| 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 | from torch import nn |
| 9 | |
| 10 | from vllm.config import ParallelConfig |
| 11 | from vllm.model_executor.models.utils import WeightsMapper |
| 12 | from vllm.platforms import current_platform |
| 13 | |
| 14 | pytestmark = pytest.mark.skipif( |
| 15 | not current_platform.is_rocm(), reason="ROCm-specific tests" |
| 16 | ) |
| 17 | |
| 18 | |
| 19 | @pytest.mark.parametrize("hash_routing", [False, True], ids=["regular", "hash"]) |
| 20 | @pytest.mark.parametrize("vision", [False, True], ids=["text", "vision"]) |
| 21 | @torch.inference_mode() |
| 22 | def test_rocm_moe_routing_and_shared_experts_match_reference( |
| 23 | hash_routing, vision, monkeypatch, default_vllm_config, dist_init |
| 24 | ) -> None: |
| 25 | """AMD V4 routing, clamped experts and shared output agree with PyTorch.""" |
| 26 | from vllm.forward_context import set_forward_context |
| 27 | from vllm.models.deepseek_v4.amd.model import DeepseekV4MoE |
| 28 | from vllm.models.deepseek_v4.common.mm_preprocess import IMAGE_SENTINEL_BASE_ID |
| 29 | from vllm.utils.torch_utils import set_default_torch_dtype |
| 30 | from vllm.v1.worker.workspace import ( |
| 31 | init_workspace_manager, |
| 32 | is_workspace_manager_initialized, |
| 33 | ) |
| 34 | |
| 35 | monkeypatch.setenv("VLLM_ROCM_USE_AITER_MOE", "0") |
| 36 | monkeypatch.setenv("VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS", "0") |
| 37 | default_vllm_config.kernel_config.moe_backend = "triton" |
| 38 | config = SimpleNamespace( |
| 39 | hidden_size=512, |
| 40 | n_routed_experts=8, |
| 41 | num_experts_per_tok=2, |
| 42 | moe_intermediate_size=512, |
| 43 | swiglu_limit=0.05, |
| 44 | norm_topk_prob=True, |
| 45 | scoring_func="sqrtsoftplus", |
| 46 | routed_scaling_factor=1.5, |
| 47 | num_hash_layers=int(hash_routing), |
| 48 | vocab_size=IMAGE_SENTINEL_BASE_ID + 8, |
| 49 | topk_method="noaux_tc", |
| 50 | vision_n_layers=int(vision), |
| 51 | n_shared_experts=1, |
| 52 | hidden_act="silu", |
| 53 | ) |
| 54 | torch.manual_seed(42) |
| 55 | with torch.device("cuda"), set_default_torch_dtype(torch.bfloat16): |
| 56 | moe = DeepseekV4MoE( |
| 57 | SimpleNamespace( |
| 58 | model_config=SimpleNamespace(hf_config=config), |
| 59 | quant_config=None, |
| 60 | parallel_config=ParallelConfig(), |
| 61 | ), |
| 62 | prefix="model.layers.0.ffn", |
| 63 | ) |
| 64 | for parameter in moe.parameters(): |
| 65 | if parameter.is_floating_point(): |
| 66 | parameter.normal_(std=0.04) |
| 67 | hidden = torch.randn(17, config.hidden_size) * 0.1 |
| 68 | input_ids = torch.arange(17, dtype=torch.int32) |
| 69 | if vision: |
| 70 | # Include both ends of the five-token sentinel interval and the |
| 71 | # first ordinary token above it. |
| 72 | input_ids[1:4] = torch.tensor( |
| 73 | [ |
| 74 | IMAGE_SENTINEL_BASE_ID, |
| 75 | IMAGE_SENTINEL_BASE_ID + 4, |
| 76 | IMAGE_SENTINEL_BASE_ID + 5, |
| 77 | ], |
| 78 | dtype=torch.int32, |
| 79 | ) |
| 80 | if moe.gate.tid2eid is not None: |
| 81 | rows = torch.arange(config.vocab_size) |
| 82 | moe.gate.tid2eid.copy_(torch.stack((rows % 8, (rows + 3) % 8), dim=1)) |
| 83 | if moe.gate.e_score_correction_bias is not None: |
| 84 | moe.gate.e_score_correction_bias.copy_(torch.arange(8) * 0.1) |
