workbench 0.8.485__tar.gz → 0.8.487__tar.gz

{workbench-0.8.485 → workbench-0.8.487}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: workbench
-Version: 0.8.485
+Version: 0.8.487
 Summary: Workbench: A Dashboard and Python API for creating and deploying AWS SageMaker Model Pipelines
 Author-email: SuperCowPowers LLC <support@supercowpowers.com>
 License: MIT License
@@ -71,6 +71,7 @@ Requires-Dist: ray[tune]; extra == "modeling"
 Requires-Dist: optuna; extra == "modeling"
 Requires-Dist: torch; extra == "modeling"
 Requires-Dist: chemprop>=2.3.0; extra == "modeling"
+Requires-Dist: tabicl[shap]>=2.2.0; extra == "modeling"
 Provides-Extra: dev
 Requires-Dist: tox; extra == "dev"
 Requires-Dist: pytest; extra == "dev"
@@ -154,7 +155,7 @@ All the rage for the Open ADMET Challenge. ADMET Workbench supports:
 - Chemprop + Descriptors Models (MPNN + Molecular Descriptors)
 - Foundation Chemprop Models (CheMeleon Pretrained)
 
-Examples: [Deploying Chemprop Models](examples/models/chemprop.py) &nbsp;·&nbsp;
+Examples: [Deploying Chemprop Models](examples/models/chemprop_model.py) &nbsp;·&nbsp;
 [Deploying Foundation Chemprop Models](examples/models/chemprop_foundation.py)
 
 **References**

{workbench-0.8.485 → workbench-0.8.487}/README.md

@@ -58,7 +58,7 @@ All the rage for the Open ADMET Challenge. ADMET Workbench supports:
 - Chemprop + Descriptors Models (MPNN + Molecular Descriptors)
 - Foundation Chemprop Models (CheMeleon Pretrained)
 
-Examples: [Deploying Chemprop Models](examples/models/chemprop.py)  · 
+Examples: [Deploying Chemprop Models](examples/models/chemprop_model.py)  · 
 [Deploying Foundation Chemprop Models](examples/models/chemprop_foundation.py)
 
 **References**

{workbench-0.8.485 → workbench-0.8.487}/applications/aws_dashboard/Dockerfile

@@ -29,7 +29,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
         -r /tmp/requirements.lock
 
 # Layer 2: Workbench only (changes often, ~20 MB)
-ARG WORKBENCH_VERSION=0.8.484
+ARG WORKBENCH_VERSION=0.8.486
 RUN pip install --no-deps "workbench==${WORKBENCH_VERSION}"
 
 # Copy app code and configuration (after pip installs to avoid cache invalidation)

{workbench-0.8.485 → workbench-0.8.487}/applications/aws_dashboard/requirements.lock

@@ -128,7 +128,7 @@ charset-normalizer==3.4.7
     # via
     #   -c ci/.uv-constraints.txt
     #   requests
-click==8.4.1
+click==8.5.0
     # via
     #   -c ci/.uv-constraints.txt
     #   flask

{workbench-0.8.485 → workbench-0.8.487}/ci/endpoint_import_smoke.py

@@ -47,6 +47,7 @@ _FRAMEWORK_SPECIFIC = {
     "workbench.endpoints.pytorch_utils": {"torch"},
     "workbench.endpoints.chemprop_shap_utils": {"torch", "chemprop"},
     "workbench.endpoints.chemprop_utils": {"chemprop"},
+    "workbench.endpoints.tabicl_utils": {"tabicl"},
 }
 
 # Extras that don't ship to any container at runtime and aren't relevant to

{workbench-0.8.485 → workbench-0.8.487}/docs/models/chemprop_models.md

@@ -247,10 +247,10 @@ Registered foundation names (`"CheMeleon"`) resolve through three rungs, in orde
 Training jobs receive `WORKBENCH_BUCKET` from the container environment, so rung 2 works out of the box once the weights are staged. Stage them with:
 
 ```bash
-python scripts/admin/push_chemeleon_models.py
+python scripts/admin/push_foundation_models.py --model chemeleon --file chemeleon_mp.pt
 ```
 
