linear.py 49.4 KB
Newer Older
1
import itertools
2
from abc import abstractmethod
3
from typing import Dict, List, Optional, Tuple
4
5
6

import torch
import torch.nn.functional as F
7
from torch.nn.parameter import Parameter, UninitializedParameter
8

9
10
11
12
13
from vllm.distributed import (divide, get_tensor_model_parallel_rank,
                              get_tensor_model_parallel_world_size,
                              split_tensor_along_last_dim,
                              tensor_model_parallel_all_gather,
                              tensor_model_parallel_all_reduce)
14
from vllm.logger import init_logger
15
16
from vllm.model_executor.layers.quantization.base_config import (
    QuantizationConfig, QuantizeMethodBase)
17
# yapf: disable
18
from vllm.model_executor.parameter import (BasevLLMParameter,
19
                                           BlockQuantScaleParameter,
20
                                           PackedColumnParameter,
21
                                           PackedvLLMParameter,
22
23
                                           PerTensorScaleParameter,
                                           RowvLLMParameter)
24
# yapf: enable
25
26
27
28
from vllm.model_executor.utils import set_weight_attrs

logger = init_logger(__name__)

29
WEIGHT_LOADER_V2_SUPPORTED = [
30
    "CompressedTensorsLinearMethod", "AWQMarlinLinearMethod",
31
    "AWQLinearMethod", "GPTQMarlinLinearMethod", "Fp8LinearMethod",
32
    "MarlinLinearMethod", "QQQLinearMethod", "GPTQMarlin24LinearMethod",
33
    "TPUInt8LinearMethod", "GPTQLinearMethod", "FBGEMMFp8LinearMethod",
34
35
    "ModelOptFp8LinearMethod", "IPEXAWQLinearMethod", "IPEXGPTQLinearMethod",
    "HQQMarlinMethod"
36
]
37

38

39
40
41
42
43
44
45
46
def adjust_marlin_shard(param, shard_size, shard_offset):
    marlin_tile_size = getattr(param, "marlin_tile_size", None)
    if marlin_tile_size is None:
        return shard_size, shard_offset

    return shard_size * marlin_tile_size, shard_offset * marlin_tile_size


47
def adjust_bitsandbytes_4bit_shard(param: Parameter,
48
                                   shard_offsets: Dict[str, Tuple[int, int]],
49
                                   loaded_shard_id: str) -> Tuple[int, int]:
50
51
    """Adjust the quantization offsets and sizes for BitsAndBytes sharding."""

52
53
    total, _ = shard_offsets["total"]
    orig_offset, orig_size = shard_offsets[loaded_shard_id]
54
55
56
57
58
59
60
61

    quantized_total = param.data.shape[0]
    quantized_offset = orig_offset * quantized_total // total
    quantized_size = orig_size * quantized_total // total

    return quantized_size, quantized_offset


62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
def adjust_scalar_to_fused_array(param, loaded_weight, shard_id):
    """For fused modules (QKV and MLP) we have an array of length
    N that holds 1 scale for each "logical" matrix. So the param
    is an array of length N. The loaded_weight corresponds to 
    one of the shards on disk. Here, we slice the param based on 
    the shard_id for loading.
    """
    qkv_idxs = {"q": 0, "k": 1, "v": 2}

    if isinstance(shard_id, str):
        shard_id = qkv_idxs[shard_id]
    elif not isinstance(shard_id, int):
        raise ValueError(f"Unknown Shard Id {shard_id}")

    # AutoFP8 scales do not have a shape
    # compressed-tensors scales do have a shape
    if len(loaded_weight.shape) != 0:
        assert loaded_weight.shape[0] == 1
        loaded_weight = loaded_weight[0]

    return param[shard_id], loaded_weight


85
class LinearMethodBase(QuantizeMethodBase):
86
87
88
    """Base class for different (maybe quantized) linear methods."""

    @abstractmethod
89
90
    def create_weights(self, layer: torch.nn.Module,
                       input_size_per_partition: int,
James Fleming's avatar
James Fleming committed
91
                       output_partition_sizes: List[int], input_size: int,
92
93
                       output_size: int, params_dtype: torch.dtype,
                       **extra_weight_attrs):
94
95
        """Create weights for a linear layer. 
           The weights will be set as attributes of the layer.
96

97
98
99
100
101
102
103
104
105
106
        Args:
            layer: The layer that is using the LinearMethodBase factory.
            input_size_per_partition: Size of the weight input dim on rank X.
            output_partition_sizes: Sizes of the output dim of each logical 
                weight on rank X. E.g., output_partition_sizes for QKVLinear
                is a list contains the width of Wq, Wk, Wv on rank X.
            input_size: Size of the input dim of the weight across all ranks.
            output_size: Size of the output dim of the weight across all ranks.
            params_dtype: Datatype of the parameters.
        """
107
108
109
        raise NotImplementedError

    @abstractmethod
110
111
112
113
    def apply(self,
              layer: torch.nn.Module,
              x: torch.Tensor,
              bias: Optional[torch.Tensor] = None) -> torch.Tensor:
114
115
        """Apply the weights in layer to the input tensor.
        Expects create_weights to have been called before on the layer."""
116
117
118
119
        raise NotImplementedError


class UnquantizedLinearMethod(LinearMethodBase):
120
    """Linear method without quantization."""
121

122
123
    def create_weights(self, layer: torch.nn.Module,
                       input_size_per_partition: int,
James Fleming's avatar
James Fleming committed
124
                       output_partition_sizes: List[int], input_size: int,
125
126
                       output_size: int, params_dtype: torch.dtype,
                       **extra_weight_attrs):
127
        weight = Parameter(torch.empty(sum(output_partition_sizes),
CHU Tianxiang's avatar
CHU Tianxiang committed
128
                                       input_size_per_partition,
129
130
131
                                       dtype=params_dtype),
                           requires_grad=False)
        set_weight_attrs(weight, {"input_dim": 1, "output_dim": 0})
132
133
        layer.register_parameter("weight", weight)
        set_weight_attrs(weight, extra_weight_attrs)
134