| 85 | if moe.gate.bias_vl is not None: |
| 86 | moe.gate.bias_vl.copy_(torch.arange(8) * -0.1) |
| 87 | |
| 88 | scores = torch.nn.functional.softplus( |
| 89 | hidden.float() @ moe.gate.weight.float().T |
| 90 | ).sqrt() |
| 91 | bias = moe.gate.e_score_correction_bias |
| 92 | selection_scores = scores if bias is None else scores + bias |
| 93 | expected_ids = selection_scores.topk(2, dim=-1).indices |
| 94 | if hash_routing: |
| 95 | expected_ids = moe.gate.tid2eid[input_ids.long()].long() |
| 96 | if vision: |
| 97 | image_mask = (input_ids >= IMAGE_SENTINEL_BASE_ID) & ( |
| 98 | input_ids < IMAGE_SENTINEL_BASE_ID + 5 |
| 99 | ) |
| 100 | expected_ids[image_mask] = ( |
| 101 | (scores + moe.gate.bias_vl).topk(2, dim=-1).indices[image_mask] |
| 102 | ) |
| 103 | expected_weights = scores.gather(1, expected_ids) |
| 104 | expected_weights *= config.routed_scaling_factor / expected_weights.sum( |
| 105 | dim=-1, keepdim=True |
| 106 | ) |
| 107 | router_logits, _ = moe.gate(hidden) |
| 108 | weights, ids = moe.experts.router.select_experts( |
| 109 | hidden, router_logits, topk_indices_dtype=torch.int32, input_ids=input_ids |
| 110 | ) |
| 111 | torch.testing.assert_close(ids.long(), expected_ids) |
| 112 | torch.testing.assert_close(weights, expected_weights, rtol=1e-5, atol=1e-6) |
| 113 | |
| 114 | def expert_reference(x, gate_up_weight, down_weight): |
| 115 | gate, up = ( |
| 116 | torch.nn.functional.linear(x, gate_up_weight).float().chunk(2, dim=-1) |
| 117 | ) |
| 118 | gate = gate.clamp(max=config.swiglu_limit) |
| 119 | up = up.clamp(-config.swiglu_limit, config.swiglu_limit) |
| 120 | activated = (torch.nn.functional.silu(gate) * up).to(x.dtype) |
| 121 | return torch.nn.functional.linear(activated, down_weight).float() |
| 122 | |
| 123 | routed = moe.experts.routed_experts |
| 124 | expected = torch.zeros_like(hidden, dtype=torch.float32) |
| 125 | for expert_id in range(config.n_routed_experts): |
| 126 | token, slot = torch.where(expected_ids == expert_id) |
| 127 | expected[token] += ( |
| 128 | expert_reference( |
| 129 | hidden[token], routed.w13_weight[expert_id], routed.w2_weight[expert_id] |
| 130 | ) |
| 131 | * expected_weights[token, slot, None] |
| 132 | ) |
| 133 | shared = moe.shared_experts |
| 134 | expected += expert_reference( |
| 135 | hidden, shared.gate_up_proj.weight, shared.down_proj.weight |
| 136 | ) |
| 137 | moe.experts._quant_method.process_weights_after_loading(routed) |
| 138 | if not is_workspace_manager_initialized(): |
| 139 | init_workspace_manager(torch.device("cuda:0")) |
| 140 | with set_forward_context(None, default_vllm_config, num_tokens=hidden.shape[0]): |
| 141 | actual = moe(hidden, input_ids) |
| 142 | torch.testing.assert_close(actual.float(), expected, rtol=0.02, atol=5e-5) |
| 143 | |
| 144 | |
| 145 | def test_rocm_packed_kv_cache_auto_uses_ds_mla_layout() -> None: |
| 146 | from vllm.config import CacheConfig |
| 147 | from vllm.models.deepseek_v4.attention import _resolve_dsv4_kv_cache_dtype |
| 148 | |
| 149 | cache_config = CacheConfig() |
| 150 | |
| 151 | resolved_dtype, torch_dtype = _resolve_dsv4_kv_cache_dtype( |
| 152 | use_fp8_ds_mla_layout=True, |
| 153 | kv_cache_dtype=cache_config.cache_dtype, |
| 154 | cache_config=cache_config, |
