linear.py 35.6 KB
Newer Older
1
from abc import abstractmethod
2
from typing import List, Optional
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
from vllm.model_executor.utils import set_weight_attrs

from vllm.logger import init_logger
zhuwenwen's avatar
zhuwenwen committed
19
import os
20
21
22
23

logger = init_logger(__name__)


24
25
26
27
28
29
30
31
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


32
class LinearMethodBase(QuantizeMethodBase):
33
34
35
    """Base class for different (maybe quantized) linear methods."""

    @abstractmethod
36
37
    def create_weights(self, layer: torch.nn.Module,
                       input_size_per_partition: int,
James Fleming's avatar
James Fleming committed
38
                       output_partition_sizes: List[int], input_size: int,
39
40
                       output_size: int, params_dtype: torch.dtype,
                       **extra_weight_attrs):
41
42
43
44
45
46
47
48
49
50
51
52
53
        """Create weights for a linear layer. 
           The weights will be set as attributes of the layer.
        
        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.
        """
54
55
56
        raise NotImplementedError

    @abstractmethod
57
58
59
60
    def apply(self,
              layer: torch.nn.Module,
              x: torch.Tensor,
              bias: Optional[torch.Tensor] = None) -> torch.Tensor:
61
62
        """Apply the weights in layer to the input tensor.
        Expects create_weights to have been called before on the layer."""
63
64
65
66
67
68
69
70
71
72
73
74
75
        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
zhuwenwen's avatar
zhuwenwen committed
76
        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
77

78
79
    def create_weights(self, layer: torch.nn.Module,
                       input_size_per_partition: int,
James Fleming's avatar
James Fleming committed
80
                       output_partition_sizes: List[int], input_size: int,
81
82
                       output_size: int, params_dtype: torch.dtype,
                       **extra_weight_attrs):
83
        weight = Parameter(torch.empty(sum(output_partition_sizes),
CHU Tianxiang's avatar
CHU Tianxiang committed
84
                                       input_size_per_partition,
85
86
87
                                       dtype=params_dtype),
                           requires_grad=False)
        set_weight_attrs(weight, {"input_dim": 1, "output_dim": 0})
88
89
        layer.register_parameter("weight", weight)
        set_weight_attrs(weight, extra_weight_attrs)
90

91
92
93
94
    def apply(self,
              layer: torch.nn.Module,
              x: torch.Tensor,
              bias: Optional[torch.Tensor] = None) -> torch.Tensor:
95
        weight = layer.weight
96
        if self.separate_bias_add:
97
            if bias is not None:
98
99
                return F.linear(x, weight) + bias
            return F.linear(x, weight)
zhuwenwen's avatar
zhuwenwen committed
100
        
zhuwenwen's avatar
zhuwenwen committed
101
        if self.use_llama_nn:
zhuwenwen's avatar
zhuwenwen committed
102
            weight = weight.reshape(weight.shape[1], -1) 
zhuwenwen's avatar
zhuwenwen committed
103
104
105
106
            if bias is not None:
                return torch.matmul(x, weight) + bias
            else:
                return torch.matmul(x, weight) 
zhuwenwen's avatar
zhuwenwen committed
107
108
        else:
            return F.linear(x, weight, bias)
109
110


111
112
class LinearBase(torch.nn.Module):
    """Base linear layer.
113
114
115
116
117
118
119

    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.
120
        quant_config: Quantization configure.
121
122
123
124
125
126
127
128
    """

    def __init__(
        self,
        input_size: int,
        output_size: int,
        skip_bias_add: bool = False,
        params_dtype: Optional[torch.dtype] = None,
129
        quant_config: Optional[QuantizationConfig] = None,
130
131
132
133
134
135
136
137
138
139
    ):
        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
140
        if quant_config is None:
141
142
            self.quant_method: Optional[
                QuantizeMethodBase] = UnquantizedLinearMethod()
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
        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.
    """

162
163
164
165
166
167
168
    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):
169
170
171
        super().__init__(input_size, output_size, skip_bias_add, params_dtype,
                         quant_config)

172
173
        # All the linear layer supports quant method.
        assert self.quant_method is not None
174
175
176
177
        self.quant_method.create_weights(self, self.input_size,
                                         [self.output_size], self.input_size,
                                         self.output_size, self.params_dtype)

178
179
        if bias:
            self.bias = Parameter(
180
                torch.empty(self.output_size, dtype=self.params_dtype))
181
182
183
184
185
186
            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
187
        assert self.quant_method is not None
188
        output = self.quant_method.apply(self, x, bias)
189
190
191
        output_bias = self.bias if self.skip_bias_add else None
        return output, output_bias

