| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import importlib.metadata |
| 4 | import importlib.util |
| 5 | |
| 6 | import pytest |
| 7 | import torch |
| 8 | from packaging import version |
| 9 | |
| 10 | from tests.quantization.utils import load_model_without_vllm_runner |
| 11 | from vllm.config import set_current_vllm_config |
| 12 | from vllm.forward_context import set_forward_context |
| 13 | from vllm.model_executor.layers.attention import Attention |
| 14 | from vllm.model_executor.model_loader import get_model_loader |
| 15 | from vllm.model_executor.model_loader.default_loader import DefaultModelLoader |
| 16 | from vllm.platforms import current_platform |
| 17 | |
| 18 | if current_platform.is_rocm(): |
| 19 | from vllm.platforms.rocm import on_gfx950 |
| 20 | else: |
| 21 | |
| 22 | def on_gfx950() -> bool: |
| 23 | return False |
| 24 | |
| 25 | |
| 26 | DEVICE_TYPE = current_platform.device_type |
| 27 | DTYPE = ["bfloat16"] |
| 28 | |
| 29 | TORCHAO_AVAILABLE = importlib.util.find_spec("torchao") is not None |
| 30 | TORCHAO_VERSION_0_18_AVAILABLE = TORCHAO_AVAILABLE and version.parse( |
| 31 | importlib.metadata.version("torchao") |
| 32 | ) >= version.parse("0.18.0") |
| 33 | |
| 34 | |
| 35 | # Fix for leaking `torch.set_float32_matmul_precision("high")` |
| 36 | # https://github.com/pytorch/ao/blob/v0.17.0/torchao/quantization/utils.py#L696 |
| 37 | @pytest.fixture(autouse=True) |
| 38 | def restore_float32_matmul_precision(): |
| 39 | original_precision = torch.get_float32_matmul_precision() |
| 40 | try: |
| 41 | yield |
| 42 | finally: |
| 43 | torch.set_float32_matmul_precision(original_precision) |
| 44 | |
| 45 | |
| 46 | @pytest.mark.skipif( |
| 47 | current_platform.is_rocm() and current_platform.is_fp8_fnuz(), |
| 48 | reason="Only fp8_fnuz supported on CDNA3 architecture", |
| 49 | ) |
| 50 | @pytest.mark.skipif(not TORCHAO_AVAILABLE, reason="torchao is not available") |
| 51 | def test_pre_quantized_model(monkeypatch, dist_init, workspace_init): |
| 52 | model, vllm_config = load_model_without_vllm_runner( |
| 53 | "torchao-testing/opt-125m-Float8WeightOnlyConfig-v2-0.15.0", |
| 54 | quantization="torchao", |
| 55 | dtype="bfloat16", |
| 56 | ) |
| 57 | |
| 58 | monkeypatch.setattr(Attention, "forward", lambda _, q, k, v: q.contiguous()) |
| 59 | input_ids = torch.tensor([1, 2, 3, 4], device=DEVICE_TYPE) |
| 60 | positions = torch.arange(input_ids.numel(), device=DEVICE_TYPE) |
| 61 | with ( |
| 62 | set_current_vllm_config(vllm_config), |
| 63 | set_forward_context(None, vllm_config, num_tokens=input_ids.numel()), |
| 64 | ): |
| 65 | hidden_states = model(input_ids, positions, None) |
| 66 | logits = model.compute_logits(hidden_states) |
| 67 | assert torch.isfinite(logits).all() |
| 68 | |
| 69 | |
| 70 | @pytest.mark.skipif(not TORCHAO_AVAILABLE, reason="torchao is not available") |
| 71 | @pytest.mark.skip( |
| 72 | reason="since torchao nightly is only compatible with torch nightly" |
| 73 | "currently https://github.com/pytorch/ao/issues/2919, we'll have to skip " |
| 74 | "torchao tests that requires newer versions (0.14.0.dev+) for now" |
| 75 | ) |
| 76 | def test_opt_125m_awq_int4wo_model_loading_with_params(vllm_runner): |
| 77 | torch._dynamo.reset() |
| 78 | model_name = "torchao-testing/opt-125m-AWQConfig-Int4WeightOnlyConfig-v2-0.14.0.dev" |
| 79 | with vllm_runner( |
| 80 | model_name=model_name, |
