Support replication axis in dump comparator (#19282)

2026-02-25 09:48:43 +08:00
parent 2e2b18e870
commit b7af58b9af
10 changed files with 778 additions and 62 deletions
@@ -148,7 +148,7 @@ class TestCpZigzagTpE2E:
            if isinstance(plan, ReorderPlan):
                current = execute_reorder_plan(plan, current)
            else:
-                current = execute_unshard_plan(plan, current)
+                current, _ = execute_unshard_plan(plan, current)

        assert len(current) == 1
        assert torch.allclose(current[0], full_tensor)
@@ -5,12 +5,16 @@ import torch

 from sglang.srt.debug_utils.comparator.aligner.unshard.executor import (
    _apply_unshard,
+    _verify_replicated_group,
    execute_unshard_plan,
 )
 from sglang.srt.debug_utils.comparator.aligner.unshard.planner import (
    compute_unshard_plan,
 )
-from sglang.srt.debug_utils.comparator.aligner.unshard.types import AxisInfo
+from sglang.srt.debug_utils.comparator.aligner.unshard.types import (
+    AxisInfo,
+    PickParams,
+)
 from sglang.srt.debug_utils.comparator.dims import ParallelAxis, parse_dims
 from sglang.test.ci.ci_register import register_cpu_ci

@@ -29,9 +33,10 @@ class TestExecuteUnshardPlan:
        plans = compute_unshard_plan(dim_specs, parallel_infos)
        assert len(plans) == 1

-        result = execute_unshard_plan(plans[0], shards)
+        result, warnings = execute_unshard_plan(plans[0], shards)
        assert len(result) == 1
        assert torch.allclose(result[0], full_tensor)
+        assert warnings == []

    def test_scrambled_world_ranks_correct_result(self) -> None:
        full_tensor = torch.randn(4, 8)
@@ -54,9 +59,10 @@ class TestExecuteUnshardPlan:
            shards[1],  # world_rank=3, axis_rank=1
        ]

-        result = execute_unshard_plan(plans[0], tensors_ordered_by_world_rank)
+        result, warnings = execute_unshard_plan(plans[0], tensors_ordered_by_world_rank)
        assert len(result) == 1
        assert torch.allclose(result[0], full_tensor)
+        assert warnings == []

    def test_single_step_reduces_tensor_count(self) -> None:
        """8 tensors with 2 groups of 4 produce 2 output tensors."""
@@ -85,10 +91,10 @@ class TestExecuteUnshardPlan:
            for tp_rank in range(4):
                tensors.append(source[tp_rank])

-        intermediate = execute_unshard_plan(plans[0], tensors)
+        intermediate, _ = execute_unshard_plan(plans[0], tensors)
        assert len(intermediate) == 4

-        final = execute_unshard_plan(plans[1], intermediate)
+        final, _ = execute_unshard_plan(plans[1], intermediate)
        assert len(final) == 1

    def test_cp_tp_concat(self) -> None:
@@ -116,7 +122,7 @@ class TestExecuteUnshardPlan:

        current = tensors
        for plan in plans:
-            current = execute_unshard_plan(plan, current)
+            current, _ = execute_unshard_plan(plan, current)

        assert len(current) == 1
        assert torch.allclose(current[0], full_tensor)
@@ -157,7 +163,7 @@ class TestExecuteUnshardPlan:

        current = tensors
        for plan in plans:
-            current = execute_unshard_plan(plan, current)
+            current, _ = execute_unshard_plan(plan, current)

        assert len(current) == 1
        assert torch.allclose(current[0], full_tensor)
@@ -169,7 +175,12 @@ class TestExecuteUnshardPlan:
            pass

        with pytest.raises(ValueError, match="Unsupported unshard"):
-            _apply_unshard(_FakeParams(), [torch.randn(2, 2)])
+            _apply_unshard(
+                _FakeParams(),
+                [torch.randn(2, 2)],
+                axis=ParallelAxis.TP,
+                group_index=0,
+            )