| 155 | ) |
| 156 | |
| 157 | assert resolved_dtype == "fp8_ds_mla" |
| 158 | assert torch_dtype is torch.uint8 |
| 159 | assert cache_config.cache_dtype == "fp8_ds_mla" |
| 160 | |
| 161 | |
| 162 | def test_rocm_packed_kv_cache_rejects_unquantized_dtype() -> None: |
| 163 | from vllm.config import CacheConfig |
| 164 | from vllm.models.deepseek_v4.attention import _resolve_dsv4_kv_cache_dtype |
| 165 | |
| 166 | cache_config = CacheConfig(cache_dtype="bfloat16") |
| 167 | |
| 168 | with pytest.raises(ValueError, match="only supports fp8 kv-cache"): |
| 169 | _resolve_dsv4_kv_cache_dtype( |
| 170 | use_fp8_ds_mla_layout=True, |
| 171 | kv_cache_dtype=cache_config.cache_dtype, |
| 172 | cache_config=cache_config, |
| 173 | ) |
| 174 | |
| 175 | |
| 176 | def test_vl_mapper_preserves_rocm_weight_mapping() -> None: |
| 177 | from vllm.models.deepseek_v4.amd.model import _make_deepseek_v4_weights_mapper |
| 178 | from vllm.models.deepseek_v4.common.vl_model import ( |
| 179 | _make_deepseek_v4_vl_weights_mapper, |
| 180 | ) |
| 181 | |
| 182 | text_mapper = _make_deepseek_v4_weights_mapper("fp4", fuse_shared_experts=True) |
| 183 | mapper = _make_deepseek_v4_vl_weights_mapper(text_mapper, image_enabled=True) |
| 184 | |
| 185 | assert mapper._map_name("layers.3.attn.wq_a.input_scale") == ( |
| 186 | "language_model.model.layers.3.attn.wq_a.input_scale_2" |
| 187 | ) |
| 188 | assert mapper._map_name("layers.3.ffn.shared_experts.w2.weight") == ( |
| 189 | "language_model.model.layers.3.ffn.shared_experts.w2.weight" |
| 190 | ) |
| 191 | assert mapper._map_name("head.weight") == "language_model.lm_head.weight" |
| 192 | |
| 193 | |
| 194 | def test_rocm_moe_wires_vision_routing_on_hash_and_regular_layers( |
| 195 | monkeypatch: pytest.MonkeyPatch, |
| 196 | ) -> None: |
| 197 | from vllm.models.deepseek_v4.amd import model as rocm_model |
| 198 | from vllm.models.deepseek_v4.common.mm_preprocess import IMAGE_SENTINEL_BASE_ID |
| 199 | |
| 200 | captured: list[dict] = [] |
| 201 | |
| 202 | class FakeGate(nn.Module): |
| 203 | def __init__(self, **kwargs) -> None: |
| 204 | super().__init__() |
| 205 | |
| 206 | def fake_factory(**kwargs): |
| 207 | captured.append(kwargs) |
| 208 | return nn.Identity() |
| 209 | |
| 210 | monkeypatch.setattr(rocm_model, "GateLinear", FakeGate) |
| 211 | monkeypatch.setattr(rocm_model, "FusedMoEFactory", fake_factory) |
| 212 | monkeypatch.setattr(rocm_model, "get_tensor_model_parallel_world_size", lambda: 1) |
| 213 | monkeypatch.setattr(rocm_model, "get_tensor_model_parallel_rank", lambda: 0) |
| 214 | |
| 215 | config = SimpleNamespace( |
| 216 | hidden_size=16, |
| 217 | n_routed_experts=4, |
| 218 | num_experts_per_tok=2, |
| 219 | moe_intermediate_size=8, |
| 220 | swiglu_limit=None, |
| 221 | norm_topk_prob=True, |
| 222 | scoring_func="sqrtsoftplus", |
| 223 | num_hash_layers=1, |
| 224 | vocab_size=32, |
| 225 | topk_method="noaux_tc", |
| 226 | vision_n_layers=1, |
| 227 | n_shared_experts=None, |
| 228 | ) |
| 229 | vllm_config = SimpleNamespace( |
| 230 | model_config=SimpleNamespace(hf_config=config), |
| 231 | quant_config=None, |
| 232 | parallel_config=ParallelConfig(), |
| 233 | ) |
| 234 | |
| 235 | hash_moe = rocm_model.DeepseekV4MoE(vllm_config, prefix="model.layers.0.ffn") |
