| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import tempfile |
| 5 | from unittest.mock import Mock |
| 6 | |
| 7 | import huggingface_hub.constants |
| 8 | import pytest |
| 9 | from huggingface_hub.utils import LocalEntryNotFoundError |
| 10 | |
| 11 | from vllm.model_executor.model_loader.weight_utils import ( |
| 12 | download_weights_from_hf, |
| 13 | drop_checkpoint_cache, |
| 14 | maybe_remap_kv_scale_name, |
| 15 | ) |
| 16 | |
| 17 | |
| 18 | def test_download_weights_from_hf(): |
| 19 | with tempfile.TemporaryDirectory() as tmpdir: |
| 20 | # assert LocalEntryNotFoundError error is thrown |
| 21 | # if offline is set and model is not cached |
| 22 | huggingface_hub.constants.HF_HUB_OFFLINE = True |
| 23 | with pytest.raises(LocalEntryNotFoundError): |
| 24 | download_weights_from_hf( |
| 25 | "facebook/opt-125m", |
| 26 | allow_patterns=["*.safetensors", "*.bin"], |
| 27 | cache_dir=tmpdir, |
| 28 | ) |
| 29 | |
| 30 | # download the model |
| 31 | huggingface_hub.constants.HF_HUB_OFFLINE = False |
| 32 | download_weights_from_hf( |
| 33 | "facebook/opt-125m", |
| 34 | allow_patterns=["*.safetensors", "*.bin"], |
| 35 | cache_dir=tmpdir, |
| 36 | ) |
| 37 | |
| 38 | # now it should work offline |
| 39 | huggingface_hub.constants.HF_HUB_OFFLINE = True |
| 40 | assert ( |
| 41 | download_weights_from_hf( |
| 42 | "facebook/opt-125m", |
| 43 | allow_patterns=["*.safetensors", "*.bin"], |
| 44 | cache_dir=tmpdir, |
| 45 | ) |
| 46 | is not None |
| 47 | ) |
| 48 | |
| 49 | |
| 50 | class TestMaybeRemapKvScaleName: |
| 51 | """Tests for maybe_remap_kv_scale_name covering all checkpoint formats.""" |
| 52 | |
| 53 | PARAMS_DICT = { |
| 54 | "model.layers.0.self_attn.attn.k_scale": None, |
| 55 | "model.layers.0.self_attn.attn.v_scale": None, |
| 56 | "model.layers.0.self_attn.attn.q_scale": None, |
| 57 | "model.layers.0.self_attn.qkv_proj.weight": None, |
| 58 | } |
| 59 | |
| 60 | def test_qkv_proj_k_scale(self): |
| 61 | """Qwen3-MoE / llm-compressor format: qkv_proj.k_scale -> attn.k_scale |
| 62 | Regression test for https://github.com/vllm-project/vllm/issues/25047""" |
| 63 | result = maybe_remap_kv_scale_name( |
| 64 | "model.layers.0.self_attn.qkv_proj.k_scale", self.PARAMS_DICT |
| 65 | ) |
| 66 | assert result == "model.layers.0.self_attn.attn.k_scale" |
| 67 | |
| 68 | def test_qkv_proj_v_scale(self): |
| 69 | """Qwen3-MoE / llm-compressor format: qkv_proj.v_scale -> attn.v_scale |
| 70 | Regression test for https://github.com/vllm-project/vllm/issues/25047""" |
| 71 | result = maybe_remap_kv_scale_name( |
| 72 | "model.layers.0.self_attn.qkv_proj.v_scale", self.PARAMS_DICT |
| 73 | ) |
| 74 | assert result == "model.layers.0.self_attn.attn.v_scale" |
| 75 | |
| 76 | def test_modelopt_k_proj_k_scale(self): |
| 77 | """ModelOpt format: k_proj.k_scale -> attn.k_scale.""" |
| 78 | result = maybe_remap_kv_scale_name( |
| 79 | "model.layers.0.self_attn.k_proj.k_scale", self.PARAMS_DICT |
| 80 | ) |
| 81 | assert result == "model.layers.0.self_attn.attn.k_scale" |
| 82 | |
| 83 | def test_modelopt_v_proj_v_scale(self): |
| 84 | """ModelOpt format: v_proj.v_scale -> attn.v_scale.""" |
