linear.py 32.7 KB
Newer Older
1
from abc import abstractmethod
2
from typing import Dict, List, Optional, Tuple
3
4
5
6
7

import torch
import torch.nn.functional as F
from torch.nn.parameter import Parameter

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
18
19
20
from vllm.model_executor.utils import set_weight_attrs

logger = init_logger(__name__)


21
22
23
24
25
26
27
28
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


29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
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


44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
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


67
class LinearMethodBase(QuantizeMethodBase):
68
69
70
    """Base class for different (maybe quantized) linear methods."""

    @abstractmethod
71
72
    def create_weights(self, layer: torch.nn.Module,
                       input_size_per_partition: int,
James Fleming's avatar
James Fleming committed
73
                       output_partition_sizes: List[int], input_size: int,
74
75
                       output_size: int, params_dtype: torch.dtype,
                       **extra_weight_attrs):
76
77
        """Create weights for a linear layer. 
           The weights will be set as attributes of the layer.
78

79
80
81
82
83
84
85
86
87
88
        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.
        """
89
90
91
        raise NotImplementedError

    @abstractmethod
92
93
94
95
    def apply(self,
              layer: torch.nn.Module,
              x: torch.Tensor,
              bias: Optional[torch.Tensor] = None) -> torch.Tensor:
96
97
        """Apply the weights in layer to the input tensor.
        Expects create_weights to have been called before on the layer."""
98
99
100
101
102
103
104
105
106
107
108
109
110
111
        raise NotImplementedError


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

    Args:
        separate_bias_add: If true, add bias separately after matrix
                           multiplication.
    """

    def __init__(self, separate_bias_add: bool = False):
        self.separate_bias_add = separate_bias_add

112
113
    def create_weights(self, layer: torch.nn.Module,
                       input_size_per_partition: int,
James Fleming's avatar
James Fleming committed
114
                       output_partition_sizes: List[int], input_size: int,
115
116
                       output_size: int, params_dtype: torch.dtype,
                       **extra_weight_attrs):
117
        weight = Parameter(torch.empty(sum(output_partition_sizes),
CHU Tianxiang's avatar
CHU Tianxiang committed
118
                                       input_size_per_partition,
119
120
121
                                       dtype=params_dtype),
                           requires_grad=False)
        set_weight_attrs(weight, {"input_dim": 1, "output_dim": 0})
122
123
        layer.register_parameter("weight", weight)
        set_weight_attrs(weight, extra_weight_attrs)
124

125
126
127
128
    def apply(self,
              layer: torch.nn.Module,
              x: torch.Tensor,
              bias: Optional[torch.Tensor] = None) -> torch.Tensor:
129
        weight = layer.weight
130
        if self.separate_bias_add:
131
            if bias is not None:
132
133
134
135
136
                return F.linear(x, weight) + bias
            return F.linear(x, weight)
        return F.linear(x, weight, bias)


137
138
class LinearBase(torch.nn.Module):
    """Base linear layer.
139
140
141
142
143
144
145

    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.
146
        quant_config: Quantization configure.
147
148
149
150
151
152
153
154
    """

    def __init__(
        self,
        input_size: int,
        output_size: int,
        skip_bias_add: bool = False,
        params_dtype: Optional[torch.dtype] = None,
155
        quant_config: Optional[QuantizationConfig] = None,
156
157
158
159
160
161
162
163
164
165
    ):
        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
166
        if quant_config is None:
167
168
            self.quant_method: Optional[
                QuantizeMethodBase] = UnquantizedLinearMethod()
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
        else:
            self.quant_method = quant_config.get_quant_method(self)

    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.
    """

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

198
199
        # All the linear layer supports quant method.
        assert self.quant_method is not None
200
201
202
203
        self.quant_method.create_weights(self, self.input_size,
                                         [self.output_size], self.input_size,
                                         self.output_size, self.params_dtype)

204
205
        if bias:
            self.bias = Parameter(
206
                torch.empty(self.output_size, dtype=self.params_dtype))
207
208
209
210
211
212
            set_weight_attrs(self.bias, {"output_dim": 0})
        else:
            self.register_parameter("bias", None)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        bias = self.bias if not self.skip_bias_add else None
213
        assert self.quant_method is not None
214
        output = self.quant_method.apply(self, x, bias)
215
216
217
        output_bias = self.bias if self.skip_bias_add else None
        return output, output_bias

