fix: resolve audited dispatch, kernel, and rollout bugs

- re-register the linear family with the operator dispatcher (ASTR_OPS / op_backend / resolve)
- fix bf16 gemv misaligned-address faults and element mispairing for offset weights
- reject misaligned bf16_swiglu inputs with a clear error and fall back in the backend gate
- make the rollout reuse decision, validation, and return atomic under one policy snapshot
- add the documented post-scoring rollout version check
- derive live+1 under the scheduler lock in optimizer_step via apply_weight_update(None, ...)
- reject rollout_max_policy_lag below rollout_interval - 1 at config time
- sync gemv stream-test inputs before switching streams; drop dead loader imports

tests/extension/test_gemv.py

@@ -130,6 +130,7 @@ def test_bf16_gemv_small_batch_fuses_bias():
 def test_bf16_gemv_uses_current_stream():
     x = torch.randn(1536, device="cuda", dtype=torch.bfloat16)
     weight = torch.randn(256, 1536, device="cuda", dtype=torch.bfloat16)
+    torch.cuda.synchronize()
     stream = torch.cuda.Stream()
     with torch.cuda.stream(stream):
         actual = bf16_gemv(x, weight)
@@ -171,6 +172,43 @@ def test_bf16_gemv_handles_unaligned_k(n, k):
     torch.testing.assert_close(actual3, F.linear(x3, weight), rtol=0.02, atol=0.5)
 
 
+@skip_no_gemv
+@pytest.mark.parametrize("m", [1, 2, 3, 4])
+def test_bf16_gemv_accepts_complementary_misalignment(m):
+    """Misaligned weight rows plus an x base chosen so the vectorized branch
+    is entered with a non-16B-aligned ``x`` pointer (regression: the branch
+    guard checked ``x + whead`` alignment but the uint4 view was rooted at
+    ``x`` itself, faulting with a misaligned-address CUDA error)."""
+    torch.manual_seed(37)
+    n, k = 256, 1536
+    # offset 5 elements = +10 bytes: weight rows land at 10 % 16 (whead=3)
+    # and x at 10 % 16, so (x + 2*whead) % 16 == 0 selects the fast path.
+    big_w = torch.randn(n * k + 8, device="cuda", dtype=torch.bfloat16)
+    weight = big_w[5 : 5 + n * k].view(n, k)
+    big_x = torch.randn(m * k + 8, device="cuda", dtype=torch.bfloat16)
+    x = big_x[5 : 5 + m * k].view(m, k) if m > 1 else big_x[5 : 5 + k]
+    assert (x.data_ptr() & 15) == 10 and (weight.data_ptr() & 15) == 10
+
+    actual = bf16_gemv(x, weight)
+    expected = F.linear(x, weight)
+    torch.testing.assert_close(actual, expected, rtol=0.02, atol=0.5 if m > 1 else 0.25)
+
+
+@skip_no_gemv
+def test_bf16_gemv_scalar_path_handles_misaligned_weight_only():
+    """Weight rows misaligned while x stays 16B-aligned take the scalar-x
+    middle and must stay exact."""
+    torch.manual_seed(41)
+    n, k = 256, 1536
+    big_w = torch.randn(n * k + 8, device="cuda", dtype=torch.bfloat16)
+    weight = big_w[5 : 5 + n * k].view(n, k)
+    x = torch.randn(2, k, device="cuda", dtype=torch.bfloat16)
+    assert (weight.data_ptr() & 15) == 10 and (x.data_ptr() & 15) == 0
+
+    actual = bf16_gemv(x, weight)
+    torch.testing.assert_close(actual, F.linear(x, weight), rtol=0.02, atol=0.5)
+
+
 @skip_no_gemv
 def test_bf16_gemv_small_batch_cuda_graph_replay():
     torch.manual_seed(31)

tests/extension/test_linear_dispatch.py

@@ -7,6 +7,7 @@ import torch.nn.functional as F
 
 from astrai.extension import is_available, linear
 from astrai.extension.backend import linear as public_linear
+from astrai.extension.dispatch import explain, op_backend, resolve
 
