"vllm/config/vllm.py" did not exist on "0ff8ebb2d700b2e39457a661ef979b0da2ad73b3"
linear.py 44.6 KB
Newer Older
1
from abc import abstractmethod
2
from typing import Dict, List, Optional, Tuple
3
4
5

import torch
import torch.nn.functional as F
6
from torch.nn.parameter import Parameter, UninitializedParameter
7

8
9
10
11
12
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)
13
from vllm.logger import init_logger
14
15
from vllm.model_executor.layers.quantization.base_config import (
    QuantizationConfig, QuantizeMethodBase)
16
17
from vllm.model_executor.parameter import (BasevLLMParameter,
                                           PackedvLLMParameter)
18
19
20
21
from vllm.model_executor.utils import set_weight_attrs

logger = init_logger(__name__)

22
23
WEIGHT_LOADER_V2_SUPPORTED = ["CompressedTensorsLinearMethod"]

24

25
26
27
28
29
30
31
32
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


33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
def adjust_bitsandbytes_shard(param: Parameter,
                              qkv_offsets: Dict[str, Tuple[int, int]],
                              loaded_shard_id: str) -> Tuple[int, int]:
    """Adjust the quantization offsets and sizes for BitsAndBytes sharding."""

    total, _ = qkv_offsets["total"]
    orig_offset, orig_size = qkv_offsets[loaded_shard_id]

    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


48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
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


71
class LinearMethodBase(QuantizeMethodBase):
72
73
74
    """Base class for different (maybe quantized) linear methods."""

    @abstractmethod
75
76
    def create_weights(self, layer: torch.nn.Module,
                       input_size_per_partition: int,
James Fleming's avatar
James Fleming committed
77
                       output_partition_sizes: List[int], input_size: int,
78
79
                       output_size: int, params_dtype: torch.dtype,
                       **extra_weight_attrs):
80
81
        """Create weights for a linear layer. 
           The weights will be set as attributes of the layer.
82

83
84
85
86
87
88
89
90
91
92
        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.
        """
93
94
95
        raise NotImplementedError

    @abstractmethod
96
97
98
99
    def apply(self,
              layer: torch.nn.Module,
              x: torch.Tensor,
              bias: Optional[torch.Tensor] = None) -> torch.Tensor:
100
101
        """Apply the weights in layer to the input tensor.
        Expects create_weights to have been called before on the layer."""
102
103
104
105
        raise NotImplementedError


class UnquantizedLinearMethod(LinearMethodBase):
106
    """Linear method without quantization."""
107

108
109
    def create_weights(self, layer: torch.nn.Module,
                       input_size_per_partition: int,
James Fleming's avatar
James Fleming committed
110
                       output_partition_sizes: List[int], input_size: int,
111
112
                       output_size: int, params_dtype: torch.dtype,
                       **extra_weight_attrs):
113
        weight = Parameter(torch.empty(sum(output_partition_sizes),
CHU Tianxiang's avatar
CHU Tianxiang committed
114
                                       input_size_per_partition,
115
116
117
                                       dtype=params_dtype),
                           requires_grad=False)
        set_weight_attrs(weight, {"input_dim": 1, "output_dim": 0})
118
119
        layer.register_parameter("weight", weight)
        set_weight_attrs(weight, extra_weight_attrs)
120

121
122
123
124
    def apply(self,
              layer: torch.nn.Module,
              x: torch.Tensor,
              bias: Optional[torch.Tensor] = None) -> torch.Tensor:
125
126

        return F.linear(x, layer.weight, bias)
127
128


129
130
class LinearBase(torch.nn.Module):
    """Base linear layer.
131
132
133
134
135
136
137

    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.
138
        quant_config: Quantization configure.
139
140
141
142
143
144
145
146
    """

    def __init__(
        self,
        input_size: int,
        output_size: int,
        skip_bias_add: bool = False,
        params_dtype: Optional[torch.dtype] = None,
147
        quant_config: Optional[QuantizationConfig] = None,
148
        prefix: str = "",
149
150
151
152
153
154
155
156
157
158
    ):
        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
159
        if quant_config is None:
160
161
            self.quant_method: Optional[
                QuantizeMethodBase] = UnquantizedLinearMethod()
162
        else:
163
164
            self.quant_method = quant_config.get_quant_method(self,
                                                              prefix=prefix)
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179