| 236 | regular_moe = rocm_model.DeepseekV4MoE(vllm_config, prefix="model.layers.1.ffn") |
| 237 | |
| 238 | assert hash_moe.gate.tid2eid is not None |
| 239 | assert regular_moe.gate.tid2eid is None |
| 240 | for moe, factory_kwargs in zip((hash_moe, regular_moe), captured, strict=True): |
| 241 | assert moe.gate.e_score_correction_bias is not None |
| 242 | assert moe.gate.bias_vl is not None |
| 243 | assert factory_kwargs["bias_vl"] is moe.gate.bias_vl |
| 244 | assert factory_kwargs["image_sentinel_lo"] == IMAGE_SENTINEL_BASE_ID |
| 245 | |
| 246 | |
| 247 | def test_rocm_fse_rejects_data_parallel_deployment( |
| 248 | monkeypatch: pytest.MonkeyPatch, |
| 249 | ) -> None: |
| 250 | from vllm.models.deepseek_v4.amd.model import _fuse_shared_experts_enabled |
| 251 | |
| 252 | config = SimpleNamespace(n_shared_experts=1) |
| 253 | |
| 254 | monkeypatch.setenv("VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS", "1") |
| 255 | with pytest.raises(ValueError, match="data parallelism"): |
| 256 | _fuse_shared_experts_enabled(config, ParallelConfig(data_parallel_size=2)) |
| 257 | |
| 258 | # Expert parallelism already disables FSE: no raise, FSE stays off. |
| 259 | assert not _fuse_shared_experts_enabled( |
| 260 | config, ParallelConfig(data_parallel_size=2, enable_expert_parallel=True) |
| 261 | ) |
| 262 | assert _fuse_shared_experts_enabled(config, ParallelConfig()) |
| 263 | |
| 264 | |
| 265 | def test_rocm_mtp_forwards_input_ids_for_vision_routing( |
| 266 | monkeypatch: pytest.MonkeyPatch, |
| 267 | ) -> None: |
| 268 | from vllm.models.deepseek_v4.amd import mtp as rocm_mtp |
| 269 | |
| 270 | hidden_size = 4 |
| 271 | hc_mult = 2 |
| 272 | |
| 273 | class FakeNorm(nn.Module): |
| 274 | def __init__(self) -> None: |
| 275 | super().__init__() |
| 276 | self.weight = nn.Parameter(torch.ones(hidden_size)) |
| 277 | self.variance_epsilon = 1e-6 |
| 278 | |
| 279 | class FakeMTPBlock(nn.Module): |
| 280 | use_fused_mhc = False |
| 281 | |
| 282 | def __init__(self) -> None: |
| 283 | super().__init__() |
| 284 | self.seen_input_ids: torch.Tensor | None = None |
| 285 | |
| 286 | def forward( |
| 287 | self, |
| 288 | *, |
| 289 | positions: torch.Tensor, |
| 290 | x: torch.Tensor, |
| 291 | input_ids: torch.Tensor | None, |
| 292 | ): |
| 293 | self.seen_input_ids = input_ids |
| 294 | return x, None, None, None |
| 295 | |
| 296 | def passthrough_mtp_input( |
| 297 | inputs_embeds: torch.Tensor, |
| 298 | positions: torch.Tensor, |
| 299 | previous_hidden_states: torch.Tensor, |
| 300 | *args, |
| 301 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 302 | return inputs_embeds, previous_hidden_states |
| 303 | |
| 304 | monkeypatch.setattr( |
| 305 | rocm_mtp, "_FUSED_MTP_INPUT_RMSNORM_KERNEL", passthrough_mtp_input |
| 306 | ) |
| 307 | |
| 308 | layer = object.__new__(rocm_mtp.DeepSeekV4MultiTokenPredictorLayer) |
| 309 | nn.Module.__init__(layer) |
| 310 | layer.config = SimpleNamespace(hidden_size=hidden_size) |
| 311 | layer.hc_mult = hc_mult |
| 312 | layer.enorm = FakeNorm() |
| 313 | layer.hnorm = FakeNorm() |
| 314 | layer.e_proj = nn.Identity() |
| 315 | layer.h_proj = nn.Identity() |
| 316 | layer.mtp_block = FakeMTPBlock() |
| 317 | |
| 318 | input_ids = torch.tensor([11, 12]) |