 # The package attribute ``linear`` is the dispatched function; reach the
 # module object explicitly for monkeypatching its private helpers.
@@ -166,3 +167,60 @@ def test_dispatched_linear_cuda_graph_replay(monkeypatch):
         graph.replay()
         expected = F.linear(x, weight)
     torch.testing.assert_close(actual, expected, rtol=0.02, atol=0.25)
+
+
+def test_linear_family_is_registered_with_shared_dispatcher():
+    from astrai.extension.dispatch import _FAMILIES
+
+    assert "linear" in _FAMILIES
+    x = torch.randn(2, 8)
+    weight = torch.randn(4, 8)
+    resolution = resolve("linear", x, weight)
+    assert resolution.record.family == "linear"
+    assert resolution.origin in ("chain", "fallback")
+    assert "linear" in explain("linear", x, weight)
+
+
+@skip_no_gemv
+def test_ops_env_override_forces_torch_for_capable_call(monkeypatch):
+    """ASTR_OPS=linear=torch must keep working after the M-band rewrite
+    (regression: the family was silently dropped from the dispatcher, so
+    the override warned, fell through, and the gemv kernel still ran)."""
+    monkeypatch.setenv("ASTR_OPS", "linear=torch")
+    x = torch.randn(2, 1536, device="cuda", dtype=torch.bfloat16)
+    weight = torch.randn(1536, 1536, device="cuda", dtype=torch.bfloat16)
+    with torch.no_grad():
+        assert not _routes_to_gemv(monkeypatch, x, weight)
+        torch.testing.assert_close(
+            linear(x, weight), F.linear(x, weight), rtol=0.02, atol=0.25
+        )
+
+
+@skip_no_gemv
+def test_ops_env_override_forces_gemv(monkeypatch):
+    monkeypatch.setenv("ASTR_OPS", "linear=gemv")
+    x = torch.randn(1, 1536, device="cuda", dtype=torch.bfloat16)
+    weight = torch.randn(256, 1536, device="cuda", dtype=torch.bfloat16)
+    with torch.no_grad():
+        # M=1 is outside the auto band but inside the forced gemv record.
+        assert _routes_to_gemv(monkeypatch, x, weight)
+
+
+@skip_no_gemv
+def test_op_backend_context_selects_torch(monkeypatch):
+    monkeypatch.setenv("ASTRAI_GEMV", "1")
+    x = torch.randn(2, 1536, device="cuda", dtype=torch.bfloat16)
+    weight = torch.randn(256, 1536, device="cuda", dtype=torch.bfloat16)
+    with torch.no_grad(), op_backend(linear="torch"):
+        assert not _routes_to_gemv(monkeypatch, x, weight)
+        torch.testing.assert_close(
+            linear(x, weight), F.linear(x, weight), rtol=0.02, atol=0.25
+        )
+    # The override is scoped: the forced mode applies again afterwards.
+    with torch.no_grad():
+        assert _routes_to_gemv(monkeypatch, x, weight)
+
+
+def test_op_backend_rejects_unknown_linear_handle():
+    with pytest.raises(ValueError, match="Unknown linear implementation"):
+        op_backend(linear="nonexistent").__enter__()

tests/extension/test_swiglu.py

@@ -97,3 +97,16 @@ def test_bf16_swiglu_uses_current_stream_and_cuda_graph():
 def test_bf16_swiglu_rejects_unsupported_inputs(make_args, error):
     with pytest.raises(RuntimeError, match=error):
         bf16_swiglu(*make_args())
+
+
+@skip_no_swiglu
+def test_bf16_swiglu_rejects_misaligned_storage_with_clear_error():
+    """Contiguous-but-offset views must fail the wrapper's TORCH_CHECK with
+    an actionable message instead of a sticky CUDA misaligned-address error
+    (regression: the kernel casts x directly to uint4 without checking)."""
+    k = 1536
+    base = torch.randn(k + 1, device="cuda", dtype=torch.bfloat16)
+    x = base[1:]  # +2 bytes: contiguous but not 16B-aligned
+    weights = torch.randn(8, k, device="cuda", dtype=torch.bfloat16)
+    with pytest.raises(RuntimeError, match="16-byte aligned"):
+        bf16_swiglu(x, weights, weights)