-The script verifies size and checksum against the registry in `src/workbench/training/foundation_models.py`, confirms the file loads as a ChemProp checkpoint, then uploads the `.pt` plus a `SOURCE.json` provenance sidecar. Run with `--dry-run` to preview and `--force` to replace an existing object. Until an install stages the weights, jobs fall back to the public origin — which is exactly the dependency staging removes.
+The script verifies size and checksum against the registry in `src/workbench/training/foundation_models.py`, confirms the file loads as the expected checkpoint, then uploads the `.pt` plus a `SOURCE.json` provenance sidecar. Run with `--dry-run` to preview and `--force` to replace an existing object. Until an install stages the weights, jobs fall back to the public origin — which is exactly the dependency staging removes.
 
 ## Hyperparameters
 
@@ -468,7 +468,7 @@ All ChemProp models include built-in uncertainty quantification:
 - All options compose — you can use Foundation + Multi-Task + Descriptors together
 
 !!! note "Examples"
-    Full code listings are in the repository under [`examples/models/chemprop.py`](https://github.com/SuperCowPowers/workbench/blob/main/examples/models/chemprop.py) and [`examples/models/chemprop_foundation.py`](https://github.com/SuperCowPowers/workbench/blob/main/examples/models/chemprop_foundation.py).
+    Full code listings are in the repository under [`examples/models/chemprop_model.py`](https://github.com/SuperCowPowers/workbench/blob/main/examples/models/chemprop_model.py) and [`examples/models/chemprop_foundation.py`](https://github.com/SuperCowPowers/workbench/blob/main/examples/models/chemprop_foundation.py).
 
 ---
 

{workbench-0.8.485 → workbench-0.8.487}/examples/models/chemprop_foundation.py

@@ -9,7 +9,7 @@ to predict Mordred molecular descriptors. This gives the MPNN a strong prior for
 molecular representation learning.
 
 Weights resolve from the local cache, then the Workbench bucket, then the public
-origin URL. Stage them in your bucket with scripts/admin/push_chemeleon_models.py
+origin URL. Stage them in your bucket with scripts/admin/push_foundation_models.py
 so training jobs never reach the public internet.
 
 References:

workbench-0.8.487/examples/models/tabicl_local.py

@@ -0,0 +1,48 @@
+"""TabICL trained and served on this machine: no AWS account, no SageMaker images.
+
+Needs the modeling extra: pip install "workbench[modeling]"
+"""
+
+from workbench.api import ModelFramework, ModelType, PublicData
+from workbench.local import DataSource
+
+feature_list = [
+    "molwt",
+    "mollogp",
+    "molmr",
+    "heavyatomcount",
+    "numhacceptors",
+    "numhdonors",
+    "numheteroatoms",
+    "numrotatablebonds",
+    "numvalenceelectrons",
+    "numaromaticrings",
+    "numsaturatedrings",
+    "numaliphaticrings",
+    "ringcount",
+    "tpsa",
+    "labuteasa",
+    "balabanj",
+    "bertzct",
+]
+
+# Public AqSol data -> local DataSource -> local FeatureSet (column names are stored lowercase)
+df = PublicData().get("comp_chem/aqsol/aqsol_public_data")
+ds = DataSource(df, name="aqsol_local")
+fs = ds.to_features("aqsol_local_features", id_column="ID")
+
+# TabICL Regression Model: fold models give the cross-fold metrics, one cached model serves
+model = fs.to_model(
+    "aqsol-local-reg-tabicl",
+    model_type=ModelType.UQ_REGRESSOR,
+    model_framework=ModelFramework.TABICL,
+    target_column="solubility",
+    feature_list=feature_list,
+)
+metrics = model.get_inference_metrics()
+print(metrics)
+
+# A local Endpoint loads the model in this process
+end = model.to_endpoint()
+preds = end.inference(fs.pull_dataframe().head(20))
+print(preds[["id", "solubility", "prediction", "prediction_std", "confidence"]].head())

