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
from vllm.model_executor.utils import set_weight_attrs
gaoqiong's avatar
gaoqiong committed
17

zhuwenwen's avatar
zhuwenwen committed
18
import os
19
20
21
22

logger = init_logger(__name__)


23
24
25
26
27
28
29
30
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


31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
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


46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
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


69
class LinearMethodBase(QuantizeMethodBase):
70
71
72
    """Base class for different (maybe quantized) linear methods."""

    @abstractmethod
73
74
    def create_weights(self, layer: torch.nn.Module,
                       input_size_per_partition: int,
James Fleming's avatar
James Fleming committed
75
                       output_partition_sizes: List[int], input_size: int,
76
77
                       output_size: int, params_dtype: torch.dtype,
                       **extra_weight_attrs):
78
79
        """Create weights for a linear layer. 
           The weights will be set as attributes of the layer.
80

81
82
83
84
85
86
87
88
89
90
        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.
        """
91
92
93
        raise NotImplementedError

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


class UnquantizedLinearMethod(LinearMethodBase):
104
    """Linear method without quantization."""
105
106
    
    def __init__(self):
zhuwenwen's avatar
zhuwenwen committed
107
        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
108

109
110
    def create_weights(self, layer: torch.nn.Module,
                       input_size_per_partition: int,
James Fleming's avatar
James Fleming committed
111
                       output_partition_sizes: List[int], input_size: int,
112
113
                       output_size: int, params_dtype: torch.dtype,
                       **extra_weight_attrs):
114
        weight = Parameter(torch.empty(sum(output_partition_sizes),
CHU Tianxiang's avatar
CHU Tianxiang committed
115
                                       input_size_per_partition,
116
117
118
                                       dtype=params_dtype),
                           requires_grad=False)
        set_weight_attrs(weight, {"input_dim": 1, "output_dim": 0})
119
120
        layer.register_parameter("weight", weight)
        set_weight_attrs(weight, extra_weight_attrs)
121

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

zhuwenwen's avatar
zhuwenwen committed
127
        if self.use_llama_nn:
zhuwenwen's avatar
zhuwenwen committed
128
            if bias is not None:
zhuwenwen's avatar
zhuwenwen committed
129
                if len(x.shape) == 2: 
130
                    return torch.addmm(bias, x, layer.weight)
zhuwenwen's avatar
zhuwenwen committed
131
                else:
132
                    return torch.matmul(x, layer.weight) + bias
zhuwenwen's avatar
zhuwenwen committed
133
            else:
134
                return torch.matmul(x, layer.weight)
zhuwenwen's avatar
zhuwenwen committed
135
        else:
136
            return F.linear(x, layer.weight, bias)
137

138

139
140
class LinearBase(torch.nn.Module):
    """Base linear layer.
141
142
143
144
145
146
147

    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.
148
        quant_config: Quantization configure.
149
150
151
152
153
154
155
156
    """

    def __init__(
        self,
        input_size: int,
        output_size: int,
        skip_bias_add: bool = False,
        params_dtype: Optional[torch.dtype] = None,
157
        quant_config: Optional[QuantizationConfig] = None,
158
159
160
161
162
163
164
165
166
167
    ):
        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
168
        if quant_config is None:
169
170
            self.quant_method: Optional[
                QuantizeMethodBase] = UnquantizedLinearMethod()
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
        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.
    """

190
191
192
193
194
195
196
    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):
197
198
199
        super().__init__(input_size, output_size, skip_bias_add, params_dtype,
                         quant_config)

200
201
        # All the linear layer supports quant method.
        assert self.quant_method is not None
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
207
        if bias:
            self.bias = Parameter(
208
                torch.empty(self.output_size, dtype=self.params_dtype))
209
210
211
212
213
214
            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
215
        assert self.quant_method is not None
216
        output = self.quant_method.apply(self, x, bias)
217
218
219
        output_bias = self.bias if self.skip_bias_add else None
        return output, output_bias