135
136
137
138
    def apply(self,
              layer: torch.nn.Module,
              x: torch.Tensor,
              bias: Optional[torch.Tensor] = None) -> torch.Tensor:
139
140

        return F.linear(x, layer.weight, bias)
141
142


143
144
class LinearBase(torch.nn.Module):
    """Base linear layer.
145
146
147
148
149
150
151

    Args:
        input_size: input dimension of the linear layer.
        output_size: output dimension of the linear layer.
        bias: If true, add bias.
        skip_bias_add: If true, skip adding bias but instead return it.
        params_dtype: Data type for the parameters.
152
        quant_config: Quantization configure.
153
154
155
156
157
158
159
160
    """

    def __init__(
        self,
        input_size: int,
        output_size: int,
        skip_bias_add: bool = False,
        params_dtype: Optional[torch.dtype] = None,
161
        quant_config: Optional[QuantizationConfig] = None,
162
        prefix: str = "",
163
164
165
166
167
168
169
170
171
172
    ):
        super().__init__()

        # Keep input parameters
        self.input_size = input_size
        self.output_size = output_size
        self.skip_bias_add = skip_bias_add
        if params_dtype is None:
            params_dtype = torch.get_default_dtype()
        self.params_dtype = params_dtype
173
        if quant_config is None:
174
175
            self.quant_method: Optional[
                QuantizeMethodBase] = UnquantizedLinearMethod()
176
        else:
177
178
            self.quant_method = quant_config.get_quant_method(self,
                                                              prefix=prefix)
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        raise NotImplementedError


class ReplicatedLinear(LinearBase):
    """Replicated linear layer.

    Args:
        input_size: input dimension of the linear layer.
        output_size: output dimension of the linear layer.
        bias: If true, add bias.
        skip_bias_add: If true, skip adding bias but instead return it.
        params_dtype: Data type for the parameters.
        quant_config: Quantization configure.
194
195
        prefix: The name of the layer in the state dict, including all parents
                        (e.g. model.layers.0.qkv_proj)
196
197
    """

198
199
200
201
202
203
    def __init__(self,
                 input_size: int,
                 output_size: int,
                 bias: bool = True,
                 skip_bias_add: bool = False,
                 params_dtype: Optional[torch.dtype] = None,
204
                 quant_config: Optional[QuantizationConfig] = None,
205
206
207
208
209
210
211
                 prefix: str = ""):
        super().__init__(input_size,
                         output_size,
                         skip_bias_add,
                         params_dtype,
                         quant_config,
                         prefix=prefix)
212

213
214
        # All the linear layer supports quant method.
        assert self.quant_method is not None
215
216
217
218
219
        self.quant_method.create_weights(self,
                                         self.input_size, [self.output_size],
                                         self.input_size,
                                         self.output_size,
                                         self.params_dtype,
220
                                         weight_loader=self.weight_loader)
221

222
223
        if bias:
            self.bias = Parameter(
224
                torch.empty(self.output_size, dtype=self.params_dtype))
225
226
227
228
            set_weight_attrs(self.bias, {
                "output_dim": 0,
                "weight_loader": self.weight_loader,
            })
229
230
231
        else:
            self.register_parameter("bias", None)

232
233
234
235
236
237
238
239
240
    def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor):
        # If the weight on disk does not have a shape, give it one
        # (such scales for AutoFp8).
        if len(loaded_weight.shape) == 0:
            loaded_weight = loaded_weight.reshape(1)

        assert param.size() == loaded_weight.size()
        param.data.copy_(loaded_weight)

241
242
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        bias = self.bias if not self.skip_bias_add else None
243
        assert self.quant_method is not None
244
        output = self.quant_method.apply(self, x, bias)
245
246
247
        output_bias = self.bias if self.skip_bias_add else None
        return output, output_bias

248
249
250
251
252
253
    def extra_repr(self) -> str:
        s = f"in_features={self.input_size}"
        s += f", output_features={self.output_size}"
        s += f", bias={self.bias is not None}"
        return s

254

255
class ColumnParallelLinear(LinearBase):
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
    """Linear layer with column parallelism.

    The linear layer is defined as Y = XA + b. A is parallelized along
    its second dimension as A = [A_1, ..., A_p].

    Args:
        input_size: first dimension of matrix A.
        output_size: second dimension of matrix A.
        bias: If true, add bias.
        gather_output: If true, call all-gather on output and make Y available
                       to all GPUs, otherwise, every GPU will have its output
                       which is Y_i = XA_i
        skip_bias_add: This was added to enable performance optimizations where
                       bias can be fused with other element-wise operations. we
                       skip adding bias but instead return it.
        params_dtype: Data type for the parameters.
272
        quant_config: Quantization configure.
James Fleming's avatar
James Fleming committed
273
274
        output_sizes: list of output sizes packed into one output, like for QKV
                       the list would be size 3.
275
276
        prefix: The name of the layer in the state dict, including all parents
                        (e.g. model.layers.0.qkv_proj) 
277
278
    """

279
280
281
282
283
284
285
286
    def __init__(self,
                 input_size: int,
                 output_size: int,
                 bias: bool = True,
                 gather_output: bool = False,
                 skip_bias_add: bool = False,
                 params_dtype: Optional[torch.dtype] = None,
                 quant_config: Optional[QuantizationConfig] = None,
287
                 output_sizes: Optional[List[int]] = None,
288
                 prefix: str = ""):
289
        super().__init__(input_size, output_size, skip_bias_add, params_dtype,
290
                         quant_config, prefix)
291
292

        self.gather_output = gather_output
293

294
295
        # Divide the weight matrix along the last dimension.
        tp_size = get_tensor_model_parallel_world_size()
296
297
298
299
300
301
302
303
304
305
        assert self.quant_method is not None
        self.output_size_per_partition = divide(self.output_size, tp_size)
        self.output_partition_sizes = [self.output_size_per_partition]
        # If QKV or MergedColumn, use output size of each partition.
        if hasattr(self, "output_sizes"):
            self.output_partition_sizes = [
                divide(output_size, tp_size)
                for output_size in self.output_sizes
            ]