workbench-0.8.487/examples/models/tabicl_model.py

@@ -0,0 +1,35 @@
+from workbench.api import FeatureSet, Model, ModelType, ModelFramework, Endpoint
+
+# Grab a FeatureSet
+my_features = FeatureSet("aqsol_features")
+model = Model("aqsol-regression")
+feature_list = model.features()
+target = model.target()
+
+# Recreate Flag in case you want to recreate the artifacts
+recreate = True
+
+# TabICL Regression Model (a tabular foundation model: in-context learning, no gradient training)
+if recreate or not Model("aqsol-reg-tabicl").exists():
+    feature_set = FeatureSet("aqsol_features")
+    m = feature_set.to_model(
+        name="aqsol-reg-tabicl",
+        model_type=ModelType.UQ_REGRESSOR,
+        model_framework=ModelFramework.TABICL,
+        feature_list=feature_list,
+        target_column=target,
+        description="TabICL Regression Model for AQSol",
+        tags=["tabicl", "molecular descriptors"],
+    )
+    m.set_owner("BW")
+
+# Create an Endpoint for the Regression Model. Serverless is refused when the model's
+# measured serving memory is over 5 GB; pass serverless=False for a real-time endpoint.
+if recreate or not Endpoint("aqsol-reg-tabicl").exists():
+    m = Model("aqsol-reg-tabicl")
+    end = m.to_endpoint(tags=["tabicl", "molecular descriptors"])
+    end.set_owner("BW")
+
+    # Run inference on the endpoint
+    end.test_inference()
+    end.cross_fold_inference()

{workbench-0.8.485 → workbench-0.8.487}/lambda_layers/build_deploy.sh

@@ -26,7 +26,7 @@ WORKBENCH_ROOT="$(cd "$SCRIPT_DIR/.." &> /dev/null && pwd)"
 AWS_ACCOUNT_ID="507740646243"
 REGION_LIST=("us-east-1" "us-west-2")
 PYTHON_VERSION="3.12"
-WORKBENCH_VERSION=0.8.484   # keep in lockstep with the image's ARG WORKBENCH_VERSION
+WORKBENCH_VERSION=0.8.486   # keep in lockstep with the image's ARG WORKBENCH_VERSION
 LAMBDA_ARCH="manylinux2014_x86_64"
 
 GREEN='\033[0;32m'

{workbench-0.8.485 → workbench-0.8.487}/ml_pipelines/OpenADMET/pxr/phase1/pxr_tabicl_spike_phase1.py

@@ -17,9 +17,18 @@ Arms:
 TabICL arms are averaged over SEEDS before the paired bootstrap, so the comparison
 carries row noise only; the per-seed spread is reported alongside.
 
-Without the readouts TabICL loses (RAE 0.687 vs chemprop 0.591, +0.096 [+0.056, +0.142]),
-and stacking chemprop's own OOF pEC50 prediction only recovers chemprop (0.609,
-+0.019 [-0.004, +0.043]).
+Results (phase1_test, paired RAE delta vs chemprop with 95% CI):
+
+    arm               RAE    MAE    delta
+    chemprop          0.591  0.472
+    chemprop_readout  0.569  0.454  -0.022 [-0.049, +0.008]  wins 93% of resamples
+    tabicl            0.687  0.548  +0.096 [+0.056, +0.135]
+    tabicl_readout    0.569  0.455  -0.021 [-0.052, +0.010]  wins 90% of resamples
+
+The primary-screen readouts carry the gain: they lift TabICL by 0.118 RAE to a tie with
+chemprop + readouts. On the same inputs the two models land in the same place, so TabICL
+is a competitive framework option rather than a winner here. Stacking chemprop's own OOF
+pEC50 prediction into TabICL instead of the readouts only recovers chemprop (0.609).
 
