluation.nemo_evaluator_config.config.params.temperature=0.6 \\
+    -o evaluation.nemo_evaluator_config.config.params.top_p=0.95
+```
+"""
+
+
 def get_eval_factory_config(
-    cfg: DictConfig, user_task_config: DictConfig, task_definition: dict
+    cfg: DictConfig,
+    user_task_config: DictConfig,
 ) -> dict:
     """Extract config fields for eval factory.
 
     This function extracts the config field similar to how overrides are handled.
+
+    It applies task-level overrides to the global overrides.
     """
+
+    if cfg.evaluation.get("overrides") or user_task_config.get("overrides"):
+        raise ValueError(_MIGRATION_MESSAGE)
+
+    logger.debug("Getting nemo evaluator merged config")
     # Extract config fields similar to overrides - convert to basic Python types first
     # Support both new and old format for backward compatibility
     cfg_config = cfg.evaluation.get("nemo_evaluator_config") or cfg.evaluation.get(
@@ -80,66 +142,146 @@ def get_eval_factory_config(
         user_config = OmegaConf.to_container(user_config, resolve=True)
 
     # Merge the configs
-    config_fields = copy.deepcopy(cfg_config or {})
-    config_fields.update(user_config or {})
+    merged_nemo_evaluator_config: dict = OmegaConf.to_container(
+        OmegaConf.merge(cfg_config, user_config)
+    )
+
+    logger.debug(
+        "Merged nemo evaluator config, not final",
+        source_global_cfg=cfg_config,
+        source_task_config=user_config,
+        result=merged_nemo_evaluator_config,
+    )
 
-    return config_fields
+    return merged_nemo_evaluator_config
 
 
 def get_eval_factory_command(
-    cfg: DictConfig, user_task_config: DictConfig, task_definition: dict
+    cfg: DictConfig,
+    user_task_config: DictConfig,
+    task_definition: dict,
 ) -> CmdAndReadableComment:
-    config_fields = get_eval_factory_config(cfg, user_task_config, task_definition)
-
-    overrides = copy.deepcopy(dict(cfg.evaluation.get("overrides", {})))
-    overrides.update(dict(user_task_config.get("overrides", {})))
-    # NOTE(dfridman): Temporary fix to make sure that the overrides arg is not split into multiple lines.
-    # Consider passing a JSON object on Eval Factory side
-    overrides = {
-        k: (v.strip("\n") if isinstance(v, str) else v) for k, v in overrides.items()
-    }
-    overrides_str = ",".join([f"{k}={v}" for k, v in overrides.items()])
-    model_url = get_endpoint_url(cfg, user_task_config, task_definition)
+    # This gets the eval_factory_config merged from both top-level and task-level.
+    merged_nemo_evaluator_config = get_eval_factory_config(
+        cfg,
+        user_task_config,
+    )
 
-    model_id = get_served_model_name(cfg)
-    model_type = task_definition["endpoint_type"]
-    eval_type = task_definition["task"]
+    # We now prepare the config to be passed to `nemo-evaluator` command.
+    _set_nested_optionally_overriding(
+        merged_nemo_evaluator_config,
+        ["target", "api_endpoint", "url"],
+        get_endpoint_url(
+            cfg,
+            merged_nemo_evaluator_config=merged_nemo_evaluator_config,
+            endpoint_type=task_definition["endpoint_type"],
+        ),
+    )
+    _set_nested_optionally_overriding(
+        merged_nemo_evaluator_config,
+        ["target", "api_endpoint", "model_id"],
+        get_served_model_name(cfg),
+    )
+    _set_nested_optionally_overriding(
+        merged_nemo_evaluator_config,
+        ["target", "api_endpoint", "type"],
+        task_definition["endpoint_type"],
+    )
+    _set_nested_optionally_overriding(
+        merged_nemo_evaluator_config,
+        ["config", "type"],
+        task_definition["task"],
+    )
+    _set_nested_optionally_overriding(
+        merged_nemo_evaluator_config,
+        ["config", "output_dir"],
+        "/results",
+    )
+    # FIXME(martas): update to api_key_name after 25.12 is released
+    _set_nested_optionally_overriding(
+        merged_nemo_evaluator_config,
+        ["target", "api_endpoint", "api_key"],
+        "API_KEY",
+    )
+    _set_nested_optionally_overriding(
+        merged_nemo_evaluator_config,
+        [
+            "metadata",
+            "launcher_resolved_config",
+        ],
+        OmegaConf.to_container(cfg, resolve=True),
+    )
+    _set_nested_optionally_overriding(
+        merged_nemo_evaluator_config,
+        ["metadata", "versioning"],
+        get_versions(),
+    )
 
