linear.py 31.7 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
63
        """Apply the weights in layer to the input tensor.

        Expects create_weights to have been called before on the layer."""
64
65
66
67
68
69
70
71
72
73
74
75
76
        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
77
        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
78

79
80
    def create_weights(self, layer: torch.nn.Module,
                       input_size_per_partition: int,
James Fleming's avatar
James Fleming committed
81
                       output_partition_sizes: List[int], input_size: int,
82
83
                       output_size: int, params_dtype: torch.dtype,
                       **extra_weight_attrs):
James Fleming's avatar
James Fleming committed
84
        output_size_per_partition = sum(output_partition_sizes)
CHU Tianxiang's avatar
CHU Tianxiang committed
85
86
        weight = Parameter(torch.empty(output_size_per_partition,
                                       input_size_per_partition,
87
88
89
                                       dtype=params_dtype),
                           requires_grad=False)
        set_weight_attrs(weight, {"input_dim": 1, "output_dim": 0})
90
91
        layer.register_parameter("weight", weight)
        set_weight_attrs(weight, extra_weight_attrs)
92

93
94
95
96
    def apply(self,
              layer: torch.nn.Module,
              x: torch.Tensor,
              bias: Optional[torch.Tensor] = None) -> torch.Tensor:
97
        weight = layer.weight
98
        if self.separate_bias_add:
99
            if bias is not None:
100
101
                return F.linear(x, weight) + bias
            return F.linear(x, weight)
zhuwenwen's avatar
zhuwenwen committed
102
        
zhuwenwen's avatar
zhuwenwen committed
103
        if self.use_llama_nn:
zhuwenwen's avatar
zhuwenwen committed
104
            weight = weight.reshape(weight.shape[1], -1) 
zhuwenwen's avatar
zhuwenwen committed
105
106
107
108
            if bias is not None:
                return torch.matmul(x, weight) + bias
            else:
                return torch.matmul(x, weight) 
zhuwenwen's avatar
zhuwenwen committed
109
110
        else:
            return F.linear(x, weight, bias)
111
112


113
114
class LinearBase(torch.nn.Module):
    """Base linear layer.
115
116
117
118
119
120
121

    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.
122
        quant_config: Quantization configure.
123
124
125
126
127
128
129
130
    """

    def __init__(
        self,
        input_size: int,
        output_size: int,
        skip_bias_add: bool = False,
        params_dtype: Optional[torch.dtype] = None,
131
        quant_config: Optional[QuantizationConfig] = None,
132
133
134
135
136
137
138
139
140
141
    ):
        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
142
        if quant_config is None:
143
144
            self.quant_method: Optional[
                QuantizeMethodBase] = UnquantizedLinearMethod()
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
        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.
    """

    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,
    ):
        super().__init__(input_size, output_size, skip_bias_add, params_dtype,
                         quant_config)

176
177
        # All the linear layer supports quant method.
        assert self.quant_method is not None
178
179
180
181
        self.quant_method.create_weights(self, self.input_size,
                                         [self.output_size], self.input_size,
                                         self.output_size, self.params_dtype)

182
183
        if bias:
            self.bias = Parameter(
184
                torch.empty(self.output_size, dtype=self.params_dtype))
185
186
187
188
189
190
            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
191
        assert self.quant_method is not None
192
        output = self.quant_method.apply(self, x, bias)
193
194
195
        output_bias = self.bias if self.skip_bias_add else None
        return output, output_bias

196
197
198
199
200
201
    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

202

203
class ColumnParallelLinear(LinearBase):
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
    """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.
220
        quant_config: Quantization configure.
James Fleming's avatar
James Fleming committed
221
222
        output_sizes: list of output sizes packed into one output, like for QKV
                       the list would be size 3.
223
224
225
226
227
228
229
230
231
232
    """

