mplang-nightly 0.1.dev146__tar.gz → 0.1.dev148__tar.gz

{mplang_nightly-0.1.dev146 → mplang_nightly-0.1.dev148}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mplang-nightly
-Version: 0.1.dev146
+Version: 0.1.dev148
 Summary: Multi-Party Programming Language
 Author-email: SecretFlow Team <secretflow-contact@service.alipay.com>
 License:                                  Apache License

{mplang_nightly-0.1.dev146 → mplang_nightly-0.1.dev148}/mplang/backend/base.py

@@ -12,20 +12,21 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-"""Backend kernel registry & per-participant runtime (explicit op->kernel binding).
+"""Backend kernel registry: mapping kernel_id -> implementation.
 
-This version decouples *kernel implementation registration* from *operation binding*.
+This module provides a lightweight registry for backend kernel implementations.
+It does not track or decide which kernel handles a given semantic operation;
+that policy (op -> kernel_id) is managed externally by each ``RuntimeContext``.
 
-Concepts:
-    * kernel_id: unique identifier of a concrete backend implementation.
-    * op_type: semantic operation name carried by ``PFunction.fn_type``.
-    * bind_op(op_type, kernel_id): performed by higher layer (see ``backend.context``)
-        to select which implementation handles an op. Runtime dispatch is now a 2-step:
-        pfunc.fn_type -> active kernel_id -> KernelSpec.fn
+Exposed primitives:
+* ``@kernel_def(kernel_id)``: decorator to register a kernel implementation.
+* ``get_kernel_spec(kernel_id)``: look up a previously registered kernel.
+* ``cur_kctx()`` / ``KernelContext``: execution context available only
+    inside a kernel body (rank, world_size, per-backend state pockets, and a
+    runtime-wide cache shared by kernels of the same runtime instance).
 
-The previous implicit "import == register+bind" coupling is removed. Kernel
-modules only call ``@kernel_def(kernel_id)``. Default bindings are established
-centrally (lazy) the first time a runtime executes a kernel.
+No global op binding table exists here; callers resolve an op to a kernel_id
+before invoking the kernel function.
 """
 
 from __future__ import annotations
@@ -38,12 +39,10 @@ from typing import Any
 __all__ = [
     "KernelContext",
     "KernelSpec",
-    "bind_op",
     "cur_kctx",
-    "get_kernel_for_op",
+    "get_kernel_spec",
+    "kernel_exists",
     "list_kernels",
-    "list_ops",
-    "unbind_op",
 ]
 
 
@@ -116,9 +115,6 @@ class KernelSpec:
 # All registered kernel implementations: kernel_id -> spec
 _KERNELS: dict[str, KernelSpec] = {}
 
-# Active op bindings: op_type -> kernel_id
-_BINDINGS: dict[str, str] = {}
-
 
 def kernel_def(kernel_id: str, /, **meta: Any) -> Callable[[KernelFn], KernelFn]:
     """Decorator to register a concrete kernel implementation.
@@ -136,34 +132,11 @@ def kernel_def(kernel_id: str, /, **meta: Any) -> Callable[[KernelFn], KernelFn]
     return _decorator
 
 
-def bind_op(op_type: str, kernel_id: str, *, force: bool = True) -> None:
-    """Bind an op_type to a registered kernel implementation.
-
-    Args:
-        op_type: Semantic operation name.
-        kernel_id: Previously registered kernel identifier.
-        force: If False and op_type already bound, keep existing binding.
-               If True (default), overwrite.
-    """
-    if kernel_id not in _KERNELS:
+def get_kernel_spec(kernel_id: str) -> KernelSpec:
+    """Return KernelSpec for a registered kernel_id (no op binding lookup)."""
+    spec = _KERNELS.get(kernel_id)
+    if spec is None:
         raise KeyError(f"kernel_id {kernel_id} not registered")
-    if not force and op_type in _BINDINGS:
-        return
-    _BINDINGS[op_type] = kernel_id
-
-
-def unbind_op(op_type: str) -> None:
-    _BINDINGS.pop(op_type, None)
-
-
-def get_kernel_for_op(op_type: str) -> KernelSpec:
-    kid = _BINDINGS.get(op_type)
-    if kid is None:
-        # Tests expect NotImplementedError for unsupported operations
-        raise NotImplementedError(f"no backend kernel registered for op {op_type}")
-    spec = _KERNELS.get(kid)
-    if spec is None:  # inconsistent state
-        raise RuntimeError(f"active kernel_id {kid} missing spec")
     return spec
 