| 81 | quantization="torchao", |
| 82 | dtype="bfloat16", |
| 83 | pt_load_map_location=f"{DEVICE_TYPE}:0", |
| 84 | ) as llm: |
| 85 | output = llm.generate_greedy(["The capital of France is"], max_tokens=4) |
| 86 | |
| 87 | assert output |
| 88 | |
| 89 | |
| 90 | @pytest.mark.skipif(not TORCHAO_AVAILABLE, reason="torchao is not available") |
| 91 | @pytest.mark.skipif( |
| 92 | on_gfx950() and not TORCHAO_VERSION_0_18_AVAILABLE, |
| 93 | reason="requires torchao>=0.18.0 on gfx950", |
| 94 | ) |
| 95 | def test_online_quant_config_dict_json(vllm_runner, enable_pickle): |
| 96 | """Testing online quantization, load_weights integration point, |
| 97 | with config dict serialized to json string |
| 98 | """ |
| 99 | torch._dynamo.reset() |
| 100 | model_name = "facebook/opt-125m" |
| 101 | |
| 102 | import json |
| 103 | |
| 104 | from torchao.core.config import config_to_dict |
| 105 | from torchao.quantization import Float8DynamicActivationFloat8WeightConfig, PerRow |
| 106 | |
| 107 | torchao_quant_config = Float8DynamicActivationFloat8WeightConfig( |
| 108 | granularity=PerRow() |
| 109 | ) |
| 110 | hf_overrides = { |
| 111 | "quantization_config_dict_json": json.dumps( |
| 112 | config_to_dict(torchao_quant_config) |
| 113 | ) |
| 114 | } |
| 115 | with vllm_runner( |
| 116 | model_name=model_name, |
| 117 | dtype="bfloat16", |
| 118 | pt_load_map_location=f"{DEVICE_TYPE}:0", |
| 119 | quantization="torchao", |
| 120 | hf_overrides=hf_overrides, |
| 121 | enforce_eager=True, |
| 122 | ) as llm: |
| 123 | output = llm.generate_greedy(["The capital of France is"], max_tokens=4) |
| 124 | |
| 125 | load_config = llm.llm.llm_engine.vllm_config.load_config |
| 126 | model_config = llm.llm.llm_engine.vllm_config.model_config |
| 127 | |
| 128 | def load_weights(model): |
| 129 | model_loader = get_model_loader(load_config) |
| 130 | assert isinstance(model_loader, DefaultModelLoader) |
| 131 | weights_iterator = model_loader.get_all_weights(model_config, model) |
| 132 | model.load_weights(weights_iterator) |
| 133 | |
| 134 | llm.apply_model(load_weights) |
| 135 | |
| 136 | reload_output = llm.generate_greedy(["The capital of France is"], max_tokens=4) |
| 137 | assert output[0][0] == reload_output[0][0] |
| 138 | |
| 139 | |
| 140 | @pytest.mark.skipif(not TORCHAO_AVAILABLE, reason="torchao is not available") |
| 141 | @pytest.mark.skipif( |
| 142 | on_gfx950() and not TORCHAO_VERSION_0_18_AVAILABLE, |
| 143 | reason="requires torchao>=0.18.0 on gfx950", |
| 144 | ) |
| 145 | def test_online_quant_config_file(vllm_runner): |
| 146 | """Testing on the fly quantization, load_weights integration point, |
| 147 | with config file |
| 148 | """ |
| 149 | torch._dynamo.reset() |
| 150 | model_name = "facebook/opt-125m" |
| 151 | import json |
| 152 | from tempfile import NamedTemporaryFile |
| 153 | |
| 154 | from torchao.core.config import config_to_dict |
| 155 | from torchao.quantization import Float8DynamicActivationFloat8WeightConfig, PerRow |
| 156 | |
| 157 | config = Float8DynamicActivationFloat8WeightConfig(granularity=PerRow()) |
| 158 | |
| 159 | with NamedTemporaryFile(mode="w", delete=False) as f: |
| 160 | f.write(json.dumps(config_to_dict(config))) |
| 161 | # close the file to save it |
| 162 | f.close() |
| 163 | config_file_name = str(f.name) |
| 164 | |
| 165 | hf_overrides = {"quantization_config_file": config_file_name} |
