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>
97 lines
3.7 KiB
Python
97 lines
3.7 KiB
Python
"""GPT-3 QKV projection benchmark: sharded across PEs via pe_accel_v1.
|
||
|
||
GPT-3 architecture:
|
||
d_model = 12288 (hidden dimension)
|
||
n_heads = 96 (attention heads)
|
||
d_head = 128 (dimension per head)
|
||
|
||
Sharding strategy (column-wise across all PEs):
|
||
X : (seq_len, d_model) -- replicated to all PEs
|
||
W_Q/K/V : (d_model, d_model) -- column-wise sharded across cubes × PEs
|
||
out_Q/K/V: (seq_len, d_model) -- column-wise sharded across cubes × PEs
|
||
|
||
Each PE computes:
|
||
Q_slice = X @ W_Q_slice : (seq_len, d_model) @ (d_model, cols_per_pe) -> (seq_len, cols_per_pe)
|
||
K_slice, V_slice: same
|
||
|
||
PE count is configurable via N_CUBES × N_PE_PER_CUBE (DPPolicy override).
|
||
topology.yaml is unchanged.
|
||
|
||
Run:
|
||
kernbench run gpt3_qkv
|
||
"""
|
||
from kernbench.policy.placement.dp import DPPolicy
|
||
|
||
# -- PE configuration (DPPolicy overrides — does not change topology.yaml) -----
|
||
N_SIPS = 1
|
||
N_CUBES = 16 # cubes per SIP
|
||
N_PE_PER_CUBE = 8 # PEs per cube
|
||
N_PES = N_CUBES * N_PE_PER_CUBE # 128 total
|
||
|
||
# -- GPT-3 architecture -------------------------------------------------------
|
||
GPT3_D_MODEL = 12288
|
||
SEQ_LEN = 32
|
||
COLS_PER_PE = GPT3_D_MODEL // N_PES # 12288 / 128 = 96
|
||
DTYPE = "f16"
|
||
|
||
|
||
def _gpt3_qkv_kernel(x_ptr, wq_ptr, wk_ptr, wv_ptr,
|
||
out_q_ptr, out_k_ptr, out_v_ptr,
|
||
seq_len, d_model, cols_per_pe, tl, DTYPE="f16"):
|
||
"""GPT-3 QKV sharded: each PE uses program_id to index its VA slice."""
|
||
pid = tl.program_id(0)
|
||
bpe = 2 # f16
|
||
|
||
M = int(seq_len)
|
||
K = int(d_model)
|
||
N = int(cols_per_pe)
|
||
|
||
w_slice = K * N * bpe
|
||
out_slice = M * N * bpe
|
||
|
||
x = tl.load(int(x_ptr), shape=(M, K), dtype=DTYPE)
|
||
wq = tl.ref(int(wq_ptr) + pid * w_slice, shape=(K, N), dtype=DTYPE)
|
||
wk = tl.ref(int(wk_ptr) + pid * w_slice, shape=(K, N), dtype=DTYPE)
|
||
wv = tl.ref(int(wv_ptr) + pid * w_slice, shape=(K, N), dtype=DTYPE)
|
||
|
||
hq = tl.composite(op="gemm", a=x, b=wq,
|
||
out_ptr=int(out_q_ptr) + pid * out_slice)
|
||
hk = tl.composite(op="gemm", a=x, b=wk,
|
||
out_ptr=int(out_k_ptr) + pid * out_slice)
|
||
hv = tl.composite(op="gemm", a=x, b=wv,
|
||
out_ptr=int(out_v_ptr) + pid * out_slice)
|
||
|
||
tl.wait(hq)
|
||
tl.wait(hk)
|
||
tl.wait(hv)
|
||
|
||
|
||
def run(torch):
|
||
"""Run the GPT-3 QKV benchmark."""
|
||
M = SEQ_LEN
|
||
K = GPT3_D_MODEL
|
||
N = COLS_PER_PE
|
||
|
||
# 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_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_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")
|
||
wk = torch.empty((K, GPT3_D_MODEL), dtype=DTYPE, dp=dp_sharded, name="wk")
|
||
wv = torch.empty((K, GPT3_D_MODEL), dtype=DTYPE, dp=dp_sharded, name="wv")
|
||
out_q = torch.empty((M, GPT3_D_MODEL), dtype=DTYPE, dp=dp_sharded, name="out_q")
|
||
out_k = torch.empty((M, GPT3_D_MODEL), dtype=DTYPE, dp=dp_sharded, name="out_k")
|
||
out_v = torch.empty((M, GPT3_D_MODEL), dtype=DTYPE, dp=dp_sharded, name="out_v")
|
||
|
||
torch.launch("gpt3_qkv", _gpt3_qkv_kernel,
|
||
x, wq, wk, wv, out_q, out_k, out_v,
|
||
M, K, N)
|