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>
This commit is contained in:
2026-04-14 13:02:19 -07:00
parent 787409ced1
commit 357cab525b
20 changed files with 549 additions and 328 deletions
+16 -16
View File
@@ -48,8 +48,8 @@ def test_from_numpy_creates_host_tensor():
assert h._handle is None
# Submit a no-op so run_bench has at least one handle.
torch.zeros((1, 8), dtype="f16",
dp=DPPolicy(sip="replicate", cube="replicate", pe="replicate",
num_sips=1, num_cubes=1, num_pes=1),
dp=DPPolicy(cube="replicate", pe="replicate",
num_cubes=1, num_pes=1),
name="dummy")
_run_with(body)
@@ -63,8 +63,8 @@ def test_copy_and_numpy_single_pe():
a single-PE (no real sharding) tensor."""
def body(torch):
dp = DPPolicy(sip="replicate", cube="replicate", pe="replicate",
num_sips=1, num_cubes=1, num_pes=1)
dp = DPPolicy(cube="replicate", pe="replicate",
num_cubes=1, num_pes=1)
t = torch.zeros((1, 16), dtype="f16", dp=dp, name="t")
src = np.arange(16, dtype=np.float16).reshape(1, 16)
t.copy_(torch.from_numpy(src))
@@ -83,8 +83,8 @@ def test_copy_and_numpy_multi_pe_column_wise():
def body(torch):
n_pe = 8
dp = DPPolicy(sip="replicate", cube="replicate", pe="column_wise",
num_sips=1, num_cubes=1, num_pes=n_pe)
dp = DPPolicy(cube="replicate", pe="column_wise",
num_cubes=1, num_pes=n_pe)
t = torch.zeros((1, n_pe * 4), dtype="f16", dp=dp, name="t")
src = np.arange(n_pe * 4, dtype=np.float16).reshape(1, n_pe * 4)
t.copy_(torch.from_numpy(src))
@@ -107,8 +107,8 @@ def test_copy_and_numpy_multi_cube():
n_pe_per_cube = 8
n_cubes = 2
total = n_cubes * n_pe_per_cube # 16
dp = DPPolicy(sip="replicate", cube="column_wise", pe="column_wise",
num_sips=1, num_cubes=n_cubes)
dp = DPPolicy(cube="column_wise", pe="column_wise",
num_cubes=n_cubes)
t = torch.zeros((1, total * 4), dtype="f16", dp=dp, name="t")
src = np.arange(total * 4, dtype=np.float16).reshape(1, total * 4)
t.copy_(torch.from_numpy(src))
@@ -126,8 +126,8 @@ def test_copy_shape_mismatch_raises():
"""copy_ with mismatched shapes raises ValueError."""
def body(torch):
dp = DPPolicy(sip="replicate", cube="replicate", pe="replicate",
num_sips=1, num_cubes=1, num_pes=1)
dp = DPPolicy(cube="replicate", pe="replicate",
num_cubes=1, num_pes=1)
t = torch.zeros((1, 8), dtype="f16", dp=dp, name="t")
src = np.zeros((1, 16), dtype=np.float16)
with pytest.raises(ValueError, match="copy_ shape mismatch"):
@@ -143,8 +143,8 @@ def test_setitem_getitem_single_pe():
"""Scalar and slice assignment on a single-PE tensor round-trips."""
def body(torch):
dp = DPPolicy(sip="replicate", cube="replicate", pe="replicate",
num_sips=1, num_cubes=1, num_pes=1)
dp = DPPolicy(cube="replicate", pe="replicate",
num_cubes=1, num_pes=1)
t = torch.zeros((1, 8), dtype="f16", dp=dp, name="t")
# Scalar broadcast
@@ -169,8 +169,8 @@ def test_setitem_getitem_multi_pe_shard_aligned():
def body(torch):
n_pe = 8
n_elem = 4 # per shard
dp = DPPolicy(sip="replicate", cube="replicate", pe="column_wise",
num_sips=1, num_cubes=1, num_pes=n_pe)
dp = DPPolicy(cube="replicate", pe="column_wise",
num_cubes=1, num_pes=n_pe)
t = torch.zeros((1, n_pe * n_elem), dtype="f16", dp=dp, name="t")
# Write each shard with its rank value
@@ -197,8 +197,8 @@ def test_setitem_cross_shard_raises():
def body(torch):
n_pe = 4
n_elem = 4
dp = DPPolicy(sip="replicate", cube="replicate", pe="column_wise",
num_sips=1, num_cubes=1, num_pes=n_pe)
dp = DPPolicy(cube="replicate", pe="column_wise",
num_cubes=1, num_pes=n_pe)
t = torch.zeros((1, n_pe * n_elem), dtype="f16", dp=dp, name="t")
with pytest.raises(NotImplementedError, match="spans multiple shards"):
t[0, 2:6] = 1.0 # crosses shard 0 (0:4) and shard 1 (4:8)