 
@@ -171,5 +144,6 @@ def list_kernels() -> list[str]:
     return sorted(_KERNELS.keys())
 
 
-def list_ops() -> list[str]:
-    return sorted(_BINDINGS.keys())
+def kernel_exists(kernel_id: str) -> bool:
+    """Return True if a kernel_id has been registered."""
+    return kernel_id in _KERNELS

{mplang_nightly-0.1.dev146 → mplang_nightly-0.1.dev148}/mplang/backend/context.py

@@ -15,11 +15,10 @@
 from __future__ import annotations
 
 from collections.abc import Mapping
-from dataclasses import dataclass, field
 from typing import Any
 
 from mplang.backend import base
-from mplang.backend.base import KernelContext, bind_op, get_kernel_for_op
+from mplang.backend.base import KernelContext, get_kernel_spec, kernel_exists
 from mplang.core.dtype import UINT8, DType
 from mplang.core.pfunc import PFunction
 from mplang.core.table import TableLike, TableType
@@ -100,30 +99,57 @@ _DEFAULT_BINDINGS: dict[str, str] = {
 # --- RuntimeContext ---
 
 
-@dataclass
 class RuntimeContext:
-    rank: int
-    world_size: int
-    bindings: Mapping[str, str] | None = None  # optional overrides
-    state: dict[str, dict[str, Any]] = field(default_factory=dict)
-    cache: dict[str, Any] = field(default_factory=dict)
-    stats: dict[str, Any] = field(default_factory=dict)
-
-    def __post_init__(self) -> None:
+    """Per-runtime execution context with isolated op->kernel bindings.
+
+    Parameters
+    ----------
+    rank : int
+        Local rank of this participant.
+    world_size : int
+        Total number of participants.
+    initial_bindings : Mapping[str, str] | None, optional
+        Optional partial overrides applied on top of the default binding table
+        during construction (override semantics, not replace). After
+        initialization, all (re)binding must go through ``bind_op`` /
+        ``rebind_op``.
+    state / cache / stats : dict, optional
+        Mutable pockets reused across kernel invocations. If omitted, new
+        dictionaries are created.
+    """
+
+    __slots__ = ("_ibindings", "cache", "rank", "state", "stats", "world_size")
+
+    def __init__(
+        self,
+        rank: int,
+        world_size: int,
+        initial_bindings: Mapping[str, str] | None = None,
+        *,
+        state: dict[str, dict[str, Any]] | None = None,
+        cache: dict[str, Any] | None = None,
+        stats: dict[str, Any] | None = None,
+    ) -> None:
         _ensure_impl_imported()
-        if self.bindings is not None:
-            for op, kid in self.bindings.items():
-                bind_op(op, kid)
-        else:
-            for op, kid in _DEFAULT_BINDINGS.items():
-                bind_op(op, kid)
-        # Initialize stats pocket
+        self.rank = rank
+        self.world_size = world_size
+        # Merge defaults with user overrides (override semantics)
+        self._ibindings: dict[str, str] = {
+            **_DEFAULT_BINDINGS,
+            **(initial_bindings or {}),
+        }
+        self.state = state if state is not None else {}
+        self.cache = cache if cache is not None else {}
+        self.stats = stats if stats is not None else {}
         self.stats.setdefault("op_calls", {})
 
