Align packaged GRPO model dtype with trainer precision

.agents/skills/fine-tuning/reference/grpo-training.md

@@ -198,6 +198,14 @@ end-to-end provider deadline or public cancellation qualification. The first
 executable candidate requires identical model and tokenizer revisions because
 both consume one verified local snapshot.
 
+The packaged runtime resolves `GRPOConfig` precision before loading the model.
+When the retained model dtype is `null`, it passes the effective FP16 or BF16
+dtype to Unsloth's loader. Pinned Unsloth checks that loaded dtype against
+trainer precision during `GRPOTrainer` construction; allowing the loader to
+auto-select BF16 on BF16-capable hardware while the recipe requests FP16 can
+fail at `TRAINER_SETUP`. A closed `TypeError` marker alone cannot establish
+that this was the historical cause, and it does not authorize replay.
+
 GRPO has a separate fixed 131-member offline inventory and 13-member outer
 closure. Existing SFT inventories remain unchanged. Maintenance uses the
 checked-in command, offline first because the outer closure binds its bytes:

.claude/skills/fine-tuning/reference/grpo-training.md

@@ -198,6 +198,14 @@ end-to-end provider deadline or public cancellation qualification. The first
 executable candidate requires identical model and tokenizer revisions because
 both consume one verified local snapshot.
 
+The packaged runtime resolves `GRPOConfig` precision before loading the model.
+When the retained model dtype is `null`, it passes the effective FP16 or BF16
+dtype to Unsloth's loader. Pinned Unsloth checks that loaded dtype against
+trainer precision during `GRPOTrainer` construction; allowing the loader to
+auto-select BF16 on BF16-capable hardware while the recipe requests FP16 can
+fail at `TRAINER_SETUP`. A closed `TypeError` marker alone cannot establish
+that this was the historical cause, and it does not authorize replay.
+
 GRPO has a separate fixed 131-member offline inventory and 13-member outer
 closure. Existing SFT inventories remain unchanged. Maintenance uses the
 checked-in command, offline first because the outer closure binds its bytes:

.skills/fine-tuning/reference/grpo-training.md

@@ -198,6 +198,14 @@ end-to-end provider deadline or public cancellation qualification. The first
 executable candidate requires identical model and tokenizer revisions because
 both consume one verified local snapshot.
 
+The packaged runtime resolves `GRPOConfig` precision before loading the model.
+When the retained model dtype is `null`, it passes the effective FP16 or BF16
+dtype to Unsloth's loader. Pinned Unsloth checks that loaded dtype against
+trainer precision during `GRPOTrainer` construction; allowing the loader to
+auto-select BF16 on BF16-capable hardware while the recipe requests FP16 can
+fail at `TRAINER_SETUP`. A closed `TypeError` marker alone cannot establish
+that this was the historical cause, and it does not authorize replay.
+
 GRPO has a separate fixed 131-member offline inventory and 13-member outer
 closure. Existing SFT inventories remain unchanged. Maintenance uses the
 checked-in command, offline first because the outer closure binds its bytes:

Trainers/grpo/runtime_v1.py

@@ -468,10 +468,16 @@ def mark(value):
     args = build_grpo_config(config, output_root / 'checkpoints',
                            config_class=deps['config_class'], bf16_supported=deps['bf16_supported'])
     helper = deps['model_loader']
+    # Unsloth's patched GRPOTrainer checks the loaded model dtype against the
+    # effective GRPOConfig precision. Its loader otherwise chooses BF16 on a
+    # BF16-capable GPU even when the retained recipe requested FP16 training.
+    model_dtype = config['model'].get('dtype')
+    if model_dtype is None:
+        model_dtype = 'bfloat16' if args.bf16 else 'float16' if args.fp16 else None
     mark("MODEL_LIBRARY_LOAD")
     model, tokenizer, is_vl = helper.load_model_and_tokenizer(
         str(model_snapshot), max_seq_length=config['model']['max_seq_length'],
-        dtype=config['model'].get('dtype'), load_in_4bit=config['model'].get('load_in_4bit', True),
+        dtype=model_dtype, load_in_4bit=config['model'].get('load_in_4bit', True),
         offline=True)
     if is_vl:
         raise ValueError("retained GRPO requires text model")

tests/trainers/grpo/test_runtime_v1.py

@@ -102,6 +102,38 @@ def test_real_flow_offline_rewards_metrics_and_mandatory_roles(tmp_path, monkeyp
     assert 'parameter_updated' not in result['training_lineage'].read_text()
 