    def test_cp_tp_ep_three_axis_concat(self) -> None:
        """CP=2 + TP=2 + EP=2: three-step unshard reconstructs original tensor."""
@@ -205,7 +216,7 @@ class TestExecuteUnshardPlan:

        current = tensors
        for plan in plans:
-            current = execute_unshard_plan(plan, current)
+            current, _ = execute_unshard_plan(plan, current)

        assert len(current) == 1
        assert torch.allclose(current[0], full_tensor)
@@ -253,11 +264,211 @@ class TestExecuteUnshardPlan:

        current = tensors
        for plan in plans:
-            current = execute_unshard_plan(plan, current)
+            current, _ = execute_unshard_plan(plan, current)

        assert len(current) == 1
        assert torch.allclose(current[0], full_tensor)


+class TestPickOperation:
+    def test_pick_single_group(self) -> None:
+        """PickParams picks the first tensor from a single group."""
+        tensor = torch.randn(4, 8)
+        dim_specs = parse_dims("h d")
+        parallel_infos = [
+            {ParallelAxis.TP: AxisInfo(axis_rank=0, axis_size=2)},
+            {ParallelAxis.TP: AxisInfo(axis_rank=1, axis_size=2)},
+        ]
+
+        plans = compute_unshard_plan(dim_specs, parallel_infos)
+        assert len(plans) == 1
+        assert isinstance(plans[0].params, PickParams)
+
+        result, warnings = execute_unshard_plan(plans[0], [tensor, tensor.clone()])
+        assert len(result) == 1
+        assert torch.allclose(result[0], tensor)
+        assert warnings == []
+
+    def test_pick_multiple_groups(self) -> None:
+        """PickParams with multiple groups picks one from each."""
+        dim_specs = parse_dims("h(tp)")
+        parallel_infos: list[dict[ParallelAxis, AxisInfo]] = [
+            {
+                ParallelAxis.CP: AxisInfo(axis_rank=0, axis_size=2),
+                ParallelAxis.TP: AxisInfo(axis_rank=0, axis_size=2),
+            },
+            {
+                ParallelAxis.CP: AxisInfo(axis_rank=1, axis_size=2),
+                ParallelAxis.TP: AxisInfo(axis_rank=0, axis_size=2),
+            },
+            {
+                ParallelAxis.CP: AxisInfo(axis_rank=0, axis_size=2),
+                ParallelAxis.TP: AxisInfo(axis_rank=1, axis_size=2),
+            },
+            {
+                ParallelAxis.CP: AxisInfo(axis_rank=1, axis_size=2),
+                ParallelAxis.TP: AxisInfo(axis_rank=1, axis_size=2),
+            },
+        ]
+
+        plans = compute_unshard_plan(dim_specs, parallel_infos)
+        pick_plans = [p for p in plans if isinstance(p.params, PickParams)]
+        assert len(pick_plans) == 1
+        assert pick_plans[0].axis == ParallelAxis.CP
+
+        tensor = torch.randn(4)
+        tensors = [tensor.clone() for _ in range(4)]
+
+        result, warnings = execute_unshard_plan(pick_plans[0], tensors)
+        assert len(result) == 2
+        assert warnings == []
+
+    def test_replicated_tp_sharded_cp_e2e(self) -> None:
+        """CP2 TP2, dims='b s(cp) d': replicated TP pick + sharded CP concat round-trip."""
+        torch.manual_seed(42)
+        full_tensor = torch.randn(4, 8, 16)
+        cp_chunks = list(full_tensor.chunk(2, dim=1))
+
+        tensors: list[torch.Tensor] = []
+        parallel_infos: list[dict[ParallelAxis, AxisInfo]] = []
+        for cp_rank in range(2):
+            for tp_rank in range(2):
+                tensors.append(cp_chunks[cp_rank].clone())
+                parallel_infos.append(
+                    {
+                        ParallelAxis.CP: AxisInfo(axis_rank=cp_rank, axis_size=2),
+                        ParallelAxis.TP: AxisInfo(axis_rank=tp_rank, axis_size=2),
+                    }
+                )
+
+        dim_specs = parse_dims("b s(cp) d")
+        plans = compute_unshard_plan(dim_specs, parallel_infos)
+        assert len(plans) == 2
+
+        current = tensors
+        for plan in plans:
+            current, _ = execute_unshard_plan(plan, current)
+
+        assert len(current) == 1
+        assert torch.allclose(current[0], full_tensor)
+
+    def test_fully_replicated_e2e(self) -> None:
+        """CP2 TP2, dims='b h d': fully replicated → 2 pick steps → 1 tensor."""
+        torch.manual_seed(42)
+        full_tensor = torch.randn(4, 8, 16)
+
+        tensors: list[torch.Tensor] = []
+        parallel_infos: list[dict[ParallelAxis, AxisInfo]] = []
+        for cp_rank in range(2):
+            for tp_rank in range(2):