James Fleming's avatar
James Fleming committed
306
307
        if output_sizes is None:
            output_sizes = [output_size]
308

309
310
311
312
313
314
315
        self.quant_method.create_weights(
            layer=self,
            input_size_per_partition=self.input_size,
            output_partition_sizes=self.output_partition_sizes,
            input_size=self.input_size,
            output_size=self.output_size,
            params_dtype=self.params_dtype,
316
317
            weight_loader=(
                self.weight_loader_v2 if self.quant_method.__class__.__name__
318
                in WEIGHT_LOADER_V2_SUPPORTED else self.weight_loader))
319
320
321
322
323
324
325
326
327
328
329
330
331
332
        if bias:
            self.bias = Parameter(
                torch.empty(self.output_size_per_partition,
                            dtype=params_dtype))
            set_weight_attrs(self.bias, {
                "output_dim": 0,
                "weight_loader": self.weight_loader,
            })
        else:
            self.register_parameter("bias", None)

    def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor):
        tp_rank = get_tensor_model_parallel_rank()
        output_dim = getattr(param, "output_dim", None)
333
334
335
336
337
338
339
340
341
342
343

        # Special case for GGUF
        is_gguf_weight = getattr(param, "is_gguf_weight", False)
        is_gguf_weight_type = getattr(param, "is_gguf_weight_type", False)
        if is_gguf_weight_type:
            param.weight_type = loaded_weight.item()

        # Materialize GGUF UninitializedParameter
        if is_gguf_weight and isinstance(param, UninitializedParameter):
            param.materialize(loaded_weight.shape, dtype=loaded_weight.dtype)

344
345
        use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit", False)

346
        param_data = param.data
347
348
349
        # bitsandbytes loads the weights of the specific portion
        # no need to narrow here
        if output_dim is not None and not use_bitsandbytes_4bit:
350
351
352
353
            shard_size = param_data.shape[output_dim]
            start_idx = tp_rank * shard_size
            loaded_weight = loaded_weight.narrow(output_dim, start_idx,
                                                 shard_size)
354
355
356
357
358

        # Special case for loading scales off disk, which often do not
        # have a shape (such as in the case of AutoFP8).
        if len(loaded_weight.shape) == 0:
            loaded_weight = loaded_weight.reshape(1)
359

360
361
362
        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)

363
    def weight_loader_v2(self, param: Parameter, loaded_weight: torch.Tensor):
364
365
366
367
368
        # Special case for loading scales off disk, which often do not
        # have a shape (such as in the case of AutoFP8).
        if len(loaded_weight.shape) == 0:
            assert loaded_weight.numel() == 1
            loaded_weight = loaded_weight.reshape(1)
369
370
        param.load_column_parallel_weight(loaded_weight=loaded_weight)

371
372
373
374
    def forward(self, input_):
        bias = self.bias if not self.skip_bias_add else None

        # Matrix multiply.
375
        assert self.quant_method is not None
376
        output_parallel = self.quant_method.apply(self, input_, bias)
377
378
379
380
381
382
383
384
        if self.gather_output:
            # All-gather across the partitions.
            output = tensor_model_parallel_all_gather(output_parallel)
        else:
            output = output_parallel
        output_bias = self.bias if self.skip_bias_add else None
        return output, output_bias

385
386
387
388
389
390
391
392
    def extra_repr(self) -> str:
        s = f"in_features={self.input_size}"
        s += f", output_features={self.output_size_per_partition}"
        s += f", bias={self.bias is not None}"
        s += f", tp_size={get_tensor_model_parallel_world_size()}"
        s += f", gather_output={self.gather_output}"
        return s

393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411

class MergedColumnParallelLinear(ColumnParallelLinear):
    """Packed linear layers with column parallelism.

    Similar to ColumnParallelLinear, but the weight matrix is concatenated
    along the output dimension. When the weight matrix is loaded, the
    different partitions are sharded separately.

    Args:
        input_size: input dimension of the linear layer.
        output_sizes: list of output dimensions of the linear layer.
        bias: If true, add bias.
        gather_output: If true, call all-gather on output and make the output
                       available to all GPUs, otherwise, every GPU will have
                       its own output.
        skip_bias_add: This was added to enable performance optimizations where
                       bias can be fused with other element-wise operations. we
                       skip adding bias but instead return it.
        params_dtype: Data type for the parameters.
412
        quant_config: Quantization configure.
413
414
        prefix: The name of the layer in the state dict, including all parents
                        (e.g. model.layers.0.qkv_proj)
415
416
    """

417
418
419
420
421
422
423
    def __init__(self,
                 input_size: int,
                 output_sizes: List[int],
                 bias: bool = True,
                 gather_output: bool = False,
                 skip_bias_add: bool = False,
                 params_dtype: Optional[torch.dtype] = None,
424
                 quant_config: Optional[QuantizationConfig] = None,
425
                 prefix: str = ""):
426
427
428
        self.output_sizes = output_sizes
        tp_size = get_tensor_model_parallel_world_size()
        assert all(output_size % tp_size == 0 for output_size in output_sizes)
429
430
431
432
433
434
        super().__init__(input_size=input_size,
                         output_size=sum(output_sizes),
                         bias=bias,
                         gather_output=gather_output,
                         skip_bias_add=skip_bias_add,
                         params_dtype=params_dtype,
435
436
                         quant_config=quant_config,
                         prefix=prefix)
437
438
439
440
441

    def weight_loader(self,
                      param: Parameter,
                      loaded_weight: torch.Tensor,
                      loaded_shard_id: Optional[int] = None):
James Fleming's avatar
James Fleming committed
442

443
444
445
446
447
448
449
450
451
        # Special case for GGUF
        # initialize GGUF param after we know the quantize type
        is_gguf_weight = getattr(param, "is_gguf_weight", False)
        is_gguf_weight_type = getattr(param, "is_gguf_weight_type", False)
        if is_gguf_weight_type:
            param.data[loaded_shard_id].copy_(loaded_weight)
            param.shard_weight_type[loaded_shard_id] = loaded_weight.item()
            return

