| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """GLM-5.3-Flash sequence-parallel MoE layout. |
| 4 | |
| 5 | With DP > 1, TP > 1 and expert parallelism, ``Glm5NextModel`` shards the token |
| 6 | dimension across the TP group once at the model entry and every layer runs its |
| 7 | MLP on that shard. A module that still does tensor-parallel collectives there |
| 8 | sums the partial results of *different* tokens held by different ranks, so the |
| 9 | dense MLP of the first ``first_k_dense_replace`` layers must hold replicated |
| 10 | weights and run without any collective, like the shared experts already do. |
| 11 | """ |
| 12 | |
| 13 | from types import SimpleNamespace |
| 14 | |
| 15 | import pytest |
| 16 | import torch |
| 17 | from torch import nn |
| 18 | from transformers import Glm5NextTextConfig |
| 19 | |
| 20 | from vllm.model_executor import parameter |
| 21 | from vllm.model_executor.layers import linear |
| 22 | from vllm.models.glm5next.common import model as glm_model |
| 23 | |
| 24 | HIDDEN, INTERMEDIATE, TP_SIZE = 8, 16, 2 |
| 25 | |
| 26 | |
| 27 | class _Attention(nn.Module): |
| 28 | def __init__(self, *args, **kwargs) -> None: |
| 29 | super().__init__() |
| 30 | self.o_proj = SimpleNamespace(reduce_results=True) |
| 31 | |
| 32 | |
| 33 | def _fake_tensor_parallel_world(monkeypatch) -> None: |
| 34 | """Pretend to be rank 1 of a TP=2 group whose all-reduce must never run.""" |
| 35 | |
| 36 | def _forbidden_all_reduce(*args, **kwargs): |
| 37 | raise AssertionError("tensor-parallel all-reduce on a sequence-parallel shard") |
| 38 | |
| 39 | for module in (linear, parameter): |
| 40 | monkeypatch.setattr( |
| 41 | module, "get_tensor_model_parallel_world_size", lambda: TP_SIZE |
| 42 | ) |
| 43 | monkeypatch.setattr(module, "get_tensor_model_parallel_rank", lambda: 1) |
| 44 | monkeypatch.setattr( |
| 45 | linear, "tensor_model_parallel_all_reduce", _forbidden_all_reduce |
| 46 | ) |
| 47 | |
| 48 | |
| 49 | def _reference_mlp(mlp: glm_model.Glm5NextMLP, x: torch.Tensor) -> torch.Tensor: |
| 50 | gate_up = x @ mlp.gate_up_proj.weight.T |
| 51 | gate, up = gate_up.split(INTERMEDIATE, dim=-1) |
| 52 | return (torch.nn.functional.silu(gate) * up) @ mlp.down_proj.weight.T |
| 53 | |
| 54 | |
| 55 | def _assert_runs_locally_on_shard(mlp: glm_model.Glm5NextMLP) -> None: |
| 56 | device = "cuda" if torch.cuda.is_available() else "cpu" |
| 57 | mlp.to(device) |
| 58 | assert mlp.gate_up_proj.weight.shape == (2 * INTERMEDIATE, HIDDEN) |
| 59 | assert mlp.down_proj.weight.shape == (HIDDEN, INTERMEDIATE) |
| 60 | torch.manual_seed(0) |
| 61 | for param in mlp.parameters(): |
| 62 | param.data.normal_() |
| 63 | shard = torch.randn(3, HIDDEN, device=device) |
| 64 | torch.testing.assert_close(mlp(shard), _reference_mlp(mlp, shard)) |
| 65 | |
| 66 | |
| 67 | def test_dense_mlp_runs_locally_on_the_sequence_parallel_shard( |
| 68 | monkeypatch, default_vllm_config |
| 69 | ): |
| 70 | _fake_tensor_parallel_world(monkeypatch) |
| 71 | |
| 72 | sharded = glm_model.Glm5NextMLP(HIDDEN, INTERMEDIATE, "silu") |
| 73 | assert sharded.down_proj.tp_size == TP_SIZE |
| 74 | assert sharded.gate_up_proj.weight.shape == (2 * INTERMEDIATE // TP_SIZE, HIDDEN) |
| 75 | |
| 76 | _assert_runs_locally_on_shard( |
| 77 | glm_model.Glm5NextMLP(HIDDEN, INTERMEDIATE, "silu", is_sequence_parallel=True) |
| 78 | ) |
| 79 | |
| 80 | |
| 81 | @pytest.mark.parametrize("use_sequence_parallel_moe", [True, False]) |
| 82 | def test_dense_layer_mlp_follows_the_sequence_parallel_layout( |
| 83 | monkeypatch, default_vllm_config, use_sequence_parallel_moe: bool |
| 84 | ): |
| 85 | _fake_tensor_parallel_world(monkeypatch) |
| 86 | monkeypatch.setattr(glm_model, "Glm5NextLinearAttention", _Attention) |
| 87 | monkeypatch.setattr(glm_model, "Glm5NextMLAAttention", _Attention) |
| 88 | |
| 89 | config = Glm5NextTextConfig( |
| 90 | hidden_size=HIDDEN, |
| 91 | intermediate_size=INTERMEDIATE, |
| 92 | num_hidden_layers=2, |
| 93 | first_k_dense_replace=1, |
| 94 | n_routed_experts=4, |
| 95 | layer_types=["linear_attention", "deepseek_sparse_attention"], |
| 96 | mhc=False, |
| 97 | ) |
| 98 | vllm_config = SimpleNamespace( |
| 99 | cache_config=None, |
| 100 | quant_config=None, |
| 101 | parallel_config=SimpleNamespace( |
| 102 | use_sequence_parallel_moe=use_sequence_parallel_moe |
| 103 | ), |
| 104 | kernel_config=SimpleNamespace(enable_jit_warmup=False), |
| 105 | ) |
| 106 | |
| 107 | layer = glm_model.Glm5NextDecoderLayer(vllm_config, config, layer_idx=0) |
| 108 | |
| 109 | assert isinstance(layer.mlp, glm_model.Glm5NextMLP) |
| 110 | assert layer.self_attn.o_proj.reduce_results is not use_sequence_parallel_moe |
| 111 | if use_sequence_parallel_moe: |
| 112 | _assert_runs_locally_on_shard(layer.mlp) |
| 113 | else: |
| 114 | assert layer.mlp.down_proj.tp_size == TP_SIZE |