| 85 | result = maybe_remap_kv_scale_name( |
| 86 | "model.layers.0.self_attn.v_proj.v_scale", self.PARAMS_DICT |
| 87 | ) |
| 88 | assert result == "model.layers.0.self_attn.attn.v_scale" |
| 89 | |
| 90 | def test_deprecated_kv_scale(self): |
| 91 | """Old format: kv_scale -> attn.k_scale (deprecated).""" |
| 92 | result = maybe_remap_kv_scale_name( |
| 93 | "model.layers.0.self_attn.kv_scale", self.PARAMS_DICT |
| 94 | ) |
| 95 | assert result == "model.layers.0.self_attn.attn.k_scale" |
| 96 | |
| 97 | def test_default_bare_k_scale(self): |
| 98 | """Default format: .k_scale -> .attn.k_scale.""" |
| 99 | result = maybe_remap_kv_scale_name( |
| 100 | "model.layers.0.self_attn.k_scale", self.PARAMS_DICT |
| 101 | ) |
| 102 | assert result == "model.layers.0.self_attn.attn.k_scale" |
| 103 | |
| 104 | def test_non_scale_name_unchanged(self): |
| 105 | """Non-scale names should be returned unchanged.""" |
| 106 | name = "model.layers.0.self_attn.qkv_proj.weight" |
| 107 | result = maybe_remap_kv_scale_name(name, self.PARAMS_DICT) |
| 108 | assert result == name |
| 109 | |
| 110 | def test_nvfp4_modelopt_k_proj_k_scale(self): |
| 111 | """ModelOpt NVFP4 format (e.g. nvidia/Qwen3-30B-A3B-NVFP4): |
| 112 | k_proj.k_scale -> attn.k_scale. |
| 113 | Validates that NVFP4 checkpoints are not broken by this change.""" |
| 114 | result = maybe_remap_kv_scale_name( |
| 115 | "model.layers.0.self_attn.k_proj.k_scale", self.PARAMS_DICT |
| 116 | ) |
| 117 | assert result == "model.layers.0.self_attn.attn.k_scale" |
| 118 | |
| 119 | def test_nvfp4_modelopt_v_proj_v_scale(self): |
| 120 | """ModelOpt NVFP4 format (e.g. nvidia/Qwen3-30B-A3B-NVFP4): |
| 121 | v_proj.v_scale -> attn.v_scale. |
| 122 | Validates that NVFP4 checkpoints are not broken by this change.""" |
| 123 | result = maybe_remap_kv_scale_name( |
| 124 | "model.layers.0.self_attn.v_proj.v_scale", self.PARAMS_DICT |
| 125 | ) |
| 126 | assert result == "model.layers.0.self_attn.attn.v_scale" |
| 127 | |
| 128 | def test_qwen3_vl_moe_qkv_proj_k_scale(self): |
| 129 | """Qwen3-VL-MoE uses the same fused qkv_proj naming as Qwen3-MoE. |
| 130 | Regression test for qwen3_vl_moe.py fix (same bug as #25047).""" |
| 131 | result = maybe_remap_kv_scale_name( |
| 132 | "model.layers.0.self_attn.qkv_proj.k_scale", self.PARAMS_DICT |
| 133 | ) |
| 134 | assert result == "model.layers.0.self_attn.attn.k_scale" |
| 135 | |
| 136 | def test_qwen3_vl_moe_qkv_proj_v_scale(self): |
| 137 | """Qwen3-VL-MoE uses the same fused qkv_proj naming as Qwen3-MoE. |
| 138 | Regression test for qwen3_vl_moe.py fix (same bug as #25047).""" |
| 139 | result = maybe_remap_kv_scale_name( |
| 140 | "model.layers.0.self_attn.qkv_proj.v_scale", self.PARAMS_DICT |
| 141 | ) |
| 142 | assert result == "model.layers.0.self_attn.attn.v_scale" |
| 143 | |
| 144 | def test_nvfp4_weight_scale_not_remapped(self): |
| 145 | """NVFP4 weight_scale should not be touched by remap (not a kv scale).""" |
| 146 | name = "model.layers.0.self_attn.k_proj.weight_scale" |
| 147 | result = maybe_remap_kv_scale_name(name, self.PARAMS_DICT) |
| 148 | assert result == name |
