fix: install trl>=0.16 last with --upgrade to beat unsloth dep pins

unsloth can resolve trl to an older version when installed together,
leaving the conda base's old TRL (which imports GuidedDecodingParams
from vllm.sampling_params — removed in vllm 0.12). Installing TRL
separately at the end with --upgrade ensures 0.16+ wins.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

training/Dockerfile

@@ -4,11 +4,10 @@ RUN apt-get update && apt-get install -y git && rm -rf /var/lib/apt/lists/*
 
 RUN pip install --upgrade pip
 
-# Pin vLLM to version TRL supports, install unsloth
+# Install vllm, unsloth, and base deps first
 RUN pip install --no-cache-dir \
     "vllm==0.12.0" \
     unsloth \
-    "trl>=0.16.0" \
     transformers \
     accelerate \
     datasets \
@@ -16,6 +15,10 @@ RUN pip install --no-cache-dir \
     peft \
     wandb
 
+# Install TRL last with --upgrade so it is not pinned down by unsloth deps
+# trl>=0.16 dropped the GuidedDecodingParams import that broke on vllm 0.12
+RUN pip install --upgrade --no-cache-dir "trl>=0.16.0"
+
 WORKDIR /app
 COPY train.py .
 COPY train_rfc005.py .