 
+@pytest.mark.parametrize(('fp16', 'bf16', 'supported', 'expected'), [
+    (True, False, True, 'float16'),
+    (False, True, True, 'bfloat16'),
+    (False, True, False, 'float16'),
+])
+def test_model_load_precision_matches_effective_trainer_precision(
+        tmp_path, fp16, bf16, supported, expected):
+    values = inputs(tmp_path)
+    values['config']['training'].update(fp16=fp16, bf16=bf16)
+    values['workload']['config_sha256'] = runtime._sha(runtime._canonical(values['config']))
+    events = []
+    deps = dependencies(events)()
+    deps['bf16_supported'] = supported
+    def load(*args, **kwargs):
+        events.append(('model', args, kwargs))
+        return SimpleNamespace(config=SimpleNamespace(dtype=kwargs['dtype'])), Tokenizer(), False
+    deps['model_loader'].load_model_and_tokenizer = load
+    trainer_base = deps['trainer_class']
+    class PrecisionCheckingTrainer(trainer_base):
+        def __init__(self, **kwargs):
+            dtype = kwargs['model'].config.dtype
+            precision = kwargs['args']
+            if (dtype == 'float16' and precision.bf16) or (dtype != 'float16' and precision.fp16):
+                raise TypeError('incompatible model and trainer precision')
+            super().__init__(**kwargs)
+    deps['trainer_class'] = PrecisionCheckingTrainer
+    runtime.run_grpo(**values, _dependency_factory=lambda: deps)
+    assert events[0][2]['dtype'] == expected
+    assert (events[1][1]['args'].fp16, events[1][1]['args'].bf16) == (
+        expected == 'float16', expected == 'bfloat16')
+
+
 @pytest.mark.parametrize('change', ['dataset', 'reward', 'config', 'model', 'method', 'nonfinite', 'bool', 'custom', 'pivot', 'empty'])
 def test_rejects_before_dependency_effects(tmp_path, change):
     values = inputs(tmp_path)

