fix: address third-party deprecation warnings

script/brev/README.md

@@ -10,17 +10,16 @@ NeMo Safe Synthesizer without setting up CUDA, drivers, or Python locally.
 Nothing in this directory is executed by the repo or by CI. A Brev Launchable is
 configured in the Brev web console, and the setup script is pasted into a form field
 there. This directory exists so that configuration is versioned and reviewable rather
-than living only in a browser. When you change `setup.sh`, you must also paste the new
-contents into the console for the change to take effect.
+than living only in a browser. When you change `setup.sh` or `welcome.md`, you must also
+update it in the console for the change to take effect.
 
 ### Files
 
 - `setup.sh`: Pasted into the Launchable's Setup Script field. Installs the CUDA
   build of Safe Synthesizer into a dedicated venv, registers it as the default Jupyter
   kernel, and drops the tutorial notebooks in `$HOME`.
-- `welcome.md`: Becomes the customer's `$HOME/README.md`. Fetched at provisioning
-  time from the same tarball as the tutorials, not baked into `setup.sh` -- it would
-  otherwise consume a tenth of the 16 KiB script budget.
+- `welcome.md`: Added to the Launchable's Source files so it renders on the Launchable
+  webpage and appears as the customer's `$HOME/welcome.md`.
 
 ### Console configuration
 
@@ -33,7 +32,7 @@ Launchable with these settings.
 | Software | Install Jupyter on the host | Enabled |
 | Software | Run a Setup Script | Enabled, contents of `setup.sh` |
 | Software | Image ID | Leave blank |
-| Source | Code source | No code files (`setup.sh` downloads the tutorials itself) |
+| Source | Code source | `welcome.md` |
 | Hardware | GPU | 1× 80 GiB VRAM, single GPU |
 | Hardware | Disk | 200 GiB or more -- not resizable after creation |
 | Network | Ports | 8888, named `jupyter` |
@@ -67,8 +66,8 @@ there. Everything operational is a dotfile, which the browser hides by default.
 ```text
 $HOME/
   tutorials/                    the three tutorial notebooks and their datasets
-  README.md                     where to start, rendered on double-click
-                                (SETUP-IN-PROGRESS.md until setup finishes)
+  welcome.md                    where to start, rendered on double-click
+  SETUP-IN-PROGRESS.md          present only while setup is running or after failure
 
   .nss-venv/                    cu129 venv, registered as the default kernel
   .cache/huggingface/           model cache (Hugging Face's default location)
@@ -91,12 +90,12 @@ hard way on a real instance.
   them -- and they have to match the release being installed, not this repo's `main`.
   The script resolves the latest version from the PyPI JSON API, fetches that tag's
   `pyproject.toml`, and reads the CUDA index URLs out of it, then pins the install to
-  that exact version so the two cannot drift. Selection is keyed on the URL containing
-  `cu129`, not on the index name: the flashinfer entry was renamed
-  `flashinfer-jit-cache` → `flashinfer-jit-cache-cu129` between 0.1.8 and 0.1.9, so
-  names are not stable across releases. The parse runs inside a process substitution and
-  therefore cannot fail the script, so the count of discovered indexes is what validates
-  it.
+  that exact version so the two cannot drift. Selection uses the CUDA extra in each
+  index's name or URL and includes indexes referenced by `[tool.uv.sources]` for that
+  extra. The source lookup matters for variant-neutral indexes such as
+  `https://flashinfer.ai/whl/`, while the URL lookup handles names that changed between
+  releases. The parse runs inside a process substitution and therefore cannot fail the
+  script, so the count of discovered indexes is what validates it.
 - uv is installed from a checksum-verified tarball, not `curl | sh`. The
   `astral.sh/install.sh` path logs `no checksums to verify`, so nothing validated what
   it downloaded. The script fetches the pinned release tarball, compares it against the
@@ -106,12 +105,11 @@ hard way on a real instance.
   accepts connections well before this script finishes, so a user who opens it early
   would otherwise see an empty or half-populated file browser and assume the Launchable
   is broken. `SETUP-IN-PROGRESS.md` is written before any slow work, rewritten by the
-  `ERR` trap if provisioning fails, and replaced by `README.md` on success.