192
193
194
195
196
197
    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

198

199
class ColumnParallelLinear(LinearBase):
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
    """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.
216
        quant_config: Quantization configure.
James Fleming's avatar
James Fleming committed
217
218
        output_sizes: list of output sizes packed into one output, like for QKV
                       the list would be size 3.
219
220
    """

221
222
223
224
225
226
227
228
229
    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):
230
231
        super().__init__(input_size, output_size, skip_bias_add, params_dtype,
                         quant_config)
232
233

        self.gather_output = gather_output
234

235
236
        # Divide the weight matrix along the last dimension.
        tp_size = get_tensor_model_parallel_world_size()
237
238
239
240
241
242
243
244
245
246
        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
247
248
        if output_sizes is None:
            output_sizes = [output_size]
249
250
251
252
253
254
255
256
        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)
257
258
259
260
261
262
263
264
265
266
        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)
zhuwenwen's avatar
zhuwenwen committed
267
        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
268
269

    def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor):
270
271
272
273
        # Special case for Fp8 scales.
        fp8_scales_shard_indexer = getattr(param, "fp8_scales_shard_indexer",
                                           None)

274
275
276
277
278
279
280
281
        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)
282
283
284
285
286
287
        # Special case for Fp8 scales.
        elif fp8_scales_shard_indexer is not None:
            param_data, loaded_weight = fp8_scales_shard_indexer(param_data,
                                                                 loaded_weight,
                                                                 shard_id=0)

288
        assert param_data.shape == loaded_weight.shape
zhuwenwen's avatar
zhuwenwen committed
289
        if self.use_llama_nn:
zhuwenwen's avatar
zhuwenwen committed
290
291
            loaded_weight = loaded_weight.transpose(0, 1)
            loaded_weight = loaded_weight.reshape(param_data.shape[0],-1)
292
293
294
295
296
297
        param_data.copy_(loaded_weight)

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

        # Matrix multiply.
298
        assert self.quant_method is not None
299
        output_parallel = self.quant_method.apply(self, input_, bias)
300
301
302
303
304
305
306
307
        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

308
309
310
311
312
313
314
315
    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

316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334

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.
335
        quant_config: Quantization configure.
336
337
    """

338
339
340
341
342
343
344
345
    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):
346
347
348
        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)
349
350
351
352
353
354
355
        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)
zhuwenwen's avatar
zhuwenwen committed
356
        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
357
358
359
360
361
362


    def weight_loader(self,
                      param: Parameter,
                      loaded_weight: torch.Tensor,
                      loaded_shard_id: Optional[int] = None):
James Fleming's avatar
James Fleming committed
363

364
365
        param_data = param.data
        output_dim = getattr(param, "output_dim", None)
366
        # Special case for AQLM codebooks.
James Fleming's avatar
James Fleming committed
367
        is_metadata = getattr(param, "is_metadata", False)
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387

        param_shard_splitter = getattr(param, "shard_splitter", None)

        if output_dim is not None and param_shard_splitter is not None:
            raise NotImplementedError(
                "We do not currently support output_dim != None and "
                "shard_splitter != None for a parameter. Please open an issue."
            )
        # If a parameter has defined a shard_splitter to be used for
        # the weight, it should be applied before the weight is
        # loaded/copied to the parameter. The shard_splitter applies
        # logic by using the loaded_shard_id to ensure that the loaded
        # param is loaded to the correct location
        # within the parameter defined by the linear method.
        if loaded_shard_id is None and param_shard_splitter is not None:
            raise NotImplementedError(
                "We do not currently support loaded_shard_id == None and "
                "shard_splitter != None for a parameter. Please open an issue."
            )

388
389
390
391
        # Special case for Fp8 scales.
        fp8_scales_shard_indexer = getattr(param, "fp8_scales_shard_indexer",
                                           None)

392
393
394
395
396
397
398
399
400
401
402
403
404
        if loaded_shard_id is None:
            # Loaded weight is already packed.
            if output_dim is None:
                assert param_data.shape == loaded_weight.shape
                param_data.copy_(loaded_weight)
                return
            current_shard_offset = 0
            shard_offsets = []
            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:
405
                # Special case for Quantization.
406
407
408
409
410
                # 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
411
                    # Special case for Marlin.
412
413
414
                    shard_size, shard_offset = adjust_marlin_shard(
                        param, shard_size, shard_offset)