| 149 | |
| 150 | def test_nvfp4_input_scale_not_remapped(self): |
| 151 | """NVFP4 input_scale should not be touched by remap (not a kv scale).""" |
| 152 | name = "model.layers.0.self_attn.k_proj.input_scale" |
| 153 | result = maybe_remap_kv_scale_name(name, self.PARAMS_DICT) |
| 154 | assert result == name |
| 155 | |
| 156 | def test_missing_target_returns_none(self): |
| 157 | """If remapped name not in params_dict, return None.""" |
| 158 | empty_params: dict[str, None] = {} |
| 159 | result = maybe_remap_kv_scale_name( |
| 160 | "model.layers.0.self_attn.qkv_proj.k_scale", empty_params |
| 161 | ) |
| 162 | assert result is None |
| 163 | |
| 164 | |
| 165 | def test_checkpoint_weight_mapper_discards_serialized_g_idx(): |
| 166 | from vllm.model_executor.layers.quantization.base_config import ( |
| 167 | QuantizationConfig, |
| 168 | ) |
| 169 | |
| 170 | mapper = QuantizationConfig.get_checkpoint_weight_mapper() |
| 171 | |
| 172 | assert mapper._map_name("model.layers.0.g_idx") is None |
| 173 | assert mapper._map_name("model.layers.0.qweight") == "model.layers.0.qweight" |
| 174 | |
| 175 | |
| 176 | class TestKvCacheScaleMapper: |
| 177 | """The `WeightsMapper` returned by `get_cache_scale_mapper` replaces the |
| 178 | per-model `maybe_remap_kv_scale_name` calls. It must remap the same set of |
| 179 | checkpoint formats (the non-`params_dict`-dependent ones) and be idempotent |
| 180 | so it composes safely with a model's own qkv/gate_up `hf_to_vllm_mapper`.""" |
| 181 | |
| 182 | def _mapper(self): |
| 183 | # `get_cache_scale_mapper` does not use `self`; call it on the base |
| 184 | # class to get the default (non-config-specific) mapper. |
| 185 | from vllm.model_executor.layers.quantization.base_config import ( |
| 186 | QuantizationConfig, |
| 187 | ) |
| 188 | |
| 189 | return QuantizationConfig.get_cache_scale_mapper() |
| 190 | |
| 191 | def _map(self, name: str) -> str | None: |
| 192 | return self._mapper()._map_name(name) |
| 193 | |
| 194 | @pytest.mark.parametrize( |
| 195 | "name,expected", |
| 196 | [ |
| 197 | # Qwen3-MoE / llm-compressor fused qkv_proj |
| 198 | ( |
| 199 | "model.layers.0.self_attn.qkv_proj.k_scale", |
| 200 | "model.layers.0.self_attn.attn.k_scale", |
| 201 | ), |
| 202 | ( |
| 203 | "model.layers.0.self_attn.qkv_proj.v_scale", |
| 204 | "model.layers.0.self_attn.attn.v_scale", |
| 205 | ), |
| 206 | # ModelOpt / NVFP4 k_proj/v_proj |
| 207 | ( |
| 208 | "model.layers.0.self_attn.k_proj.k_scale", |
| 209 | "model.layers.0.self_attn.attn.k_scale", |
| 210 | ), |
| 211 | ( |
| 212 | "model.layers.0.self_attn.v_proj.v_scale", |
| 213 | "model.layers.0.self_attn.attn.v_scale", |
| 214 | ), |
| 215 | # deprecated fused kv_scale and bare scales |
| 216 | ( |
| 217 | "model.layers.0.self_attn.kv_scale", |
| 218 | "model.layers.0.self_attn.attn.k_scale", |
| 219 | ), |
| 220 | ( |
| 221 | "model.layers.0.self_attn.k_scale", |
| 222 | "model.layers.0.self_attn.attn.k_scale", |
| 223 | ), |
| 224 | # NemotronH mixer |
| 225 | ( |
| 226 | "model.layers.0.mixer.k_proj.k_scale", |
| 227 | "model.layers.0.mixer.attn.k_scale", |
