ADR-0026: DPPolicy intra-device only + ShardSpec structural coords

DPPolicy no longer carries a cross-SIP axis. SIP-level placement is
solely controlled by torch.ahbm.set_device(rank) (ADR-0024); DPPolicy
itself describes only the cube × PE layout within one SIP. ShardSpec
switches to structural (sip, cube, pe) coordinates; the flat pe_index
field/property is fully removed — silent drift between global-flat and
SIP-local interpretations was a foot-gun flagged by ADR-0024 D11.

Breaking API (explicit TypeError / AttributeError):
- DPPolicy(sip=...) / DPPolicy(num_sips=...) -> TypeError
- ShardSpec.pe_index -> AttributeError
- ShardSpec(pe_index=...) -> TypeError
- resolve_dp_policy now takes target_sip= (required), no num_sips.

Downstream migration:
- PE allocator dict keyed by (sip, cube, pe) tuples, in both
  _ensure_allocators and _free_tensor. deploy_tensor uses tuple lookup.
- _create_tensor passes target_sip=current_sip; post-hoc pe_index
  shifting removed entirely.
- launch._compute_local_shape drops the dp.sip branch.
- Internal resolvers (column_wise / row_wise / replicate / tiled_*)
  return _LocalPeShard (cube-local identifier) instead of ShardSpec —
  resolve_dp_policy lifts them to full structural coords.

Tests:
- New tests/test_adr0026_dppolicy_intra_device.py (12 tests) pins the
  contract end-to-end.
- test_sip_parallel.py rewritten: SIP composition now modeled as two
  resolve_dp_policy(target_sip=...) calls (ADR-0024 launcher style).
- Call-site migration: test_tensor, test_va_integration, test_va_offset,
  test_runtime_api_tensor, test_tl_recv_async, test_ccl_* and benches
  gemm_single_pe, gpt3_qkv, va_offset_verify, ccl_allreduce (legacy
  branch) all use intra-device DPPolicy and structural ShardSpec.

Result: 523 passed, 1 strict xfail (ring_default_ws — unchanged
ADR-0024 Phase B blocker; architectural fix deferred to ADR-0027).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

benches/ccl_allreduce.py

@@ -32,29 +32,26 @@ def _derive_dp(spec: dict, world_size: int) -> DPPolicy:
     """Legacy DPPolicy for world_size > SIP count (rank = flat PE index).
 
     Used only in the ccl.yaml-override path so the existing matrix tests
-    with explicit world_size (8, 16, 7 etc.) keep working. The new
-    ADR-0024 TP path (rank = SIP) uses a per-rank DPPolicy inside the
-    worker instead.
+    with explicit world_size (8, 16, 7 etc.) keep working. ADR-0026:
+    DPPolicy is intra-device only, so this legacy path now always stays
+    within a single SIP and distributes the override world_size across
+    that SIP's cubes and PEs.
     """
-    sips = int(spec["system"]["sips"]["count"])
-    cm = spec["sip"]["cube_mesh"]
     pl = spec["cube"]["pe_layout"]
     pes_per_cube = int(pl["pe_per_corner"]) * len(pl["corners"])
+    cm = spec["sip"]["cube_mesh"]
     cubes_per_sip = int(cm["w"]) * int(cm["h"])
-    total = sips * cubes_per_sip * pes_per_cube
-    if world_size == total:
-        return DPPolicy(sip="column_wise", cube="column_wise", pe="column_wise")
     if world_size <= pes_per_cube:
         return DPPolicy(
-            sip="replicate", cube="replicate", pe="column_wise",
-            num_sips=1, num_cubes=1, num_pes=world_size,
+            cube="replicate", pe="column_wise",
+            num_cubes=1, num_pes=world_size,
         )
     if world_size <= cubes_per_sip * pes_per_cube:
         return DPPolicy(
-            sip="replicate", cube="column_wise", pe="column_wise",
-            num_sips=1, num_cubes=world_size // pes_per_cube,
+            cube="column_wise", pe="column_wise",
+            num_cubes=world_size // pes_per_cube,
         )
-    return DPPolicy(sip="column_wise", cube="column_wise", pe="column_wise")
+    return DPPolicy(cube="column_wise", pe="column_wise")
 
 
 def worker(rank: int, world_size: int, torch) -> None:

benches/gemm_single_pe.py

@@ -3,7 +3,7 @@
 Full host-to-PE pipeline:
   Host → PCIE_EP → IO_CPU → M_CPU → PE_CPU → SchedulerV2 → PE_DMA → HBM
 