-    create_file_cmd = _yaml_to_echo_command(
-        yaml.safe_dump(config_fields), "config_ef.yaml"
+    # Now get the pre_cmd either from `evaluation.pre_cmd` or task-level pre_cmd. Note the
+    # order -- task level wins.
+    pre_cmd: str = (
+        user_task_config.get("pre_cmd") or cfg.evaluation.get("pre_cmd") or ""
     )
-    eval_command = f"""cmd=$([[ $(command -v nemo-evaluator) ]] && echo 'nemo-evaluator' || echo 'eval-factory') && $cmd run_eval --model_id {model_id} --model_type {model_type} --eval_type {eval_type} --model_url {model_url} --api_key_name API_KEY --output_dir /results --run_config config_ef.yaml"""
 
-    if overrides:
-        eval_command = f"{eval_command} --overrides {overrides_str}"
+    is_potentially_unsafe = False
+    if pre_cmd:
+        logger.warning(
+            "Found non-empty pre_cmd that might be a security risk if executed. "
+            "Setting `is_potentially_unsafe` to `True`",
+            pre_cmd=pre_cmd,
+        )
+        is_potentially_unsafe = True
+        _set_nested_optionally_overriding(
+            merged_nemo_evaluator_config,
+            ["metadata", "pre_cmd"],
+            pre_cmd,
+        )
+
+    create_pre_script_cmd = _str_to_echo_command(pre_cmd, filename="pre_cmd.sh")
+
+    create_yaml_cmd = _str_to_echo_command(
+        yaml.safe_dump(merged_nemo_evaluator_config), "config_ef.yaml"
+    )
+
+    # NOTE: we use `source` to allow tricks like exports etc (if needed) -- it runs in the same
+    # shell as the command.
+    eval_command = (
+        "cmd=$(command -v nemo-evaluator >/dev/null 2>&1 && echo nemo-evaluator || echo eval-factory) "
+        + "&& source pre_cmd.sh "
+        + "&& $cmd run_eval --run_config config_ef.yaml"
+    )
 
     # We return both the command and the debugging base64-decoded strings, useful
     # for exposing when building scripts.
     return CmdAndReadableComment(
-        cmd=create_file_cmd.cmd + " && " + eval_command, debug=create_file_cmd.debug
+        cmd=create_pre_script_cmd.cmd
+        + " && "
+        + create_yaml_cmd.cmd
+        + " && "
+        + eval_command,
+        debug=create_pre_script_cmd.debug + "\n\n" + create_yaml_cmd.debug,
+        is_potentially_unsafe=is_potentially_unsafe,
     )
 
 
 def get_endpoint_url(
-    cfg: DictConfig, user_task_config: DictConfig, task_definition: dict
+    cfg: DictConfig,
+    merged_nemo_evaluator_config: dict,
+    endpoint_type: str,
 ) -> str:
     def apply_url_override(url: str) -> str:
         """Apply user URL override if provided."""
-        nemo_evaluator_config_url = user_task_config.get(
-            "nemo_evaluator_config", {}
-        ).get("target.api_endpoint.url", None)
-
-        override_url = user_task_config.get("overrides", {}).get(
-            "config.target.api_endpoint.url", None
-        )
-        return (
-            override_url
-            if override_url is not None
-            else nemo_evaluator_config_url
-            if nemo_evaluator_config_url is not None
-            else url
+        nemo_evaluator_config_url = (
+            merged_nemo_evaluator_config.get("target", {})
+            .get("api_endpoint", {})
+            .get("url", None)
         )
 