-- The welcome text lives in `welcome.md`, not a heredoc. It is pulled from the
-  same tarball as the tutorials, so the two always match, and it is staged as a dotfile
-  until the final step so it never appears while setup is still running. The fetch is
-  non-fatal: `script/brev/` exists in no released tag, so it resolves only from the
-  `main` fallback until a release includes it.
+  `ERR` trap if provisioning fails, and removed on success.
+- The welcome text lives in the Launchable's Source configuration, not a heredoc or
+  release tarball. Brev renders it on the Launchable webpage and copies it to
+  `$HOME/welcome.md`; keeping the console copy synchronized with this directory is a
+  manual deployment step.
 - The setup script has a 16 KiB limit. Brev rejects anything larger, which is why
   the script carries short comments pointing here rather than full explanations. Check
   `wc -c script/brev/setup.sh` before pasting.

script/brev/setup.sh

@@ -21,11 +21,10 @@ readonly REPO_URL="https://github.com/NVIDIA-NeMo/Safe-Synthesizer"
 
 : "${HOME:?HOME is not set}"
 
-# $HOME is the file browser root: only tutorials/ and README.md are visible.
+# $HOME is the file browser root: only customer-facing files stay visible.
 readonly TUTORIALS_DIR="${HOME}/tutorials"
-readonly README_FILE="${HOME}/README.md"
+readonly WELCOME_FILE="${HOME}/welcome.md"
 readonly WAIT_FILE="${HOME}/SETUP-IN-PROGRESS.md"
-readonly WELCOME_STAGED="${HOME}/.nss-welcome.md"
 
 readonly BIN_DIR="${HOME}/.local/bin"
 readonly VENV_DIR="${HOME}/.nss-venv"
@@ -61,7 +60,7 @@ NeMo Safe Synthesizer is still installing -- roughly 5-10 minutes from when
 the instance started. Files appear as it progresses, so a partly-filled file
 browser is expected. Nothing here is ready to run yet.
 
-When setup finishes, this file is replaced by README.md. Refresh to check.
+When setup finishes, this file disappears. Open welcome.md to get started.
 EOF
 
 export PATH="${BIN_DIR}:${PATH}"
@@ -123,8 +122,8 @@ else
   NSS_VERSION="$(curl -fsSL https://pypi.org/pypi/nemo-safe-synthesizer/json \
     | "${VENV_DIR}/bin/python" -c 'import json, sys; print(json.load(sys.stdin)["info"]["version"])')"
 
-  # Indexes come from the installed release's pyproject. Match both generated
-  # names and URLs because static index names do not enforce the CUDA suffix.
+  # Indexes come from the installed release's pyproject. Match CUDA names and
+  # URLs plus source-mapped indexes whose names are variant-neutral.
   pyproject="$(mktemp)"
   curl -fsSL "${REPO_URL}/raw/v${NSS_VERSION}/pyproject.toml" -o "${pyproject}"
   index_args=()
@@ -140,15 +139,29 @@ import sys
 import tomllib
 
 with open(sys.argv[1], "rb") as handle:
-    indexes = tomllib.load(handle)["tool"]["uv"]["index"]
+    uv_config = tomllib.load(handle)["tool"]["uv"]
 
+indexes = uv_config["index"]
 cuda_extra = os.environ["CUDA_EXTRA"]
+
+# Some indexes carry no CUDA variant in their name or URL. Source config is
+# not wheel metadata, so collect indexes mapped to packages for this extra.
+source_indexes = {
+    entry["index"]
+    for value in uv_config.get("sources", {}).values()
+    for entry in (value if isinstance(value, list) else [value])
+    if isinstance(entry, dict)
+    and entry.get("extra") == cuda_extra
+    and "index" in entry
+}
+
 print(
     "\n".join(
         index["url"]
         for index in indexes
         if index["name"].endswith(f"-{cuda_extra}")
         or f"/{cuda_extra}" in index["url"]
+        or index["name"] in source_indexes
     )
 )
 PY
@@ -208,13 +221,6 @@ else
       # Written last; the guard keys on this, so partial runs are redone.
       : >"${TUTORIALS_DIR}/.fetched"
       log "tutorials extracted from ${ref}"
-      # Same tarball as the tutorials. Non-fatal -- see README.