     def run_kernel(self, pfunc: PFunction, arg_list: list[Any]) -> list[Any]:
         fn_type = pfunc.fn_type
-        spec = get_kernel_for_op(fn_type)
-        fn = spec.fn
+        kid = self._ibindings.get(fn_type)
+        if kid is None:
+            raise NotImplementedError(f"no backend kernel registered for op {fn_type}")
+        spec = get_kernel_spec(kid)
+        fn = spec.fn  # kernel implementation
         if len(arg_list) != len(pfunc.ins_info):
             raise ValueError(
                 f"kernel {fn_type} arg count mismatch: got {len(arg_list)}, expect {len(pfunc.ins_info)}"
@@ -186,17 +212,32 @@ class RuntimeContext:
 
     # ---- explicit (re)binding API ----
     def bind_op(self, op_type: str, kernel_id: str, *, force: bool = False) -> None:
-        """Bind an operation to a kernel at runtime.
+        """Bind an operation to a kernel for THIS context only.
 
-        force=False (default) preserves any existing binding to avoid accidental
-        silent overrides. Use ``rebind_op`` or ``force=True`` to intentionally
-        change a binding.
+        force=False (default) keeps existing binding (no silent override).
         """
-        base.bind_op(op_type, kernel_id, force=force)
+        if not kernel_exists(kernel_id):
+            raise KeyError(f"kernel_id {kernel_id} not registered")
+        if not force and op_type in self._ibindings:
+            return
+        self._ibindings[op_type] = kernel_id
 
     def rebind_op(self, op_type: str, kernel_id: str) -> None:
         """Force rebind an operation to a different kernel (shorthand)."""
-        base.bind_op(op_type, kernel_id, force=True)
+        self.bind_op(op_type, kernel_id, force=True)
+
+    # Introspection helpers
+    def list_bound_ops(self) -> list[str]:  # pragma: no cover - convenience
+        return sorted(self._ibindings.keys())
+
+    def get_binding(self, op_type: str) -> str | None:  # pragma: no cover
+        return self._ibindings.get(op_type)
+
+    def __repr__(self) -> str:  # pragma: no cover - debug aid
+        return (
+            f"RuntimeContext(rank={self.rank}, world_size={self.world_size}, "
+            f"bound_ops={len(self._ibindings)})"
+        )
 
 
 def _validate_table_arg(

{mplang_nightly-0.1.dev146 → mplang_nightly-0.1.dev148}/mplang/core/expr/evaluator.py

@@ -196,6 +196,28 @@ class EvalSemantic:
                     "uniform_cond: predicate is not uniform across parties"
                 )
 
+    # ------------------------------ While helpers ------------------------------
+    def _check_while_predicate(self, cond_result: list[Any]) -> Any:
+        """Validate while_loop predicate evaluation result.
+
+        Ensures the condition function returns exactly one value and that value
+        is non-None. Returns the boolean predicate value for convenience.
+
+        Raises:
+            AssertionError: If condition function returns != 1 value.
+            RuntimeError: If the single predicate value is None.
+        """
+        assert len(cond_result) == 1, (
+            f"Condition function must return a single value, got {cond_result}"
+        )
+        cond_value = cond_result[0]
+        if cond_value is None:
+            raise RuntimeError(
+                "while_loop condition produced None on rank "
+                f"{self.rank}; ensure the predicate yields a boolean for every party."
+            )
+        return cond_value
+
 
 class RecursiveEvaluator(EvalSemantic, ExprVisitor):
     """Recursive visitor-based evaluator."""
@@ -307,12 +329,8 @@ class RecursiveEvaluator(EvalSemantic, ExprVisitor):
             cond_env = dict(zip(expr.cond_fn.params, state, strict=True))
             cond_evaluator = self._fork(cond_env)
             cond_result = expr.cond_fn.body.accept(cond_evaluator)
-
-            assert len(cond_result) == 1, (
-                f"Condition function must return a single value, got {cond_result}"
-            )
-
-            if not cond_result[0]:
+            cond_value = self._check_while_predicate(cond_result)
+            if not cond_value:
                 break
 