452
453
454
455
456
457
458
        if is_gguf_weight:
            tp_size = get_tensor_model_parallel_world_size()
            tp_rank = get_tensor_model_parallel_rank()

            output_dim = getattr(param, "output_dim", None)
            shard_size = loaded_weight.size(output_dim) // tp_size
            start_idx = tp_rank * shard_size
459

460
461
462
463
464
465
466
467
468
            loaded_weight = loaded_weight.narrow(output_dim, start_idx,
                                                 shard_size)

            param.shard_id.append(loaded_shard_id)
            param.shard_id_map[loaded_shard_id] = len(param.data_container)
            param.data_container.append(loaded_weight)
            if len(param.data_container) == 2:
                self.qweight = param.materialize_nested()
            return
469

470
471
        param_data = param.data
        output_dim = getattr(param, "output_dim", None)
472
        # Special case for AQLM codebooks.
James Fleming's avatar
James Fleming committed
473
        is_metadata = getattr(param, "is_metadata", False)
474
475
        # Special case for per-tensor scale to load scalar into fused array.
        needs_scalar_to_array = getattr(param, "needs_scalar_to_array", False)
476

477
        if loaded_shard_id is None:
478
479
            # Loaded weight is already fused on disk (mlp).
            # (e.g., Phi-3's gate_up_proj).
480
            if output_dim is None:
481
                if needs_scalar_to_array:
482
483
                    param_data, loaded_weight = adjust_scalar_to_fused_array(
                        param_data, loaded_weight, 0)
484

485
486
487
488
                assert param_data.shape == loaded_weight.shape
                param_data.copy_(loaded_weight)
                return
            current_shard_offset = 0
489
490
            use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit",
                                            False)
491
            shard_offsets: List[Tuple[int, int, int]] = []
492
493
494
495
496
            for i, output_size in enumerate(self.output_sizes):
                shard_offsets.append((i, current_shard_offset, output_size))
                current_shard_offset += output_size
            packed_dim = getattr(param, "packed_dim", None)
            for shard_id, shard_offset, shard_size in shard_offsets:
497
                # Special case for Quantization.
498
499
500
501
502
                # If quantized, we need to adjust the offset and size to account
                # for the packing.
                if packed_dim == output_dim:
                    shard_size = shard_size // param.pack_factor
                    shard_offset = shard_offset // param.pack_factor
503
                    # Special case for Marlin.
504
505
                    shard_size, shard_offset = adjust_marlin_shard(
                        param, shard_size, shard_offset)
506

507
                if use_bitsandbytes_4bit:
508
509
510
511
512
513
514
515
516
                    index = list(itertools.accumulate([0] + self.output_sizes))
                    orig_offsets = {
                        str(i): (index[i], size)
                        for i, size in enumerate(self.output_sizes)
                    }
                    orig_offsets["total"] = (self.output_size, 0)
                    shard_size, shard_offset = adjust_bitsandbytes_4bit_shard(
                        param, orig_offsets, str(shard_id))

517
518
519
520
521
522
523
524
525
526
527
                loaded_weight_shard = loaded_weight.narrow(
                    output_dim, shard_offset, shard_size)
                self.weight_loader(param, loaded_weight_shard, shard_id)
            return

        assert loaded_shard_id < len(self.output_sizes)
        tp_rank = get_tensor_model_parallel_rank()
        tp_size = get_tensor_model_parallel_world_size()
        if output_dim is not None:
            shard_offset = sum(self.output_sizes[:loaded_shard_id]) // tp_size
            shard_size = self.output_sizes[loaded_shard_id] // tp_size
528
            # Special case for quantization.
529
530
531
532
533
534
            # If quantized, we need to adjust the offset and size to account
            # for the packing.
            packed_dim = getattr(param, "packed_dim", None)
            if packed_dim == output_dim:
                shard_size = shard_size // param.pack_factor
                shard_offset = shard_offset // param.pack_factor
535
                # Special case for Marlin.
536
537
538
                shard_size, shard_offset = adjust_marlin_shard(
                    param, shard_size, shard_offset)

539
540
541
            use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit",
                                            False)
            if use_bitsandbytes_4bit:
542
543
544
545
                shard_size = loaded_weight.shape[output_dim]
                shard_offset = loaded_weight.shape[output_dim] * \
                    loaded_shard_id

546
547
548
            param_data = param_data.narrow(output_dim, shard_offset,
                                           shard_size)
            start_idx = tp_rank * shard_size
549
550
551
552
553
            # bitsandbytes loads the weights of the specific portion
            # no need to narrow here
            if not use_bitsandbytes_4bit:
                loaded_weight = loaded_weight.narrow(output_dim, start_idx,
                                                     shard_size)
554
        # Special case for AQLM codebooks.
James Fleming's avatar
James Fleming committed
555
556
557
558
559
        elif is_metadata:
            # metadata indicates fixed size concatenated along dim 0
            shard_size = loaded_weight.shape[0]
            shard_offset = loaded_shard_id * shard_size
            param_data = param_data.narrow(0, shard_offset, shard_size)
560

561
562
563
        # Special case for per-tensor scales in fused case.
        elif needs_scalar_to_array:
            param_data, loaded_weight = adjust_scalar_to_fused_array(
564
565
                param_data, loaded_weight, loaded_shard_id)

566
        else:
CHU Tianxiang's avatar
CHU Tianxiang committed
567
568
569
570
571
572
            ignore_warning = getattr(param, "ignore_warning", False)
            if not ignore_warning:
                logger.warning(
                    "Loading a weight without `output_dim` attribute in "
                    "MergedColumnParallelLinear, assume the weight is "
                    "the same for all partitions.")
573

574
575
576
        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)

577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
    def _load_fused_module_from_checkpoint(self, param: BasevLLMParameter,
                                           loaded_weight: torch.Tensor):
        """
        Handle special case for models where MLP layers are already
        fused on disk. In this case, we have no shard id. This function
        determmines the shard id by splitting these layers and then calls
        the weight loader using the shard id.

