ywkang
998cc85762
Major changes:
PE-level IPCQ infrastructure:
- New PE_IPCQ component: ring-buffer control plane with 4-direction
neighbor mapping, head/tail pointers, backpressure (poll/sleep).
- PE_DMA extended with vc_comm channel for IPCQ outbound/inbound DMA,
including in-flight data snapshot (D9) and op_log recording at
outbound time for Phase 2 replay correctness.
- IpcqDmaToken piggyback model: data + metadata travel together,
atomic visibility at receiver (invariant I6).
- Credit return fast path: bottleneck-BW latency, no fabric vc_comm.
Phase 2 data execution (ADR-0020 integration):
- op_log extended: DmaWriteCmd now captures src_space/src_addr for
Phase 2 dma_write copy; ipcq_copy ops recorded at outbound time.
- DataExecutor replays dma_write + ipcq_copy in t_start order.
- Engine._flush_data_phase: incremental cursor-based replay after
each engine.wait() so host reads see post-Phase-2 data.
- KernelRunner Phase 1 writes disabled when op_log is active to
prevent stale data from corrupting the MemoryStore snapshot.
TLContext / kernel API:
- tl.send(dir, src=TensorHandle), tl.recv(dir, shape, dtype),
tl.recv_async, tl.wait(RecvFuture), copy_to_dst mode.
- TensorHandle operator overloading (add/sub/mul/div) via thread-local
active TLContext → MathCmd dispatch through PE_MATH.
- PE-local scratch allocator for math output handles.
- tl.load returns space="hbm" handles for correct Phase 2 addressing.
- Additional math functions: maximum, minimum, fma, clamp, softmax, cdiv.
Unified ccl_allreduce bench (PyTorch-compat host code):
- Single benches/ccl_allreduce.py with run() + worker(rank, ws, torch)
split matching real PyTorch DDP worker pattern.
- torch.distributed facade: init_process_group, get_world_size,
get_rank, get_backend, all_reduce, barrier — only real PyTorch names.
- AhbmCCLBackend: eager install_ipcq at init, all_reduce dispatches
kernel via tensor shard metadata (n_elem from shards[0].nbytes).
- world_size derived from topology spec (sips × cubes × pes_per_cube)
with optional algorithm-level override in ccl.yaml.
Tensor API (PyTorch-compat surface):
- Tensor.numpy(): gather-aware (all shards via VA-based addressing).
- Tensor.copy_(source): scatter from host tensor into sharded target.
- RuntimeContext.from_numpy(arr): host-side staging tensor.
- Tensor.data property fixed to use numpy() (was shards[0]-only).
Algorithm modules moved to src/kernbench/ccl/algorithms/:
- ring_allreduce, mesh_allreduce, tree_allreduce, hello_send.
- Each module exports kernel_args(world_size, n_elem) helper.
- ccl.yaml module paths updated to kernbench.ccl.algorithms.*.
Dead code removed:
- 7 per-variant bench files (ccl_allreduce_{tcm,hbm,sram}, etc.).
- _run_ccl_bench greenlet-per-SIP scheduler.
- benches.loader.is_ccl_bench + run_rank detection.
- benches/ccl/ directory.
Tests:
- New test_ccl_allreduce_matrix.py: 7 parametrized cases
(ring×3 buffers, ring 8/16, mesh 4, tree 7).
- New test_runtime_api_tensor.py: copy_/numpy/from_numpy unit tests.
- Existing tests updated for new import paths + world_size_override.
Docs:
- Korean ccl-author-guide.md and ADR-0023 paths updated.
- New English versions: ccl-author-guide.en.md, ADR-0023.en.md.
502 tests pass.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
137 lines
4.9 KiB
Python
137 lines
4.9 KiB
Python
"""Tests for the pytorch-compat Tensor API extensions.
|
|
|
|
Covers the new ``torch.from_numpy`` factory and ``Tensor.numpy``,
|
|
``Tensor.copy_`` methods used by the unified ``ccl_allreduce`` bench.
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
import numpy as np
|
|
import pytest
|
|
|
|
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
|
|
|
|
|
|
def _engine_factory(topology, device):
|
|
return GraphEngine(getattr(topology, "topology_obj", topology), enable_data=True)
|
|
|
|
|
|
def _run_with(bench_body):
|
|
topo = resolve_topology("topology.yaml")
|
|
return run_bench(
|
|
topology=topo,
|
|
bench_fn=bench_body,
|
|
device=resolve_device("all"),
|
|
engine_factory=_engine_factory,
|
|
)
|
|
|
|
|
|
# ── from_numpy ──────────────────────────────────────────────────────
|
|
|
|
|
|
def test_from_numpy_creates_host_tensor():
|
|
"""torch.from_numpy returns a kernbench Tensor with the array stored
|
|
in its host buffer (not deployed to any PE)."""
|
|
|
|
def body(torch):
|
|
arr = np.arange(8, dtype=np.float16).reshape(1, 8)
|
|
h = torch.from_numpy(arr)
|
|
# Host tensor has shape/dtype matching the array.
|
|
assert h.shape == (1, 8)
|
|
assert h.dtype == "f16"
|
|
# numpy() round-trips the host buffer.
|
|
assert np.array_equal(h.numpy(), arr)
|
|
# No deploy → no real shards.
|
|
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),
|
|
name="dummy")
|
|
|
|
_run_with(body)
|
|
|
|
|
|
# ── single-PE replicated tensor ─────────────────────────────────────
|
|
|
|
|
|
def test_copy_and_numpy_single_pe():
|
|
"""copy_ from a numpy array, then numpy() round-trips correctly on
|
|
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)
|
|
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))
|
|
gathered = t.numpy()
|
|
assert gathered.shape == (1, 16)
|
|
assert np.array_equal(gathered, src)
|
|
|
|
_run_with(body)
|
|
|
|
|
|
# ── multi-PE column-wise sharding (1 cube) ──────────────────────────
|
|
|
|
|
|
def test_copy_and_numpy_multi_pe_column_wise():
|
|
"""copy_ splits across 8 PEs in one cube, numpy() reassembles."""
|
|
|
|
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)
|
|
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))
|
|
gathered = t.numpy()
|
|
assert gathered.shape == (1, n_pe * 4)
|
|
assert np.array_equal(gathered, src)
|
|
# Sanity: there really were 8 shards.
|
|
assert len(t._handle.shards) == n_pe
|
|
|
|
_run_with(body)
|
|
|
|
|
|
# ── multi-cube sharding ─────────────────────────────────────────────
|
|
|
|
|
|
def test_copy_and_numpy_multi_cube():
|
|
"""copy_ across 2 cubes (16 PEs total), numpy() reassembles."""
|
|
|
|
def body(torch):
|
|
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)
|
|
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))
|
|
gathered = t.numpy()
|
|
assert np.array_equal(gathered, src)
|
|
assert len(t._handle.shards) == total
|
|
|
|
_run_with(body)
|
|
|
|
|
|
# ── shape mismatch raises ───────────────────────────────────────────
|
|
|
|
|
|
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)
|
|
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"):
|
|
t.copy_(torch.from_numpy(src))
|
|
|
|
_run_with(body)
|