 Needs openadmet_pxr_readout and pxr-reg-chemprop-readout-phase1 (see pipelines.json).
 

workbench-0.8.487/ml_pipelines/OpenADMET/pxr/phase1/pxr_tabicl_uq_phase1.py

@@ -0,0 +1,138 @@
+"""PXR phase-1: TabICL's native intervals vs the UQModel V1 intervals built on them.
+
+The tabicl template feeds only the width of TabICL's predictive distribution
+(`prediction_std`) into UQModel V1, which then builds its own conformal intervals.
+This scores both sets of intervals on the 253 `phase1_test` rows, so the choice of
+which to publish as the `q_*` columns rests on coverage and width, not assumption.
+
+Mirrors the template: scaffold-split fold models give out-of-fold predictions for V1's
+calibration, one cached model on all `train` rows scores the held-out rows. Features are
+the scoreboard's best arm (CheMeleon PCA256 + primary-screen readouts).
+
+Result (253 rows; interval score = width + miss penalty, lower is better):
+
+    nominal   native coverage / score   V1 coverage / score
+    50%       54.5% / 1.518             56.1% / 1.499
+    68%       71.1% / 1.904             70.4% / 1.894
+    80%       79.4% / 2.307             80.2% / 2.320
+    95%       94.1% / 3.425             92.1% / 3.833
+
+A tie through 80%; at 95% the native intervals cover better and score better, a
+difference of about five compounds. V1's error model leans on TabICL's width
+(prediction_std importance 0.48 of five features), so the native signal carries through.
+The template keeps V1 for the `q_*` columns: same contract as every other framework.
+
+Runs locally (not via ml_pipeline_launcher), ~6 minutes on CPU:
+
+    uv run python pxr_tabicl_uq_phase1.py
+"""
+
+import numpy as np
+import pandas as pd
+import torch
+from chemprop import data
+from sklearn.decomposition import PCA
+from sklearn.preprocessing import StandardScaler
+
+from workbench.api import FeatureSet
+from workbench.endpoints.tabicl_utils import predict_with_std
+from workbench.endpoints.uq_regression import fit_regression_uq, uq_query_df
+from workbench.training.chemprop_core import load_foundation_weights
+from workbench.training.splits import get_split_indices
+from workbench.training.tabicl_core import fit_tabicl
+
+fs_name = "openadmet_pxr_readout"
+id_col, target = "molecule_name", "pec50"
+READOUTS = ["lfc_8um_readout", "lfc_33um_readout"]
+PCA_DIMS = 256
+HYPERPARAMETERS = {"n_estimators": 8, "batch_size": 1, "seed": 42, "pca_components": None}
+
+# Nominal coverage -> (lower, upper) quantile columns, as UQModel V1 names them
+LEVELS = {0.50: ("q_25", "q_75"), 0.68: ("q_16", "q_84"), 0.80: ("q_10", "q_90"), 0.95: ("q_025", "q_975")}
+ALPHAS = {
+    "q_025": 0.025,
+    "q_10": 0.10,
+    "q_16": 0.16,
+    "q_25": 0.25,
+    "q_75": 0.75,
+    "q_84": 0.84,
+    "q_90": 0.90,
+    "q_975": 0.975,
+}
+
+
+def chemeleon_embed(smiles: list[str], batch_size: int = 256) -> np.ndarray:
+    """Frozen CheMeleon MPNN + mean aggregation -> one 2048-d vector per molecule."""
+    mp, agg = load_foundation_weights("CheMeleon")
+    mp.eval()
+    dataset = data.MoleculeDataset([data.MoleculeDatapoint.from_smi(s) for s in smiles])
+    loader = data.build_dataloader(dataset, batch_size=batch_size, shuffle=False)
+    with torch.no_grad():
+        return np.vstack([agg(mp(batch.bmg), batch.bmg.batch).numpy() for batch in loader])
+
+
+def interval_report(name: str, intervals: pd.DataFrame, y: np.ndarray) -> list[dict]:
+    """Coverage, mean width, and interval score (width + miss penalty; lower is better) per level."""