| 319 | positions = torch.tensor([3, 4]) |
| 320 | inputs_embeds = torch.arange(8, dtype=torch.float32).view(2, hidden_size) |
| 321 | previous_hidden_states = torch.arange(16, dtype=torch.float32).view(2, -1) |
| 322 | |
| 323 | output = layer( |
| 324 | input_ids, |
| 325 | positions, |
| 326 | previous_hidden_states, |
| 327 | inputs_embeds, |
| 328 | ) |
| 329 | |
| 330 | assert layer.mtp_block.seen_input_ids is input_ids |
| 331 | expected = previous_hidden_states.view(2, hc_mult, hidden_size) |
| 332 | expected = expected + inputs_embeds.unsqueeze(-2) |
| 333 | torch.testing.assert_close(output, expected.flatten(1)) |
| 334 | |
| 335 | |
| 336 | def test_rocm_compute_logits_local_skips_gather() -> None: |
| 337 | from vllm.models.deepseek_v4.amd.model import DeepseekV4ForCausalLM |
| 338 | |
| 339 | calls: list[tuple[nn.Module, torch.Tensor, bool]] = [] |
| 340 | |
| 341 | def logits_processor( |
| 342 | lm_head: nn.Module, hidden_states: torch.Tensor, *, skip_gather: bool = False |
| 343 | ) -> torch.Tensor: |
| 344 | calls.append((lm_head, hidden_states, skip_gather)) |
| 345 | return hidden_states + 1 |
| 346 | |
| 347 | model = object.__new__(DeepseekV4ForCausalLM) |
| 348 | nn.Module.__init__(model) |
| 349 | model.lm_head = nn.Identity() |
| 350 | model.logits_processor = logits_processor |
| 351 | hidden_states = torch.tensor([4.0]) |
| 352 | |
| 353 | result = model.compute_logits_local(hidden_states) |
| 354 | |
| 355 | assert torch.equal(result, torch.tensor([5.0])) |
| 356 | assert calls == [(model.lm_head, hidden_states, True)] |
| 357 | |
| 358 | |
| 359 | class _FakeLanguageModel(nn.Module): |
| 360 | finalizes_weights_during_load = False |
| 361 | |
| 362 | def __init__(self) -> None: |
| 363 | super().__init__() |
| 364 | self.tensor_a = nn.Parameter(torch.zeros(1)) |
| 365 | self.tensor_c = nn.Parameter(torch.zeros(1)) |
| 366 | self.finalized_values: list[tuple[float, float]] = [] |
| 367 | |
| 368 | def process_weights_after_loading(self) -> None: |
| 369 | self.finalized_values.append((self.tensor_a.item(), self.tensor_c.item())) |
| 370 | |
| 371 | def compute_logits_local(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 372 | return hidden_states + 1 |
| 373 | |
| 374 | |
| 375 | def test_vl_wrapper_streams_then_delegates_finalization() -> None: |
| 376 | from vllm.models.deepseek_v4.common.vl_model import ( |
| 377 | DeepseekV4ForConditionalGeneration, |
| 378 | ) |
| 379 | |
| 380 | model = object.__new__(DeepseekV4ForConditionalGeneration) |
| 381 | nn.Module.__init__(model) |
| 382 | model.language_model = _FakeLanguageModel() |
| 383 | model.vision = nn.Module() |
| 384 | model.vision.tensor_b = nn.Parameter(torch.zeros(1)) |
| 385 | model.hf_to_vllm_mapper = WeightsMapper() |
| 386 | |
| 387 | def interleaved_weights(): |
| 388 | yield "language_model.tensor_a", torch.tensor([1.0]) |
| 389 | assert model.language_model.tensor_a.item() == 1.0 |
| 390 | yield "vision.tensor_b", torch.tensor([2.0]) |
| 391 | assert model.vision.tensor_b.item() == 2.0 |
| 392 | yield "language_model.tensor_c", torch.tensor([3.0]) |
| 393 | |
| 394 | loaded = model.load_weights(interleaved_weights()) |
| 395 | |
| 396 | assert loaded == { |
| 397 | "language_model.tensor_a", |