    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.
180
181
        prefix: The name of the layer in the state dict, including all parents
                        (e.g. model.layers.0.qkv_proj)
182
183
    """

184
185
186
187
188
189
    def __init__(self,
                 input_size: int,
                 output_size: int,
                 bias: bool = True,
                 skip_bias_add: bool = False,
                 params_dtype: Optional[torch.dtype] = None,
190
                 quant_config: Optional[QuantizationConfig] = None,
191
192
193
194
195
196
197
                 prefix: str = ""):
        super().__init__(input_size,
                         output_size,
                         skip_bias_add,
                         params_dtype,
                         quant_config,
                         prefix=prefix)
198

199
200
        # All the linear layer supports quant method.
        assert self.quant_method is not None
201
202
203
204
205
        self.quant_method.create_weights(self,
                                         self.input_size, [self.output_size],
                                         self.input_size,
                                         self.output_size,
                                         self.params_dtype,
206
                                         weight_loader=self.weight_loader,
207
                                         prefix=prefix)
208

209
210
        if bias:
            self.bias = Parameter(
211
                torch.empty(self.output_size, dtype=self.params_dtype))
212
213
214
215
            set_weight_attrs(self.bias, {
                "output_dim": 0,
                "weight_loader": self.weight_loader,
            })
216
217
218
        else:
            self.register_parameter("bias", None)

219
220
221
222
223
224
225
226
227
    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)

228
229
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        bias = self.bias if not self.skip_bias_add else None
230
        assert self.quant_method is not None
231
        output = self.quant_method.apply(self, x, bias)
232
233
234
        output_bias = self.bias if self.skip_bias_add else None
        return output, output_bias

235
236
237
238
239
240
    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

241

242
class ColumnParallelLinear(LinearBase):
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
    """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.
259
        quant_config: Quantization configure.
James Fleming's avatar
James Fleming committed
260
261
        output_sizes: list of output sizes packed into one output, like for QKV
                       the list would be size 3.
262
263
        prefix: The name of the layer in the state dict, including all parents
                        (e.g. model.layers.0.qkv_proj) 
264
265
    """

266
267
268
269
270
271
272
273
    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,
274
                 output_sizes: Optional[List[int]] = None,
275
                 prefix: str = ""):
276
        super().__init__(input_size, output_size, skip_bias_add, params_dtype,
277
                         quant_config, prefix)
278
279

        self.gather_output = gather_output
280

281
282
        # Divide the weight matrix along the last dimension.
        tp_size = get_tensor_model_parallel_world_size()
283
284
285
286
287
288
289
290
291
292
        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
293
294
        if output_sizes is None:
            output_sizes = [output_size]
295

296
297
298
299
300
301
302
        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,
303
304
305
            weight_loader=(
                self.weight_loader_v2 if self.quant_method.__class__.__name__
                in WEIGHT_LOADER_V2_SUPPORTED else self.weight_loader),
306
            prefix=prefix)
307
308
309
310
311
312
313
314
315
316
317
318
319
320
        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)
321
322
323
324
325
326
327
328
329
330
331

        # 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)

332
333
334
335
336
337
        param_data = param.data
        if output_dim is not None:
            shard_size = param_data.shape[output_dim]
            start_idx = tp_rank * shard_size
            loaded_weight = loaded_weight.narrow(output_dim, start_idx,
                                                 shard_size)
338
339
340
341
342

        # 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)
343

344
345
346
        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)

347
348
349
    def weight_loader_v2(self, param: Parameter, loaded_weight: torch.Tensor):
        param.load_column_parallel_weight(loaded_weight=loaded_weight)

350
351
352
353
    def forward(self, input_):
        bias = self.bias if not self.skip_bias_add else None