415
416
417
418
419
420
421
422
423
424
425
                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
426
            # Special case for quantization.
427
428
429
430
431
432
            # 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
433
                # Special case for Marlin.
434
435
436
                shard_size, shard_offset = adjust_marlin_shard(
                    param, shard_size, shard_offset)

zhuwenwen's avatar
zhuwenwen committed
437
            if self.use_llama_nn:
zhuwenwen's avatar
zhuwenwen committed
438
439
440
441
442
                param_data_ = param_data.narrow(output_dim, shard_offset,
                                            shard_size)
            else:
                param_data = param_data.narrow(output_dim, shard_offset,
                                            shard_size)
443
444
445
            start_idx = tp_rank * shard_size
            loaded_weight = loaded_weight.narrow(output_dim, start_idx,
                                                 shard_size)
446
        # Special case for AQLM codebooks.
James Fleming's avatar
James Fleming committed
447
448
449
450
451
        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)
452
453
454
455
456
457
458

        # If a param_shard_splitter is defined by the LinearMethod, use it.
        elif param_shard_splitter is not None:
            logical_widths = getattr(param, "logical_widths", None)
            param_data, loaded_weight = param_shard_splitter(
                param_data, loaded_weight, loaded_shard_id, logical_widths)

459
460
461
462
463
        # Special case for Fp8 scales.
        elif fp8_scales_shard_indexer is not None:
            param_data, loaded_weight = fp8_scales_shard_indexer(
                param_data, loaded_weight, loaded_shard_id)

464
        else:
CHU Tianxiang's avatar
CHU Tianxiang committed
465
466
467
468
469
470
            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.")
471
472
473
474
475
476
477

        if fp8_scales_shard_indexer is None:
            if len(param_data.shape) == 0:
                param_data = param_data.reshape(1)

            if len(loaded_weight.shape) == 0:
                loaded_weight = loaded_weight.reshape(1)
zhuwenwen's avatar
zhuwenwen committed
478
                
zhuwenwen's avatar
zhuwenwen committed
479
        if self.use_llama_nn:
zhuwenwen's avatar
zhuwenwen committed
480
481
            assert param_data_.shape == loaded_weight.shape
            param_data_.copy_(loaded_weight)
zhuwenwen's avatar
zhuwenwen committed
482
            if loaded_shard_id == 1 and len(param_data.shape) == 2:
zhuwenwen's avatar
zhuwenwen committed
483
484
485
486
487
                param_data = param_data.transpose(0, 1)
                param.data = param_data.reshape(param_data.shape[1], -1)
        else:
            assert param_data.shape == loaded_weight.shape
            param_data.copy_(loaded_weight)
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511



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.
512
        quant_config: Quantization configure.
513
514
    """

515
516
517
518
519
520
521
522
523
    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):
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
        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
543
544
545
546
        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
547
        ]
548
549
550
551
552
553
554
        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)
zhuwenwen's avatar
zhuwenwen committed
555
        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
556
557
558
559
560
561
562

    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)
563
        # Special case for AQLM codebooks.
James Fleming's avatar
James Fleming committed
564
        is_metadata = getattr(param, "is_metadata", False)
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584

        param_shard_splitter = getattr(param, "shard_splitter", None)

        if output_dim is not None and param_shard_splitter is not None:
            raise NotImplementedError(
                "We do not currently support output_dim != None and "
                "shard_splitter != None for a parameter. Please open an issue."
            )
        # If a parameter has defined a shard_splitter to be used for
        # the weight, it should be applied before the weight is
        # loaded/copied to the parameter. The shard_splitter applies
        # logic by using the loaded_shard_id to ensure that the loaded
        # param is loaded to the correct location
        # within the parameter defined by the linear method.
        if loaded_shard_id is None and param_shard_splitter is not None:
            raise NotImplementedError(
                "We do not currently support loaded_shard_id == None and "
                "shard_splitter != None for a parameter. Please open an issue."
            )

585
586
587
        # Special case for Fp8 scales.
        fp8_scales_shard_indexer = getattr(param, "fp8_scales_shard_indexer",
                                           None)
588

589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
        if loaded_shard_id is None:
            # Loaded weight is already packed.
            if output_dim is None:
                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:
605
                # Special case for Quantized Weights.
606
607
608
609
610
                # 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
611

612
                    # Special case for Marlin.
613
614
615
                    shard_size, shard_offset = adjust_marlin_shard(
                        param, shard_size, shard_offset)

616
617
618
619
620
621
622
                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"]
623
624

        # If output dim is defined, use the default loading process.
625
626
627
628
629
630
631
632
633
634
635
        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
636
            # Special case for Quantized Weights.
637
638
639
640
641
642
            # 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
643

