utils.py 16.5 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
"""Utilities for selecting and loading models."""
4

5
6
7
import inspect
import warnings
from contextlib import contextmanager
8
from dataclasses import dataclass, field
9

zhuwenwen's avatar
zhuwenwen committed
10
import os
11
12
import torch
from torch import nn
13
from typing_extensions import assert_never
14

15
from vllm.attention.layer import Attention, MLAAttention
16
from vllm.config import ModelConfig, VllmConfig, set_current_vllm_config
17
from vllm.logger import init_logger
18
from vllm.model_executor.layers.quantization.base_config import (
19
20
21
    QuantizationConfig,
    QuantizeMethodBase,
)
22
from vllm.model_executor.models.interfaces import SupportsQuant
23
from vllm.utils.platform_utils import is_pin_memory_available
24
from vllm import envs
25

26
27
logger = init_logger(__name__)

28

29
30
31
32
def initialize_model(
    vllm_config: VllmConfig,
    *,
    prefix: str = "",
33
34
    model_class: type[nn.Module] | None = None,
    model_config: ModelConfig | None = None,
35
36
) -> nn.Module:
    """Initialize a model with the given configurations."""
37
38
    if model_config is None:
        model_config = vllm_config.model_config
39
40
41
42
43
44
45
46
47
48
    if model_class is None:
        model_class, _ = get_model_architecture(model_config)

    if vllm_config.quant_config is not None:
        configure_quant_config(vllm_config.quant_config, model_class)

    signatures = inspect.signature(model_class.__init__)
    all_params = [param.name for param in signatures.parameters.values()]
    if "vllm_config" in all_params and "prefix" in all_params:
        # new-style model class
49
        with set_current_vllm_config(vllm_config, check_compile=True, prefix=prefix):
50
51
            return model_class(vllm_config=vllm_config, prefix=prefix)

52
53
54
55
56
57
58
    msg = (
        "vLLM model class should accept `vllm_config` and `prefix` as "
        "input arguments. Possibly you have an old-style model class"
        " registered from out of tree and it is used for new vLLM version. "
        "Check https://docs.vllm.ai/en/latest/design/arch_overview.html "
        "for the design and update the model class accordingly."
    )
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
    warnings.warn(msg, DeprecationWarning, stacklevel=2)

    logger.warning(
        "Trying to guess the arguments for old-style model class %s",
        model_class,
    )
    # try to be compatible with old-style model class
    kwargs = {}
    if "prefix" in all_params:
        kwargs["prefix"] = prefix
    if "config" in all_params:
        kwargs["config"] = model_config.hf_config
    if "cache_config" in all_params:
        kwargs["cache_config"] = vllm_config.cache_config
    if "quant_config" in all_params:
        kwargs["quant_config"] = vllm_config.quant_config
    if "lora_config" in all_params:
        kwargs["lora_config"] = vllm_config.lora_config
    if "scheduler_config" in all_params:
        kwargs["scheduler_config"] = vllm_config.scheduler_config
79
    with set_current_vllm_config(vllm_config, check_compile=True, prefix=prefix):
80
81
82
        return model_class(**kwargs)


83
84
85
def process_weights_after_loading(
    model: nn.Module, model_config: ModelConfig, target_device: torch.device
) -> None:
86
87
88
89
90
91
92
93
    if getattr(model, "process_weights_after_loading_already_called", False):
        # In case `process_weights_after_loading` is called multiple times
        # we'll skip it at later times
        logger.debug_once(
            "process_weights_after_loading already called for model %s", model
        )
        return

94
95
    # to avoid circular dependency
    from vllm.model_executor.model_loader.online_quantization import (
96
97
98
99
        maybe_save_metadata_and_attributes_for_weight_reloading,
    )

    maybe_save_metadata_and_attributes_for_weight_reloading(model, model_config)
100

101
102
103
104
105
106
107
108
109
110
111
    for _, module in model.named_modules():
        quant_method = getattr(module, "quant_method", None)
        if isinstance(quant_method, QuantizeMethodBase):
            # When quant methods need to process weights after loading
            # (for repacking, quantizing, etc), they expect parameters
            # to be on the global target device. This scope is for the
            # case where cpu offloading is used, where we will move the
            # parameters onto device for processing and back off after.
            with device_loading_context(module, target_device):
                quant_method.process_weights_after_loading(module)

112
113
    # Initialize post-load attention weights for both Attention and MLA.
    # NOTE: Happens after other modules so we can easily decompress weights.
114
    for _, module in model.named_modules():
115
        if isinstance(module, (Attention, MLAAttention)) and hasattr(
116
117
            module, "process_weights_after_loading"
        ):
118
119
120
121
122
123
            # TODO(lucas): see if there is a way to unify the signatures
            # of process_weights_after_loading
            module.process_weights_after_loading(model_config.dtype)


