[test] Fix SFT tokenization tests for VLM changes

- _build_trainer: set is_vlm/processor so _load_and_tokenize (which now
  reads them) works on the bare object.__new__ trainer.
- vlm_processor fixture: importorskip torchvision so the VLM tests skip
  cleanly under the base dev extra instead of erroring (torchvision ships
  via the vllm-bearing extras used in CI).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: s-chundi

tests/train/test_sft_tokenization.py

@@ -30,6 +30,10 @@ def tokenizer():
 
 @pytest.fixture(scope="module")
 def vlm_processor():
+    # The Qwen3-VL processor pulls in a video processor that needs torchvision,
+    # which only ships with the vllm-bearing extras (skyrl-train/fsdp), not the
+    # base dev extra. Skip rather than error when it is unavailable.
+    pytest.importorskip("torchvision")
     proc = AutoProcessor.from_pretrained("Qwen/Qwen3-VL-2B-Instruct")
     if proc.tokenizer.pad_token is None:
         proc.tokenizer.pad_token = proc.tokenizer.eos_token
@@ -874,6 +878,8 @@ def _build_trainer(tokenizer, sft_cfg):
     trainer = object.__new__(SFTTrainer)
     trainer.sft_cfg = sft_cfg
     trainer.tokenizer = tokenizer
+    trainer.is_vlm = False
+    trainer.processor = None
     return trainer