fix(tests): three environment-specific failures in the v0.72.4 suite

Down from 34 CI failures to 0 after the trl cap; these three are the remainder,
all in the new tests and all device- or version-specific rather than product
bugs:

- test_trl_itself_still_refuses_batch_one built a KTOConfig without bf16=False.
  Newer TRL configs default bf16 on, so a CPU-only runner raised "Your setup
  doesn't support bf16/gpu" before ever reaching the batch-size check the test
  exists to pin. Reproduced locally with CUDA_VISIBLE_DEVICES=-1.

- test_kto_streams_at_batch_two runs a real trainer.train(). That works on CUDA
  (dev box and CI) but fails CPU-only under torch 2.13/trl 0.24 with "Tensor on
  device cpu is not on the expected device meta!". Streaming exists to bound
  VRAM, so a streamed model on CPU is a test convenience rather than a
  configuration — the stance v0.72.3 already took for PEFT re-dispatch — so the
  full-step test is gated to the production device. KTO's schema gate, setup,
  reference behaviour and layer-read accounting are all still checked on CPU.

- test_a_reference_using_loss_reads_more_layers_than_sft[kto] hit MPS on the
  macOS runners; its sibling tests already carry the MPS skip and this one had
  been missed.

Verified: 9 passed with CUDA hidden, 67 passed on CUDA.

The published benchmarks record notes the CPU limitation, and that it was
invisible on a CUDA dev box — a locally-green suite is a weak signal whenever a
path forks on device.

benchmarks/gate-v0.72.4-preference-losses.md

@@ -276,6 +276,21 @@ All four saved adapters: **120 tensors, 0 keys carrying the streaming wrapper
 segment, 60/60 non-zero `lora_B`** — i.e. ordinary LoRA adapters that load into any
 non-streaming model.
 
+### One CPU-only limitation, found by CI rather than locally
+
+A full KTO **training step** over a streamed model runs on CUDA (verified here and
+on the dev box) but fails on a CPU-only runner under newer torch/TRL with
+`Tensor on device cpu is not on the expected device meta!`. Streaming exists to
+bound VRAM, so a streamed model on CPU is a test convenience rather than a real
+configuration — the same stance v0.72.3 took on PEFT's re-dispatch — and that test
+is therefore gated to the production device. Everything else about KTO (the schema
+gate, `setup()`, the reference behaviour, the layer-read accounting) is still
+exercised on CPU.
+
+Worth stating plainly: this was invisible on the dev box, which has CUDA. The
+locally-green suite is a weaker signal than it looks whenever a code path forks on
+device.
+
 Ten rejected configurations each named its own reason (rollout tasks, unsupported
 task, KTO at batch 1, unsloth backend, 8-bit, no adapter, DoRA, `batch_size: auto`,
 packing).

tests/test_v07204.py

@@ -962,6 +962,10 @@ class TestKtoBatchIsRefusedEarly:
             report_to=[],
             max_length=32,
             max_prompt_length=16,
+            # Newer TRL configs enable bf16 by default, which a CPU-only runner
+            # rejects with "Your setup doesn't support bf16/gpu" BEFORE reaching
+            # the batch-size check this test exists to pin.
+            bf16=False,
         )
         with pytest.raises(ValueError, match="batch size"):
             KTOTrainer(