@contextmanager
124
def device_loading_context(module: torch.nn.Module, target_device: torch.device):
125
126
127
128
129
    if target_device.type == "cpu":
        # If target is CPU, no need to move anything
        yield module
        return

130
    original_device_states: dict[str, torch.device] = {}
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164

    # Store original device states and move parameters to GPU if they're on CPU
    for name, p in module.named_parameters():
        if p.device.type == "cpu":
            original_device_states[name] = p.device
            p.data = p.data.to(target_device)
        # Parameters already on target device are not touched

    try:
        yield module

    finally:
        # Restore parameters to their original devices, ignoring new parameters
        pin_memory = is_pin_memory_available()
        for name, p in module.named_parameters():
            if name in original_device_states:
                original_device: torch.device = original_device_states[name]
                if original_device.type == "cpu":
                    # `torch.empty_like` does not support `pin_memory` argument
                    cpu_data = torch.empty_strided(
                        size=p.data.size(),
                        stride=p.data.stride(),
                        dtype=p.data.dtype,
                        layout=p.data.layout,
                        device="cpu",
                        pin_memory=pin_memory,
                    )
                    cpu_data.copy_(p.data)
                    p.data = cpu_data
                else:
                    p.data = p.data.to(original_device)
        # New parameters or parameters already on target device are untouched


165
_MODEL_ARCH_BY_HASH = dict[int, tuple[type[nn.Module], str]]()
166
167
168
"""Caches the outputs of `_get_model_architecture`."""


169
def _get_model_architecture(model_config: ModelConfig) -> tuple[type[nn.Module], str]:
170
    from vllm.model_executor.models.adapters import as_embedding_model, as_seq_cls_model
Cyrus Leung's avatar
Cyrus Leung committed
171

172
    architectures = getattr(model_config.hf_config, "architectures", [])
zhuwenwen's avatar
zhuwenwen committed
173
    visions = getattr(model_config.hf_config, "visual", []) or getattr(model_config.hf_config, "vision_config", [])
174
175
    # TODO: 'Qwen2_5_VLForConditionalGeneration', 
    support_nn_architectures = ['LlamaForCausalLM', 'QWenLMHeadModel', 'Qwen2ForCausalLM', 'Qwen2VLForConditionalGeneration', 'Qwen2MoeForCausalLM', 'Qwen3ForCausalLM', 'Qwen3MoeForCausalLM',
zhuwenwen's avatar
zhuwenwen committed
176
                                'ChatGLMModel', 'Glm4ForCausalLM', 'ChatGLMForConditionalGeneration', 'BaichuanForCausalLM', 'BloomForCausalLM', 'TeleChat2ForCausalLM', 'MixtralForCausalLM', 'FalconForCausalLM',
zhuwenwen's avatar
zhuwenwen committed
177
                                'MedusaModel', 'MLPSpeculatorPreTrainedModel', 'DeepseekV2ForCausalLM', 'DeepseekV3ForCausalLM', 'DeepSeekMTPModel']  
178
    if any(arch in architectures for arch in support_nn_architectures): 
179
180
181
182
183
184
185
186
187
188
        if not envs.VLLM_USE_NN:
            if os.getenv('LLAMA_NN') != '0': 
                if (architectures == ['QWenLMHeadModel'] or architectures == ['ChatGLMModel'] ) and visions != []:
                    os.environ['LLAMA_NN'] = '0'
                else:
                    os.environ['LLAMA_NN'] = '1'
            if (architectures == ['BloomForCausalLM'] or architectures == ['FalconForCausalLM']) or os.getenv('LM_NN') == '0':
                os.environ['LM_NN'] = '0'
            else:
                os.environ['LM_NN'] = '1'
189
                