-      if tar -xzf "${tarball}" -C "${tarball_dir}" --strip-components=3 \
-        "${top}/script/brev/welcome.md" 2>/dev/null; then
-        mv "${tarball_dir}/welcome.md" "${WELCOME_STAGED}"
-      else
-        log "WARNING: welcome.md not present in ${ref}"
-      fi
       fetched=1
       break
     fi
@@ -343,19 +349,23 @@ if [[ "${registered}" -ne 1 ]]; then
   log "WARNING: kernel not registered; notebooks may open on the wrong Python"
 fi
 
+# Pre-compile third-party packages that emit SyntaxWarnings on first import so
+# the warnings go into the setup log rather than appearing in notebook output.
+log "pre-compiling packages"
+"${VENV_DIR}/bin/python" -W ignore::SyntaxWarning \
+  -c "import torchao, range_regex" 2>/dev/null || true
+
 # Smoke check -- fail provisioning loudly rather than handing over a broken VM.
 
 log "verifying install"
 "${VENV_DIR}/bin/safe-synthesizer" --version
 "${VENV_DIR}/bin/python" \
   -c "import torch; print('cuda available:', torch.cuda.is_available())"
 
-# Hand over: swap the "please wait" file for the welcome text.
+# Hand over: the Source-provided welcome stays visible after setup completes.
 
-if [[ -f "${WELCOME_STAGED}" ]]; then
-  mv "${WELCOME_STAGED}" "${README_FILE}"
-else
-  log "WARNING: no welcome.md staged; skipping ${README_FILE}"
+if [[ ! -f "${WELCOME_FILE}" ]]; then
+  log "WARNING: ${WELCOME_FILE} is missing; check the Launchable Source files"
 fi
 rm -f "${WAIT_FILE}"
 
@@ -365,7 +375,7 @@ cat <<EOF
 
   Safe Synthesizer is ready.
 
-  Start here : ${README_FILE}
+  Start here : ${WELCOME_FILE}
   Tutorials  : ${TUTORIALS_DIR}/
   Kernel     : "Safe Synthesizer" -- already the default
   Setup log  : ${LOG_FILE}

src/nemo_safe_synthesizer/config/training.py

@@ -4,14 +4,16 @@
 from __future__ import annotations
 
 import importlib
+import math
+import warnings
 from enum import StrEnum
 from typing import (
     TYPE_CHECKING,
     Annotated,
     Literal,
 )
 
-from pydantic import Field
+from pydantic import Field, model_validator
 
 from ..configurator.parameters import (
     Parameters,
@@ -151,6 +153,19 @@ def _mxfp4_config() -> QuantizationConfigMixin:
 ValueGTZero = ValueValidator(lambda p: range_validator(p, lambda v: v >= 0))
 
 
+def is_valid_warmup(value: float) -> bool:
+    """Whether a warmup setting is a usable ratio or step count.
+
+    Mirrors how transformers interprets ``warmup_steps``: ``0`` disables warmup,
+    values below 1 are a ratio of total training steps, and values of 1 or more
+    are an absolute step count. Fractional values of 1 or more are rejected
+    because transformers truncates them (``1.5`` silently becomes ``1``), and
+    non-finite values are rejected because they raise ``OverflowError`` once
+    converted to an integer.
+    """
+    return math.isfinite(value) and value >= 0 and (value < 1 or float(value).is_integer())
+
+
 class TrainingHyperparams(Parameters):
     """Hyperparameters that control the training process behavior.
 
@@ -210,15 +225,43 @@ class TrainingHyperparams(Parameters):
         ),
     ] = 0.01
 
-    warmup_ratio: Annotated[
+    warmup_steps: Annotated[
         float,
-        ValueValidator(value_func=lambda v: v > 0),
+        ValueValidator(value_func=is_valid_warmup),
         Field(
-            title="warmup_ratio",
-            description="Ratio of total training steps used for a linear warmup from 0 to the learning rate. Must be > 0.",
+            title="warmup_steps",
+            description=(
+                "Linear warmup from 0 to the learning rate. "
+                "A whole number of 1 or more sets the exact number of warmup steps; "
+                "a float in (0, 1) is treated as a ratio of total training steps; "
+                "0 disables warmup. "
+                "Must be finite and >= 0, and cannot be fractional at or above 1."
