| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from types import SimpleNamespace |
| 5 | |
| 6 | import pytest |
| 7 | import torch |
| 8 | |
| 9 | from vllm.lora.utils import get_supported_lora_modules |
| 10 | from vllm.model_executor.layers.quantization.modelopt import ModelOptNvFp4Config |
| 11 | from vllm.models.inkling.nvidia import moe |
| 12 | from vllm.models.inkling.nvidia.model import _TmlForCausalLMBase |
| 13 | from vllm.platforms import current_platform |
| 14 | |
| 15 | |
| 16 | def test_gate_loads_directly_into_padded_runtime_weight() -> None: |
| 17 | gate = moe.InklingGate( |
| 18 | d_model=4, |
| 19 | n_routed_experts=5, |
| 20 | n_shared_experts=2, |
| 21 | experts_per_token=2, |
| 22 | route_scale=1.0, |
| 23 | ) |
| 24 | loaded = torch.arange(28, dtype=gate.weight.dtype).reshape(7, 4) |
| 25 | |
| 26 | gate.weight.weight_loader(gate.weight, loaded) |
| 27 | |
| 28 | assert gate.weight.shape == (8, 4) |
| 29 | torch.testing.assert_close(gate.weight[:7], loaded) |
| 30 | torch.testing.assert_close(gate.weight[7], torch.zeros(4)) |
| 31 | |
| 32 | |
| 33 | @pytest.mark.skipif(not current_platform.is_cuda(), reason="requires CUDA") |
| 34 | @pytest.mark.parametrize(("num_tokens", "expected_calls"), [(1, 1), (64, 1), (65, 0)]) |
| 35 | def test_gate_uses_ll_bf16_gemm_through_token_limit( |
| 36 | monkeypatch, num_tokens, expected_calls |
| 37 | ) -> None: |
| 38 | gate = moe.InklingGate( |
| 39 | d_model=8, |
| 40 | n_routed_experts=5, |
| 41 | n_shared_experts=2, |
| 42 | experts_per_token=2, |
| 43 | route_scale=1.0, |
| 44 | ).to(device="cuda", dtype=torch.bfloat16) |
| 45 | hidden_states = torch.randn(num_tokens, 8, device="cuda", dtype=torch.bfloat16) |
| 46 | calls = [] |
| 47 | |
| 48 | def fake_ll_bf16_gemm(x, weight): |
| 49 | calls.append((x, weight)) |
| 50 | return torch.ones(num_tokens, 8, device="cuda", dtype=torch.float32) |
| 51 | |
| 52 | monkeypatch.setattr( |
| 53 | moe.current_platform, "has_device_capability", lambda capability: True |
| 54 | ) |
| 55 | monkeypatch.setattr(moe.ll_bf16, "is_available", lambda: True) |
| 56 | monkeypatch.setattr(moe.ll_bf16, "ll_bf16_gemm", fake_ll_bf16_gemm) |
| 57 | |
| 58 | logits = gate.compute_logits(hidden_states) |
| 59 | |
| 60 | assert len(calls) == expected_calls |
| 61 | if calls: |
| 62 | assert calls[0][0] is hidden_states |
| 63 | assert calls[0][1] is gate.weight |
| 64 | assert logits.shape == (num_tokens, 8) |
| 65 | assert logits.dtype == torch.float32 |
| 66 | |
| 67 | |
| 68 | def test_gate_is_not_a_lora_target() -> None: |
| 69 | model = torch.nn.Module() |
| 70 | model.gate = moe.InklingGate( |
| 71 | d_model=4, |
| 72 | n_routed_experts=5, |
| 73 | n_shared_experts=2, |
| 74 | experts_per_token=2, |
| 75 | route_scale=1.0, |
| 76 | ) |
| 77 | |
| 78 | assert "gate" not in get_supported_lora_modules(model) |
| 79 | |
| 80 | |
| 81 | def test_custom_embedding_is_not_a_lora_target() -> None: |
| 82 | model = torch.nn.Module() |
| 83 | model.embedding_modules = _TmlForCausalLMBase.embedding_modules |
| 84 | |
| 85 | supported = get_supported_lora_modules(model) |
| 86 | |
| 87 | assert "embed_tokens" not in supported |
| 88 | assert "lm_head" in supported |
| 89 | |
| 90 | |
| 91 | def test_inkling_mapper_maps_modelopt_exclusions() -> None: |
| 92 | quant_config = ModelOptNvFp4Config.from_config( |
| 93 | { |
| 94 | "quantization": { |
| 95 | "quant_algo": "NVFP4", |