| 228 | ), |
| 229 | # already in vLLM form -> unchanged (idempotent) |
| 230 | ( |
| 231 | "model.layers.0.self_attn.attn.k_scale", |
| 232 | "model.layers.0.self_attn.attn.k_scale", |
| 233 | ), |
| 234 | # non-kv scales must not be touched |
| 235 | ( |
| 236 | "model.layers.0.self_attn.k_proj.weight_scale", |
| 237 | "model.layers.0.self_attn.k_proj.weight_scale", |
| 238 | ), |
| 239 | ( |
| 240 | "model.layers.0.self_attn.k_proj.input_scale", |
| 241 | "model.layers.0.self_attn.k_proj.input_scale", |
| 242 | ), |
| 243 | # regular weights untouched |
| 244 | ( |
| 245 | "model.layers.0.self_attn.q_proj.weight", |
| 246 | "model.layers.0.self_attn.q_proj.weight", |
| 247 | ), |
| 248 | ], |
| 249 | ) |
| 250 | def test_remap(self, name, expected): |
| 251 | assert self._map(name) == expected |
| 252 | |
| 253 | @pytest.mark.parametrize( |
| 254 | "name", |
| 255 | [ |
| 256 | "model.layers.0.self_attn.k_scale", |
| 257 | "model.layers.0.self_attn.k_proj.k_scale", |
| 258 | "model.layers.0.self_attn.qkv_proj.v_scale", |
| 259 | "model.layers.0.mixer.k_proj.k_scale", |
| 260 | ], |
| 261 | ) |
| 262 | def test_idempotent(self, name): |
| 263 | once = self._map(name) |
| 264 | assert once is not None |
| 265 | assert self._map(once) == once |
| 266 | |
| 267 | def test_composes_with_qkv_mapper(self): |
| 268 | """Applied together with a model's qkv/gate_up mapper, the regex scale |
| 269 | rules run before the substr rename, so scales are normalized to `.attn.` |
| 270 | and regular projections are still fused correctly.""" |
| 271 | from vllm.model_executor.models.utils import WeightsMapper |
| 272 | |
| 273 | model_mapper = WeightsMapper( |
| 274 | orig_to_new_substr={ |
| 275 | ".q_proj": ".qkv_proj.q", |
| 276 | ".k_proj": ".qkv_proj.k", |
| 277 | ".v_proj": ".qkv_proj.v", |
| 278 | } |
| 279 | ) |
| 280 | # AutoWeightsLoader does `mapper |= cache_scale_mapper` |
| 281 | combined = model_mapper | self._mapper() |
| 282 | |
| 283 | assert ( |
| 284 | combined._map_name("model.layers.0.self_attn.q_proj.weight") |
| 285 | == "model.layers.0.self_attn.qkv_proj.q.weight" |
| 286 | ) |
| 287 | assert ( |
| 288 | combined._map_name("model.layers.0.self_attn.k_proj.k_scale") |
| 289 | == "model.layers.0.self_attn.attn.k_scale" |
| 290 | ) |
| 291 | assert ( |
| 292 | combined._map_name("model.layers.0.self_attn.k_scale") |
| 293 | == "model.layers.0.self_attn.attn.k_scale" |
| 294 | ) |
| 295 | |
| 296 | |
| 297 | @pytest.mark.parametrize("fails", [False, True]) |
| 298 | def test_drop_checkpoint_cache_releases_local_weight_files( |
| 299 | tmp_path, monkeypatch, fails |
| 300 | ): |
| 301 | """Evict only weights; cache eviction errors must not prevent loading.""" |
| 302 | for name in ("weights.safetensors", "config.json"): |
| 303 | (tmp_path / name).touch() |
| 304 | fadvise = Mock(side_effect=OSError("cache eviction failed") if fails else None) |
| 305 | monkeypatch.setattr("os.posix_fadvise", fadvise) |
| 306 | drop_checkpoint_cache(str(tmp_path)) |
| 307 | fadvise.assert_called_once() |
| 308 | |
| 309 | |
| 310 | if __name__ == "__main__": |
| 311 | test_download_weights_from_hf() |