644
                # Special case for Marlin.
645
646
647
                shard_size, shard_offset = adjust_marlin_shard(
                    param, shard_size, shard_offset)

zhuwenwen's avatar
zhuwenwen committed
648
            if self.use_llama_nn:
zhuwenwen's avatar
zhuwenwen committed
649
650
651
652
                param_data_ = param_data.narrow(output_dim, shard_offset,
                                           shard_size)
            else:
                param_data = param_data.narrow(output_dim, shard_offset,
653
                                           shard_size)
zhuwenwen's avatar
zhuwenwen committed
654
            if loaded_shard_id == "q":
655
656
657
                shard_id = tp_rank
            else:
                shard_id = tp_rank // self.num_kv_head_replicas
658
659
660
            start_idx = shard_id * shard_size
            loaded_weight = loaded_weight.narrow(output_dim, start_idx,
                                                 shard_size)
661
        # Special case for for AQLM codebooks.
James Fleming's avatar
James Fleming committed
662
663
664
665
666
667
        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)
668
669
670
671
672
673
        # If a param_shard_splitter is defined by the LinearMethod, use it.
        elif param_shard_splitter is not None:
            logical_widths = getattr(param, "logical_widths", None)
            param_data, loaded_weight = param_shard_splitter(
                param_data, loaded_weight, loaded_shard_id, logical_widths)

674
675
676
677
        # Special case for Fp8 scales.
        elif fp8_scales_shard_indexer is not None:
            param_data, loaded_weight = fp8_scales_shard_indexer(
                param_data, loaded_weight, loaded_shard_id)
678
        else:
CHU Tianxiang's avatar
CHU Tianxiang committed
679
680
681
682
683
684
            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.")
685
686
687
688
689
        if len(param_data.shape) == 0:
            param_data = param_data.reshape(1)

        if len(loaded_weight.shape) == 0:
            loaded_weight = loaded_weight.reshape(1)
zhuwenwen's avatar
zhuwenwen committed
690
        
zhuwenwen's avatar
zhuwenwen committed
691
        if self.use_llama_nn:
zhuwenwen's avatar
zhuwenwen committed
692
693
            assert param_data_.shape == loaded_weight.shape
            param_data_.copy_(loaded_weight)
zhuwenwen's avatar
zhuwenwen committed
694
            if loaded_shard_id == "v" and len(param_data.shape) == 2:
zhuwenwen's avatar
zhuwenwen committed
695
696
697
698
699
                param_data = param_data.transpose(0, 1) 
                param.data = param_data.reshape(param_data.shape[1], -1) 
        else:
            assert param_data.shape == loaded_weight.shape
            param_data.copy_(loaded_weight)
700
701


702
class RowParallelLinear(LinearBase):
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
    """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.
725
        quant_config: Quantization configure.
726
727
    """

728
729
730
731
732
733
734
735
736
    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):
737
738
739
        super().__init__(input_size, output_size, skip_bias_add, params_dtype,
                         quant_config)

740
741
742
743
744
745
        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)
746
        assert self.quant_method is not None
747
748
749
750
751
752
753
754
        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)
755
756
757
758
759
760
        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(
761
                torch.empty(self.output_size, dtype=params_dtype))
762
763
764
765
766
767
            set_weight_attrs(self.bias, {
                "output_dim": 0,
                "weight_loader": self.weight_loader,
            })
        else:
            self.register_parameter("bias", None)
zhuwenwen's avatar
zhuwenwen committed
768
        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
769
770

    def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor):
771
772
773
774
        # Special case for Fp8 scales.
        fp8_scales_shard_indexer = getattr(param, "fp8_scales_shard_indexer",
                                           None)

775
776
777
778
779
780
781
782
        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)
783

784
785
786
787
788
789
        # Special case for Fp8 scales.
        elif fp8_scales_shard_indexer is not None:
            param_data, loaded_weight = fp8_scales_shard_indexer(param_data,
                                                                 loaded_weight,
                                                                 shard_id=0)

790
791
792
        if fp8_scales_shard_indexer is None and len(loaded_weight.shape) == 0:
            loaded_weight = loaded_weight.reshape(1)

793
        assert param_data.shape == loaded_weight.shape
zhuwenwen's avatar
zhuwenwen committed
794
        if self.use_llama_nn:
zhuwenwen's avatar
zhuwenwen committed
795
796
            loaded_weight = loaded_weight.transpose(0, 1)
            loaded_weight=loaded_weight.reshape(param_data.shape[0],-1)
797
798
799
800
801
802
803
804
805
806
807
808
809
        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.
810
        assert self.quant_method is not None
811
        output_parallel = self.quant_method.apply(self, input_parallel)
812
813
814
815
816
817
818
819
820
821
822
823
        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
824
825
826
827
828
829
830
831

    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