tests/extension/test_swiglu_dispatch.py

@@ -95,3 +95,22 @@ def test_auto_uses_unfused_chain_until_shape_is_qualified(monkeypatch):
         actual = swiglu(x, up_weight, gate_weight)
         expected = reference_swiglu(x, up_weight, gate_weight)
     torch.testing.assert_close(actual, expected, rtol=0.03, atol=0.1)
+
+
+@skip_no_swiglu
+def test_mode_one_falls_back_for_misaligned_storage(monkeypatch):
+    """Contiguous-but-offset views must route to the unfused torch chain
+    even with ASTRAI_SWIGLU=1 instead of reaching the uint4-only kernel
+    (regression: the fused primitive faulted with a misaligned-address
+    CUDA error for such inputs)."""
+    monkeypatch.setenv("ASTRAI_SWIGLU", "1")
+    k = 1536
+    x_base = torch.randn(2 * k + 8, device="cuda", dtype=torch.bfloat16) * 0.1
+    x = x_base[1 : 1 + 2 * k].view(2, k)
+    assert x.is_contiguous() and (x.data_ptr() & 15) != 0
+    up_weight = torch.randn(64, k, device="cuda", dtype=torch.bfloat16) * 0.02
+    gate_weight = torch.randn_like(up_weight)
+    with torch.no_grad():
+        actual = swiglu(x, up_weight, gate_weight)
+        expected = reference_swiglu(x, up_weight, gate_weight)
+    torch.testing.assert_close(actual, expected, rtol=0.03, atol=0.01)

tests/inference/test_scheduler.py

@@ -583,6 +583,32 @@ def test_scheduler_applies_weight_mutation_and_version_atomically(device):
         with pytest.raises(RuntimeError, match="optimizer failed"):
             scheduler.apply_weight_update(2, failed_mutation)
         assert scheduler.policy_version == 1
+
+        # None derives live+1 under the lock: no read-compute-write race
+        # on the current version for advance-by-one callers.
+        assert scheduler.apply_weight_update(None, mutate) == "updated"
+        assert scheduler.policy_version == 2
+    finally:
+        scheduler.stop()
+
+
+def test_scheduler_atomic_advance_survives_interleaved_publish(device):
+    """A concurrent publish between reading the live version and applying
+    the update must not fail ``require_advance`` (regression: callers
+    computed live+1 outside the lock, a TOCTOU that raised spuriously)."""
+    scheduler, _tok, _model = _make_real_scheduler(device)
+
+    try:
+        # Simulate the race directly: a version read that goes stale before
+        # apply_weight_update acquires the lock. With None the scheduler
+        # re-derives live+1 inside the critical section.
+        stale_read = scheduler.policy_version + 1
+        scheduler.update_weights(1)
+        assert stale_read == 1  # now equals live -> explicit form would raise
+        with pytest.raises(ValueError, match="must advance"):
+            scheduler.apply_weight_update(stale_read, lambda: "ok")
+        assert scheduler.apply_weight_update(None, lambda: "ok") == "ok"
+        assert scheduler.policy_version == 2
     finally:
         scheduler.stop()
 