+                tensors.append(full_tensor.clone())
+                parallel_infos.append(
+                    {
+                        ParallelAxis.CP: AxisInfo(axis_rank=cp_rank, axis_size=2),
+                        ParallelAxis.TP: AxisInfo(axis_rank=tp_rank, axis_size=2),
+                    }
+                )
+
+        dim_specs = parse_dims("b h d")
+        plans = compute_unshard_plan(dim_specs, parallel_infos)
+        assert len(plans) == 2
+        assert all(isinstance(p.params, PickParams) for p in plans)
+
+        current = tensors
+        for plan in plans:
+            current, _ = execute_unshard_plan(plan, current)
+
+        assert len(current) == 1
+        assert torch.allclose(current[0], full_tensor)
+
+
+class TestVerifyReplicatedGroup:
+    def test_warns_on_mismatch(self) -> None:
+        """_verify_replicated_group produces warning when replicas differ."""
+        tensor_a = torch.ones(4)
+        tensor_b = torch.ones(4) + 0.1
+
+        warnings = _verify_replicated_group(
+            [tensor_a, tensor_b],
+            axis=ParallelAxis.TP,
+            group_index=0,
+        )
+        assert len(warnings) == 1
+        assert warnings[0].axis == "tp"
+        assert warnings[0].group_index == 0
+        assert warnings[0].differing_index == 1
+        assert warnings[0].baseline_index == 0
+        assert warnings[0].max_abs_diff == pytest.approx(0.1, abs=1e-5)
+
+    def test_no_warn_when_identical(self) -> None:
+        """_verify_replicated_group produces no warning for identical replicas."""
+        tensor = torch.randn(4, 8)
+
+        warnings = _verify_replicated_group(
+            [tensor, tensor.clone()],
+            axis=ParallelAxis.TP,
+            group_index=0,
+        )
+        assert warnings == []
+
+    def test_multiple_mismatches(self) -> None:
+        """_verify_replicated_group reports each differing replica."""
+        baseline = torch.zeros(4)
+        other_a = torch.ones(4)
+        other_b = torch.ones(4) * 2
+
+        warnings = _verify_replicated_group(
+            [baseline, other_a, other_b],
+            axis=ParallelAxis.CP,
+            group_index=1,
+        )
+        assert len(warnings) == 2
+        assert warnings[0].differing_index == 1
+        assert warnings[1].differing_index == 2
+        assert warnings[1].max_abs_diff == pytest.approx(2.0, abs=1e-5)
+
+    def test_execute_returns_warnings(self) -> None:
+        """execute_unshard_plan returns warnings for replicated mismatch."""
+        dim_specs = parse_dims("h d")
+        parallel_infos = [
+            {ParallelAxis.TP: AxisInfo(axis_rank=0, axis_size=2)},
+            {ParallelAxis.TP: AxisInfo(axis_rank=1, axis_size=2)},
+        ]
+        plans = compute_unshard_plan(dim_specs, parallel_infos)
+
+        tensor_a = torch.zeros(4)
+        tensor_b = torch.ones(4)
+
+        result, warnings = execute_unshard_plan(plans[0], [tensor_a, tensor_b])
+        assert len(result) == 1
+        assert len(warnings) == 1
+        assert torch.allclose(result[0], tensor_a)
+
+    def test_atol_boundary_within(self) -> None:
+        """Difference exactly at atol (1e-6) → torch.allclose passes → no warning."""
+        baseline = torch.zeros(4)
+        other = torch.full((4,), 1e-6)
+
+        warnings = _verify_replicated_group(
+            [baseline, other],
+            axis=ParallelAxis.TP,
+            group_index=0,
+        )
+        assert warnings == []
+
+    def test_atol_boundary_exceeded(self) -> None:
+        """Difference just above atol (1e-6 + 1e-9) → torch.allclose fails → warning."""
+        baseline = torch.zeros(4)
+        other = torch.full((4,), 1e-6 + 1e-9)
+
+        warnings = _verify_replicated_group(
+            [baseline, other],
+            axis=ParallelAxis.TP,
+            group_index=0,
+        )
+        assert len(warnings) == 1
+        assert warnings[0].differing_index == 1
+
+
 if __name__ == "__main__":
    sys.exit(pytest.main([__file__]))
@@ -5,7 +5,11 @@ import pytest
 from sglang.srt.debug_utils.comparator.aligner.unshard.planner import (
    compute_unshard_plan,
 )
-from sglang.srt.debug_utils.comparator.aligner.unshard.types import AxisInfo
+from sglang.srt.debug_utils.comparator.aligner.unshard.types import (
+    AxisInfo,
+    ConcatParams,
+    PickParams,
+)
 from sglang.srt.debug_utils.comparator.dims import ParallelAxis, parse_dims
 from sglang.test.ci.ci_register import register_cpu_ci