        # Matrix multiply.
354
        assert self.quant_method is not None
355
        output_parallel = self.quant_method.apply(self, input_, bias)
356
357
358
359
360
361
362
363
        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

364
365
366
367
368
369
370
371
    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

372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390

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.
391
        quant_config: Quantization configure.
392
393
        prefix: The name of the layer in the state dict, including all parents
                        (e.g. model.layers.0.qkv_proj)
394
395
    """

396
397
398
399
400
401
402
    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,
403
                 quant_config: Optional[QuantizationConfig] = None,
404
                 prefix: str = ""):
405
406
407
        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)
408
409
410
411
412
413
        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,
414
415
                         quant_config=quant_config,
                         prefix=prefix)
416
417
418
419
420

    def weight_loader(self,
                      param: Parameter,
                      loaded_weight: torch.Tensor,
                      loaded_shard_id: Optional[int] = None):
James Fleming's avatar
James Fleming committed
421

422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
        # 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

        if is_gguf_weight and isinstance(param, UninitializedParameter):
            from gguf.constants import GGML_QUANT_SIZES

            ori_shape = param.tensor_shape
            weight_types = self.qweight_type.shard_weight_type.values()
            row_size = []
            for weight_type in weight_types:
                block_size, type_size = GGML_QUANT_SIZES[weight_type]
                row_size.append(ori_shape[1] // block_size * type_size)
            q_shape = (ori_shape[0], max(row_size))
            param.materialize(q_shape, dtype=loaded_weight.dtype)

443
444
        param_data = param.data
        output_dim = getattr(param, "output_dim", None)
445
        # Special case for AQLM codebooks.
James Fleming's avatar
James Fleming committed
446
        is_metadata = getattr(param, "is_metadata", False)
447
448
        # Special case for per-tensor scale to load scalar into fused array.
        needs_scalar_to_array = getattr(param, "needs_scalar_to_array", False)
449

450
        if loaded_shard_id is None:
451
            # Loaded weight is already fused on disk (qkv/mlp).
452
            if output_dim is None:
453
                if needs_scalar_to_array:
454
455
                    param_data, loaded_weight = adjust_scalar_to_fused_array(
                        param_data, loaded_weight, 0)
456

457
458
459
460
                assert param_data.shape == loaded_weight.shape
                param_data.copy_(loaded_weight)
                return
            current_shard_offset = 0
461
            shard_offsets: List[Tuple[int, int, int]] = []
462
463
464
465
466
            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:
467
                # Special case for Quantization.
468
469
470
471
472
                # 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
473
                    # Special case for Marlin.
474
475
476
                    shard_size, shard_offset = adjust_marlin_shard(
                        param, shard_size, shard_offset)

477
478
479
480
481
482
483
484
485
486
487
                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
488
            # Special case for quantization.
489
490
491
492
493
494
            # 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
495
                # Special case for Marlin.
496
497
498
                shard_size, shard_offset = adjust_marlin_shard(
                    param, shard_size, shard_offset)

499
500
501
502
503
504
            use_bitsandbytes = getattr(param, "use_bitsandbytes", False)
            if use_bitsandbytes:
                shard_size = loaded_weight.shape[output_dim]
                shard_offset = loaded_weight.shape[output_dim] * \
                    loaded_shard_id

505
506
507
508
509
510
511
            if is_gguf_weight:
                shard_size = loaded_weight.shape[output_dim]
                shard_offset = loaded_weight.shape[output_dim] * \
                    loaded_shard_id
                param.shard_id.append(loaded_shard_id)
                param.shard_size[loaded_shard_id] = loaded_weight.shape

512
513
514
515
516
            param_data = param_data.narrow(output_dim, shard_offset,
                                           shard_size)
            start_idx = tp_rank * shard_size
            loaded_weight = loaded_weight.narrow(output_dim, start_idx,
                                                 shard_size)
517
        # Special case for AQLM codebooks.
James Fleming's avatar
James Fleming committed
518
519
520
521
522
        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)
523