tests/trainer/test_online_e2e.py

@@ -144,3 +144,38 @@ def test_online_rollout_end_to_end(
     checkpoint = Checkpoint.load(checkpoint_dir)
     assert checkpoint.meta["policy_version"] == 2
     assert len(created_reference_models) == 1
+
+
+def _minimal_online_config(**overrides):
+    """A TrainConfig for online GRPO that only needs field overrides."""
+    defaults = dict(
+        strategy="online_grpo",
+        model_fn=lambda: torch.nn.Linear(2, 2),
+        dataset=InstructionDataset(),
+        optimizer_fn=lambda m: torch.optim.SGD(m.parameters(), lr=0.0),
+        scheduler_fn=lambda o: SchedulerFactory.create(
+            "cosine", o, warmup_steps=1, lr_decay_steps=4, min_rate=0.05
+        ),
+        reward_model_fn=LengthRewardModel,
+    )
+    defaults.update(overrides)
+    return TrainConfig(**defaults)
+
+
+def test_online_config_rejects_contradictory_policy_lag():
+    """rollout_max_policy_lag below rollout_interval - 1 guarantees a fatal
+    RolloutVersionError mid-training; it must fail at config time instead."""
+    with pytest.raises(ValueError, match="rollout_max_policy_lag=0"):
+        _minimal_online_config(rollout_interval=3, rollout_max_policy_lag=0)
+
+    # lag == interval - 1 (including the derived default) stays valid.
+    config = _minimal_online_config(rollout_interval=3, rollout_max_policy_lag=2)
+    assert config.rollout_max_policy_lag == 2
+    config = _minimal_online_config(rollout_interval=3)
+    assert config.rollout_max_policy_lag is None
+
+    # Offline strategies never consult the rollout window.
+    config = _minimal_online_config(
+        strategy="sft", rollout_interval=3, rollout_max_policy_lag=0
+    )
+    assert config.rollout_max_policy_lag == 0

tests/trainer/test_online_strategy.py

@@ -62,6 +62,9 @@ class _RecordingRunner:
 
     def apply_weight_update(self, policy_version, update):
         result = update()
+        if policy_version is None:
+            # Mirror the scheduler: None derives live+1 under the lock.
+            policy_version = self.policy_version + 1
         self.update_weights(policy_version)
         return result
 

tests/trainer/test_rollout.py

@@ -459,7 +459,7 @@ def test_rollout_runner_publishes_cache_before_concurrent_policy_update(device):
         nonlocal validation_calls
         validation_calls += 1
         original_validate(result, live_version=live_version)
-        if validation_calls == 2:
+        if validation_calls == 3:
             final_validation_started.set()
             assert allow_final_validation_to_finish.wait(timeout=5)
 
@@ -503,6 +503,58 @@ def test_rollout_runner_derives_default_policy_lag_from_interval(device):
     assert runner.max_policy_lag == 3
 
 
+def _interleave_before_snapshot(runner, callback):
+    """Wrap ``with_policy_snapshot`` so ``callback`` runs just before a
+    named snapshot callback enters the generator/scheduler locks."""
+    original_snapshot = runner.generator.with_policy_snapshot
+
+    def wrapper(inspect):
+        if inspect.__name__ == "reuse":
+            callback()
+        return original_snapshot(inspect)
+
+    runner.generator.with_policy_snapshot = wrapper
+
+
+def test_rollout_runner_reuse_reads_cache_inside_the_snapshot(device):
+    """The reuse decision must observe the cache under the policy snapshot
+    (regression: the cache was read outside the lock, so a concurrent
+    commit between the read and the lock silently handed the trainer a
+    stale rollout — a lost update)."""
+    import dataclasses
+
+    runner, _ = _make_runner(device, rollout_interval=100)
+    batch = _make_instruction_batch(n=1)
+    first, _ = runner(batch)
+    assert first.policy_version == 0
+
+    def concurrent_refresh():
+        runner.update_weights(1)
+        runner._cache = dataclasses.replace(first, policy_version=1)
+        runner._steps_since_rollout = 0
+
+    _interleave_before_snapshot(runner, concurrent_refresh)
+    result, fresh = runner(batch)
+    assert fresh is False
+    assert result is not first
+    assert result.policy_version == 1
+
+
+def test_rollout_runner_recovers_when_cache_cleared_before_reuse_snapshot(device):
+    """A cache clear between the reuse decision and the snapshot must
+    trigger a fresh rollout instead of an assertion failure (regression:
+    ``assert cached is not None`` fired because the object was captured
+    outside the lock)."""
+    runner, _ = _make_runner(device, rollout_interval=100)
+    batch = _make_instruction_batch(n=1)
+    first, _ = runner(batch)
+
+    _interleave_before_snapshot(runner, runner.clear_cache)
+    result, fresh = runner(batch)
+    assert fresh is True
+    assert result is not first
+
+
 @pytest.mark.parametrize(
     ("kwargs", "message"),
     [