190
            if architectures in [['DeepseekV3ForCausalLM'], ['DeepSeekMTPModel']]:
191
192
                if not envs.is_set("VLLM_USE_LIGHTOP"):
                    os.environ['VLLM_USE_LIGHTOP'] = '1'
193
194
                if not envs.is_set("VLLM_USE_OPT_CAT"):
                    os.environ['VLLM_USE_OPT_CAT'] = '1'
195
196
                if not envs.is_set("VLLM_USE_FUSED_FILL_RMS_CAT"):
                    os.environ['VLLM_USE_FUSED_FILL_RMS_CAT'] = '1'
197
198
199
200
201
                # if model_config.quantization in {"slimquant_w4a8", "slimquant_w4a8_marlin", "slimquant_compressed_tensors_marlin", "compressed-tensors"}:
                #     if not envs.is_set("USE_FUSED_RMS_QUANT"):
                #         os.environ['USE_FUSED_RMS_QUANT'] = '1'
                #     if not envs.is_set("USE_FUSED_SILU_MUL_QUANT"):
                #         os.environ['USE_FUSED_SILU_MUL_QUANT'] = '1'
202
            else:
203
204
                # if not envs.is_set("VLLM_USE_PD_SPLIT"):
                #     os.environ['VLLM_USE_PD_SPLIT'] = '1'
205
                if architectures in [['Qwen3MoeForCausalLM']]:
206
207
                    if not envs.is_set("VLLM_USE_LIGHTOP_MOE_ALIGN"):
                        os.environ['VLLM_USE_LIGHTOP_MOE_ALIGN'] = '1'
208
209
210
211
212
213
                    if not envs.is_set("VLLM_USE_LIGHTOP_MOE_SUM"):
                        os.environ['VLLM_USE_LIGHTOP_MOE_SUM'] = '1'    
                    if not envs.is_set("VLLM_USE_FUSE_SILU_AND_MUL"):
                        os.environ['VLLM_USE_FUSE_SILU_AND_MUL'] = '1'
                    if not envs.is_set("VLLM_USE_OPT_RESHAPE_AND_CACHE"):
                        os.environ['VLLM_USE_OPT_RESHAPE_AND_CACHE'] = '1'
214
215
216
                # if architectures in [['Qwen3ForCausalLM']]:
                #     if not envs.is_set("VLLM_USE_OPT_RESHAPE_AND_CACHE"):
                #         os.environ['VLLM_USE_OPT_RESHAPE_AND_CACHE'] = '0'
217

218
219
220
221
            if architectures in [['DeepseekV32ForCausalLM']]:
                if not envs.is_set("VLLM_USE_V32_ENCODE"):
                    os.environ['VLLM_USE_V32_ENCODE'] = '1'
                    
222
223
224
225
            if os.getenv('GEMM_PAD') != '1': 
                os.environ['GEMM_PAD'] = '0'
            if os.getenv('FA_PAD') != '1': 
                os.environ['FA_PAD'] = '0'
226
227
228
229
230
231
        else:
            if architectures in [['DeepseekV3ForCausalLM'], ['DeepSeekMTPModel']]:
                if not envs.is_set("VLLM_USE_LIGHTOP"):
                    os.environ['VLLM_USE_LIGHTOP'] = '1'
                if not envs.is_set("VLLM_USE_OPT_CAT"):
                    os.environ['VLLM_USE_OPT_CAT'] = '1'
232
233
                if not envs.is_set("VLLM_USE_FUSED_FILL_RMS_CAT"):
                    os.environ['VLLM_USE_FUSED_FILL_RMS_CAT'] = '1'
234
235
236
237
238
                # if model_config.quantization in {"slimquant_w4a8", "slimquant_w4a8_marlin", "slimquant_compressed_tensors_marlin", "compressed-tensors"}:
                #     if not envs.is_set("USE_FUSED_RMS_QUANT"):
                #         os.environ['USE_FUSED_RMS_QUANT'] = '1'
                #     if not envs.is_set("USE_FUSED_SILU_MUL_QUANT"):
                #         os.environ['USE_FUSED_SILU_MUL_QUANT'] = '1'
239
            else:
240
241
                # if not envs.is_set("VLLM_USE_PD_SPLIT"):
                #     os.environ['VLLM_USE_PD_SPLIT'] = '1'
242
                if architectures in [['Qwen3MoeForCausalLM']]:
243
244
                    if not envs.is_set("VLLM_USE_LIGHTOP_MOE_ALIGN"):
                        os.environ['VLLM_USE_LIGHTOP_MOE_ALIGN'] = '1'
245
246
247
248
249
250
                    if not envs.is_set("VLLM_USE_LIGHTOP_MOE_SUM"):
                        os.environ['VLLM_USE_LIGHTOP_MOE_SUM'] = '1'    
                    if not envs.is_set("VLLM_USE_FUSE_SILU_AND_MUL"):
                        os.environ['VLLM_USE_FUSE_SILU_AND_MUL'] = '1'
                    if not envs.is_set("VLLM_USE_OPT_RESHAPE_AND_CACHE"):
                        os.environ['VLLM_USE_OPT_RESHAPE_AND_CACHE'] = '1'
251
252
253
254
            
            if architectures in [['DeepseekV32ForCausalLM']]:
                if not envs.is_set("VLLM_USE_V32_ENCODE"):
                    os.environ['VLLM_USE_V32_ENCODE'] = '1'
255
                    