tuner/runtime/manifests/packaged-grpo-worker-v1.json

@@ -1 +1 @@
-{"closure_digest":"56f676b31466079fb81fa402392eb9ce4634a686526946ad9280299a34780406","members":[{"path":"tuner/runtime/manifests/offline-grpo-worker-v1.json","sha256":"59b017179339d58ff24e52d77e85d7d8cf755f25fdf93217ce88f8f9129e32db","size_bytes":21344},{"path":"tuner/runtime/offline_grpo_worker.py","sha256":"aac831048855e959598a4dc45b29e614f06f8714df18e868fd6b069e3c765710","size_bytes":9323},{"path":"tuner/runtime/offline_sft_worker.py","sha256":"2304eeb00667954510c34cc4aacaa818532a301fbeabf04c2fb165d6cd1d70b5","size_bytes":36343},{"path":"tuner/runtime/packaged_grpo_child.py","sha256":"124d5b2a969d38ac2e488d58a216759108f1a55dfcc3dcaf1012791efe792e4b","size_bytes":8273},{"path":"tuner/runtime/packaged_grpo_execution.py","sha256":"a77c7d6b97ad2ded3b9631174e479e60462f1894513f90fd2fc7798828c6867d","size_bytes":14943},{"path":"tuner/runtime/packaged_grpo_verification.py","sha256":"56564c6ee6414a3a71ce7c9ef330152d812fede2a711a2329463a0b81a61b0d6","size_bytes":10319},{"path":"tuner/runtime/packaged_grpo_worker_closure.py","sha256":"81da8db5a61e9a874234d1e50c529e1056282fd0eef133b2720eb0f34f78cddc","size_bytes":3735},{"path":"tuner/runtime/packaged_sft_child.py","sha256":"ab41eb7dca28b440c043f589d297dc0d9f718ed14f6cb36c96d7bd1b7bfcdc14","size_bytes":23556},{"path":"tuner/runtime/packaged_sft_execution.py","sha256":"cab2b5032d16146422398305120f52c5a124287ecf929b640a1f17d0ea2ed648","size_bytes":80053},{"path":"tuner/runtime/packaged_training_worker.py","sha256":"9cfe13dc9f0124b4da01d05e92cf3c2c9249d075c3a508d6b93b0a7c7f0e109f","size_bytes":34986},{"path":"tuner/runtime/packaged_worker_closure.py","sha256":"e457937426e19747e867a314c1a2556ee64bd5cbc19359f535253b02403d1443","size_bytes":8705},{"path":"tuner/runtime/releases.py","sha256":"c0aa7ba37e8d35287fdd480cf0c76053a554e546c40c8bf36f5cf05d1a96d22b","size_bytes":52535},{"path":"tuner/training/packaged_grpo_compilation.py","sha256":"d186a05ffb8cc756e782b2bb9a3f66e925779104b70045bd56a5592846fec524","size_bytes":8722}],"schema_version":"synaptic-packaged-grpo-worker-closure/v1"}
+{"closure_digest":"e48763871e7e9bc50c0c2b218bbfcb885b72b56f2ae1c2c41b1e16938c198ffe","members":[{"path":"tuner/runtime/manifests/offline-grpo-worker-v1.json","sha256":"7159247a2d15faaf2654b14168c1f90359b09950f646d0e9b2e2ec683f20617b","size_bytes":21344},{"path":"tuner/runtime/offline_grpo_worker.py","sha256":"aac831048855e959598a4dc45b29e614f06f8714df18e868fd6b069e3c765710","size_bytes":9323},{"path":"tuner/runtime/offline_sft_worker.py","sha256":"2304eeb00667954510c34cc4aacaa818532a301fbeabf04c2fb165d6cd1d70b5","size_bytes":36343},{"path":"tuner/runtime/packaged_grpo_child.py","sha256":"124d5b2a969d38ac2e488d58a216759108f1a55dfcc3dcaf1012791efe792e4b","size_bytes":8273},{"path":"tuner/runtime/packaged_grpo_execution.py","sha256":"a77c7d6b97ad2ded3b9631174e479e60462f1894513f90fd2fc7798828c6867d","size_bytes":14943},{"path":"tuner/runtime/packaged_grpo_verification.py","sha256":"56564c6ee6414a3a71ce7c9ef330152d812fede2a711a2329463a0b81a61b0d6","size_bytes":10319},{"path":"tuner/runtime/packaged_grpo_worker_closure.py","sha256":"81da8db5a61e9a874234d1e50c529e1056282fd0eef133b2720eb0f34f78cddc","size_bytes":3735},{"path":"tuner/runtime/packaged_sft_child.py","sha256":"ab41eb7dca28b440c043f589d297dc0d9f718ed14f6cb36c96d7bd1b7bfcdc14","size_bytes":23556},{"path":"tuner/runtime/packaged_sft_execution.py","sha256":"cab2b5032d16146422398305120f52c5a124287ecf929b640a1f17d0ea2ed648","size_bytes":80053},{"path":"tuner/runtime/packaged_training_worker.py","sha256":"9cfe13dc9f0124b4da01d05e92cf3c2c9249d075c3a508d6b93b0a7c7f0e109f","size_bytes":34986},{"path":"tuner/runtime/packaged_worker_closure.py","sha256":"e457937426e19747e867a314c1a2556ee64bd5cbc19359f535253b02403d1443","size_bytes":8705},{"path":"tuner/runtime/releases.py","sha256":"c0aa7ba37e8d35287fdd480cf0c76053a554e546c40c8bf36f5cf05d1a96d22b","size_bytes":52535},{"path":"tuner/training/packaged_grpo_compilation.py","sha256":"d186a05ffb8cc756e782b2bb9a3f66e925779104b70045bd56a5592846fec524","size_bytes":8722}],"schema_version":"synaptic-packaged-grpo-worker-closure/v1"}