524
525
526
        # Special case for per-tensor scales in fused case.
        elif needs_scalar_to_array:
            param_data, loaded_weight = adjust_scalar_to_fused_array(
527
528
                param_data, loaded_weight, loaded_shard_id)

529
        else:
CHU Tianxiang's avatar
CHU Tianxiang committed
530
531
532
533
534
535
            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.")
536

537
538
539
        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)

540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
    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.
            if isinstance(param, PackedvLLMParameter
                          ) and param.packed_dim == param.output_dim:
                param.adjust_shard_indexes_for_packing(
                    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):
        param_data = param.data
        if loaded_shard_id is None:
            if param.output_dim is None:
                assert param_data.shape == loaded_weight.shape
                param_data.copy_(loaded_weight)
                return
            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()
        shard_offset = sum(self.output_sizes[:loaded_shard_id]) // tp_size
        shard_size = self.output_sizes[loaded_shard_id] // tp_size

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

596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617

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.
618
        quant_config: Quantization configure.
619
620
        prefix: The name of the layer in the state dict, including all parents
                        (e.g. model.layers.0.qkv_proj)
621
622
    """

623
624
625
626
627
628
629
630
    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,
631
                 quant_config: Optional[QuantizationConfig] = None,
632
                 prefix: str = ""):
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
        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
652
653
654
655
        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
656
657
        ]

658
659
660
661
662
663
        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,
664
665
                         quant_config=quant_config,
                         prefix=prefix)
666

667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
    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.
            if isinstance(param, PackedvLLMParameter
                          ) and param.packed_dim == param.output_dim:
                param.adjust_shard_indexes_for_packing(
                    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):
        param_data = param.data
        if loaded_shard_id is None:  # special case for certain models
            if param.output_dim is None:
                assert param_data.shape == loaded_weight.shape
                param_data.copy_(loaded_weight)
                return
            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)

743
744
745
746
    def weight_loader(self,
                      param: Parameter,
                      loaded_weight: torch.Tensor,
                      loaded_shard_id: Optional[str] = None):
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769

        # 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

        if is_gguf_weight and isinstance(param, UninitializedParameter):
            from gguf.constants import GGML_QUANT_SIZES

            ori_shape = param.tensor_shape
            weight_types = self.qweight_type.shard_weight_type.values()
            row_size = []
            for weight_type in weight_types:
                block_size, type_size = GGML_QUANT_SIZES[weight_type]
                row_size.append(ori_shape[1] // block_size * type_size)
            q_shape = (ori_shape[0], max(row_size))
            param.materialize(q_shape, dtype=loaded_weight.dtype)

770
771
        param_data = param.data
        output_dim = getattr(param, "output_dim", None)
772
        # Special case for AQLM codebooks.
James Fleming's avatar
James Fleming committed
773
        is_metadata = getattr(param, "is_metadata", False)
774

775
776
        # Special case for per-tensor scales in fused case.
        needs_scalar_to_array = getattr(param, "needs_scalar_to_array", False)
777

778
        if loaded_shard_id is None:
779
            # Loaded weight is already fused on disk (qkv/mlp).
780
            if output_dim is None:
781
                if needs_scalar_to_array:
782
783
                    param_data, loaded_weight = adjust_scalar_to_fused_array(
                        param_data, loaded_weight, 0)
784

785
786
787
788
789
790
791
792
793
794
795
796
797
                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),
            ]
            packed_dim = getattr(param, "packed_dim", None)
            for shard_id, shard_offset, shard_size in shard_offsets:
798
                # Special case for Quantized Weights.
799
800
801
802
803
                # 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
804

805
                    # Special case for Marlin.
806
807
808
                    shard_size, shard_offset = adjust_marlin_shard(
                        param, shard_size, shard_offset)

809
810
811
812
813
814
815
                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"]
816
817

        # If output dim is defined, use the default loading process.
818
819
820
821
822
823
824
825
826
827
828
        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
829
            # Special case for Quantized Weights.
830
831
832
833
834
835
            # 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
836

837
                # Special case for Marlin.
838
839
840
                shard_size, shard_offset = adjust_marlin_shard(
                    param, shard_size, shard_offset)