@@ -228,7 +232,7 @@ class TestComputeUnshardPlan:
        assert len(plans[2].groups) == 1
        assert len(plans[2].groups[0]) == 2

-    def test_replicated_axis_raises(self) -> None:
+    def test_sharded_axis_missing_from_rank_raises(self) -> None:
        """A world_rank missing a sharded axis raises ValueError."""
        dim_specs = parse_dims("s(cp) h(tp)")
        parallel_infos = [
@@ -238,7 +242,159 @@ class TestComputeUnshardPlan:
            },
            {
                ParallelAxis.CP: AxisInfo(axis_rank=1, axis_size=2),
-                # missing TP — replicated
+                # missing TP — sharded axis absent from rank
+            },
+        ]
+        with pytest.raises(ValueError, match="missing parallel_info"):
+            compute_unshard_plan(dim_specs, parallel_infos)
+
+
+class TestReplicatedAxes:
+    def test_replicated_tp_with_sharded_cp(self) -> None:
+        """CP2 TP2, dims='b s(cp) d' → PickPlan(TP) + ConcatPlan(CP)."""
+        dim_specs = parse_dims("b s(cp) d")
+        parallel_infos: list[dict[ParallelAxis, AxisInfo]] = [
+            {
+                ParallelAxis.CP: AxisInfo(axis_rank=0, axis_size=2),
+                ParallelAxis.TP: AxisInfo(axis_rank=0, axis_size=2),
+            },
+            {
+                ParallelAxis.CP: AxisInfo(axis_rank=0, axis_size=2),
+                ParallelAxis.TP: AxisInfo(axis_rank=1, axis_size=2),
+            },
+            {
+                ParallelAxis.CP: AxisInfo(axis_rank=1, axis_size=2),
+                ParallelAxis.TP: AxisInfo(axis_rank=0, axis_size=2),
+            },
+            {
+                ParallelAxis.CP: AxisInfo(axis_rank=1, axis_size=2),
+                ParallelAxis.TP: AxisInfo(axis_rank=1, axis_size=2),
+            },
+        ]
+        plans = compute_unshard_plan(dim_specs, parallel_infos)
+
+        assert len(plans) == 2
+        assert plans[0].axis == ParallelAxis.TP
+        assert isinstance(plans[0].params, PickParams)
+        assert len(plans[0].groups) == 2
+        for group in plans[0].groups:
+            assert len(group) == 2
+
+        assert plans[1].axis == ParallelAxis.CP
+        assert isinstance(plans[1].params, ConcatParams)
+        assert plans[1].params.dim == 1
+
+    def test_fully_replicated(self) -> None:
+        """CP2 TP2, dims='b h d' → PickPlan(CP) + PickPlan(TP)."""
+        dim_specs = parse_dims("b h d")
+        parallel_infos: list[dict[ParallelAxis, AxisInfo]] = [
+            {
+                ParallelAxis.CP: AxisInfo(axis_rank=0, axis_size=2),
+                ParallelAxis.TP: AxisInfo(axis_rank=0, axis_size=2),
+            },
+            {
+                ParallelAxis.CP: AxisInfo(axis_rank=0, axis_size=2),
+                ParallelAxis.TP: AxisInfo(axis_rank=1, axis_size=2),
+            },
+            {
+                ParallelAxis.CP: AxisInfo(axis_rank=1, axis_size=2),
+                ParallelAxis.TP: AxisInfo(axis_rank=0, axis_size=2),
+            },
+            {
+                ParallelAxis.CP: AxisInfo(axis_rank=1, axis_size=2),
+                ParallelAxis.TP: AxisInfo(axis_rank=1, axis_size=2),
+            },
+        ]
+        plans = compute_unshard_plan(dim_specs, parallel_infos)
+
+        assert len(plans) == 2
+        assert all(isinstance(p.params, PickParams) for p in plans)
+        axes = {p.axis for p in plans}
+        assert axes == {ParallelAxis.CP, ParallelAxis.TP}
+
+    def test_multiple_replicated_one_sharded(self) -> None:
+        """CP2 TP2 EP2, dims='h(tp)' → PickPlan(CP) + PickPlan(EP) + ConcatPlan(TP)."""
+        dim_specs = parse_dims("h(tp)")
+        parallel_infos: list[dict[ParallelAxis, AxisInfo]] = []
+        for cp_rank in range(2):
+            for ep_rank in range(2):
+                for tp_rank in range(2):
+                    parallel_infos.append(