             # Call body function with same arguments
@@ -445,8 +463,8 @@ class IterativeEvaluator(EvalSemantic):
                     cond_vals = self._iter_eval_graph(
                         node.cond_fn.body, {**env, **cond_env}
                     )
-                    assert len(cond_vals) == 1
-                    if not bool(cond_vals[0]):
+                    cond_val = self._check_while_predicate(cond_vals)
+                    if not bool(cond_val):
                         break
                     body_env = dict(zip(node.body_fn.params, state, strict=True))
                     new_state = self._iter_eval_graph(

{mplang_nightly-0.1.dev146 → mplang_nightly-0.1.dev148}/mplang/core/primitive.py

@@ -483,6 +483,20 @@ def uniform_cond(
     if pred_ty.dtype != BOOL:
         raise TypeError(f"uniform_cond predicate must be boolean, got {pred_ty.dtype}")
 
+    # Static pmask rule:
+    # If predicate has a static pmask (not None), it must equal the current trace
+    # context mask. Otherwise some parties would execute a branch without a
+    # defined predicate value (unsafe). To run on a subset either:
+    #   1. Trace the entire uniform_cond under a subset TraceContext (ctx.fork(mask=...))
+    #   2. Broadcast / lift predicate to full mask (e.g. pshfl_s)
+    # Pred pmask None => dynamic: defer to runtime uniformity (if verify_uniform=True).
+    pred_pmask = pred_ty.pmask
+    if pred_pmask is not None and pred_pmask != cur_tracer.mask:
+        raise ValueError(
+            "uniform_cond predicate static pmask mismatch: predicate pmask="
+            f"{pred_pmask} trace mask={cur_tracer.mask}. Trace under a subset "
+            "context (ctx.fork(mask=...)) or broadcast predicate (pshfl_s) to all parties."
+        )
     # Step 1: Trace both branches in separate contexts
     then_tracer = cur_tracer.fork()
     then_tfn = trace(then_tracer, then_fn, *args)
@@ -706,6 +720,22 @@ def while_loop(
             f"Condition function must return a boolean scalar, got dtype {cond_out_var.mptype.dtype}"
         )
 
+    # Static pmask rule:
+    # If the predicate's pmask is statically known it must match the trace context
+    # mask. Otherwise some parties in this context would lack a boolean to drive
+    # control flow (previously could lead to hang via None). To restrict to a subset:
+    #   1. Trace the entire while_loop under a subset context (ctx.fork(mask=submask)), or
+    #   2. Broadcast predicate to full mask (e.g. pshfl_s) before while_loop.
+    # Dynamic predicates (pmask=None) are allowed; runtime guard (evaluator) raises
+    # if any participating party observes None.
+    pred_pmask = cond_out_var.mptype.pmask
+    if pred_pmask is not None and pred_pmask != cur_tracer.mask:
+        raise ValueError(
+            "while_loop predicate static pmask mismatch: predicate pmask="
+            f"{pred_pmask} trace mask={cur_tracer.mask}. Trace under subset context "
+            "or broadcast predicate to all parties."
+        )
+
     # Validate body returns same number of leaves and same dtype/shape per leaf
     if len(body_tfn.out_vars) != len(cond_tfn.in_vars):
         raise ValueError(

