deps: bnb>=0.48.1 for vllm 0.18 4-bit gate

scripts/job_train.sh

@@ -27,10 +27,14 @@ pip install -q hydra-core omegaconf wandb 'huggingface_hub>=0.28' tyro
 # Core ML deps.
 # NOTE: trl 1.2 imports `is_trackio_available` from transformers, which is not
 # present in 4.49.0. Use a newer transformers line in jobs.
-# bitsandbytes 0.45.1 only ships up to cuda124 binary; vllm 0.18.0 pulls torch
-# with the cu128 wheel, so we MUST use a bnb release that ships libbitsandbytes_cuda128.so
-# (0.47.0 is the first stable line that does). Confirmed against PyPI Apr 2026.
-pip install -q transformers==4.56.2 accelerate==1.5.2 peft==0.14.0 'bitsandbytes>=0.47.0,<0.48'
+# bitsandbytes pin notes:
+#  - 0.45.1: ships cu124 only — breaks under vllm's cu128 torch.
+#  - 0.47.x: cu128 + cu124 dual-wheel, but vllm's BitsAndBytesLinearMethod
+#    gates on `bnb >= 0.48.1` (vllm/model_executor/layers/quantization/
+#    bitsandbytes.py:_check_bitsandbytes_version) and raises ImportError
+#    during `LLM(...)` init when used with a 4-bit model.
+#  - >=0.48.1: satisfies both vllm's gate and our cu128 wheel needs.
+pip install -q transformers==4.56.2 accelerate==1.5.2 peft==0.14.0 'bitsandbytes>=0.48.1'
 pip install -q datasets==4.7.0
 # Keep TRL pinned for OpenEnv path but bypass resolver deadlock with datasets pin.
 pip install -q --no-deps trl==1.2.0