        An example of a model with these fused layers:
        https://huggingface.co/microsoft/Phi-3-mini-4k-instruct
        """

        current_shard_offset = 0
        shard_offsets: List[Tuple[int, int, int]] = []
        for i, output_size in enumerate(self.output_sizes):
            shard_offsets.append((i, current_shard_offset, output_size))
            current_shard_offset += output_size

        for shard_id, shard_offset, shard_size in shard_offsets:
            # Special case for Quantization.
            # If quantized, we need to adjust the offset and size to account
            # for the packing.
599
600
            if isinstance(param, (PackedColumnParameter, PackedvLLMParameter
                                  )) and param.packed_dim == param.output_dim:
601
602
                shard_size, shard_offset = \
                    param.adjust_shard_indexes_for_packing(
603
604
605
606
607
608
609
610
611
612
613
614
                    shard_size=shard_size, shard_offset=shard_offset)

            loaded_weight_shard = loaded_weight.narrow(param.output_dim,
                                                       shard_offset,
                                                       shard_size)
            self.weight_loader_v2(param, loaded_weight_shard, shard_id)

    def weight_loader_v2(self,
                         param: BasevLLMParameter,
                         loaded_weight: torch.Tensor,
                         loaded_shard_id: Optional[int] = None):
        if loaded_shard_id is None:
615
616
617
618
            if isinstance(param, PerTensorScaleParameter):
                param.load_merged_column_weight(loaded_weight=loaded_weight,
                                                shard_id=0)
                return
619
            elif type(param) in (RowvLLMParameter, BasevLLMParameter):
620
                param.load_merged_column_weight(loaded_weight=loaded_weight)
621
                return
622
            # TODO: @dsikka - move to parameter.py
623
624
625
626
627
628
            self._load_fused_module_from_checkpoint(param, loaded_weight)
            return

        assert loaded_shard_id < len(self.output_sizes)

        tp_size = get_tensor_model_parallel_world_size()
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646

        if isinstance(param, BlockQuantScaleParameter):
            from vllm.model_executor.layers.quantization.fp8 import (
                Fp8LinearMethod, Fp8MoEMethod)
            assert self.quant_method is not None
            assert isinstance(self.quant_method,
                              (Fp8LinearMethod, Fp8MoEMethod))
            weight_block_size = self.quant_method.quant_config.weight_block_size
            assert weight_block_size is not None
            block_n, _ = weight_block_size[0], weight_block_size[1]
            shard_offset = (
                (sum(self.output_sizes[:loaded_shard_id]) + block_n - 1) //
                block_n) // tp_size
            shard_size = ((self.output_sizes[loaded_shard_id] + block_n - 1) //
                          block_n // tp_size)
        else:
            shard_offset = sum(self.output_sizes[:loaded_shard_id]) // tp_size
            shard_size = self.output_sizes[loaded_shard_id] // tp_size
647
648
649
650
651
652

        param.load_merged_column_weight(loaded_weight=loaded_weight,
                                        shard_id=loaded_shard_id,
                                        shard_offset=shard_offset,
                                        shard_size=shard_size)

653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674

class QKVParallelLinear(ColumnParallelLinear):
    """Linear layers for the attention's QKV transformation.

    Linear layers for the linear transformation of the query, key, and value
    vectors in the attention layer. The weight matrix is concatenated along
    the output dimension. The layer is parallelized along the head dimension.
    When the number of key/value heads is smaller than the number of query
    heads (e.g., multi-query/grouped-query attention), the key/value head may
    be replicated while the query heads are partitioned.

    Args:
        hidden_size: input hidden state size of the transformer.
        head_size: size of each attention head.
        total_num_heads: total number of attention query heads.
        total_num_kv_heads: total number of attention key/value heads. If
                            None, assume total_num_kv_heads = total_num_heads.
        bias: If true, add bias.
        skip_bias_add: This was added to enable performance optimizations where
                       bias can be fused with other element-wise operations. we
                       skip adding bias but instead return it.
        params_dtype: Data type for the parameters.
675
        quant_config: Quantization configure.
676
677
        prefix: The name of the layer in the state dict, including all parents
                        (e.g. model.layers.0.qkv_proj)
678
679
    """

680
681
682
683
684
685
686
687
    def __init__(self,
                 hidden_size: int,
                 head_size: int,
                 total_num_heads: int,
                 total_num_kv_heads: Optional[int] = None,
                 bias: bool = True,
                 skip_bias_add: bool = False,
                 params_dtype: Optional[torch.dtype] = None,
688
                 quant_config: Optional[QuantizationConfig] = None,
689
                 prefix: str = ""):
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
        self.hidden_size = hidden_size
        self.head_size = head_size
        self.total_num_heads = total_num_heads
        if total_num_kv_heads is None:
            total_num_kv_heads = total_num_heads
        self.total_num_kv_heads = total_num_kv_heads
        # Divide the weight matrix along the last dimension.
        tp_size = get_tensor_model_parallel_world_size()
        self.num_heads = divide(self.total_num_heads, tp_size)
        if tp_size >= self.total_num_kv_heads:
            self.num_kv_heads = 1
            self.num_kv_head_replicas = divide(tp_size,
                                               self.total_num_kv_heads)
        else:
            self.num_kv_heads = divide(self.total_num_kv_heads, tp_size)
            self.num_kv_head_replicas = 1
        input_size = self.hidden_size
        output_size = (self.num_heads +
                       2 * self.num_kv_heads) * tp_size * self.head_size
709
710
711
712
        self.output_sizes = [
            self.num_heads * self.head_size * tp_size,  # q_proj
            self.num_kv_heads * self.head_size * tp_size,  # k_proj
            self.num_kv_heads * self.head_size * tp_size,  # v_proj 
James Fleming's avatar
James Fleming committed
713
714
        ]