220
221
222
223
224
225
    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

226

227
class ColumnParallelLinear(LinearBase):
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
    """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.
244
        quant_config: Quantization configure.
James Fleming's avatar
James Fleming committed
245
246
        output_sizes: list of output sizes packed into one output, like for QKV
                       the list would be size 3.
247
248
    """

249
250
251
252
253
254
255
256
257
    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):
258
259
        super().__init__(input_size, output_size, skip_bias_add, params_dtype,
                         quant_config)
260
261

        self.gather_output = gather_output
262

263
264
        # Divide the weight matrix along the last dimension.
        tp_size = get_tensor_model_parallel_world_size()
265
266
267
268
269
270
271
272
273
274
        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
275
276
        if output_sizes is None:
            output_sizes = [output_size]
277
278
279
280
281
282
283
284
        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)
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
        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)
305
306
307
308
309

        # 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)
310

311
312
313
314
315
316
317
        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.
318
        assert self.quant_method is not None
319
        output_parallel = self.quant_method.apply(self, input_, bias)
320
321
322
323
324
325
326
327
        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

328
329
330
331
332
333
334
335
    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

336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354

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.
355
        quant_config: Quantization configure.
356
357
    """

358
359
360
361
362
363
364
365
    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):
366
367
368
        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)
369
370
371
372
373
374
375
        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)
376
377
378
379
380

    def weight_loader(self,
                      param: Parameter,
                      loaded_weight: torch.Tensor,
                      loaded_shard_id: Optional[int] = None):
James Fleming's avatar
James Fleming committed
381

382
383
        param_data = param.data
        output_dim = getattr(param, "output_dim", None)
384
        # Special case for AQLM codebooks.
James Fleming's avatar
James Fleming committed
385
        is_metadata = getattr(param, "is_metadata", False)
386
387
        # Special case for per-tensor scale to load scalar into fused array.
        needs_scalar_to_array = getattr(param, "needs_scalar_to_array", False)
388

389
        if loaded_shard_id is None:
390
            # Loaded weight is already fused on disk (qkv/mlp).
391
            if output_dim is None:
392
                if needs_scalar_to_array:
393
394
                    param_data, loaded_weight = adjust_scalar_to_fused_array(
                        param_data, loaded_weight, 0)
395

396
397
398
399
                assert param_data.shape == loaded_weight.shape
                param_data.copy_(loaded_weight)
                return
            current_shard_offset = 0
400
            shard_offsets: List[Tuple[int, int, int]] = []
401
402
403
404
405
            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:
406
                # Special case for Quantization.
407
408
409
410
411
                # 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
412
                    # Special case for Marlin.
413
414
415
                    shard_size, shard_offset = adjust_marlin_shard(
                        param, shard_size, shard_offset)

416
417
418
419
420
421
422
423
424
425
426
                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
427
            # Special case for quantization.
428
429
430
431
432
433
            # 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
434
                # Special case for Marlin.
435
436
                shard_size, shard_offset = adjust_marlin_shard(
                    param, shard_size, shard_offset)
gaoqiong's avatar
gaoqiong committed
437

438
439
440
441
442
            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
443

gaoqiong's avatar
gaoqiong committed
444
445
            param_data = param_data.narrow(output_dim, shard_offset,
                                           shard_size)
446
447
448
            start_idx = tp_rank * shard_size
            loaded_weight = loaded_weight.narrow(output_dim, start_idx,
                                                 shard_size)
449
        # Special case for AQLM codebooks.
James Fleming's avatar
James Fleming committed
450
451
452
453
454
        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)
455

456
457
458
        # Special case for per-tensor scales in fused case.
        elif needs_scalar_to_array:
            param_data, loaded_weight = adjust_scalar_to_fused_array(
459
460
                param_data, loaded_weight, loaded_shard_id)