| 166 | with vllm_runner( |
| 167 | model_name=model_name, |
| 168 | dtype="bfloat16", |
| 169 | pt_load_map_location=f"{DEVICE_TYPE}:0", |
| 170 | quantization="torchao", |
| 171 | hf_overrides=hf_overrides, |
| 172 | enforce_eager=True, |
| 173 | ) as llm: |
| 174 | output = llm.generate_greedy(["The capital of France is"], max_tokens=4) |
| 175 | |
| 176 | assert output |
| 177 | |
| 178 | |
| 179 | @pytest.mark.skipif(not TORCHAO_AVAILABLE, reason="torchao is not available") |
| 180 | @pytest.mark.skipif( |
| 181 | on_gfx950() and not TORCHAO_VERSION_0_18_AVAILABLE, |
| 182 | reason="requires torchao>=0.18.0 on gfx950", |
| 183 | ) |
| 184 | def test_reload_weights(): |
| 185 | import json |
| 186 | |
| 187 | from torchao.core.config import config_to_dict |
| 188 | from torchao.quantization import Float8DynamicActivationFloat8WeightConfig, PerRow |
| 189 | |
| 190 | from vllm import LLM, SamplingParams |
| 191 | |
| 192 | torchao_quant_config = Float8DynamicActivationFloat8WeightConfig( |
| 193 | granularity=PerRow() |
| 194 | ) |
| 195 | |
| 196 | hf_overrides = { |
| 197 | "quantization_config_dict_json": json.dumps( |
| 198 | config_to_dict(torchao_quant_config) |
| 199 | ) |
| 200 | } |
| 201 | |
| 202 | llm = LLM( |
| 203 | model="Qwen/Qwen3-0.6B", |
| 204 | dtype="bfloat16", |
| 205 | load_format="dummy", |
| 206 | enforce_eager=True, |
| 207 | quantization="torchao", |
| 208 | hf_overrides=hf_overrides, |
| 209 | ) |
| 210 | # Update load format from `dummy` to `auto` |
| 211 | llm.collective_rpc( |
| 212 | "update_config", args=({"load_config": {"load_format": "auto"}},) |
| 213 | ) |
| 214 | # Now reload real weights inplace |
| 215 | llm.collective_rpc("reload_weights") |
| 216 | prompts = [ |
| 217 | "Hello, my name is", |
| 218 | "The president of the United States is", |
| 219 | "The capital of France is", |
| 220 | "The future of AI is", |
| 221 | ] |
| 222 | # Create a sampling params object. |
| 223 | sampling_params = SamplingParams(temperature=0, top_p=0.95) |
| 224 | outputs = llm.generate(prompts, sampling_params) |
| 225 | # make sure it runs |
| 226 | for output in outputs: |
| 227 | generated_text = output.outputs[0].text |
| 228 | assert generated_text |
| 229 | # can also uncomment locally to make sure the generated |
| 230 | # output makes sense |
| 231 | # prompt = output.prompt |
| 232 | # print(f"Prompt: {prompt!r}") |
| 233 | # print(f"Output: {generated_text!r}") |
| 234 | # print("-" * 60) |
| 235 | |
| 236 | |
| 237 | @pytest.mark.skipif(not TORCHAO_AVAILABLE, reason="torchao is not available") |
| 238 | @pytest.mark.skip( |
| 239 | reason="since torchao nightly is only compatible with torch nightly" |
| 240 | "currently https://github.com/pytorch/ao/issues/2919, we'll have to skip " |
| 241 | "torchao tests that requires newer versions (0.15.0.dev+) for now" |
| 242 | ) |
| 243 | def test_safetensors_model_loading_with_params(vllm_runner): |
| 244 | torch._dynamo.reset() |
| 245 | # using this model to test safetensors loading with file sharding |
| 246 | model_name = "torchao-testing/Qwen3-8B-INT4-0.15.0dev-safetensors" |
| 247 | with vllm_runner(model_name=model_name, dtype="bfloat16") as llm: |
| 248 | output = llm.generate_greedy(["The capital of France is"], max_tokens=4) |
| 249 | |
| 250 | assert output |
| 251 | |
| 252 | |
| 253 | @pytest.mark.skipif(not TORCHAO_AVAILABLE, reason="torchao is not available") |