715
716
717
718
719
720
        super().__init__(input_size=input_size,
                         output_size=output_size,
                         bias=bias,
                         gather_output=False,
                         skip_bias_add=skip_bias_add,
                         params_dtype=params_dtype,
721
722
                         quant_config=quant_config,
                         prefix=prefix)
723

724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
    def _get_shard_offset_mapping(self, loaded_shard_id: str):
        shard_offset_mapping = {
            "q": 0,
            "k": self.num_heads * self.head_size,
            "v": (self.num_heads + self.num_kv_heads) * self.head_size,
            "total": (self.num_heads + 2 * self.num_kv_heads) * self.head_size
        }
        return shard_offset_mapping.get(loaded_shard_id)

    def _get_shard_size_mapping(self, loaded_shard_id: str):
        shard_size_mapping = {
            "q": self.num_heads * self.head_size,
            "k": self.num_kv_heads * self.head_size,
            "v": self.num_kv_heads * self.head_size,
        }
        return shard_size_mapping.get(loaded_shard_id)

    def _load_fused_module_from_checkpoint(self, param: BasevLLMParameter,
                                           loaded_weight: torch.Tensor):
        """
        Handle special case for models where QKV layers are already 
        fused on disk. In this case, we have no shard id. This function
        determmines the shard id by splitting these layers and then calls
        the weight loader using the shard id.

        An example of a model with these fused layers:
        https://huggingface.co/microsoft/Phi-3-mini-4k-instruct
        """
        shard_offsets = [
            # (shard_id, shard_offset, shard_size)
            ("q", 0, self.total_num_heads * self.head_size),
            ("k", self.total_num_heads * self.head_size,
             self.total_num_kv_heads * self.head_size),
            ("v",
             (self.total_num_heads + self.total_num_kv_heads) * self.head_size,
             self.total_num_kv_heads * self.head_size),
        ]

        for shard_id, shard_offset, shard_size in shard_offsets:
            # Special case for Quantization.
            # If quantized, we need to adjust the offset and size to account
            # for the packing.
766
767
            if isinstance(param, (PackedColumnParameter, PackedvLLMParameter
                                  )) and param.packed_dim == param.output_dim:
768
769
                shard_size, shard_offset = \
                    param.adjust_shard_indexes_for_packing(
770
771
772
773
774
775
776
777
778
779
780
781
                    shard_size=shard_size, shard_offset=shard_offset)

            loaded_weight_shard = loaded_weight.narrow(param.output_dim,
                                                       shard_offset,
                                                       shard_size)
            self.weight_loader_v2(param, loaded_weight_shard, shard_id)

    def weight_loader_v2(self,
                         param: BasevLLMParameter,
                         loaded_weight: torch.Tensor,
                         loaded_shard_id: Optional[str] = None):
        if loaded_shard_id is None:  # special case for certain models
782
            if isinstance(param, PerTensorScaleParameter):
783
                param.load_qkv_weight(loaded_weight=loaded_weight, shard_id=0)
784
                return
785
786
            elif type(param) in (RowvLLMParameter, BasevLLMParameter):
                param.load_qkv_weight(loaded_weight=loaded_weight)
787
                return
788
            # TODO: @dsikka - move to parameter.py
789
790
791
792
793
794
795
796
797
798
799
800
801
802
            self._load_fused_module_from_checkpoint(param, loaded_weight)
            return

        assert loaded_shard_id in ["q", "k", "v"]

        shard_offset = self._get_shard_offset_mapping(loaded_shard_id)
        shard_size = self._get_shard_size_mapping(loaded_shard_id)

        param.load_qkv_weight(loaded_weight=loaded_weight,
                              num_heads=self.num_kv_head_replicas,
                              shard_id=loaded_shard_id,
                              shard_offset=shard_offset,
                              shard_size=shard_size)

803
804
805
806
    def weight_loader(self,
                      param: Parameter,
                      loaded_weight: torch.Tensor,
                      loaded_shard_id: Optional[str] = None):
807
808
809
810
811
812
813
814
815
816
817

        # Special case for GGUF
        # initialize GGUF param after we know the quantize type
        is_gguf_weight = getattr(param, "is_gguf_weight", False)
        is_gguf_weight_type = getattr(param, "is_gguf_weight_type", False)
        if is_gguf_weight_type and loaded_shard_id is not None:
            idx_map = {"q": 0, "k": 1, "v": 2}
            param.data[idx_map[loaded_shard_id]].copy_(loaded_weight)
            param.shard_weight_type[loaded_shard_id] = loaded_weight.item()
            return

818
819
820
821
822
823
824
825
826
827
        if is_gguf_weight:
            tp_size = get_tensor_model_parallel_world_size()
            tp_rank = get_tensor_model_parallel_rank()

            output_dim = getattr(param, "output_dim", None)
            shard_size = loaded_weight.size(output_dim) // tp_size
            start_idx = tp_rank * shard_size

            loaded_weight = loaded_weight.narrow(output_dim, start_idx,
                                                 shard_size)
828

829
830
831
832
833
834
            param.shard_id.append(loaded_shard_id)
            param.shard_id_map[loaded_shard_id] = len(param.data_container)
            param.data_container.append(loaded_weight)
            if len(param.data_container) == 3:
                self.qweight = param.materialize_nested()
            return
835

836
837
        param_data = param.data
        output_dim = getattr(param, "output_dim", None)
838
        # Special case for AQLM codebooks.
James Fleming's avatar
James Fleming committed
839
        is_metadata = getattr(param, "is_metadata", False)
840

841
842
        # Special case for per-tensor scales in fused case.
        needs_scalar_to_array = getattr(param, "needs_scalar_to_array", False)
843

844
        if loaded_shard_id is None:
845
846
            # Loaded weight is already fused on disk (qkv).
            # (e.g., Phi-3's qkv_proj).
847
            if output_dim is None:
848
                if needs_scalar_to_array:
849
850
                    param_data, loaded_weight = adjust_scalar_to_fused_array(
                        param_data, loaded_weight, 0)
851