+                        {
+                            ParallelAxis.CP: AxisInfo(axis_rank=cp_rank, axis_size=2),
+                            ParallelAxis.EP: AxisInfo(axis_rank=ep_rank, axis_size=2),
+                            ParallelAxis.TP: AxisInfo(axis_rank=tp_rank, axis_size=2),
+                        }
+                    )
+
+        plans = compute_unshard_plan(dim_specs, parallel_infos)
+
+        assert len(plans) == 3
+        pick_plans = [p for p in plans if isinstance(p.params, PickParams)]
+        concat_plans = [p for p in plans if isinstance(p.params, ConcatParams)]
+        assert len(pick_plans) == 2
+        assert len(concat_plans) == 1
+        assert concat_plans[0].axis == ParallelAxis.TP
+
+        replicated_axes = {p.axis for p in pick_plans}
+        assert replicated_axes == {ParallelAxis.CP, ParallelAxis.EP}
+
+    def test_replicated_scrambled_ranks(self) -> None:
+        """Scrambled world_rank order with replicated axis."""
+        dim_specs = parse_dims("h(tp)")
+        parallel_infos: list[dict[ParallelAxis, AxisInfo]] = [
+            {
+                ParallelAxis.CP: AxisInfo(axis_rank=1, axis_size=2),
+                ParallelAxis.TP: AxisInfo(axis_rank=0, axis_size=2),
+            },
+            {
+                ParallelAxis.CP: AxisInfo(axis_rank=0, axis_size=2),
+                ParallelAxis.TP: AxisInfo(axis_rank=1, axis_size=2),
+            },
+            {
+                ParallelAxis.CP: AxisInfo(axis_rank=0, axis_size=2),
+                ParallelAxis.TP: AxisInfo(axis_rank=0, axis_size=2),
+            },
+            {
+                ParallelAxis.CP: AxisInfo(axis_rank=1, axis_size=2),
+                ParallelAxis.TP: AxisInfo(axis_rank=1, axis_size=2),
+            },
+        ]
+        plans = compute_unshard_plan(dim_specs, parallel_infos)
+
+        assert len(plans) == 2
+        assert plans[0].axis == ParallelAxis.CP
+        assert isinstance(plans[0].params, PickParams)
+        assert plans[1].axis == ParallelAxis.TP
+        assert isinstance(plans[1].params, ConcatParams)
+
+    def test_replicated_axis_inconsistent_size_raises(self) -> None:
+        """Replicated axis with inconsistent sizes raises ValueError."""
+        dim_specs = parse_dims("h(tp)")
+        parallel_infos: list[dict[ParallelAxis, AxisInfo]] = [
+            {
+                ParallelAxis.CP: AxisInfo(axis_rank=0, axis_size=2),
+                ParallelAxis.TP: AxisInfo(axis_rank=0, axis_size=2),
+            },
+            {
+                ParallelAxis.CP: AxisInfo(axis_rank=0, axis_size=4),
+                ParallelAxis.TP: AxisInfo(axis_rank=1, axis_size=2),
+            },
+        ]
+        with pytest.raises(ValueError, match="Inconsistent axis_size"):
+            compute_unshard_plan(dim_specs, parallel_infos)
+
+    def test_replicated_axis_missing_from_rank_raises(self) -> None:
+        """A rank missing a replicated axis that other ranks have raises ValueError."""
+        dim_specs = parse_dims("h(tp)")
+        parallel_infos: list[dict[ParallelAxis, AxisInfo]] = [
+            {
+                ParallelAxis.CP: AxisInfo(axis_rank=0, axis_size=2),
+                ParallelAxis.TP: AxisInfo(axis_rank=0, axis_size=2),
+            },
+            {
+                # missing CP — replicated axis absent from this rank
+                ParallelAxis.TP: AxisInfo(axis_rank=1, axis_size=2),
            },
        ]
        with pytest.raises(ValueError, match="missing parallel_info"):
@@ -6,6 +6,7 @@ import pytest
 from sglang.srt.debug_utils.comparator.output_types import (
    ComparisonRecord,
    ConfigRecord,
+    ReplicatedMismatchWarning,
    SkipRecord,
    SummaryRecord,
    parse_record_json,
@@ -118,5 +119,71 @@ class TestRecordTypes:
            assert restored == record