| 254 | @pytest.mark.skip( |
| 255 | reason="since torchao nightly is only compatible with torch nightly" |
| 256 | "currently https://github.com/pytorch/ao/issues/2919, we'll have to skip " |
| 257 | "torchao tests that requires newer versions (0.14.0.dev+) for now" |
| 258 | ) |
| 259 | def test_opt_125m_module_fqn_to_config_regex_model(vllm_runner): |
| 260 | torch._dynamo.reset() |
| 261 | model_name = "torchao-testing/opt-125m-ModuleFqnToConfig-v1-regex-0.14.0.dev" |
| 262 | with vllm_runner( |
| 263 | model_name=model_name, dtype="bfloat16", pt_load_map_location=f"{DEVICE_TYPE}:0" |
| 264 | ) as llm: |
| 265 | output = llm.generate_greedy(["The capital of France is"], max_tokens=4) |
| 266 | |
| 267 | assert output |
| 268 | |
| 269 | |
| 270 | @pytest.mark.skipif(not TORCHAO_AVAILABLE, reason="torchao is not available") |
| 271 | @pytest.mark.skip( |
| 272 | reason="since torchao nightly is only compatible with torch nightly" |
| 273 | "currently https://github.com/pytorch/ao/issues/2919, we'll have to skip " |
| 274 | "torchao tests that requires newer versions (0.14.0.dev+) for now" |
| 275 | ) |
| 276 | def test_opt_125m_int4wo_model_running_preshuffled_kernel(vllm_runner, monkeypatch): |
| 277 | """We load a model with Int4Tensor (plain format) linear weights |
| 278 | and verify that the weight is updated to Int4PreshuffledTensor |
| 279 | after loading in vllm |
| 280 | """ |
| 281 | from torchao.quantization import Int4PreshuffledTensor |
| 282 | from torchao.utils import _is_fbgemm_gpu_genai_available, is_sm_at_least_90 |
| 283 | |
| 284 | torch._dynamo.reset() |
| 285 | monkeypatch.setenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "1") |
| 286 | model_name = "torchao-testing/opt-125m-Int4WeightOnlyConfig-v2-0.14.0.dev" |
| 287 | # Note: using enforce_eager=True because the `bf16i4bf16_shuffled` doesn't |
| 288 | # have meta kernel implemented yet, can remove this flag after that is implemented |
| 289 | with vllm_runner( |
| 290 | model_name=model_name, |
| 291 | quantization="torchao", |
| 292 | dtype="bfloat16", |
| 293 | pt_load_map_location=f"{DEVICE_TYPE}:0", |
| 294 | enforce_eager=True, |
| 295 | ) as llm: |
| 296 | |
| 297 | def has_int4_preshuffled_tensor_weight(model): |
| 298 | return isinstance( |
| 299 | model.model.decoder.layers[0].self_attn.qkv_proj.weight, |
| 300 | Int4PreshuffledTensor, |
| 301 | ) |
| 302 | |
| 303 | def get_weight_attrs(model): |
| 304 | weight = model.model.decoder.layers[0].self_attn.qkv_proj.weight |
| 305 | return [ |
| 306 | weight.requires_grad, |
| 307 | weight.input_dim, |
| 308 | weight.output_dim, |
| 309 | hasattr(weight, "weight_loader"), |
| 310 | ] |
| 311 | |
| 312 | llm_engine = llm.get_llm().llm_engine |
| 313 | has_int4_preshuffled_tensor = any( |
| 314 | llm_engine.apply_model(has_int4_preshuffled_tensor_weight) |
| 315 | ) |
| 316 | weight_attrs = llm_engine.apply_model(get_weight_attrs)[0] |
| 317 | |
| 318 | # making sure we are using Int4PreshuffledTensor on H100 GPU, when |
| 319 | # fbgemm_gpu_genai |
| 320 | # library is installed, otherwise it should be using Int4Tensor |
| 321 | if _is_fbgemm_gpu_genai_available() and is_sm_at_least_90(): |
| 322 | assert has_int4_preshuffled_tensor |
| 323 | else: |
| 324 | assert not has_int4_preshuffled_tensor |
| 325 | |
| 326 | assert weight_attrs == [False, 1, 0, True] |