852
853
854
855
856
857
858
859
860
861
862
                assert param_data.shape == loaded_weight.shape
                param_data.copy_(loaded_weight)
                return
            shard_offsets = [
                # (shard_id, shard_offset, shard_size)
                ("q", 0, self.total_num_heads * self.head_size),
                ("k", self.total_num_heads * self.head_size,
                 self.total_num_kv_heads * self.head_size),
                ("v", (self.total_num_heads + self.total_num_kv_heads) *
                 self.head_size, self.total_num_kv_heads * self.head_size),
            ]
863
864
865
            use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit",
                                            False)

866
867
            packed_dim = getattr(param, "packed_dim", None)
            for shard_id, shard_offset, shard_size in shard_offsets:
868
                # Special case for Quantized Weights.
869
870
871
872
873
                # If quantized, we need to adjust the offset and size to account
                # for the packing.
                if packed_dim == output_dim:
                    shard_size = shard_size // param.pack_factor
                    shard_offset = shard_offset // param.pack_factor
874

875
                    # Special case for Marlin.
876
877
878
                    shard_size, shard_offset = adjust_marlin_shard(
                        param, shard_size, shard_offset)

879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
                if use_bitsandbytes_4bit:
                    orig_qkv_offsets = {
                        "q": (0, self.total_num_heads * self.head_size),
                        "k": (self.total_num_heads * self.head_size,
                              self.total_num_kv_heads * self.head_size),
                        "v":
                        ((self.total_num_heads + self.total_num_kv_heads) *
                         self.head_size,
                         self.total_num_kv_heads * self.head_size),
                        "total":
                        ((self.total_num_heads + 2 * self.total_num_kv_heads) *
                         self.head_size, 0)
                    }

                    shard_size, shard_offset = adjust_bitsandbytes_4bit_shard(
                        param, orig_qkv_offsets, shard_id)

896
897
898
899
900
901
902
                loaded_weight_shard = loaded_weight.narrow(
                    output_dim, shard_offset, shard_size)
                self.weight_loader(param, loaded_weight_shard, shard_id)
            return

        tp_rank = get_tensor_model_parallel_rank()
        assert loaded_shard_id in ["q", "k", "v"]
903
904

        # If output dim is defined, use the default loading process.
905
906
907
908
909
910
911
912
913
914
915
        if output_dim is not None:
            if loaded_shard_id == "q":
                shard_offset = 0
                shard_size = self.num_heads * self.head_size
            elif loaded_shard_id == "k":
                shard_offset = self.num_heads * self.head_size
                shard_size = self.num_kv_heads * self.head_size
            elif loaded_shard_id == "v":
                shard_offset = (self.num_heads +
                                self.num_kv_heads) * self.head_size
                shard_size = self.num_kv_heads * self.head_size
916
            # Special case for Quantized Weights.
917
918
919
920
921
922
            # If quantized, we need to adjust the offset and size to account
            # for the packing.
            packed_dim = getattr(param, "packed_dim", None)
            if packed_dim == output_dim:
                shard_size = shard_size // param.pack_factor
                shard_offset = shard_offset // param.pack_factor
923

924
                # Special case for Marlin.
925
926
927
                shard_size, shard_offset = adjust_marlin_shard(
                    param, shard_size, shard_offset)

928
929
930
            use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit",
                                            False)
            if use_bitsandbytes_4bit:
931
932
933
934
935
936
937
938
939
940
941
                orig_qkv_offsets = {
                    "q": (0, self.num_heads * self.head_size),
                    "k": (self.num_heads * self.head_size,
                          self.num_kv_heads * self.head_size),
                    "v":
                    ((self.num_heads + self.num_kv_heads) * self.head_size,
                     self.num_kv_heads * self.head_size),
                    "total":
                    ((self.num_heads + 2 * self.num_kv_heads) * self.head_size,
                     0)
                }
942
                shard_size, shard_offset = adjust_bitsandbytes_4bit_shard(
943
944
                    param, orig_qkv_offsets, loaded_shard_id)

945
946
            param_data = param_data.narrow(output_dim, shard_offset,
                                           shard_size)
947
948
949
950
            if loaded_shard_id == "q":
                shard_id = tp_rank
            else:
                shard_id = tp_rank // self.num_kv_head_replicas
951
            start_idx = shard_id * shard_size
952
953
954
955
956
957
958

            # bitsandbytes loads the weights of the specific portion
            # no need to narrow here
            if not use_bitsandbytes_4bit:
                loaded_weight = loaded_weight.narrow(output_dim, start_idx,
                                                     shard_size)

959
        # Special case for for AQLM codebooks.
James Fleming's avatar
James Fleming committed
960
961
962
963
964
965
        elif is_metadata:
            # metadata indicates fixed size concatenated along dim 0
            shard_size = loaded_weight.shape[0]
            shard_index = ["q", "k", "v"].index(loaded_shard_id)
            param_data = param_data.narrow(0, shard_index * shard_size,
                                           shard_size)
966
967
968
        # Special case for per-tensor scales in fused case.
        elif needs_scalar_to_array:
            param_data, loaded_weight = adjust_scalar_to_fused_array(
969
                param_data, loaded_weight, loaded_shard_id)
970
        else:
CHU Tianxiang's avatar
CHU Tianxiang committed
971
972
973
974
975
976
            ignore_warning = getattr(param, "ignore_warning", False)
            if not ignore_warning:
                logger.warning(
                    "Loading a weight without `output_dim` attribute in "
                    "QKVParallelLinear, assume the weight is the same "
                    "for all partitions.")
977

978
979
980
981
        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)


982
class RowParallelLinear(LinearBase):
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
    """Linear layer with row parallelism.