+        if nemo_evaluator_config_url:
+            return nemo_evaluator_config_url
+
+        return url
+
     if cfg.deployment.type == "none":
         # For deployment: none, use target URL regardless of executor type
         if OmegaConf.is_missing(cfg.target.api_endpoint, "url"):
@@ -160,9 +302,16 @@ def get_endpoint_url(
 
     else:
         # Local executor - use localhost
-        task_endpoint_type = task_definition["endpoint_type"]
-        endpoint_uri = cfg.deployment.endpoints[task_endpoint_type]
-        endpoint_url = f"http://127.0.0.1:{cfg.deployment.port}{endpoint_uri}"
+        endpoint_uri = cfg.deployment.endpoints[endpoint_type]
+
+        # Use HAProxy port if multiple_instances is enabled
+        if cfg.deployment.get("multiple_instances", False):
+            proxy_config = cfg.execution.get("proxy", {}).get("config", {})
+            port = proxy_config.get("haproxy_port", 5009)
+        else:
+            port = cfg.deployment.port
+
+        endpoint_url = f"http://127.0.0.1:{port}{endpoint_uri}"
         return endpoint_url
 
 

nemo_evaluator_launcher/common/logging_utils.py

@@ -61,8 +61,9 @@ import structlog
 # both are unset, default would be used.
 _LOG_LEVEL_ENV_VAR = "NEMO_EVALUATOR_LOG_LEVEL"
 _DEFAULT_LOG_LEVEL = "WARNING"
-_SENSITIVE_KEY_SUBSTRINGS = {
-    # Keep minimal, broad substrings (normalized: lowercased, no spaces/_/-)
+_SENSITIVE_KEY_SUBSTRINGS_NORMALIZED = {
+    # Keep minimal, broad substrings
+    # NOTE: normalized: lowercased, no spaces/_/-
     "authorization",  # covers proxy-authorization, etc.
     "apikey",  # covers api_key, api-key, x-api-key, nvidia_api_key, ...
     "accesskey",  # covers access_key / access-key
@@ -73,6 +74,10 @@ _SENSITIVE_KEY_SUBSTRINGS = {
     "pwd",  # common shorthand
     "passwd",  # common variant
 }
+_ALLOWLISTED_KEYS_SUBSTRINGS = {
+    # NOTE: non-normalized (for allowlisting we want more control)
+    "_tokens",  # This likely would allow us to not redact useful stuff like `limit_tokens`, `max_new_tokens`
+}
 
 
 def _mask(val: object) -> str:
@@ -91,8 +96,11 @@ def _normalize(name: object) -> str:
 
 
 def _is_sensitive_key(key: object) -> bool:
-    k = _normalize(key)
-    return any(substr in k for substr in _SENSITIVE_KEY_SUBSTRINGS)
+    k_norm = _normalize(key)
+    k_non_norm = str(key)
+    return any(
+        substr in k_norm for substr in _SENSITIVE_KEY_SUBSTRINGS_NORMALIZED
+    ) and not any(substr in k_non_norm for substr in _ALLOWLISTED_KEYS_SUBSTRINGS)
 
 
 def _redact_mapping(m: dict) -> dict:
@@ -263,6 +271,9 @@ def _configure_structlog() -> None:
         structlog.processors.UnicodeDecoder(),
     ]
 
+    # Check if stderr is a TTY to determine if colors should be enabled
+    colors_enabled = sys.stderr.isatty()
+
     logging.config.dictConfig(
         {
             "version": 1,
@@ -273,7 +284,7 @@ def _configure_structlog() -> None:
                     "()": "structlog.stdlib.ProcessorFormatter",
                     "processors": [
                         *shared_processors,
-                        MainConsoleRenderer(colors=True),
+                        MainConsoleRenderer(colors=colors_enabled),
                     ],
                 },
                 # Formatter for plain file output