| 327 | output = llm.generate_greedy(["The capital of France is"], max_tokens=32) |
| 328 | |
| 329 | assert output |
| 330 | |
| 331 | |
| 332 | @pytest.mark.skipif(not TORCHAO_AVAILABLE, reason="torchao is not available") |
| 333 | @pytest.mark.skip( |
| 334 | reason="since torchao nightly is only compatible with torch nightly" |
| 335 | "currently https://github.com/pytorch/ao/issues/2919, we'll have to skip " |
| 336 | "torchao tests that requires newer versions (0.14.0.dev+) for now" |
| 337 | ) |
| 338 | def test_opt_125m_int4wo_model_running_preshuffled_kernel_online_quant( |
| 339 | vllm_runner, monkeypatch |
| 340 | ): |
| 341 | """We load a bf16 model and online quantize the model to int4, then verify that |
| 342 | the weights are updated to Int4PreshuffledTensor after online quantization |
| 343 | """ |
| 344 | from torchao.quantization import Int4PreshuffledTensor |
| 345 | from torchao.utils import _is_fbgemm_gpu_genai_available, is_sm_at_least_90 |
| 346 | |
| 347 | torch._dynamo.reset() |
| 348 | model_name = "facebook/opt-125m" |
| 349 | |
| 350 | monkeypatch.setenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "1") |
| 351 | |
| 352 | import json |
| 353 | |
| 354 | from torchao.core.config import config_to_dict |
| 355 | from torchao.quantization import Int4WeightOnlyConfig |
| 356 | |
| 357 | torchao_quant_config = Int4WeightOnlyConfig( |
| 358 | group_size=128, int4_packing_format="plain" |
| 359 | ) |
| 360 | hf_overrides = { |
| 361 | "quantization_config_dict_json": json.dumps( |
| 362 | config_to_dict(torchao_quant_config) |
| 363 | ) |
| 364 | } |
| 365 | |
| 366 | # Note: using enforce_eager=True because the `bf16i4bf16_shuffled` doesn't |
| 367 | # have meta kernel implemented yet, can remove this flag after that is implemented |
| 368 | with vllm_runner( |
| 369 | model_name=model_name, |
| 370 | quantization="torchao", |
| 371 | dtype="bfloat16", |
| 372 | pt_load_map_location=f"{DEVICE_TYPE}:0", |
| 373 | hf_overrides=hf_overrides, |
| 374 | enforce_eager=True, |
| 375 | ) as llm: |
| 376 | |
| 377 | def has_int4_preshuffled_tensor_weight(model): |
| 378 | return isinstance( |
| 379 | model.model.decoder.layers[0].self_attn.qkv_proj.weight, |
| 380 | Int4PreshuffledTensor, |
| 381 | ) |
| 382 | |
| 383 | def get_weight_attrs(model): |
| 384 | weight = model.model.decoder.layers[0].self_attn.qkv_proj.weight |
| 385 | return [ |
| 386 | weight.requires_grad, |
| 387 | weight.input_dim, |
| 388 | weight.output_dim, |
| 389 | hasattr(weight, "weight_loader"), |
| 390 | ] |
| 391 | |
| 392 | llm_engine = llm.get_llm().llm_engine |
| 393 | has_int4_preshuffled_tensor = any( |
| 394 | llm_engine.apply_model(has_int4_preshuffled_tensor_weight) |
| 395 | ) |
| 396 | weight_attrs = llm_engine.apply_model(get_weight_attrs)[0] |
| 397 | |
| 398 | # making sure we are using Int4PreshuffledTensor on H100 GPU, when |
| 399 | # fbgemm_gpu_genai |
| 400 | # library is installed, otherwise it should be using Int4Tensor |
| 401 | if _is_fbgemm_gpu_genai_available() and is_sm_at_least_90(): |
| 402 | assert has_int4_preshuffled_tensor |
| 403 | else: |
| 404 | assert not has_int4_preshuffled_tensor |
| 405 | |
| 406 | assert weight_attrs == [False, 1, 0, True] |
| 407 | output = llm.generate_greedy(["The capital of France is"], max_tokens=4) |
| 408 | |
| 409 | assert output |
| 410 | |
| 411 | |
| 412 | if __name__ == "__main__": |
| 413 | pytest.main([__file__]) |