+    rows = []
+    for level, (lo_col, hi_col) in LEVELS.items():
+        lo, hi = intervals[lo_col].to_numpy(), intervals[hi_col].to_numpy()
+        miss = np.maximum(lo - y, 0) + np.maximum(y - hi, 0)
+        rows.append(
+            {
+                "intervals": name,
+                "nominal": level,
+                "coverage": float(np.mean((y >= lo) & (y <= hi))),
+                "mean_width": float(np.mean(hi - lo)),
+                "interval_score": float(np.mean((hi - lo) + (2 / (1 - level)) * miss)),
+            }
+        )
+    return rows
+
+
+# Data: train rows fit and calibrate, phase1_test rows score
+df = FeatureSet(fs_name).pull_dataframe()[[id_col, "smiles", target, "split"] + READOUTS]
+train = df[df["split"] == "train"].reset_index(drop=True)
+test = df[df["split"] == "phase1_test"].reset_index(drop=True)
+print(f"{len(train)} train / {len(test)} phase1_test rows")
+
+# CheMeleon embedding -> standardize -> PCA (fit on train only), plus the readouts
+emb_train, emb_test = chemeleon_embed(train["smiles"].tolist()), chemeleon_embed(test["smiles"].tolist())
+scaler = StandardScaler().fit(emb_train)
+pca = PCA(n_components=PCA_DIMS, random_state=0).fit(scaler.transform(emb_train))
+features = [f"pc_{i}" for i in range(PCA_DIMS)] + READOUTS
+X_train = pd.DataFrame(pca.transform(scaler.transform(emb_train)), columns=features[:PCA_DIMS])
+X_test = pd.DataFrame(pca.transform(scaler.transform(emb_test)), columns=features[:PCA_DIMS])
+X_train[READOUTS], X_test[READOUTS] = train[READOUTS], test[READOUTS]
+
+# Out-of-fold predictions and spreads from uncached fold models (V1's calibration set)
+folds = get_split_indices(train, n_splits=5, strategy="scaffold", target_column=None, test_size=0.2, random_state=42)
+oof_pred, oof_std = np.full(len(train), np.nan), np.full(len(train), np.nan)
+for fold_idx, (train_idx, val_idx) in enumerate(folds):
+    print(f"Fold {fold_idx + 1}/{len(folds)}: context {len(train_idx)}, val {len(val_idx)}")
+    fold_model, _ = fit_tabicl(HYPERPARAMETERS, X_train.iloc[train_idx], train[target].iloc[train_idx], kv_cache=False)
+    oof_pred[val_idx], oof_std[val_idx] = predict_with_std(fold_model, X_train.iloc[val_idx])
+
+# Served model on all train rows: held-out mean, spread, and native quantiles
+model, _ = fit_tabicl(HYPERPARAMETERS, X_train, train[target], kv_cache="repr")
+out = model.predict(X_test, output_type=["mean", "quantiles"], alphas=sorted(ALPHAS.values()))
+native = pd.DataFrame(out["quantiles"], columns=sorted(ALPHAS, key=ALPHAS.get))
+pred, std = np.asarray(out["mean"]), ((native["q_84"] - native["q_16"]) / 2).to_numpy()
+
+# UQModel V1, calibrated on the out-of-fold rows exactly as the template does
+prox_df = train[[id_col, target, "smiles"]].copy()
+prox_df["in_model"] = True
+uq_model = fit_regression_uq(
+    per_target={
+        target: {"ids": train[id_col].tolist(), "y_true": train[target].values, "y_pred": oof_pred, "y_std": oof_std}
+    },
+    prox_df=prox_df,
+    id_column=id_col,
+    features=features,
+    active_version="v1",
+)["uq_model"]
+v1 = uq_model.predict(uq_query_df(uq_model, test), pred, std)
+
+y = test[target].to_numpy()
+report = pd.DataFrame(interval_report("tabicl_native", native, y) + interval_report("uq_v1", v1, y))
+print(f"\nphase1_test MAE {np.abs(pred - y).mean():.3f}; intervals on {len(y)} rows (interval_score: lower is better)")
+print(report.to_string(index=False, float_format="%.3f"))