    The linear layer is defined as Y = XA + b. A is parallelized along
    its first dimension and X along its second dimension as:
               -   -
              | A_1 |
              | .   |
          A = | .   |        X = [X_1, ..., X_p]
              | .   |
              | A_p |
               -   -
    Arguments:
        input_size: first dimension of matrix A.
        output_size: second dimension of matrix A.
        bias: If true, add bias. Note that bias is not parallelized.
        input_is_parallel: If true, we assume that the input is already
                           split across the GPUs and we do not split
                           again.
        skip_bias_add: This was added to enable performance optimization where
                       bias can be fused with other element-wise operations.
                       We skip adding bias but instead return it.
        params_dtype: Data type for the parameters.
1005
        quant_config: Quantization configure.
1006
1007
    """

1008
1009
1010
1011
1012
1013
1014
1015
    def __init__(self,
                 input_size: int,
                 output_size: int,
                 bias: bool = True,
                 input_is_parallel: bool = True,
                 skip_bias_add: bool = False,
                 params_dtype: Optional[torch.dtype] = None,
                 reduce_results: bool = True,
1016
                 quant_config: Optional[QuantizationConfig] = None,
1017
                 prefix: str = ""):
1018
        super().__init__(input_size, output_size, skip_bias_add, params_dtype,
1019
                         quant_config, prefix)
1020

1021
1022
1023
1024
        self.input_is_parallel = input_is_parallel
        self.reduce_results = reduce_results

        # Divide the weight matrix along the last dimension.
1025
        self.tp_rank = get_tensor_model_parallel_rank()
1026
1027
        self.tp_size = get_tensor_model_parallel_world_size()
        self.input_size_per_partition = divide(input_size, self.tp_size)
1028
        assert self.quant_method is not None
1029

1030
1031
1032
1033
1034
1035
1036
        self.quant_method.create_weights(
            layer=self,
            input_size_per_partition=self.input_size_per_partition,
            output_partition_sizes=[self.output_size],
            input_size=self.input_size,
            output_size=self.output_size,
            params_dtype=self.params_dtype,
1037
1038
            weight_loader=(
                self.weight_loader_v2 if self.quant_method.__class__.__name__
1039
                in WEIGHT_LOADER_V2_SUPPORTED else self.weight_loader))
1040
1041
1042
1043
1044
1045
        if not reduce_results and (bias and not skip_bias_add):
            raise ValueError("When not reduce the results, adding bias to the "
                             "results can lead to incorrect results")

        if bias:
            self.bias = Parameter(
1046
                torch.empty(self.output_size, dtype=params_dtype))
1047
1048
1049
1050
1051
1052
1053
1054
1055
            set_weight_attrs(self.bias, {
                "output_dim": 0,
                "weight_loader": self.weight_loader,
            })
        else:
            self.register_parameter("bias", None)

    def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor):
        tp_rank = get_tensor_model_parallel_rank()
1056
        tp_size = get_tensor_model_parallel_world_size()
1057
        input_dim = getattr(param, "input_dim", None)
1058
        use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit", False)
1059
1060
1061
1062
1063
1064
1065
1066
1067

        # Special case for GGUF
        is_gguf_weight = getattr(param, "is_gguf_weight", False)
        is_gguf_weight_type = getattr(param, "is_gguf_weight_type", False)
        if is_gguf_weight_type:
            param.weight_type = loaded_weight.item()

        # Materialize GGUF UninitializedParameter
        if is_gguf_weight and isinstance(param, UninitializedParameter):
1068
1069
1070
1071
            weight_shape = list(loaded_weight.shape)
            if input_dim:
                weight_shape[input_dim] = weight_shape[input_dim] // tp_size
            param.materialize(tuple(weight_shape), dtype=loaded_weight.dtype)
1072

1073
        param_data = param.data
1074
1075
1076
        # bitsandbytes loads the weights of the specific portion
        # no need to narrow here
        if input_dim is not None and not use_bitsandbytes_4bit:
1077
1078
1079
1080
            shard_size = param_data.shape[input_dim]
            start_idx = tp_rank * shard_size
            loaded_weight = loaded_weight.narrow(input_dim, start_idx,
                                                 shard_size)
1081

1082
1083
1084
        # Special case for loading scales off disk, which often do not
        # have a shape (such as in the case of AutoFP8).
        if len(loaded_weight.shape) == 0:
1085
1086
            loaded_weight = loaded_weight.reshape(1)

1087
1088
1089
        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)

1090
1091
    def weight_loader_v2(self, param: BasevLLMParameter,
                         loaded_weight: torch.Tensor):
1092
1093
1094
1095
1096
1097
1098

        # Special case for loading scales off disk, which often do not
        # have a shape (such as in the case of AutoFP8).
        if len(loaded_weight.shape) == 0:
            assert loaded_weight.numel() == 1
            loaded_weight = loaded_weight.reshape(1)

1099
1100
        param.load_row_parallel_weight(loaded_weight=loaded_weight)

1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
    def forward(self, input_):
        if self.input_is_parallel:
            input_parallel = input_
        else:
            tp_rank = get_tensor_model_parallel_rank()
            splitted_input = split_tensor_along_last_dim(
                input_, num_partitions=self.tp_size)
            input_parallel = splitted_input[tp_rank].contiguous()

        # Matrix multiply.
1111
        assert self.quant_method is not None
1112
1113
1114
1115
1116
1117
        # Only fuse bias add into GEMM for rank 0 (this ensures that
        # bias will not get added more than once in TP>1 case)
        bias_ = None if (self.tp_rank > 0 or self.skip_bias_add) else self.bias
        output_parallel = self.quant_method.apply(self,
                                                  input_parallel,
                                                  bias=bias_)
1118
        if self.reduce_results and self.tp_size > 1:
1119
            output = tensor_model_parallel_all_reduce(output_parallel)
1120
        else:
1121
1122
1123
            output = output_parallel

        output_bias = self.bias if self.skip_bias_add else None
1124
1125

        return output, output_bias
1126
1127
1128
1129
1130
1131
1132
1133

    def extra_repr(self) -> str:
        s = f"input_features={self.input_size_per_partition}"
        s += f", output_features={self.output_size}"
        s += f", bias={self.bias is not None}"
        s += f", tp_size={self.tp_size}"
        s += f", reduce_results={self.reduce_results}"
        return s