Sync docker to dataset volume

docker/Dockerfile.l4

@@ -0,0 +1,77 @@
+# syntax=docker/dockerfile:1.7
+# ZOMBIE-WEIGHTS L4 image: two isolated uv/Python-3.12 environments sharing one HF cache.
+#
+#   /opt/venv_core       torch 2.13 (from base), transformers <5.22, trl <1.15,
+#                        transformer-lens==3.8.0, bnb, CuratorKIT, CircuitKIT,
+#                        SafeTune, AuditKIT — NO torchao
+#   /opt/venv_aligntune  aligntune 0.1.35, trl==1.7.1, torchao 0.18, torchvision 0.28
+#
+# Build:
+#   docker build -f docker/Dockerfile.l4 -t ghcr.io/divyanshu-kumar14/zombie-weights-l4:latest .
+FROM pytorch/pytorch:2.13.0-cuda12.6-cudnn9-runtime
+
+ENV DEBIAN_FRONTEND=noninteractive \
+    HF_HOME=/root/.cache/huggingface \
+    HF_HUB_ENABLE_HF_TRANSFER=1 \
+    PYTHONUNBUFFERED=1 \
+    TOKENIZERS_PARALLELISM=false \
+    ALIGNTUNE_ENABLE_VLLM=0 \
+    UV_LINK_MODE=copy
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        git curl ca-certificates build-essential \
+    && rm -rf /var/lib/apt/lists/*
+
+# uv (pinned standalone binary; avoids PEP 668 system-pip restrictions)
+COPY --from=ghcr.io/astral-sh/uv:0.12.17 /uv /usr/local/bin/uv
+
+# ---------------------------------------------------------------------------
+# Core environment (base image Python 3.12.3 + system torch 2.13.0)
+# uv creates the venv; pip resolves against system-site-packages so torch is
+# reused instead of downloading a second CUDA stack.
+# ---------------------------------------------------------------------------
+RUN uv venv --system-site-packages --seed /opt/venv_core
+RUN /opt/venv_core/bin/pip install --no-cache-dir \
+        "transformers>=5.15,<5.22" \
+        "trl>=0.12,<1.15" \
+        "peft>=0.18,<1" \
+        "huggingface-hub>=1.5,<2" \
+        "transformer-lens==3.8.0" \
+        "bitsandbytes>=0.41" \
+        "datasets>=3.0" \
+        "accelerate>=1.4" \
+        "sentencepiece" \
+        "protobuf" \
+        "scikit-learn" \
+        "hf_transfer" \
+        "curatorkit[hf,connectors]" \
+        "circuitkit==0.2.3" \
+        "safetune==0.1.18" \
+        "auditkit==1.1.11"
+
+# ---------------------------------------------------------------------------
+# AlignTune environment (isolated; torchao only lives here)
+# ---------------------------------------------------------------------------
+RUN uv venv --system-site-packages --seed /opt/venv_aligntune
+RUN /opt/venv_aligntune/bin/pip install --no-cache-dir \
+        "torchvision==0.28.0" \
+        "transformers>=5.15,<5.22" \
+        "peft>=0.18,<1" \
+        "datasets>=4.7.0" \
+        "accelerate>=1.6" \
+        "trl==1.7.1" \
+        "torchao==0.18.0" \
+        "aligntune==0.1.35"
+
+# ---------------------------------------------------------------------------
+# Project code + hard import gates
+# ---------------------------------------------------------------------------
+COPY docker/verify_imports.py /opt/zombie-weights/docker/verify_imports.py
+COPY src/ /opt/zombie-weights/src/
+COPY scripts/ /opt/zombie-weights/scripts/
+
+RUN /opt/venv_core/bin/python /opt/zombie-weights/docker/verify_imports.py core \
+ && /opt/venv_aligntune/bin/python /opt/zombie-weights/docker/verify_imports.py aligntune
+
+WORKDIR /opt/zombie-weights
+CMD ["bash"]

docker/verify_imports.py

@@ -21,6 +21,7 @@ def _version(mod: str) -> str:
 
 
 def check_core() -> dict:
+    import circuitkit  # must be imported before transformer_lens to install Cohere/Gemma patches
     import torch
     import transformers
     import trl
@@ -28,7 +29,6 @@ def check_core() -> dict:
     import huggingface_hub
     import transformer_lens
     import curatorkit
-    import circuitkit
     import safetune
     import auditkit