| 96 | "group_size": 16, |
| 97 | "kv_cache_quant_algo": None, |
| 98 | "exclude_modules": [ |
| 99 | "model.llm.layers.2.mlp.experts", |
| 100 | "model.llm.layers.2.mlp.shared_experts", |
| 101 | ], |
| 102 | } |
| 103 | } |
| 104 | ) |
| 105 | |
| 106 | quant_config.apply_vllm_mapper( |
| 107 | _TmlForCausalLMBase.hf_to_vllm_mapper.get_rename_mapper() |
| 108 | ) |
| 109 | |
| 110 | assert quant_config.is_layer_excluded("model.layers.2.mlp.experts") |
| 111 | assert quant_config.is_layer_excluded("model.layers.2.mlp.shared_experts") |
| 112 | assert not quant_config.is_layer_excluded("model.layers.3.mlp.experts") |
| 113 | |
| 114 | |
| 115 | @pytest.mark.parametrize("projection", ["w13", "w2"]) |
| 116 | @pytest.mark.parametrize("nested", [False, True]) |
| 117 | @pytest.mark.parametrize( |
| 118 | "suffix", |
| 119 | [ |
| 120 | "input_global_scale", |
| 121 | "weight_global_scale", |
| 122 | "weight_packed", |
| 123 | "weight_scale", |
| 124 | ], |
| 125 | ) |
| 126 | def test_inkling_mapper_maps_compressed_tensors_expert_params( |
| 127 | projection: str, suffix: str, nested: bool |
| 128 | ) -> None: |
| 129 | projection_param = ( |
| 130 | f"{projection}_weight.{suffix}" if nested else f"{projection}_{suffix}" |
| 131 | ) |
| 132 | source = f"model.llm.layers.2.mlp.experts.{projection_param}" |
| 133 | |
| 134 | mapped = _TmlForCausalLMBase.hf_to_vllm_mapper.apply_list([source]) |
| 135 | |
| 136 | assert mapped == [f"model.layers.2.mlp.experts.{projection}_{suffix}"] |
| 137 | |
| 138 | |
| 139 | @pytest.mark.parametrize(("projection", "amax"), [("w13", 4.375), ("w2", 2960.0)]) |
| 140 | def test_moe_loads_calibrated_input_scale(projection: str, amax: float) -> None: |
| 141 | experts = SimpleNamespace( |
| 142 | w13_input_scale=torch.nn.Parameter(torch.empty(3, 2)), |
| 143 | w2_input_scale=torch.nn.Parameter(torch.empty(3)), |
| 144 | ) |
| 145 | layer = SimpleNamespace(experts=SimpleNamespace(routed_experts=experts)) |
| 146 | |
| 147 | loaded = moe.InklingMoE.load_expert_weight( |
| 148 | layer, |
| 149 | f"experts.{projection}_weight.input_amax", |
| 150 | torch.tensor([amax]), |
| 151 | ) |
| 152 | |
| 153 | scale = getattr(experts, f"{projection}_input_scale") |
| 154 | expected = torch.full_like(scale, amax / (448.0 * 6.0)) |
| 155 | torch.testing.assert_close(scale, expected) |
| 156 | assert loaded == [f"experts.routed_experts.{projection}_input_scale"] |
| 157 | |
| 158 | |
| 159 | @pytest.mark.parametrize("projection", ["w13", "w2"]) |
| 160 | @pytest.mark.parametrize("scale_kind", ["input", "weight"]) |
| 161 | def test_moe_loads_compressed_tensors_global_scale( |
| 162 | projection: str, scale_kind: str |
| 163 | ) -> None: |
| 164 | param = torch.nn.Parameter(torch.empty(3, 2 if projection == "w13" else 1)) |
| 165 | if projection == "w2": |
| 166 | param = torch.nn.Parameter(param.squeeze(1)) |
| 167 | experts = SimpleNamespace( |
| 168 | **{f"{projection}_{scale_kind}_global_scale": param}, |
| 169 | moe_config=SimpleNamespace(moe_parallel_config=SimpleNamespace(tp_rank=0)), |
| 170 | ) |
| 171 | layer = SimpleNamespace( |
| 172 | experts=SimpleNamespace(routed_experts=experts), |
| 173 | _local_expert_slots=lambda: {0: 0, 1: 1, 2: 2}, |
| 174 | ) |
| 175 | checkpoint_scale = torch.tensor([[1.0], [2.0], [3.0]]) |
| 176 | |
| 177 | loaded = moe.InklingMoE.load_expert_weight( |
| 178 | layer, |
| 179 | f"experts.{projection}_{scale_kind}_global_scale", |
