[Model]Fix MiniMaxM2ForCausalLM perf regression (#45935)

Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai>
This commit is contained in:
Jee Jee Li
2026-06-22 00:28:52 +08:00
committed by GitHub
parent 2cac89f9da
commit 745bba5ea8
2 changed files with 223 additions and 24 deletions
+59 -3
View File
@@ -10,8 +10,12 @@ from torch.multiprocessing import spawn
from tests.kernels.utils import opcheck
from tests.utils import ensure_current_vllm_config, init_test_distributed_environment
from vllm.distributed import cleanup_dist_env_and_memory
from vllm.model_executor.layers.minimax_rms_norm import MiniMaxText01RMSNormTP
from vllm.model_executor.layers.minimax_rms_norm import (
MiniMaxText01RMSNormTP,
rms_norm_tp,
)
from vllm.platforms import current_platform
from vllm.triton_utils import HAS_TRITON
from vllm.utils.network_utils import get_open_port
from vllm.utils.torch_utils import set_random_seed
@@ -54,8 +58,19 @@ def _worker_forward_qk(
torch.manual_seed(seed + 1000 + local_rank)
qkv = torch.randn(num_tokens, hq + hk + hk, dtype=dtype, device="cuda")
q_ref, k_ref, v_ref = qkv.clone().split([hq, hk, hk], dim=-1)
ref_q, ref_k = MiniMaxText01RMSNormTP.forward_qk(q_norm, k_norm, q_ref, k_ref)
# Reference: eager all-reduce path. ``forward_qk`` no longer all-reduces
# the variance (it is the tp==1 / already-reduced building block), so the
# multi-rank reference must use the eager path that performs the global
# variance all-reduce, matching the fused kernel below.
ref_q, ref_k = rms_norm_tp._minimax_qk_norm_tp_eager(
qkv.clone(),
q_norm.weight,
k_norm.weight,
hq,
hk,
world_size,
eps,
)
# Set up Lamport workspace.
from vllm.distributed.parallel_state import get_tp_group
@@ -150,3 +165,44 @@ def test_minimax_reduce_rms_qk(
nprocs=world_size,
join=True,
)
@pytest.mark.skipif(
not current_platform.is_cuda() or not HAS_TRITON,
reason="CUDA and Triton required",
)
@pytest.mark.parametrize("num_tokens", [1, 7, 128, 333, 2049])
@pytest.mark.parametrize("hidden_dims", [(3072, 512), (768, 256), (3000, 500)])
@pytest.mark.parametrize("dtype", [torch.bfloat16, torch.float16])
@pytest.mark.parametrize("tp_world", [1, 4, 8])
@pytest.mark.parametrize("eps", [1e-6])
@pytest.mark.parametrize("seed", [42])
def test_minimax_qk_norm_triton_fallback(
monkeypatch, num_tokens, hidden_dims, dtype, tp_world, eps, seed
):
"""Single-GPU check: Triton fallback kernels vs the pure-torch reference.
The all-reduce is a TP communication barrier, so it is monkeypatched to
identity here; both the Triton path and the reference see the same
(patched) reduction. This validates the kernel math and the folded
``/ tp_world`` scaling without needing multiple ranks -- ``hidden_dims``
are the per-rank q/k segment widths.
"""
monkeypatch.setattr(rms_norm_tp, "_all_reduce_variance", lambda v: v)
q_size, kv_size = hidden_dims
device = "cuda"
torch.manual_seed(seed)
qkv = torch.randn(num_tokens, q_size + 2 * kv_size, dtype=dtype, device=device)
q_weight = torch.randn(q_size, dtype=dtype, device=device)
k_weight = torch.randn(kv_size, dtype=dtype, device=device)
q_triton, k_triton = rms_norm_tp._minimax_qk_norm_tp_fallback(
qkv, q_weight, k_weight, q_size, kv_size, 0, tp_world, eps
)
q_ref, k_ref = rms_norm_tp._minimax_qk_norm_tp_eager(
qkv, q_weight, k_weight, q_size, kv_size, tp_world, eps
)
torch.testing.assert_close(q_triton, q_ref, atol=3e-2, rtol=3e-2)
torch.testing.assert_close(k_triton, k_ref, atol=3e-2, rtol=3e-2)