-Single PE: num_sips=1, num_cubes=1, num_pes=1 via DPPolicy override.
+Single PE: num_cubes=1, num_pes=1 via DPPolicy override.
 Both operands use tl.ref (HBM-resident); scheduler_v2 tiles and streams
 per-tile DMA internally.
 
@@ -30,7 +30,7 @@ def _gemm_kernel(a_ptr, b_ptr, out_ptr, M, K, N, tl, DTYPE="f16"):
 def run(torch):
     """Run the single-PE GEMM benchmark."""
     dp = DPPolicy(cube="replicate", pe="replicate",
-                  num_sips=1, num_cubes=1, num_pes=1)
+                  num_cubes=1, num_pes=1)
 
     a   = torch.empty((M, K), dtype=DTYPE, dp=dp, name="a")
     b   = torch.empty((K, N), dtype=DTYPE, dp=dp, name="b")

benches/gpt3_qkv.py

@@ -72,12 +72,16 @@ def run(torch):
     K = GPT3_D_MODEL
     N = COLS_PER_PE
 
-    # X: replicated across all PEs
+    # ADR-0026: DPPolicy is intra-device only. For multi-SIP execution the
+    # ADR-0024 launcher calls this bench once per SIP (each worker via
+    # torch.ahbm.set_device(rank)); here the policy describes only the
+    # cube × PE layout within a single SIP.
+    # X: replicated across all PEs within the SIP
     dp_replicate = DPPolicy(cube="replicate", pe="replicate",
-                            num_sips=N_SIPS, num_cubes=N_CUBES, num_pes=N_PE_PER_CUBE)
-    # W_Q/K/V, out_Q/K/V: column-wise sharded across all PEs
+                            num_cubes=N_CUBES, num_pes=N_PE_PER_CUBE)
+    # W_Q/K/V, out_Q/K/V: column-wise sharded across all PEs within the SIP
     dp_sharded = DPPolicy(cube="column_wise", pe="column_wise",
-                          num_sips=N_SIPS, num_cubes=N_CUBES, num_pes=N_PE_PER_CUBE)
+                          num_cubes=N_CUBES, num_pes=N_PE_PER_CUBE)
 
     x     = torch.empty((M, K), dtype=DTYPE, dp=dp_replicate, name="x")
     wq    = torch.empty((K, GPT3_D_MODEL), dtype=DTYPE, dp=dp_sharded, name="wq")

benches/va_offset_verify.py

@@ -1,7 +1,7 @@
 """VA offset verification benchmark.
 
 Verifies that Triton-style base_ptr + pid * stride addressing works correctly
-with full TP sharding (sip/cube/pe all column_wise). Each PE loads its own
+with intra-SIP TP sharding (cube/pe column_wise). Each PE loads its own
 block from a sharded tensor and stores it back.
 
 The kernel uses standard Triton patterns:
@@ -28,7 +28,7 @@ def _copy_kernel(src_ptr, dst_ptr, M, K, tl, DTYPE="f16"):
 
 def run(torch):
     """Run the VA offset verification benchmark with full TP sharding."""
-    dp = DPPolicy(sip="column_wise", cube="column_wise", pe="column_wise")
+    dp = DPPolicy(cube="column_wise", pe="column_wise")
     src = torch.zeros((M, K), dtype=DTYPE, dp=dp, name="src")
     dst = torch.empty((M, K), dtype=DTYPE, dp=dp, name="dst")