Deploy Qwen3-Coder-Next uncensored fine-tuner

Dockerfile

@@ -32,12 +32,16 @@ RUN pip install --no-cache-dir torch torchvision torchaudio --index-url https://
 # Install transformers from git main (qwen3_next model type not in any stable release yet)
 RUN pip install --no-cache-dir "git+https://github.com/huggingface/transformers.git@main"
 
-# Install flash-attention for faster training
-RUN pip install --no-cache-dir flash-attn --no-build-isolation 2>/dev/null || echo "Flash attention build failed, continuing without it"
-
-# Install causal-conv1d and flash-linear-attention for DeltaNet layers (optional but recommended)
-RUN pip install --no-cache-dir causal-conv1d 2>/dev/null || echo "causal-conv1d build failed, continuing without it"
-RUN pip install --no-cache-dir flash-linear-attention 2>/dev/null || echo "flash-linear-attention build failed, continuing without it"
+# Flash-attention, causal-conv1d, flash-linear-attention are optional perf
+# optimizations.  Building from source OOMs the HF Spaces builder (~16 GB RAM)
+# so we only attempt pre-built binary wheels (--only-binary :all:).
+# If no wheel exists for this CUDA/Python combo the install is skipped.
+RUN pip install --no-cache-dir --only-binary :all: flash-attn 2>/dev/null \
+    || echo "flash-attn: no pre-built wheel, skipping"
+RUN pip install --no-cache-dir --only-binary :all: causal-conv1d 2>/dev/null \
+    || echo "causal-conv1d: no pre-built wheel, skipping"
+RUN pip install --no-cache-dir --only-binary :all: flash-linear-attention 2>/dev/null \
+    || echo "flash-linear-attention: no pre-built wheel, skipping"
 
 # Create non-root user FIRST
 RUN useradd -m -u 1000 user