Merge qwen35_export CI fixes

Merged latest main, resolved save.py/KTO test conflicts, fixed TRL/GRPO KTO drift

tests/saving/test_qwen3_5_vlm_full_finetune_key_remap.py

@@ -0,0 +1,71 @@
+import ast
+import types
+from pathlib import Path
+
+import torch
+
+
+def _load_qwen3_5_vlm_save_helpers():
+    source = Path(__file__).parents[2] / "unsloth" / "save.py"
+    tree = ast.parse(source.read_text(encoding = "utf-8"))
+    helpers = [
+        node
+        for node in tree.body
+        if isinstance(node, ast.FunctionDef)
+        and node.name
+        in {
+            "_is_qwen3_5_vlm",
+            "_qwen3_5_vlm_state_dict_for_save",
+        }
+    ]
+    module = ast.Module(body = helpers, type_ignores = [])
+    ast.fix_missing_locations(module)
+    namespace = {}
+    exec(compile(module, str(source), "exec"), namespace)
+    return namespace
+
+
+def _qwen3_5_vlm_model():
+    return types.SimpleNamespace(
+        config = types.SimpleNamespace(
+            architectures = ["Qwen3_5ForConditionalGeneration"],
+            model_type = "qwen3_5",
+            vision_config = types.SimpleNamespace(),
+        )
+    )
+
+
+def test_qwen3_5_vlm_state_dict_uses_hf_checkpoint_namespace():
+    helpers = _load_qwen3_5_vlm_save_helpers()
+    state_dict = {
+        "language_model.model.embed_tokens.weight": torch.ones(2, 2),
+        "language_model.model.layers.0.input_layernorm.weight": torch.ones(2),
+        "language_model.lm_head.weight": torch.ones(2, 2),
+        "visual.blocks.0.norm1.weight": torch.ones(2),
+        "other.weight": torch.ones(2),
+    }
+
+    remapped = helpers["_qwen3_5_vlm_state_dict_for_save"](state_dict)
+
+    assert "model.language_model.embed_tokens.weight" in remapped
+    assert "model.language_model.layers.0.input_layernorm.weight" in remapped
+    assert "lm_head.weight" in remapped
+    assert "model.visual.blocks.0.norm1.weight" in remapped
+    assert "other.weight" in remapped
+    assert "language_model.model.embed_tokens.weight" not in remapped
+    assert "language_model.lm_head.weight" not in remapped
+    assert "visual.blocks.0.norm1.weight" not in remapped
+
+
+def test_qwen3_5_vlm_detection_requires_vision_config():
+    helpers = _load_qwen3_5_vlm_save_helpers()
+    assert helpers["_is_qwen3_5_vlm"](_qwen3_5_vlm_model())
+
+    model = types.SimpleNamespace(
+        config = types.SimpleNamespace(
+            architectures = ["Qwen3_5ForCausalLM"],
+            model_type = "qwen3_5_text",
+        )
+    )
+
+    assert not helpers["_is_qwen3_5_vlm"](model)

tests/version_compat/test_trl_grpo_pinned_symbols.py

@@ -560,6 +560,7 @@ def test_trl_kto_get_batch_logps_signature(tag: str):
         if src is None:
             continue
         checked_sources.append((path, src))
+        # Legacy: explicit get_batch_logps method.
         if has_def(src, "get_batch_logps", "func"):
             return
         # TRL 1.x: refactored into _compute_logps + selective_log_softmax.

unsloth/save.py

@@ -457,6 +457,36 @@ def _preserve_tokenizer_eos_token(
         )
 
 
+def _is_qwen3_5_vlm(model):
+    config = getattr(model, "config", None)
+    if config is None or not hasattr(config, "vision_config"):
+        return False
+    architectures = getattr(config, "architectures", None) or ()
+    return any(
+        architecture
+        in (
+            "Qwen3_5ForConditionalGeneration",
+            "Qwen3_5MoeForConditionalGeneration",
+        )
+        for architecture in architectures
+    ) or getattr(config, "model_type", None) in ("qwen3_5", "qwen3_5_moe")
+
+
+def _qwen3_5_vlm_state_dict_for_save(state_dict):
+    remapped_state_dict = {}
+    for key, value in state_dict.items():
+        if key.startswith("language_model.model."):
+            new_key = "model.language_model." + key[len("language_model.model.") :]
+        elif key.startswith("visual."):
+            new_key = "model.visual." + key[len("visual.") :]
+        elif key.startswith("language_model.lm_head."):
+            new_key = "lm_head." + key[len("language_model.lm_head.") :]
+        else:
+            new_key = key
+        remapped_state_dict[new_key] = value
+    return remapped_state_dict
+
+
 @torch.inference_mode
 def unsloth_save_model(
     model,
@@ -2983,18 +3013,24 @@ def unsloth_generic_save(
         if not is_main_process:
             return
 
-        # Honor merged_16bit by casting to the target dtype if needed
         _save_kwargs = dict(
             safe_serialization = safe_serialization,
             max_shard_size = max_shard_size,
             variant = variant,
         )
+        is_qwen3_5_vlm = _is_qwen3_5_vlm(model)
+        if ("16bit" in save_method or is_qwen3_5_vlm) and state_dict is None:
+            state_dict = model.state_dict()
         if "16bit" in save_method:
             _target_dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
-            _save_kwargs["state_dict"] = {
+            state_dict = {
                 k: v.to(dtype = _target_dtype) if v.is_floating_point() else v
-                for k, v in model.state_dict().items()
+                for k, v in state_dict.items()
             }
+        if is_qwen3_5_vlm:
+            state_dict = _qwen3_5_vlm_state_dict_for_save(state_dict)
+        if state_dict is not None:
+            _save_kwargs["state_dict"] = state_dict
 
         if push_to_hub:
             print(f"Unsloth: Pushing full fine-tuned model to '{save_directory}' ...")