+def _make_warning(**overrides) -> ReplicatedMismatchWarning:
+    defaults: dict = dict(
+        axis="tp",
+        group_index=0,
+        differing_index=1,
+        baseline_index=0,
+        max_abs_diff=0.1,
+    )
+    defaults.update(overrides)
+    return ReplicatedMismatchWarning(**defaults)
+
+
+class TestAlignWarnings:
+    def test_comparison_record_failed_when_diff_passed_but_warnings(self):
+        """ComparisonRecord with diff.passed=True but align_warnings → category=='failed'."""
+        record = ComparisonRecord(
+            name="hidden",
+            baseline=_make_tensor_info(),
+            target=_make_tensor_info(),
+            unified_shape=[4, 8],
+            shape_mismatch=False,
+            diff=_make_diff(passed=True),
+            align_warnings=[_make_warning()],
+        )
+        assert record.category == "failed"
+
+    def test_skip_record_failed_when_warnings(self):
+        """SkipRecord with align_warnings → category=='failed' instead of 'skipped'."""
+        record = SkipRecord(
+            name="x",
+            reason="no_baseline",
+            align_warnings=[_make_warning()],
+        )
+        assert record.category == "failed"
+
+    def test_align_warnings_json_round_trip(self):
+        """align_warnings survive model_dump_json → parse_record_json round-trip."""
+        warning = _make_warning(
+            axis="cp",
+            group_index=2,
+            differing_index=3,
+            baseline_index=0,
+            max_abs_diff=0.42,
+        )
+        record = ComparisonRecord(
+            name="mlp",
+            baseline=_make_tensor_info(),
+            target=_make_tensor_info(),
+            unified_shape=[4, 8],
+            shape_mismatch=False,
+            diff=_make_diff(),
+            align_warnings=[warning],
+        )
+
+        restored = parse_record_json(record.model_dump_json())
+        assert isinstance(restored, ComparisonRecord)
+        assert len(restored.align_warnings) == 1
+
+        restored_warning = restored.align_warnings[0]
+        assert restored_warning.axis == "cp"
+        assert restored_warning.group_index == 2
+        assert restored_warning.differing_index == 3
+        assert restored_warning.baseline_index == 0
+        assert restored_warning.max_abs_diff == pytest.approx(0.42)
+
+
 if __name__ == "__main__":
    sys.exit(pytest.main([__file__]))
@@ -878,6 +878,138 @@ class TestEntrypointGroupingLogical:
        assert comp.name == "hidden"