256
        # awq相关配置
zhuwenwen's avatar
zhuwenwen committed
257
        try:
258
259
            # if os.getenv('AWQ_MOE_SZ') == None:
            #     os.environ['AWQ_MOE_SZ'] = '1'
260
            if os.getenv('AWQ_PAD') == None and (torch.cuda.get_device_properties(torch.cuda.current_device()).multi_processor_count == 120):
zhuwenwen's avatar
zhuwenwen committed
261
262
263
264
265
266
                os.environ['AWQ_PAD'] = '1'
        except Exception as e:
            if os.getenv('AWQ_PAD') != '0': 
                os.environ['AWQ_PAD'] = '1'
            else:
                os.environ['AWQ_PAD'] = '0'
zhuwenwen's avatar
zhuwenwen committed
267
268
    else:
        os.environ['LLAMA_NN'] = '0'
zhuwenwen's avatar
zhuwenwen committed
269
        os.environ['LM_NN'] = '0'
270
271
        os.environ['GEMM_PAD'] = '0'
        os.environ['FA_PAD'] = '0'
zhuwenwen's avatar
zhuwenwen committed
272
        os.environ['AWQ_PAD'] = '0'
273

274
275
276
277
278
279
    model_cls, arch = model_config.registry.resolve_model_cls(
        architectures,
        model_config=model_config,
    )

    if arch == model_config._get_transformers_backend_cls():
280
281
        assert model_config.model_impl != "vllm"
        if model_config.model_impl == "auto":
282
283
284
            logger.warning_once(
                "%s has no vLLM implementation, falling back to Transformers "
                "implementation. Some features may not be supported and "
285
286
287
                "performance may not be optimal.",
                arch,
            )
288
289
290
291
292
293

    convert_type = model_config.convert_type
    if convert_type == "none":
        pass
    elif convert_type == "embed":
        logger.debug_once("Converting to embedding model.")
294
        model_cls = as_embedding_model(model_cls)
295
296
    elif convert_type == "classify":
        logger.debug_once("Converting to sequence classification model.")
297
        model_cls = as_seq_cls_model(model_cls)
298
299
    else:
        assert_never(convert_type)
300
301

    return model_cls, arch
302
303


304
305
306
307
308
309
310
311
312
313
314
def get_model_architecture(model_config: ModelConfig) -> tuple[type[nn.Module], str]:
    key = hash(
        (
            model_config.model,
            model_config.convert_type,
            model_config.runner_type,
            model_config.trust_remote_code,
            model_config.model_impl,
            tuple(getattr(model_config.hf_config, "architectures", [])),
        )
    )
315
316
317
318
319
320
321
322
    if key in _MODEL_ARCH_BY_HASH:
        return _MODEL_ARCH_BY_HASH[key]

    model_arch = _get_model_architecture(model_config)
    _MODEL_ARCH_BY_HASH[key] = model_arch
    return model_arch


323
324
325
326
def get_model_cls(model_config: ModelConfig) -> type[nn.Module]:
    return get_model_architecture(model_config)[0]


327
328
def get_architecture_class_name(model_config: ModelConfig) -> str:
    return get_model_architecture(model_config)[1]
329
330
331
332
333
334


@dataclass
class ParamMapping:
    """
    A class to handle parameter mapping for model weight loading.
335
    It creates a bidirectional mapping between packed parameters and their
336
337
    constituent parts.
    """
338

339
    packed_mapping: dict[str, list[str]]
340
    inverse_packed_mapping: dict[str, tuple[str, int]] = field(default_factory=dict)
341
342
343
344
345
346
347
348
349
350
351

    def __post_init__(self):
        for packed_name, sub_params in self.packed_mapping.items():
            # Skip self-contained cases (e.g., {"W_pack": ["W_pack"]})
            if len(sub_params) == 1 and sub_params[0] == packed_name:
                continue
            for index, param_name in enumerate(sub_params):
                self.inverse_packed_mapping[param_name] = (
                    packed_name,
                    index,
                )
352

353
    def get_sub_modules(self, module_name: str) -> tuple[str, list[str]] | None:
354
355
356
357
        for key, value in self.packed_mapping.items():
            if module_name.endswith(key):
                return key, value
        return None
358
359


360
361
362
def configure_quant_config(
    quant_config: QuantizationConfig, model_class: type[nn.Module]
):
363
364
365
366
367
368
    """
    Pass packed_modules_mapping by reference to quant_config so that
    quant_config can properly match fused modules

    Note that model attributes are passed by reference to quant_config,
    enabling them to be updated by model_class.__new__ (ex. chatglm, qwen)
369
370
371

    Once the `SupportsQuant` mixin has been added to all models, this
    function can be removed
372
    """
373
374
375
376
377
378
379
380
381
    if not issubclass(model_class, SupportsQuant):
        hf_to_vllm_mapper = getattr(model_class, "hf_to_vllm_mapper", None)
        packed_mapping = getattr(model_class, "packed_modules_mapping", None)

        # pass mappings by reference to quant_config
        if hf_to_vllm_mapper is not None:
            quant_config.apply_vllm_mapper(hf_to_vllm_mapper)
        if packed_mapping is not None:
            quant_config.packed_modules_mapping = packed_mapping