461
        else:
CHU Tianxiang's avatar
CHU Tianxiang committed
462
463
464
465
466
467
            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.")
468

gaoqiong's avatar
gaoqiong committed
469
470
        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493


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.
494
        quant_config: Quantization configure.
495
496
    """

497
498
499
500
501
502
503
504
505
    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):
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
        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
525
526
527
528
        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
529
        ]
gaoqiong's avatar
gaoqiong committed
530

531
532
533
534
535
536
537
        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)
538
539
540
541
542
543
544

    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)
545
        # Special case for AQLM codebooks.
James Fleming's avatar
James Fleming committed
546
        is_metadata = getattr(param, "is_metadata", False)
547

548
549
        # Special case for per-tensor scales in fused case.
        needs_scalar_to_array = getattr(param, "needs_scalar_to_array", False)
550

551
        if loaded_shard_id is None:
552
            # Loaded weight is already fused on disk (qkv/mlp).
553
            if output_dim is None:
554
                if needs_scalar_to_array:
555
556
                    param_data, loaded_weight = adjust_scalar_to_fused_array(
                        param_data, loaded_weight, 0)
557

558
559
560
561
562
563
564
565
566
567
568
569
570
                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:
571
                # Special case for Quantized Weights.
572
573
574
575
576
                # 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
577

578
                    # Special case for Marlin.
579
580
581
                    shard_size, shard_offset = adjust_marlin_shard(
                        param, shard_size, shard_offset)

582
583
584
585
586
587
588
                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"]
589
590

        # If output dim is defined, use the default loading process.
591
592
593
594
595
596
597
598
599
600
601
        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
602
            # Special case for Quantized Weights.
603
604
605
606
607
608
            # 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
609

610
                # Special case for Marlin.
611
612
613
                shard_size, shard_offset = adjust_marlin_shard(
                    param, shard_size, shard_offset)

614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
            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)
gaoqiong's avatar
gaoqiong committed
629
630

            param_data = param_data.narrow(output_dim, shard_offset,
zhuwenwen's avatar
zhuwenwen committed
631
                                           shard_size)
zhuwenwen's avatar
zhuwenwen committed
632
            if loaded_shard_id == "q":
633
634
635
                shard_id = tp_rank
            else:
                shard_id = tp_rank // self.num_kv_head_replicas
636
637
638
            start_idx = shard_id * shard_size
            loaded_weight = loaded_weight.narrow(output_dim, start_idx,
                                                 shard_size)
639
        # Special case for for AQLM codebooks.
James Fleming's avatar
James Fleming committed
640
641
642
643
644
645
        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)
646
647
648
        # Special case for per-tensor scales in fused case.
        elif needs_scalar_to_array:
            param_data, loaded_weight = adjust_scalar_to_fused_array(
649
                param_data, loaded_weight, loaded_shard_id)
650
        else:
CHU Tianxiang's avatar
CHU Tianxiang committed
651
652
653
654
655
656
            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.")
gaoqiong's avatar
gaoqiong committed
657
658
659

        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)
660
661


662
class RowParallelLinear(LinearBase):
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
    """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.
685
        quant_config: Quantization configure.
686
687
    """

688
689
690
691
692
693
694
695
696
    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):
697
698
699
        super().__init__(input_size, output_size, skip_bias_add, params_dtype,
                         quant_config)

700
701
702
703
704
705
        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)
706
        assert self.quant_method is not None
707
708
709
710
711
712
713
714
        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)
715
716
717
718
719
720
        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(
721
                torch.empty(self.output_size, dtype=params_dtype))
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
            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)
738

739
740
741
        # 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:
742
743
            loaded_weight = loaded_weight.reshape(1)

744
745
746
747
748
749
750
751
752
753
754
755
756
        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)

    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.
757
        assert self.quant_method is not None
758
        output_parallel = self.quant_method.apply(self, input_parallel)
759
760
761
762
763
764
765
766
767
768
769
770
        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
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}"
778
        return s