+class TestEntrypointReplicatedAxis:
+    """Test replicated-axis scenarios through the full entrypoint pipeline."""
+
+    def test_replicated_axis_identical_replicas_passed(self, tmp_path, capsys):
+        """CP2 TP2, TP replicated and identical → passed, no align_warnings."""
+        torch.manual_seed(42)
+        full_baseline = torch.randn(4, 8, 6)
+        full_target = full_baseline + torch.randn(4, 8, 6) * 0.0001
+
+        baseline_dir = tmp_path / "baseline"
+        target_dir = tmp_path / "target"
+
+        for side_dir, full_tensor in [
+            (baseline_dir, full_baseline),
+            (target_dir, full_target),
+        ]:
+            _create_replicated_tp_sharded_cp_dumps(
+                side_dir,
+                full_tensor=full_tensor,
+                name="attn_out",
+                cp_size=2,
+                tp_size=2,
+                seq_dim=1,
+                dims_str="b s(cp) d",
+            )
+
+        args = _make_args(
+            baseline_dir / _FIXED_EXP_NAME,
+            target_dir / _FIXED_EXP_NAME,
+            diff_threshold=0.01,
+        )
+
+        records = _run_and_parse(args, capsys)
+        comp = _assert_single_comparison_passed(records)
+        assert comp.align_warnings == []
+
+        summary = records[-1]
+        assert isinstance(summary, SummaryRecord)
+        assert summary.passed == 1
+
+    def test_replicated_mismatch_fails(self, tmp_path, capsys):
+        """CP2 TP2, TP replicas differ (> atol) → failed with align_warnings."""
+        torch.manual_seed(42)
+        full_baseline = torch.randn(4, 8, 6)
+        full_target = full_baseline + torch.randn(4, 8, 6) * 0.0001
+
+        baseline_dir = tmp_path / "baseline"
+        target_dir = tmp_path / "target"
+
+        for side_dir, full_tensor in [
+            (baseline_dir, full_baseline),
+            (target_dir, full_target),
+        ]:
+            _create_replicated_tp_sharded_cp_dumps(
+                side_dir,
+                full_tensor=full_tensor,
+                name="attn_out",
+                cp_size=2,
+                tp_size=2,
+                seq_dim=1,
+                dims_str="b s(cp) d",
+                tp_noise=0.5,
+            )
+
+        args = _make_args(
+            baseline_dir / _FIXED_EXP_NAME,
+            target_dir / _FIXED_EXP_NAME,
+            diff_threshold=0.01,
+        )
+
+        records = _run_and_parse(args, capsys)
+        comparisons = _get_comparisons(records)
+        assert len(comparisons) == 1
+        assert comparisons[0].category == "failed"
+        assert len(comparisons[0].align_warnings) > 0
+
+        summary = records[-1]
+        assert isinstance(summary, SummaryRecord)
+        assert summary.failed == 1
+
+    def test_summary_counts_failed_from_align_warnings_only(self, tmp_path, capsys):
+        """Diff itself passes but TP replicas differ → summary.failed=1 from align_warnings."""
+        torch.manual_seed(42)
+        full_baseline = torch.randn(4, 8, 6)
+        full_target = full_baseline + torch.randn(4, 8, 6) * 0.0001
+
+        baseline_dir = tmp_path / "baseline"
+        target_dir = tmp_path / "target"
+
+        _create_replicated_tp_sharded_cp_dumps(
+            baseline_dir,
+            full_tensor=full_baseline,
+            name="attn_out",
+            cp_size=2,
+            tp_size=2,
+            seq_dim=1,
+            dims_str="b s(cp) d",
+            tp_noise=0.5,
+        )
+        _create_replicated_tp_sharded_cp_dumps(
+            target_dir,
+            full_tensor=full_target,
+            name="attn_out",
+            cp_size=2,
+            tp_size=2,
+            seq_dim=1,
+            dims_str="b s(cp) d",
+            tp_noise=0.5,
+        )
+
+        args = _make_args(
+            baseline_dir / _FIXED_EXP_NAME,
+            target_dir / _FIXED_EXP_NAME,
+            diff_threshold=0.5,
+        )
+
+        records = _run_and_parse(args, capsys)
+        comparisons = _get_comparisons(records)
+        assert len(comparisons) == 1
+
+        comp = comparisons[0]
+        assert comp.diff is not None
+        assert comp.diff.passed
+        assert len(comp.align_warnings) > 0
+        assert comp.category == "failed"
+
+        summary = records[-1]
+        assert isinstance(summary, SummaryRecord)
+        assert summary.failed == 1
+        assert summary.passed == 0
+
+
 # --------------------------- Assertion helpers -------------------