{workbench-0.8.485 → workbench-0.8.487}/pyproject.toml

@@ -95,6 +95,7 @@ modeling = [
     "optuna",
     "torch",
     "chemprop >= 2.3.0",
+    "tabicl[shap] >= 2.2.0",
 ]
 # Dev tooling only — no runtime deps. Skipped when deriving the endpoint-import-smoke
 # forbidden list (see ci/endpoint_import_smoke.py), so nothing here constrains the

{workbench-0.8.485 → workbench-0.8.487}/sagemaker_images/base/inference/Dockerfile

@@ -33,7 +33,7 @@ COPY --from=builder /usr/local/lib/python3.12/site-packages /usr/local/lib/pytho
 COPY --from=builder /usr/local/bin /usr/local/bin
 
 # Layer 2: Workbench — rebuilds per version bump, ~20 MB delta.
-ARG WORKBENCH_VERSION=0.8.484
+ARG WORKBENCH_VERSION=0.8.486
 RUN pip install --no-deps "workbench==${WORKBENCH_VERSION}"
 
 # Add the shared serve script

{workbench-0.8.485 → workbench-0.8.487}/sagemaker_images/base/inference/requirements.lock

@@ -40,7 +40,7 @@ charset-normalizer==3.4.7
     # via
     #   -c ci/.uv-constraints.txt
     #   requests
-click==8.4.1
+click==8.5.0
     # via
     #   -c ci/.uv-constraints.txt
     #   uvicorn

{workbench-0.8.485 → workbench-0.8.487}/sagemaker_images/base/training/Dockerfile

@@ -32,5 +32,5 @@ COPY --from=builder /usr/local/lib/python3.12/site-packages /usr/local/lib/pytho
 COPY --from=builder /usr/local/bin /usr/local/bin
 
 # Layer 2: Workbench — rebuilds per version bump, ~20 MB delta.
-ARG WORKBENCH_VERSION=0.8.484
+ARG WORKBENCH_VERSION=0.8.486
 RUN pip install --no-deps "workbench==${WORKBENCH_VERSION}"

{workbench-0.8.485 → workbench-0.8.487}/sagemaker_images/ml_pipelines/Dockerfile

@@ -46,7 +46,7 @@ COPY --from=builder /usr/local/lib/python3.12/site-packages /usr/local/lib/pytho
 COPY --from=builder /usr/local/bin /usr/local/bin
 
 # Layer 2: Workbench (changes often, small layer)
-ARG WORKBENCH_VERSION=0.8.484
+ARG WORKBENCH_VERSION=0.8.486
 RUN pip install --no-deps "workbench==${WORKBENCH_VERSION}"
 
 # Layer 3: App code (tiny, changes rarely)

{workbench-0.8.485 → workbench-0.8.487}/sagemaker_images/ml_pipelines/requirements.lock

@@ -126,7 +126,7 @@ charset-normalizer==3.4.7
     # via
     #   -c ci/.uv-constraints.txt
     #   requests
-click==8.4.1
+click==8.5.0
     # via
     #   -c ci/.uv-constraints.txt
     #   flask

{workbench-0.8.485 → workbench-0.8.487}/sagemaker_images/pytorch_chem/inference/Dockerfile

@@ -43,7 +43,7 @@ COPY --from=builder /usr/local/lib/python3.12/site-packages /usr/local/lib/pytho
 COPY --from=builder /usr/local/bin /usr/local/bin
 
 # Layer 2: Workbench — rebuilds per version bump, ~20 MB delta.
-ARG WORKBENCH_VERSION=0.8.484
+ARG WORKBENCH_VERSION=0.8.486
 RUN pip install --no-deps "workbench==${WORKBENCH_VERSION}"
 
 # Add the shared serve script

{workbench-0.8.485 → workbench-0.8.487}/sagemaker_images/pytorch_chem/inference/requirements.in