    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,
233
        quant_config: Optional[QuantizationConfig] = None,
James Fleming's avatar
James Fleming committed
234
        output_sizes: Optional[List[int]] = None,
235
    ):
236
237
        super().__init__(input_size, output_size, skip_bias_add, params_dtype,
                         quant_config)
238
239

        self.gather_output = gather_output
240

241
242
243
        # Divide the weight matrix along the last dimension.
        tp_size = get_tensor_model_parallel_world_size()
        self.output_size_per_partition = divide(output_size, tp_size)
James Fleming's avatar
James Fleming committed
244
245
        if output_sizes is None:
            output_sizes = [output_size]
246
247
        # All the linear layer supports quant method.
        assert self.quant_method is not None
248
249
250
251
252
253
254
        self.quant_method.create_weights(self,
                                         self.input_size,
                                         [x // tp_size for x in output_sizes],
                                         self.input_size,
                                         self.output_size,
                                         self.params_dtype,
                                         weight_loader=self.weight_loader)
255
256
257
258
259
260
261
262
263
264
        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
265
        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
266
267

    def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor):
268
269
270
271
        # Special case for Fp8 scales.
        fp8_scales_shard_indexer = getattr(param, "fp8_scales_shard_indexer",
                                           None)

272
273
274
275
276
277
278
279
        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)
280
281
282
283
284
285
        # 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)

286
        assert param_data.shape == loaded_weight.shape
zhuwenwen's avatar
zhuwenwen committed
287
        if self.use_llama_nn:
zhuwenwen's avatar
zhuwenwen committed
288
289
            loaded_weight = loaded_weight.transpose(0, 1)
            loaded_weight = loaded_weight.reshape(param_data.shape[0],-1)
290
291
292
293
294
295
        param_data.copy_(loaded_weight)

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

        # Matrix multiply.
296
        assert self.quant_method is not None
297
        output_parallel = self.quant_method.apply(self, input_, bias)
298
299
300
301
302
303
304
305
        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

306
307
308
309
310
311
312
313
    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

314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332

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.
333
        quant_config: Quantization configure.
334
335
336
337
338
339
340
341
342
343
    """

    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,
344
        quant_config: Optional[QuantizationConfig] = None,
345
346
347
348
349
    ):
        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)
        super().__init__(input_size, sum(output_sizes), bias, gather_output,
350
                         skip_bias_add, params_dtype, quant_config,
James Fleming's avatar
James Fleming committed
351
                         self.output_sizes)
zhuwenwen's avatar
zhuwenwen committed
352
        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
353
354
355
356
357

    def weight_loader(self,
                      param: Parameter,
                      loaded_weight: torch.Tensor,
                      loaded_shard_id: Optional[int] = None):
James Fleming's avatar
James Fleming committed
358

359
360
        param_data = param.data
        output_dim = getattr(param, "output_dim", None)
361
        # Special case for AQLM codebooks.
James Fleming's avatar
James Fleming committed
362
        is_metadata = getattr(param, "is_metadata", False)
363
364
365
366
        # Special case for Fp8 scales.
        fp8_scales_shard_indexer = getattr(param, "fp8_scales_shard_indexer",
                                           None)

367
368
369
370
371
372
373
374
375
376
377
378
379
        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:
380
                # Special case for Quantization.
381
382
383
384
385
                # 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
386
                    # Special case for Marlin.
387
388
389
                    shard_size, shard_offset = adjust_marlin_shard(
                        param, shard_size, shard_offset)

390
391
392
393
394
395
396
397
398
399
400
                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
401
            # Special case for quantization.
402
403
404
405
406
407
            # 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
408
                # Special case for Marlin.
409
410
411
                shard_size, shard_offset = adjust_marlin_shard(
                    param, shard_size, shard_offset)

zhuwenwen's avatar
zhuwenwen committed
412
            if self.use_llama_nn:
zhuwenwen's avatar
zhuwenwen committed
413
414
415
416
417
                param_data_ = param_data.narrow(output_dim, shard_offset,
                                            shard_size)
            else:
                param_data = param_data.narrow(output_dim, shard_offset,
                                            shard_size)
418
419
420
            start_idx = tp_rank * shard_size
            loaded_weight = loaded_weight.narrow(output_dim, start_idx,
                                                 shard_size)
421
        # Special case for AQLM codebooks.
James Fleming's avatar
James Fleming committed
422
423
424
425
426
        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)
427
428
429
430
431
        # 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)

