ywkang
357cab525b
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>
81 lines
2.9 KiB
Python
81 lines
2.9 KiB
Python
"""Tests for recv_mode='copy_to_dst' (ADR-0023 D9.5)."""
|
|
from __future__ import annotations
|
|
|
|
import numpy as np
|
|
|
|
|
|
def test_recv_copy_to_dst_via_simpy_runner():
|
|
"""Run a kernel that uses tl.recv(..., dst_addr=..., dst_space=...).
|
|
Verify the data is moved to the dst location after recv.
|
|
"""
|
|
import importlib
|
|
|
|
from kernbench.policy.placement.dp import DPPolicy
|
|
from kernbench.runtime_api.bench_runner import run_bench
|
|
from kernbench.runtime_api.types import resolve_device
|
|
from kernbench.sim_engine.engine import GraphEngine
|
|
from kernbench.topology.builder import resolve_topology
|
|
from kernbench.common.pe_commands import TensorHandle
|
|
|
|
def kernel(t_ptr, n_elem, dst_buf_addr, tl):
|
|
rank = tl.program_id(axis=0)
|
|
ws = tl.num_programs(axis=0)
|
|
nbytes = n_elem * 2
|
|
# Each PE sends own data, then recv into a custom dst slot
|
|
current = TensorHandle(
|
|
id="loc", addr=t_ptr + rank * nbytes,
|
|
shape=(n_elem,), dtype="f16",
|
|
nbytes=nbytes, data=None, space="hbm",
|
|
)
|
|
tl.send(dir="E", src=current)
|
|
# copy_to_dst: move into a per-rank scratch HBM addr
|
|
recv = tl.recv(
|
|
dir="W", shape=(n_elem,), dtype="f16",
|
|
dst_addr=dst_buf_addr + rank * nbytes,
|
|
dst_space="hbm",
|
|
)
|
|
# Sanity: recv handle should now point to our dst addr
|
|
assert recv.addr == dst_buf_addr + rank * nbytes
|
|
assert recv.space == "hbm"
|
|
|
|
topo = resolve_topology("topology.yaml")
|
|
|
|
def run(torch):
|
|
plan = torch.install_ipcq(
|
|
algorithm="ring_allreduce_tcm", world_size_override=8,
|
|
)
|
|
a = torch.zeros(
|
|
(1, 8 * 8),
|
|
dtype="f16",
|
|
dp=DPPolicy(
|
|
cube="replicate", pe="column_wise",
|
|
num_cubes=1,
|
|
),
|
|
name="copy_in",
|
|
)
|
|
store = torch.engine.memory_store
|
|
base = a._handle.va_base or a._handle.shards[0].pa
|
|
nbytes = 8 * 2
|
|
for r in range(8):
|
|
store.write("hbm", base + r * nbytes,
|
|
np.full((8,), float(r + 1), dtype=np.float16))
|
|
|
|
# Use a separate dst region (synthetic addresses)
|
|
dst_buf = 0xC0FFEE_0000
|
|
torch.launch("ring_allreduce_tcm", kernel, a, 8, dst_buf)
|
|
|
|
# After the kernel, dst_buf + r*16 should contain rank (r-1)%8's data
|
|
for r in range(8):
|
|
arr = store.read("hbm", dst_buf + r * nbytes, shape=(8,), dtype="f16")
|
|
expected = float(((r - 1) % 8) + 1)
|
|
assert np.allclose(arr, expected), f"rank {r}: got {arr}, expected {expected}"
|
|
|
|
result = run_bench(
|
|
topology=topo, bench_fn=run,
|
|
device=resolve_device("all"),
|
|
engine_factory=lambda t, d: GraphEngine(
|
|
getattr(t, "topology_obj", t), enable_data=True
|
|
),
|
|
)
|
|
assert result.completion.ok
|