@@ -1,5 +1,5 @@
-# Deps for the chemprop inference image. Smoke-contract subset + chemprop
-# overlay + CPU torch (lockfile pins torch==X.Y.Z+cpu via the PyTorch
+# Deps for the chemprop/tabicl inference image. Smoke-contract subset + chemprop
+# + tabicl overlay + CPU torch (lockfile pins torch==X.Y.Z+cpu via the PyTorch
 # CPU index, passed in ci/lock.sh). Chemprop inference on small batches
 # (typical real-time scoring) is CPU-bound work; GPU buys no real speedup
 # at low batch sizes and ~3GB of extra image weight.
@@ -22,6 +22,7 @@ rdkit >= 2026.3.1
 fastapi
 uvicorn
 
-# Image-specific (chemprop + CUDA torch from the PyTorch index)
+# Image-specific (chemprop + tabicl + CPU torch from the PyTorch index)
 torch
 chemprop
+tabicl

{workbench-0.8.485 → workbench-0.8.487}/sagemaker_images/pytorch_chem/inference/requirements.lock

@@ -27,6 +27,7 @@ annotated-types==0.7.0
 anyio==4.15.1
     # via
     #   -c ci/.uv-constraints.txt
+    #   httpx2
     #   starlette
 astartes==1.3.3
     # via
@@ -64,9 +65,10 @@ chemprop==2.3.0
     # via
     #   -c ci/.uv-constraints.txt
     #   -r sagemaker_images/pytorch_chem/inference/requirements.in
-click==8.4.1
+click==8.5.0
     # via
     #   -c ci/.uv-constraints.txt
+    #   huggingface-hub
     #   uvicorn
 configargparse==1.7.5
     # via
@@ -84,6 +86,10 @@ dill==0.4.1
     # via
     #   -c ci/.uv-constraints.txt
     #   multiprocess
+einops==0.8.2
+    # via
+    #   -c ci/.uv-constraints.txt
+    #   tabicl
 fastapi==0.136.3
     # via
     #   -c ci/.uv-constraints.txt
@@ -91,6 +97,7 @@ fastapi==0.136.3
 filelock==3.29.3
     # via
     #   -c ci/.uv-constraints.txt
+    #   huggingface-hub
     #   torch
 frozenlist==1.8.0
     # via
@@ -100,17 +107,36 @@ frozenlist==1.8.0
 fsspec==2026.4.0
     # via
     #   -c ci/.uv-constraints.txt
+    #   huggingface-hub
     #   lightning
     #   pytorch-lightning
     #   torch
 h11==0.16.0
     # via
     #   -c ci/.uv-constraints.txt
+    #   httpcore2
     #   uvicorn
+hf-xet==1.6.0
+    # via
+    #   -c ci/.uv-constraints.txt
+    #   huggingface-hub
+httpcore2==2.13.1
+    # via
+    #   -c ci/.uv-constraints.txt
+    #   httpx2
+httpx2==2.13.1
+    # via
+    #   -c ci/.uv-constraints.txt
+    #   huggingface-hub
+huggingface-hub==2.1.1
+    # via
+    #   -c ci/.uv-constraints.txt
+    #   tabicl
 idna==3.18
     # via
     #   -c ci/.uv-constraints.txt
     #   anyio
+    #   httpx2
     #   requests
     #   yarl
 jinja2==3.1.6
@@ -198,12 +224,14 @@ numpy==2.4.6
     #   rdkit
     #   scikit-learn
     #   scipy
+    #   tabicl
     #   torchmetrics
     #   xarray
 packaging==26.2
     # via
     #   -c ci/.uv-constraints.txt
     #   awswrangler
+    #   huggingface-hub
     #   lightning
     #   lightning-utilities
     #   mordredcommunity
@@ -242,6 +270,7 @@ psutil==7.2.2
     # via
     #   -c ci/.uv-constraints.txt
     #   aimsim-core
+    #   tabicl
 pyarrow==24.0.0
     # via
     #   -c ci/.uv-constraints.txt
@@ -274,6 +303,7 @@ pytz==2026.2
 pyyaml==6.0.3
     # via
     #   -c ci/.uv-constraints.txt
+    #   huggingface-hub
     #   lightning
     #   pytorch-lightning
 rdkit==2026.3.4
@@ -304,6 +334,7 @@ scikit-learn==1.9.0
     #   aimsim-core
     #   astartes
     #   chemprop
+    #   tabicl
 scipy==1.17.1
     # via
     #   -c ci/.uv-constraints.txt
@@ -312,6 +343,7 @@ scipy==1.17.1
     #   cuik-molmaker-pin
     #   descriptastorus
     #   scikit-learn
+    #   tabicl
 setuptools==83.0.0
     # via
     #   -c ci/.uv-constraints.txt
@@ -330,6 +362,10 @@ sympy==1.14.0
     # via
     #   -c ci/.uv-constraints.txt
     #   torch
+tabicl==2.2.0
+    # via
+    #   -c ci/.uv-constraints.txt
+    #   -r sagemaker_images/pytorch_chem/inference/requirements.in
 tabulate==0.10.0
     # via
     #   -c ci/.uv-constraints.txt
@@ -345,6 +381,7 @@ torch==2.13.0+cpu
     #   chemprop
     #   lightning
     #   pytorch-lightning
+    #   tabicl
     #   torchmetrics
 torchmetrics==1.9.0
     # via
@@ -354,9 +391,16 @@ torchmetrics==1.9.0
 tqdm==4.68.2
     # via
     #   -c ci/.uv-constraints.txt
+    #   huggingface-hub
     #   lightning
     #   myerson
     #   pytorch-lightning
+    #   tabicl
+truststore==0.10.4
+    # via
+    #   -c ci/.uv-constraints.txt
+    #   httpcore2
+    #   httpx2
 typing-extensions==4.16.0
     # via
     #   -c ci/.uv-constraints.txt
@@ -365,6 +409,8 @@ typing-extensions==4.16.0
     #   anyio
     #   awswrangler
     #   fastapi
+    #   httpx2
+    #   huggingface-hub
     #   lightning
     #   lightning-utilities
     #   pydantic