| 398 | "vision.tensor_b", |
| 399 | "language_model.tensor_c", |
| 400 | } |
| 401 | assert model.language_model.finalized_values == [] |
| 402 | |
| 403 | model.process_weights_after_loading() |
| 404 | |
| 405 | assert model.language_model.finalized_values == [(1.0, 3.0)] |
| 406 | assert torch.equal( |
| 407 | model.compute_logits_local(torch.tensor([4.0])), torch.tensor([5.0]) |
| 408 | ) |
| 409 | model.process_weights_after_loading() |
| 410 | assert model.language_model.finalized_values == [(1.0, 3.0)] |
| 411 | |
| 412 | |
| 413 | class _FakeFinalizingLanguageModel(_FakeLanguageModel): |
| 414 | finalizes_weights_during_load = True |
| 415 | |
| 416 | def __init__(self) -> None: |
| 417 | super().__init__() |
| 418 | self.load_calls = 0 |
| 419 | |
| 420 | def load_weights(self, weights) -> set[str]: |
| 421 | self.load_calls += 1 |
| 422 | loaded = set() |
| 423 | for name, value in weights: |
| 424 | getattr(self, name).data.copy_(value) |
| 425 | loaded.add(name) |
| 426 | self.process_weights_after_loading() |
| 427 | return loaded |
| 428 | |
| 429 | |
| 430 | def test_vl_wrapper_groups_child_that_finalizes_during_load() -> None: |
| 431 | from vllm.models.deepseek_v4.common.vl_model import ( |
| 432 | DeepseekV4ForConditionalGeneration, |
| 433 | ) |
| 434 | |
| 435 | model = object.__new__(DeepseekV4ForConditionalGeneration) |
| 436 | nn.Module.__init__(model) |
| 437 | model.language_model = _FakeFinalizingLanguageModel() |
| 438 | model.vision = nn.Module() |
| 439 | model.vision.tensor_b = nn.Parameter(torch.zeros(1)) |
| 440 | model.hf_to_vllm_mapper = WeightsMapper() |
| 441 | |
| 442 | loaded = model.load_weights( |
| 443 | iter( |
| 444 | ( |
| 445 | ("language_model.tensor_a", torch.tensor([1.0])), |
| 446 | ("vision.tensor_b", torch.tensor([2.0])), |
| 447 | ("language_model.tensor_c", torch.tensor([3.0])), |
| 448 | ) |
| 449 | ) |
| 450 | ) |
| 451 | |
| 452 | assert loaded == { |
| 453 | "language_model.tensor_a", |
| 454 | "vision.tensor_b", |
| 455 | "language_model.tensor_c", |
| 456 | } |
| 457 | assert model.language_model.load_calls == 1 |
| 458 | assert model.language_model.finalized_values == [(1.0, 3.0)] |
| 459 | |
| 460 | # The framework's later model-level hook must not double-finalize a child |
| 461 | # which already completed this work in load_weights. |
| 462 | model.process_weights_after_loading() |
| 463 | assert model.language_model.finalized_values == [(1.0, 3.0)] |
| 464 | |
| 465 | |
| 466 | def test_vl_wrapper_dummy_load_delegates_finalization() -> None: |
| 467 | from vllm.models.deepseek_v4.common.vl_model import ( |
| 468 | DeepseekV4ForConditionalGeneration, |
| 469 | ) |
| 470 | |
| 471 | model = object.__new__(DeepseekV4ForConditionalGeneration) |
| 472 | nn.Module.__init__(model) |
| 473 | model.language_model = _FakeFinalizingLanguageModel() |
| 474 | |
| 475 | # DummyModelLoader bypasses model.load_weights(), so no finalized marker |
| 476 | # exists and the framework-level hook must still delegate to the child. |
| 477 | model.process_weights_after_loading() |
| 478 | assert model.language_model.finalized_values == [(0.0, 0.0)] |
| 479 | model.process_weights_after_loading() |
| 480 | assert model.language_model.finalized_values == [(0.0, 0.0)] |