deploy v5 (long mode, a100x4)

Dockerfile

@@ -14,15 +14,13 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
 WORKDIR /app
 
 # ── Step 1: PyTorch 2.7.0 + CUDA 12.4 ─────────────────────────────────────
-# Minimum version that satisfies ALL torchao requirements:
-#   torch.int1                              added in 2.6.0
-#   torch.utils._pytree.register_constant  added in 2.7.0  ← the blocker
-# torchvision 0.22.0 is the matching release for torch 2.7.0.
+# torch.int1 added in 2.6.0; torch.utils._pytree.register_constant in 2.7.0.
+# torchvision intentionally omitted — not needed for LLM text training and its
+# cu124 wheel for 0.22.0 caused the Step-1 build failure in prior attempts.
+# --extra-index-url keeps PyPI accessible (--index-url replaces it entirely).
 RUN pip install --upgrade pip && \
-    pip install \
-        torch==2.7.0 \
-        torchvision==0.22.0 \
-        --index-url https://download.pytorch.org/whl/cu124
+    pip install torch==2.7.0 \
+        --extra-index-url https://download.pytorch.org/whl/cu124
 
 # ── Step 2: Unsloth alone ──────────────────────────────────────────────────
 # Own pip step so unsloth controls its transformers/peft/trl pins without