841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
            use_bitsandbytes = getattr(param, "use_bitsandbytes", False)
            if use_bitsandbytes:
                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)
                }
                shard_size, shard_offset = adjust_bitsandbytes_shard(
                    param, orig_qkv_offsets, loaded_shard_id)

857
858
859
860
861
862
863
            if is_gguf_weight:
                param.shard_id.append(loaded_shard_id)
                param.shard_size[loaded_shard_id] = loaded_weight.shape
                input_dim = getattr(param, "input_dim", None)
                input_size = loaded_weight.shape[input_dim]
                param_data = param_data.narrow(input_dim, 0, input_size)

864
865
            param_data = param_data.narrow(output_dim, shard_offset,
                                           shard_size)
866
867
868
869
            if loaded_shard_id == "q":
                shard_id = tp_rank
            else:
                shard_id = tp_rank // self.num_kv_head_replicas
870
871
872
            start_idx = shard_id * shard_size
            loaded_weight = loaded_weight.narrow(output_dim, start_idx,
                                                 shard_size)
873
        # Special case for for AQLM codebooks.
James Fleming's avatar
James Fleming committed
874
875
876
877
878
879
        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)
880
881
882
        # Special case for per-tensor scales in fused case.
        elif needs_scalar_to_array:
            param_data, loaded_weight = adjust_scalar_to_fused_array(
883
                param_data, loaded_weight, loaded_shard_id)
884
        else:
CHU Tianxiang's avatar
CHU Tianxiang committed
885
886
887
888
889
890
            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.")
891

892
893
894
895
        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)


896
class RowParallelLinear(LinearBase):
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
    """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.
919
        quant_config: Quantization configure.
920
921
    """

922
923
924
925
926
927
928
929
    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,
930
                 quant_config: Optional[QuantizationConfig] = None,
931
                 prefix: str = ""):
932
        super().__init__(input_size, output_size, skip_bias_add, params_dtype,
933
                         quant_config, prefix)
934

935
936
937
938
        self.input_is_parallel = input_is_parallel
        self.reduce_results = reduce_results

        # Divide the weight matrix along the last dimension.
939
        self.tp_rank = get_tensor_model_parallel_rank()
940
941
        self.tp_size = get_tensor_model_parallel_world_size()
        self.input_size_per_partition = divide(input_size, self.tp_size)
942
        assert self.quant_method is not None
943

944
945
946
947
948
949
950
        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,
951
952
953
            weight_loader=(
                self.weight_loader_v2 if self.quant_method.__class__.__name__
                in WEIGHT_LOADER_V2_SUPPORTED else self.weight_loader),
954
            prefix=prefix)
955
956
957
958
959
960
        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(
961
                torch.empty(self.output_size, dtype=params_dtype))
962
963
964
965
966
967
968
969
970
971
            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()
        input_dim = getattr(param, "input_dim", None)
972
973
974
975
976
977
978
979
980
981
982

        # 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)

983
984
985
986
987
988
        param_data = param.data
        if input_dim is not None:
            shard_size = param_data.shape[input_dim]
            start_idx = tp_rank * shard_size
            loaded_weight = loaded_weight.narrow(input_dim, start_idx,
                                                 shard_size)
989

990
991
992
        # 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:
993
994
            loaded_weight = loaded_weight.reshape(1)

995
996
997
        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)

998
999
1000
1001
    def weight_loader_v2(self, param: BasevLLMParameter,
                         loaded_weight: torch.Tensor):
        param.load_row_parallel_weight(loaded_weight=loaded_weight)

1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
    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.
1012
        assert self.quant_method is not None
1013
1014
1015
1016
1017
1018
        # 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_)
1019
        if self.reduce_results and self.tp_size > 1:
1020
            output = tensor_model_parallel_all_reduce(output_parallel)
1021
        else:
1022
1023
1024
            output = output_parallel

        output_bias = self.bias if self.skip_bias_add else None
1025
1026

        return output, output_bias
1027
1028
1029
1030
1031
1032
1033
1034

    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