| 180 | checkpoint_scale, |
| 181 | ) |
| 182 | |
| 183 | expected = ( |
| 184 | checkpoint_scale.expand_as(param) if param.ndim == 2 else checkpoint_scale[:, 0] |
| 185 | ) |
| 186 | torch.testing.assert_close(param, expected) |
| 187 | assert loaded == [f"experts.routed_experts.{projection}_{scale_kind}_global_scale"] |
| 188 | |
| 189 | |
| 190 | @pytest.mark.parametrize(("projection", "checkpoint_rows"), [("w13", 8), ("w2", 4)]) |
| 191 | def test_moe_loads_channelwise_scale_for_tp( |
| 192 | projection: str, checkpoint_rows: int |
| 193 | ) -> None: |
| 194 | param = torch.nn.Parameter(torch.empty(2, 4, 1)) |
| 195 | experts = SimpleNamespace( |
| 196 | **{f"{projection}_weight_scale": param}, |
| 197 | moe_config=SimpleNamespace(moe_parallel_config=SimpleNamespace(tp_rank=1)), |
| 198 | ) |
| 199 | layer = SimpleNamespace( |
| 200 | experts=SimpleNamespace(routed_experts=experts), |
| 201 | _local_expert_slots=lambda: {0: 0, 2: 1}, |
| 202 | ) |
| 203 | checkpoint_scale = torch.arange(3 * checkpoint_rows).reshape(3, checkpoint_rows, 1) |
| 204 | |
| 205 | loaded = moe.InklingMoE.load_expert_weight( |
| 206 | layer, f"experts.{projection}_weight_scale", checkpoint_scale |
| 207 | ) |
| 208 | |
| 209 | expected = checkpoint_scale[[0, 2]] |
| 210 | if projection == "w13": |
| 211 | expected = expected[:, 4:].reshape(2, 2, 2, 1).transpose(1, 2).flatten(1, 2) |
| 212 | torch.testing.assert_close(param, expected.float()) |
| 213 | assert loaded == [f"experts.routed_experts.{projection}_weight_scale"] |
| 214 | |
| 215 | |
| 216 | def test_sink_down_projection_is_packed_during_load(monkeypatch) -> None: |
| 217 | monkeypatch.setattr(moe, "get_tensor_model_parallel_world_size", lambda: 2) |
| 218 | monkeypatch.setattr(moe, "get_tensor_model_parallel_rank", lambda: 1) |
| 219 | sink = moe.InklingSinkExperts(n_experts=2, d_model=3, d_mlp=8) |
| 220 | loaded = torch.arange(48, dtype=sink.w2_weight.dtype).reshape(2, 3, 8) |
| 221 | |
| 222 | sink.load_weight("w2_weight", loaded) |
| 223 | |
| 224 | expected = loaded[:, :, 4:].permute(1, 0, 2).reshape(3, 8) |
| 225 | torch.testing.assert_close(sink.w2_weight, expected) |
| 226 | |
| 227 | |
| 228 | @pytest.mark.skipif(not current_platform.is_cuda(), reason="requires CUDA") |
| 229 | def test_sink_packed_weight_forward_matches_expert_sum(monkeypatch) -> None: |
| 230 | monkeypatch.setattr(moe, "get_tensor_model_parallel_world_size", lambda: 2) |
| 231 | monkeypatch.setattr(moe, "get_tensor_model_parallel_rank", lambda: 1) |
| 232 | sink = moe.InklingSinkExperts(n_experts=2, d_model=3, d_mlp=8).to( |
| 233 | device="cuda", dtype=torch.bfloat16 |
| 234 | ) |
| 235 | torch.manual_seed(1) |
| 236 | w13 = torch.randn(2, 16, 3, dtype=torch.bfloat16) |
| 237 | w2 = torch.randn(2, 3, 8, dtype=torch.bfloat16) |
| 238 | sink.load_weight("w13_weight", w13) |
| 239 | sink.load_weight("w2_weight", w2) |
| 240 | x = torch.randn(5, 3, device="cuda", dtype=torch.bfloat16) |
| 241 | gammas = torch.randn(5, 2, device="cuda") |
| 242 | |
| 243 | output = sink(x, gammas) |
| 244 | |
| 245 | raw = torch.einsum("td,efd->tef", x, w13[:, 8:].to("cuda")) |
| 246 | hidden = torch.nn.functional.silu(raw[:, :, 0::2].float()) |
| 247 | hidden = (hidden * raw[:, :, 1::2] * gammas[:, :, None]).to(torch.bfloat16) |
| 248 | expected = torch.einsum("tef,edf->td", hidden, w2[:, :, 4:].to("cuda")) |
| 249 | torch.testing.assert_close(output, expected, rtol=0, atol=0) |