mplang_nightly-0.1.dev148/tests/backend/test_kernel_binding.py

@@ -0,0 +1,102 @@
+# Copyright 2025 Ant Group Co., Ltd.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Tests for per-RuntimeContext binding isolation
+from __future__ import annotations
+
+import pytest
+
+from mplang.backend import base
+from mplang.backend.context import RuntimeContext
+from mplang.core.dtype import INT64  # switched from INT32 to INT64 to match Python int
+from mplang.core.pfunc import PFunction
+from mplang.core.tensor import TensorType
+
+# We'll register two fake kernels for an op to test rebinding.
+# If they already exist due to other tests, we guard with try/except.
+
+
+@base.kernel_def("test.echo.v1")
+def _echo_v1(
+    pfunc: PFunction, x: int
+) -> tuple[int,]:  # pragma: no cover - executed in test
+    return (x + 1,)
+
+
+@base.kernel_def("test.echo.v2")
+def _echo_v2(
+    pfunc: PFunction, x: int
+) -> tuple[int,]:  # pragma: no cover - executed in test
+    return (x + 2,)
+
+
+def make_pfunc(op_type: str) -> PFunction:
+    # Minimal PFunction stub compatible with backend.run_kernel expectations.
+    # shape info matters only for validation; use scalar INT64 (Python int maps to int64).
+    return PFunction(
+        fn_type=op_type,
+        fn_text="",
+        ins_info=[TensorType(shape=(), dtype=INT64)],
+        outs_info=[TensorType(shape=(), dtype=INT64)],
+    )
+
+
+def test_isolated_rebind():
+    # ctx1 binds op -> v1, ctx2 binds op -> v2; they should not interfere.
+    op = "test.echo"
+    ctx1 = RuntimeContext(rank=0, world_size=1, initial_bindings={op: "test.echo.v1"})
+    ctx2 = RuntimeContext(rank=0, world_size=1, initial_bindings={op: "test.echo.v2"})
+
+    pfunc = make_pfunc(op)
+    out1 = ctx1.run_kernel(pfunc, [10])[0]
+    out2 = ctx2.run_kernel(pfunc, [10])[0]
+
+    assert out1 == 11
+    assert out2 == 12
+
+
+def test_rebind_only_affects_context():
+    op = "test.echo"
+    ctx = RuntimeContext(rank=0, world_size=1, initial_bindings={op: "test.echo.v1"})
+    pfunc = make_pfunc(op)
+    assert ctx.run_kernel(pfunc, [5])[0] == 6
+    ctx.rebind_op(op, "test.echo.v2")
+    assert ctx.run_kernel(pfunc, [5])[0] == 7
+
+
+def test_force_flag():
+    op = "test.echo"
+    ctx = RuntimeContext(rank=0, world_size=1, initial_bindings={op: "test.echo.v1"})
+    # Attempt non-force bind (should keep v1)
+    ctx.bind_op(op, "test.echo.v2", force=False)
+    pfunc = make_pfunc(op)
+    assert ctx.run_kernel(pfunc, [1])[0] == 2  # still v1 (+1)
+    # Now force
+    ctx.bind_op(op, "test.echo.v2", force=True)
+    assert ctx.run_kernel(pfunc, [1])[0] == 3
+
+
+def test_unknown_kernel_id():
+    ctx = RuntimeContext(rank=0, world_size=1)
+    with pytest.raises(KeyError):
+        ctx.bind_op("some.op", "non.existent.kernel")
+
+
+def test_missing_binding():
+    # Pick an op name unlikely in defaults
+    op = "unit.test.unbound"
+    ctx = RuntimeContext(rank=0, world_size=1)
+    pfunc = make_pfunc(op)
+    with pytest.raises(NotImplementedError):
+        ctx.run_kernel(pfunc, [0])

{mplang_nightly-0.1.dev146 → mplang_nightly-0.1.dev148}/tests/runtime/test_simulation.py

@@ -33,7 +33,14 @@ from mplang.core.dtype import FLOAT32, INT32
 from mplang.core.mask import Mask
 from mplang.core.mpobject import MPObject
 from mplang.core.mptype import MPType, Rank
-from mplang.core.primitive import constant, prank, pshfl_s, uniform_cond, while_loop
+from mplang.core.primitive import (
+    constant,
+    prank,
+    pshfl_s,
+    set_mask,
+    uniform_cond,
+    while_loop,
+)
 from mplang.core.tracer import TraceContext, TraceVar, trace
 from mplang.runtime.simulation import Simulator, SimVar
 