{workbench-0.8.485 → workbench-0.8.487}/sagemaker_images/pytorch_chem/training/Dockerfile

@@ -39,5 +39,5 @@ RUN --mount=type=cache,target=/root/.cache/uv \
         -r /tmp/requirements.lock
 
 # Layer 2: Workbench — rebuilds per version bump, ~20 MB delta.
-ARG WORKBENCH_VERSION=0.8.484
+ARG WORKBENCH_VERSION=0.8.486
 RUN pip install --no-deps "workbench==${WORKBENCH_VERSION}"

{workbench-0.8.485 → workbench-0.8.487}/sagemaker_images/pytorch_chem/training/requirements.in

@@ -1,5 +1,5 @@
-# Deps for the chemprop training image. Smoke-contract subset + chemprop +
-# shap overlay + CUDA torch (lockfile pins torch==X.Y.Z+cu130 via the PyTorch
+# Deps for the chemprop/tabicl training image. Smoke-contract subset + chemprop +
+# tabicl + shap overlay + CUDA torch (lockfile pins torch==X.Y.Z+cu130 via the PyTorch
 # cu130 index, passed in ci/lock.sh). No fastapi/uvicorn — training jobs
 # don't serve.
 #
@@ -17,9 +17,10 @@ joblib >= 1.3.2
 # republishing cuik_molmaker_pin against the newer rdkit.
 rdkit >= 2026.3.1
 
-# Image-specific (chemprop + CUDA torch + shap for SHAP feature importance during training)
+# Image-specific (chemprop + tabicl + CUDA torch + shap for SHAP feature importance during training)
 torch
 chemprop
+tabicl[shap]
 shap
 
 # Hyperparameter search (workbench.training.hpo_harness). Training-image only —