218
219
220
221
222
223
    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

224

225
class ColumnParallelLinear(LinearBase):
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
    """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.
242
        quant_config: Quantization configure.
James Fleming's avatar
James Fleming committed
243
244
        output_sizes: list of output sizes packed into one output, like for QKV
                       the list would be size 3.
245
246
    """

247
248
249
250
251
252
253
254
255
    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,
                 output_sizes: Optional[List[int]] = None):
256
257
        super().__init__(input_size, output_size, skip_bias_add, params_dtype,
                         quant_config)
258
259

        self.gather_output = gather_output
260

261
262
        # Divide the weight matrix along the last dimension.
        tp_size = get_tensor_model_parallel_world_size()
263
264
265
266
267
268
269
270
271
272
        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
273
274
        if output_sizes is None:
            output_sizes = [output_size]
275
276
277
278
279
280
281
282
        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,
            weight_loader=self.weight_loader)
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
        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)
        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)
303
304
305
306
307

        # 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)
308

309
310
311
312
313
314
315
        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)

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

        # Matrix multiply.
316
        assert self.quant_method is not None
317
        output_parallel = self.quant_method.apply(self, input_, bias)
318
319
320
321
322
323
324
325
        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

326
327
328
329
330
331
332
333
    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

334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352

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.
353
        quant_config: Quantization configure.
354
355
    """

356
357
358
359
360
361
362
363
    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,
                 quant_config: Optional[QuantizationConfig] = None):
364
365
366
        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)
367
368
369
370
371
372
373
        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,
                         quant_config=quant_config)
374
375
376
377
378

    def weight_loader(self,
                      param: Parameter,
                      loaded_weight: torch.Tensor,
                      loaded_shard_id: Optional[int] = None):
James Fleming's avatar
James Fleming committed
379

380
381
        param_data = param.data
        output_dim = getattr(param, "output_dim", None)
382
        # Special case for AQLM codebooks.
James Fleming's avatar
James Fleming committed
383
        is_metadata = getattr(param, "is_metadata", False)
384
385
        # Special case for per-tensor scale to load scalar into fused array.
        needs_scalar_to_array = getattr(param, "needs_scalar_to_array", False)
386

387
        if loaded_shard_id is None:
388
            # Loaded weight is already fused on disk (qkv/mlp).
389
            if output_dim is None:
390
391
392
                if needs_scalar_to_array is not None:
                    param_data, loaded_weight = adjust_scalar_to_fused_array(
                        param_data, loaded_weight, 0)
393

394
395
396
397
                assert param_data.shape == loaded_weight.shape
                param_data.copy_(loaded_weight)
                return
            current_shard_offset = 0
398
            shard_offsets: List[Tuple[int, int, int]] = []
399
400
401
402
403
            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:
404
                # Special case for Quantization.
405
406
407
408
409
                # 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
410
                    # Special case for Marlin.
411
412
413
                    shard_size, shard_offset = adjust_marlin_shard(
                        param, shard_size, shard_offset)

414
415
416
417
418
419
420
421
422
423
424
                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
425
            # Special case for quantization.
426
427
428
429
430
431
            # 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
432
                # Special case for Marlin.
433
434
435
                shard_size, shard_offset = adjust_marlin_shard(
                    param, shard_size, shard_offset)

436
437
438
439
440
441
            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

442
443
444
445
446
            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)
447
        # Special case for AQLM codebooks.
James Fleming's avatar
James Fleming committed
448
449
450
451
452
        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)
453

454
455
456
        # Special case for per-tensor scales in fused case.
        elif needs_scalar_to_array:
            param_data, loaded_weight = adjust_scalar_to_fused_array(
457
458
                param_data, loaded_weight, loaded_shard_id)

459
        else:
CHU Tianxiang's avatar
CHU Tianxiang committed
460
461
462
463
464
465
            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.")
466

467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)


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.
492
        quant_config: Quantization configure.
493
494
    """

495
496
497
498
499
500
501
502
503
    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,
                 quant_config: Optional[QuantizationConfig] = None):
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
        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
523
524
525
526
        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
527
528
        ]