+            ),
         ),
     ] = 0.05
 
+    warmup_ratio: Annotated[
+        float | None,
+        ValueValidator(value_func=lambda v: v is None or is_valid_warmup(v)),
+        Field(
+            title="warmup_ratio",
+            description="Deprecated. Use warmup_steps instead.",
+            exclude=True,
+        ),
+    ] = None
+
+    @model_validator(mode="after")
+    def _migrate_warmup_ratio(self) -> TrainingHyperparams:
+        if self.warmup_ratio is not None:
+            warnings.warn(
+                "warmup_ratio is deprecated and will be removed in a future release. Use warmup_steps instead.",
+                DeprecationWarning,
+                stacklevel=2,
+            )
+            if "warmup_steps" not in self.model_fields_set:
+                self.warmup_steps = self.warmup_ratio
+        return self
+
     lr_scheduler: Annotated[
         str,
         Field(

src/nemo_safe_synthesizer/evaluation/components/text_semantic_similarity.py

@@ -4,6 +4,8 @@
 from __future__ import annotations
 
 import logging
+import warnings
+from contextlib import contextmanager
 from functools import cached_property
 from typing import TYPE_CHECKING
 
@@ -36,11 +38,32 @@
 from . import multi_modal_figures as figures
 
 if TYPE_CHECKING:
+    from collections.abc import Iterator
+
     from sentence_transformers import SentenceTransformer
 
 logger = get_logger(__name__)
 
 
+@contextmanager
+def _suppress_ks_exact_fallback() -> Iterator[None]:
+    """Silence SciPy's notice that ``ks_2samp`` fell back to the asymptotic method.
+
+    ``method="auto"`` attempts the exact calculation and falls back to the
+    asymptotic approximation once the samples are large, which is the intended
+    behaviour here -- the resulting p-values are still valid. The notice is not
+    actionable, so keep it out of CLI, SDK, and notebook output rather than
+    suppressing it separately in each caller.
+    """
+    with warnings.catch_warnings():
+        warnings.filterwarnings(
+            "ignore",
+            message="ks_2samp: Exact calculation unsuccessful",
+            category=RuntimeWarning,
+        )
+        yield
+
+
 class TextSemanticSimilarityDatum(BaseModel):
     """Per-column text semantic similarity scores and PCA projections."""
 
@@ -401,12 +424,13 @@ def _get_text_semantic_similarity(
         # the minimum (most negative) difference between the empirical
         # distribution functions of the samples. The range of this statistic is
         # [0, 1], where 0 indicates no overfitting.
-        ks_test_overfitting = ks_2samp(
-            training_synth_similarity_matrix.max(axis=0),  # F(x)
-            training_similarity_matrix.max(axis=0),  # G(x)
-            alternative="less",
-            method="auto",
-        )
+        with _suppress_ks_exact_fallback():
+            ks_test_overfitting = ks_2samp(
+                training_synth_similarity_matrix.max(axis=0),  # F(x)
+                training_similarity_matrix.max(axis=0),  # G(x)
+                alternative="less",
+                method="auto",
+            )
 
         # Underfitting is measured as the extent to which the synthetic
         # data is less similar to the test data than the test data is to
@@ -417,12 +441,13 @@ def _get_text_semantic_similarity(
         # the minimum (most negative) difference between the empirical
         # distribution functions of the samples. The range of this statistic is
         # [0, 1], where 0 indicates no underfitting.
-        ks_test_underfitting = ks_2samp(
-            test_synth_similarity_matrix.max(axis=0),  # F(x)
-            test_similarity_matrix.max(axis=0),  # G(x)
-            alternative="greater",
-            method="auto",
-        )
+        with _suppress_ks_exact_fallback():
+            ks_test_underfitting = ks_2samp(
+                test_synth_similarity_matrix.max(axis=0),  # F(x)
+                test_similarity_matrix.max(axis=0),  # G(x)
+                alternative="greater",
+                method="auto",
+            )
 
         # The overall semantic similarity score combines underfitting and overfitting
         # The range of this score is [0.37, 1], where 1 indicates perfect model and