0z5a
1c3515714f
- Replace per-element loads with 128-bit uint4 vectorized loads (8 halves per access), improving every measured shape: q/k/v at M=2 from 6.0us to 5.4us, q_proj speedup 2.28-2.45x, mlp_down at M=4 2.76x, lm_head at M=1 +6-8%
- Extend kernel M support from {1,2,4,8} to all M in 1-8 via new BLOCK_M cases 3,5,6,7, since cuBLAS wmma templates pad small M to 8/16 rows and waste compute
- Keep the auto-dispatch allowlist unchanged: a 64-step greedy-walk probe on the real decode path showed mlp_down (K=6912) divergence at step 1 and argmax flips for every candidate odd-M band, the same noise class already present in the merged M=2/4 entries, so no entry has the stability evidence the gate requires
- Rejected alternatives with measurements: split-K accumulation (k/v shapes regress 6.0us to 9.2us, code removed) and MMA tiles (small M is DRAM-bound at ~1 FLOP/byte vs the ~138 needed)
- Update test_gemv M-rejection case to M=9 and test_linear_dispatch multirow fallback to M=9 for the widened range
Benchmark: 8x L20 (sm_89, CUDA 12.8), single-GPU microbench, 200 iters after 20 warmup, weights L2-resident; q(1536x1536) M=3 8.9->5.3us, kv(256x1536) M=3 8.7->3.0us, down(1536x6912) M=3 53.5->10.3us; full gate 691 passed, test_bf16_gemv_uses_current_stream passes in isolation after GPU contention rerun
186 lines
6.5 KiB
Python
186 lines
6.5 KiB
Python
import logging
|
|
|
|
import pytest
|
|
import torch
|
|
import torch.nn.functional as F
|
|
|
|
from astrai.extension import explain, is_available, linear, op_backend
|
|
from astrai.extension.backend import linear as public_linear
|
|
from astrai.model.components.linear import Linear
|
|
|
|
GEMV_AVAILABLE = (
|
|
torch.cuda.is_available()
|
|
and is_available("bf16_gemv")
|
|
and torch.cuda.get_device_capability() >= (8, 0)
|
|
)
|
|
skip_no_gemv = pytest.mark.skipif(
|
|
not GEMV_AVAILABLE,
|
|
reason="BF16 GEMV requires a built kernel and compute capability 8.0+",
|
|
)
|
|
|
|
|
|
def test_linear_backend_is_public():
|
|
assert linear is public_linear
|
|
|
|
|
|
def test_model_linear_routes_through_backend(monkeypatch):
|
|
sentinel = torch.randn(2, 4)
|
|
|
|
def fake_linear(x, weight, bias):
|
|
assert x.shape == (2, 3)
|
|
assert weight.shape == (4, 3)
|
|
assert bias is None
|
|
return sentinel
|
|
|
|
monkeypatch.setattr("astrai.model.components.linear.linear", fake_linear)
|
|
layer = Linear(3, 4)
|
|
assert layer(torch.randn(2, 3)) is sentinel
|
|
|
|
|
|
def test_cpu_and_training_calls_fall_back_to_torch(monkeypatch):
|
|
monkeypatch.setenv("ASTRAI_GEMV", "1")
|
|
x = torch.randn(2, 8, requires_grad=True)
|
|
weight = torch.randn(4, 8, requires_grad=True)
|
|
actual = linear(x, weight)
|
|
expected = F.linear(x, weight)
|
|
torch.testing.assert_close(actual, expected)
|
|
actual.sum().backward()
|
|
assert x.grad is not None
|
|
assert weight.grad is not None
|
|
|
|
|
|
def test_invalid_mode_warns_and_uses_auto(monkeypatch, caplog):
|
|
monkeypatch.setenv("ASTRAI_GEMV", "invalid-test-mode")
|
|
with caplog.at_level(logging.WARNING):
|
|
trace = explain("linear", torch.randn(1, 8), torch.randn(4, 8))
|
|
assert "using auto" in caplog.text
|
|
assert "=> torch" in trace
|
|
|
|
|
|
@skip_no_gemv
|
|
def test_mode_zero_disables_gemv(monkeypatch):
|
|
monkeypatch.setenv("ASTRAI_GEMV", "0")
|
|
x = torch.randn(1, 1536, device="cuda", dtype=torch.bfloat16)
|
|
weight = torch.randn(1536, 1536, device="cuda", dtype=torch.bfloat16)
|
|
with torch.no_grad():
|
|
assert "=> torch" in explain("linear", x, weight)
|
|
torch.testing.assert_close(linear(x, weight), F.linear(x, weight))
|
|
|
|
|
|
@skip_no_gemv
|
|
def test_mode_one_forces_capable_unmeasured_shape(monkeypatch):
|
|
monkeypatch.setenv("ASTRAI_GEMV", "1")
|
|
x = torch.randn(1, 64, device="cuda", dtype=torch.bfloat16)
|
|
weight = torch.randn(32, 64, device="cuda", dtype=torch.bfloat16)
|
|
with torch.no_grad():
|
|
assert "=> gemv" in explain("linear", x, weight)
|
|
torch.testing.assert_close(
|
|