529
530
531
532
533
534
535
        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,
                         quant_config=quant_config)
536
537
538
539
540
541
542

    def weight_loader(self,
                      param: Parameter,
                      loaded_weight: torch.Tensor,
                      loaded_shard_id: Optional[str] = None):
        param_data = param.data
        output_dim = getattr(param, "output_dim", None)
543
        # Special case for AQLM codebooks.
James Fleming's avatar
James Fleming committed
544
        is_metadata = getattr(param, "is_metadata", False)
545

546
547
        # Special case for per-tensor scales in fused case.
        needs_scalar_to_array = getattr(param, "needs_scalar_to_array", False)
548

549
        if loaded_shard_id is None:
550
            # Loaded weight is already fused on disk (qkv/mlp).
551
            if output_dim is None:
552
553
554
                if needs_scalar_to_array is not None:
                    param_data, loaded_weight = adjust_scalar_to_fused_array(
                        param_data, loaded_weight, 0)
555

556
557
558
559
560
561
562
563
564
565
566
567
568
                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:
569
                # Special case for Quantized Weights.
570
571
572
573
574
                # 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
575

576
                    # Special case for Marlin.
577
578
579
                    shard_size, shard_offset = adjust_marlin_shard(
                        param, shard_size, shard_offset)

580
581
582
583
584
585
586
                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"]
587
588

        # If output dim is defined, use the default loading process.
589
590
591
592
593
594
595
596
597
598
599
        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
600
            # Special case for Quantized Weights.
601
602
603
604
605
606
            # 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
607

608
                # Special case for Marlin.
609
610
611
                shard_size, shard_offset = adjust_marlin_shard(
                    param, shard_size, shard_offset)

612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
            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)

628
629
            param_data = param_data.narrow(output_dim, shard_offset,
                                           shard_size)
630
631
632
633
            if loaded_shard_id == "q":
                shard_id = tp_rank
            else:
                shard_id = tp_rank // self.num_kv_head_replicas
634
635
636
            start_idx = shard_id * shard_size
            loaded_weight = loaded_weight.narrow(output_dim, start_idx,
                                                 shard_size)
637
        # Special case for for AQLM codebooks.
James Fleming's avatar
James Fleming committed
638
639
640
641
642
643
        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)
644
645
646
        # Special case for per-tensor scales in fused case.
        elif needs_scalar_to_array:
            param_data, loaded_weight = adjust_scalar_to_fused_array(
647
                param_data, loaded_weight, loaded_shard_id)
648
        else:
CHU Tianxiang's avatar
CHU Tianxiang committed
649
650
651
652
653
654
            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.")
655

656
657
658
659
        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)


660
class RowParallelLinear(LinearBase):
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
    """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.
683
        quant_config: Quantization configure.
684
685
    """

686
687
688
689
690
691
692
693
694
    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,
                 quant_config: Optional[QuantizationConfig] = None):
695
696
697
        super().__init__(input_size, output_size, skip_bias_add, params_dtype,
                         quant_config)

698
699
700
701
702
703
        self.input_is_parallel = input_is_parallel
        self.reduce_results = reduce_results

        # Divide the weight matrix along the last dimension.
        self.tp_size = get_tensor_model_parallel_world_size()
        self.input_size_per_partition = divide(input_size, self.tp_size)
704
        assert self.quant_method is not None
705
706
707
708
709
710
711
712
        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,
            weight_loader=self.weight_loader)
713
714
715
716
717
718
        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(
719
                torch.empty(self.output_size, dtype=params_dtype))
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
            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)
        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)
736

737
738
739
        # 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:
740
741
            loaded_weight = loaded_weight.reshape(1)

742
743
744
745
746
747
748
749
750
751
752
753
754
755
        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)

    def forward(self, input_):
        # Set up backprop all-reduce.
        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.
756
        assert self.quant_method is not None
757
        output_parallel = self.quant_method.apply(self, input_parallel)
758
759
760
761
762
763
764
765
766
767
768
769
        if self.reduce_results and self.tp_size > 1:
            output_ = tensor_model_parallel_all_reduce(output_parallel)
        else:
            output_ = output_parallel

        if not self.skip_bias_add:
            output = output_ + self.bias if self.bias is not None else output_
            output_bias = None
        else:
            output = output_
            output_bias = self.bias
        return output, output_bias
770
771
772
773
774
775
776
777

    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