fix: remove flash_attention_2, revert to Qwen3.5-4B

Qwen3.5-4B + FA2 trips linear_attention masking in transformers.
sdpa (default) works fine; sl confirmed same approach in their sweep.
Model reverted to Qwen3.5-4B to match sl baselines.

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

src/ws/data.py

@@ -544,7 +544,7 @@ def generate_pairs(cfg: DataCfg) -> Path:
     if tok.pad_token is None:
         tok.pad_token = tok.eos_token
     model = AutoModelForCausalLM.from_pretrained(
-        cfg.model_id, torch_dtype=torch.bfloat16, device_map="cuda", attn_implementation="flash_attention_2"
+        cfg.model_id, torch_dtype=torch.bfloat16, device_map="cuda"
     )
     model.eval()
 

src/ws/eval/tinymfv_airisk.py

@@ -62,7 +62,7 @@ FRAMES: dict[str, dict[str, str | float]] = {
 
 @dataclass
 class TinyMFVAiriskCfg:
-    model: str = "Qwen/Qwen3-4B"
+    model: str = "Qwen/Qwen3.5-4B"
     behavior: str = "authority"
     adapter: str = "delora"
     out: Path = Path("out")
@@ -520,7 +520,7 @@ def run_eval(cfg: TinyMFVAiriskCfg) -> tuple[pl.DataFrame, pl.DataFrame, pl.Data
     if tok.pad_token is None:
         tok.pad_token = tok.eos_token
     tok.padding_side = "left"
-    model = AutoModelForCausalLM.from_pretrained(cfg.model, torch_dtype=torch.bfloat16, device_map="cuda", attn_implementation="flash_attention_2")
+    model = AutoModelForCausalLM.from_pretrained(cfg.model, torch_dtype=torch.bfloat16, device_map="cuda")
     model.eval()
 
     vignettes = _load_vignettes(cfg.limit)

src/ws/kl_calibrate.py

@@ -323,7 +323,7 @@ def main(cfg: KLCalibrateCfg) -> None:
         tok.pad_token = tok.eos_token
     tok.padding_side = "left"
     model = AutoModelForCausalLM.from_pretrained(
-        cfg.model, torch_dtype=torch.bfloat16, device_map="cuda", attn_implementation="flash_attention_2"
+        cfg.model, torch_dtype=torch.bfloat16, device_map="cuda"
     )
     model.eval()
 

src/ws/scripts/eval_tinymfv_calibrated.py

@@ -29,7 +29,7 @@ class EvalTinymfvCalibratedCfg:
     behavior: str = "authority"
     out: Path = Path("out")
     adapters: tuple[str, ...] = ("lora", "dora", "pissa", "delora", "oft", "ia3")
-    model: str = "Qwen/Qwen3-4B"
+    model: str = "Qwen/Qwen3.5-4B"
     bootstrap_samples: int = 256
     limit: int = 0
     batch_size: int = 16

src/ws/scripts/readme_tinymfv_table.py

@@ -103,7 +103,7 @@ def _foundation_short(behavior: str) -> dict[str, str]:
 @dataclass
 class ReadmeTinymfvCfg:
     behavior: str = "auth_care"
-    model_label: str = "Qwen3-4B"
+    model_label: str = "Qwen3.5-4B"
     out: Path = Path("out")
     adapters: tuple[str, ...] = ("lora", "dora", "pissa", "delora", "oft", "ia3")
     include_prompt_baseline: bool = True

src/ws/train.py

@@ -156,7 +156,7 @@ def train_adapter(cfg: TrainCfg, ds: Dataset) -> Path:
         tok.pad_token = tok.eos_token
 
     model = AutoModelForCausalLM.from_pretrained(
-        cfg.model_id, torch_dtype=torch.bfloat16, device_map="cuda", attn_implementation="flash_attention_2"
+        cfg.model_id, torch_dtype=torch.bfloat16, device_map="cuda"
     )
     model.config.use_cache = False