linear(x, weight), F.linear(x, weight), rtol=0.02, atol=0.25
|
|
)
|
|
|
|
|
|
@skip_no_gemv
|
|
@pytest.mark.parametrize("m", [2, 4, 8])
|
|
def test_mode_one_dispatches_supported_small_batches(monkeypatch, m):
|
|
monkeypatch.setenv("ASTRAI_GEMV", "1")
|
|
x = torch.randn(m, 1536, device="cuda", dtype=torch.bfloat16)
|
|
weight = torch.randn(256, 1536, device="cuda", dtype=torch.bfloat16)
|
|
with torch.no_grad():
|
|
assert "=> gemv" in explain("linear", x, weight)
|
|
torch.testing.assert_close(
|
|
linear(x, weight), F.linear(x, weight), rtol=0.02, atol=0.25
|
|
)
|
|
|
|
|
|
@skip_no_gemv
|
|
def test_auto_m1_falls_back_until_end_to_end_gate_passes(monkeypatch):
|
|
monkeypatch.setenv("ASTRAI_GEMV", "auto")
|
|
x = torch.randn(1, 1536, device="cuda", dtype=torch.bfloat16)
|
|
winning = torch.randn(
|
|
1536,
|
|
1536,
|
|
device="cuda",
|
|
dtype=torch.bfloat16,
|
|
requires_grad=True,
|
|
)
|
|
with torch.no_grad():
|
|
assert "=> torch" in explain("linear", x, winning)
|
|
torch.testing.assert_close(linear(x, winning), F.linear(x, winning))
|
|
|
|
|
|
@skip_no_gemv
|
|
def test_auto_selects_measured_sm89_small_batch_winner(monkeypatch):
|
|
monkeypatch.setenv("ASTRAI_GEMV", "auto")
|
|
x = torch.randn(4, 1536, device="cuda", dtype=torch.bfloat16)
|
|
weight = torch.randn(256, 1536, device="cuda", dtype=torch.bfloat16)
|
|
with torch.no_grad():
|
|
trace = explain("linear", x, weight)
|
|
if torch.cuda.get_device_capability() == (8, 9):
|
|
assert "=> auto_gemv" in trace
|
|
else:
|
|
assert "=> torch" in trace
|
|
torch.testing.assert_close(
|
|
linear(x, weight), F.linear(x, weight), rtol=0.02, atol=0.5
|
|
)
|
|
|
|
|
|
@skip_no_gemv
|
|
@pytest.mark.parametrize(
|
|
"m,n,k",
|
|
[
|
|
(2, 6912, 1536), # up/gate loses at every measured M
|
|
(2, 1536, 6912), # long-K accumulation changed checkpoint greedy output
|
|
(4, 100000, 1536), # LM head misses the 5% M=4 gate
|
|
(4, 1536, 6912), # long-K accumulation changed checkpoint greedy output
|
|
(8, 256, 1536), # remaining M=8 winners miss the 3% end-to-end gate
|
|
],
|
|
)
|
|
def test_auto_rejects_measured_small_batch_losers(monkeypatch, m, n, k):
|
|
monkeypatch.setenv("ASTRAI_GEMV", "auto")
|
|
x = torch.randn(m, k, device="cuda", dtype=torch.bfloat16)
|
|
weight = torch.randn(n, k, device="cuda", dtype=torch.bfloat16)
|
|
with torch.no_grad():
|
|
assert "=> torch" in explain("linear", x, weight)
|
|
|
|
|
|
@skip_no_gemv
|
|
def test_grad_enabled_and_unsupported_multirow_always_fall_back(monkeypatch):
|
|
monkeypatch.setenv("ASTRAI_GEMV", "1")
|
|
x = torch.randn(1, 1536, device="cuda", dtype=torch.bfloat16)
|
|
weight = torch.randn(
|
|
256, 1536, device="cuda", dtype=torch.bfloat16, requires_grad=True
|
|
)
|
|
assert "=> torch" in explain("linear", x, weight)
|
|
with torch.no_grad():
|
|
oversized = torch.randn(9, 1536, device="cuda", dtype=torch.bfloat16)
|
|
assert "=> torch" in explain("linear", oversized, weight)
|
|
|
|
|
|
@skip_no_gemv
|
|
def test_explicit_gemv_selection_respects_capability(monkeypatch):
|
|
monkeypatch.setenv("ASTRAI_GEMV", "0")
|
|
x = torch.randn(1, 1536, device="cuda", dtype=torch.bfloat16)
|
|
weight = torch.randn(256, 1536, device="cuda", dtype=torch.bfloat16)
|
|
with torch.no_grad(), op_backend(linear="gemv"):
|
|
assert "=> gemv" in explain("linear", x, weight)
|
|
torch.testing.assert_close(
|
|
linear(x, weight), F.linear(x, weight), rtol=0.02, atol=0.25
|
|
)
|
|
|
|
|
|
@skip_no_gemv
|
|
def test_dispatched_linear_cuda_graph_replay(monkeypatch):
|
|
monkeypatch.setenv("ASTRAI_GEMV", "1")
|
|
x = torch.randn(1, 1536, device="cuda", dtype=torch.bfloat16)
|
|
weight = torch.randn(256, 1536, device="cuda", dtype=torch.bfloat16)
|
|
with torch.no_grad():
|
|
for _ in range(3):
|
|
linear(x, weight)
|
|
graph = torch.cuda.CUDAGraph()
|
|
with torch.cuda.graph(graph):
|
|
actual = linear(x, weight)
|
|
x.copy_(torch.randn_like(x))
|
|
graph.replay()
|
|
expected = F.linear(x, weight)
|
|
torch.testing.assert_close(actual, expected, rtol=0.02, atol=0.25)
|