@@ -1138,6 +1270,49 @@ def _create_cp_zigzag_tp_sharded_dumps(
    return directory / _FIXED_EXP_NAME


+def _create_replicated_tp_sharded_cp_dumps(
+    directory: Path,
+    *,
+    full_tensor: torch.Tensor,
+    name: str,
+    cp_size: int,
+    tp_size: int,
+    seq_dim: int,
+    dims_str: str,
+    tp_noise: float = 0.0,
+) -> Path:
+    """Create CP-sharded + TP-replicated dump files from a full tensor.
+
+    CP direction: chunks along seq_dim (sharded).
+    TP direction: clones (replicated), with optional noise to simulate mismatch.
+    """
+    cp_chunks: list[torch.Tensor] = list(full_tensor.chunk(cp_size, dim=seq_dim))
+
+    rank: int = 0
+    for cp_rank in range(cp_size):
+        for tp_rank in range(tp_size):
+            shard = cp_chunks[cp_rank].clone()
+            if tp_noise > 0 and tp_rank > 0:
+                shard = shard + torch.randn_like(shard) * tp_noise
+
+            _create_rank_dump(
+                directory,
+                rank=rank,
+                name=name,
+                tensor=shard,
+                dims=dims_str,
+                parallel_info={
+                    "cp_rank": cp_rank,
+                    "cp_size": cp_size,
+                    "tp_rank": tp_rank,
+                    "tp_size": tp_size,
+                },
+            )
+            rank += 1
+
+    return directory / _FIXED_EXP_NAME
+
+
 def _create_tp_sharded_dumps(
    directory: Path,
    *,