@@ -1392,6 +1399,180 @@ class TestWhileLoop:
         assert results[0].values[0] == 6  # Party 0: 3 iterations
         assert results[0].values[1] == 4  # Party 1: 2 iterations
 
+    def test_while_loop_subset_state_mask(self):
+        """Loop state and control stay on subset of parties."""
+
+        cluster_spec = ClusterSpec.simple(world_size=3)
+        full_mask = Mask(0b111)
+        subset_mask = Mask(0b011)
+        trace_ctx = TraceContext(cluster_spec=cluster_spec, mask=full_mask)
+        simulator = Simulator.simple(world_size=3)
+
+        def subset_loop():
+            init_state = set_mask(constant(np.int64(0)), subset_mask)
+            threshold = set_mask(constant(np.int64(3)), subset_mask)
+            step = set_mask(constant(np.int64(1)), subset_mask)
+
+            def cond_fn(state):
+                subset_pred = simp.run(lambda val, limit: val < limit)(state, threshold)
+                return pshfl_s(subset_pred, full_mask, [Rank(0), Rank(0), Rank(0)])
+
+            def body_fn(state):
+                return simp.run(lambda val, inc: val + inc)(state, step)
+
+            return while_loop(cond_fn, body_fn, init_state)
+
+        with with_ctx(trace_ctx):
+            traced_fn = trace(trace_ctx, subset_loop)
+
+        func_expr = traced_fn.make_expr()
+        assert func_expr is not None
+        expr = func_expr.body
+        results = simulator.evaluate(expr, {})
+
+        assert len(results) == 1
+        sim_var = results[0]
+        assert isinstance(sim_var, SimVar)
+        assert sim_var.mptype.pmask == subset_mask
+
+        values = sim_var.values
+        assert len(values) == 3
+        assert values[0] == 3
+        assert values[1] == 3
+        assert values[2] is None
+
+    def test_while_loop_subset_context_mask_success(self):
+        """Trace under subset context mask; predicate pmask==context mask so no broadcast needed.
+
+        Ensures static pmask validation (design A) does NOT raise when the trace context
+        itself is the subset. Predicate pmask equals the context mask.
+        """
+        # Use a 2-party cluster because only parties 0 and 1 participate.
+        cluster_spec = ClusterSpec.simple(world_size=2)
+        subset_mask = Mask(0b11)  # parties 0 and 1
+        trace_ctx = TraceContext(cluster_spec=cluster_spec, mask=subset_mask)
+        simulator = Simulator.simple(world_size=2)
+
+        def subset_loop():
+            init_state = set_mask(constant(np.int64(0)), subset_mask)
+            threshold = set_mask(constant(np.int64(3)), subset_mask)
+            step = set_mask(constant(np.int64(1)), subset_mask)
+
+            def cond_fn(state):
+                # Returns bool with pmask=subset_mask (no broadcast)
+                return simp.run(lambda val, limit: val < limit)(state, threshold)
+
+            def body_fn(state):
+                return simp.run(lambda val, inc: val + inc)(state, step)
+
+            return while_loop(cond_fn, body_fn, init_state)
+
+        with with_ctx(trace_ctx):
+            traced_fn = trace(trace_ctx, subset_loop)
+
+        func_expr = traced_fn.make_expr()
+        assert func_expr is not None
+        expr = func_expr.body
+        results = simulator.evaluate(expr, {})
+
+        assert len(results) == 1
+        sim_var = results[0]
+        assert isinstance(sim_var, SimVar)
+        assert sim_var.mptype.pmask == subset_mask
+        values = sim_var.values
+        assert len(values) == 2
+        assert values[0] == 3
+        assert values[1] == 3
+
+    def test_while_loop_predicate_static_pmask_mismatch_error(self):
+        """Full context mask but predicate has smaller static pmask -> trace-time ValueError.
+
+        We purposely do NOT broadcast the subset predicate to full mask, expecting the
+        new static pmask validation in while_loop to raise.
+        """
+        cluster_spec = ClusterSpec.simple(world_size=3)
+        full_mask = Mask(0b111)
+        subset_mask = Mask(0b011)
+        trace_ctx = TraceContext(cluster_spec=cluster_spec, mask=full_mask)
+
+        def bad_loop():
+            init_state = set_mask(constant(np.int64(0)), subset_mask)
+            threshold = set_mask(constant(np.int64(2)), subset_mask)
+            step = set_mask(constant(np.int64(1)), subset_mask)
+
+            def cond_fn(state):
+                # Returns bool with pmask=subset_mask only; no broadcast.
+                return simp.run(lambda val, limit: val < limit)(state, threshold)
+
+            def body_fn(state):
+                return simp.run(lambda val, inc: val + inc)(state, step)
+
+            return while_loop(cond_fn, body_fn, init_state)
+
+        with with_ctx(trace_ctx):
+            with pytest.raises(
+                ValueError, match=r"while_loop predicate static pmask mismatch"
+            ):
+                trace(trace_ctx, bad_loop)
+
+    def test_while_loop_cond_body_with_aux_party(self):
+        """Loop state on subset while cond/body still invoke a third party."""
+
+        cluster_spec = ClusterSpec.simple(world_size=3)
+        full_mask = Mask(0b111)
+        subset_mask = Mask(0b011)
+        aux_mask = Mask(0b100)
+        trace_ctx = TraceContext(cluster_spec=cluster_spec, mask=full_mask)
+        simulator = Simulator.simple(world_size=3)
+
+        def cooperative_loop():
+            subset_state = set_mask(constant(np.int64(0)), subset_mask)
+            aux_state = set_mask(constant(np.int64(0)), aux_mask)
+
+            subset_limit = set_mask(constant(np.int64(6)), subset_mask)
+            subset_step = set_mask(constant(np.int64(2)), subset_mask)
+            aux_step = set_mask(constant(np.int64(1)), aux_mask)
+
+            def cond_fn(states):
+                sub_val, aux_val = states
+
+                # Auxiliary party executes a helper kernel (result ignored by others)
+                _ = simp.run(lambda val, inc: val + inc)(aux_val, aux_step)
+                subset_pred = simp.run(lambda val, limit: val < limit)(
+                    sub_val, subset_limit
+                )
+                # Broadcast predicate so every party observes the same boolean
+                return pshfl_s(subset_pred, full_mask, [Rank(0), Rank(0), Rank(0)])
+
+            def body_fn(states):
+                sub_val, aux_val = states
+
+                next_sub = simp.run(lambda val, step: val + step)(sub_val, subset_step)
+                next_aux = simp.run(lambda val, inc: val + inc)(aux_val, aux_step)
+
+                return (next_sub, next_aux)
+
+            return while_loop(cond_fn, body_fn, (subset_state, aux_state))
+
+        with with_ctx(trace_ctx):
+            traced_fn = trace(trace_ctx, cooperative_loop)
+
+        func_expr = traced_fn.make_expr()
+        assert func_expr is not None
+        expr = func_expr.body
+        results = simulator.evaluate(expr, {})
+
+        assert len(results) == 2
+        subset_result, aux_result = results
+
+        assert isinstance(subset_result, SimVar)
+        assert subset_result.mptype.pmask == subset_mask
+        assert subset_result.values == [6, 6, None]
+
+        assert isinstance(aux_result, SimVar)
+        assert aux_result.mptype.pmask == aux_mask
+        assert aux_result.values == [None, None, 3]
+
     def test_nested_while_with_conditional(self, simulator, trace_context):
         """Test: While_loop containing conditional operations.