432
        else:
CHU Tianxiang's avatar
CHU Tianxiang committed
433
434
435
436
437
438
            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.")
zhuwenwen's avatar
zhuwenwen committed
439
        
zhuwenwen's avatar
zhuwenwen committed
440
        if self.use_llama_nn:
zhuwenwen's avatar
zhuwenwen committed
441
442
            assert param_data_.shape == loaded_weight.shape
            param_data_.copy_(loaded_weight)
zhuwenwen's avatar
zhuwenwen committed
443
            if loaded_shard_id == 1 and len(param_data.shape) == 2:
zhuwenwen's avatar
zhuwenwen committed
444
445
446
447
448
                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)
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471


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.
472
        quant_config: Quantization configure.
473
474
475
476
477
478
479
480
481
482
483
    """

    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,
484
        quant_config: Optional[QuantizationConfig] = None,
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
    ):
        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
James Fleming's avatar
James Fleming committed
505
506
507
508
509
510
        output_sizes = [
            self.num_heads * tp_size * self.head_size,
            self.num_kv_heads * tp_size * self.head_size,
            self.num_kv_heads * tp_size * self.head_size
        ]

511
        super().__init__(input_size, output_size, bias, False, skip_bias_add,
512
                         params_dtype, quant_config, output_sizes)
zhuwenwen's avatar
zhuwenwen committed
513
        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
514
515
516
517
518
519
520

    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)
521
        # Special case for AQLM codebooks.
James Fleming's avatar
James Fleming committed
522
        is_metadata = getattr(param, "is_metadata", False)
523
524
525
        # Special case for Fp8 scales.
        fp8_scales_shard_indexer = getattr(param, "fp8_scales_shard_indexer",
                                           None)
526

527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
        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:
543
                # Special case for Quantized Weights.
544
545
546
547
548
                # 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
549

550
                    # Special case for Marlin.
551
552
553
                    shard_size, shard_offset = adjust_marlin_shard(
                        param, shard_size, shard_offset)

554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
                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"]
        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
572
            # Special case for Quantized Weights.
573
574
575
576
577
578
            # 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
579

580
                # Special case for Marlin.
581
582
583
                shard_size, shard_offset = adjust_marlin_shard(
                    param, shard_size, shard_offset)

zhuwenwen's avatar
zhuwenwen committed
584
            if self.use_llama_nn:
zhuwenwen's avatar
zhuwenwen committed
585
586
587
588
                param_data_ = param_data.narrow(output_dim, shard_offset,
                                           shard_size)
            else:
                param_data = param_data.narrow(output_dim, shard_offset,
589
                                           shard_size)
zhuwenwen's avatar
zhuwenwen committed
590
            if loaded_shard_id == "q":
591
592
593
                shard_id = tp_rank
            else:
                shard_id = tp_rank // self.num_kv_head_replicas
594
595
596
            start_idx = shard_id * shard_size
            loaded_weight = loaded_weight.narrow(output_dim, start_idx,
                                                 shard_size)
597
        # Special case for for AQLM codebooks.
James Fleming's avatar
James Fleming committed
598
599
600
601
602
603
        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)
604
605
606
607
        # 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)
608
        else:
CHU Tianxiang's avatar
CHU Tianxiang committed
609
610
611
612
613
614
            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.")
zhuwenwen's avatar
zhuwenwen committed
615
        
zhuwenwen's avatar
zhuwenwen committed
616
        if self.use_llama_nn:
zhuwenwen's avatar
zhuwenwen committed
617
618
            assert param_data_.shape == loaded_weight.shape
            param_data_.copy_(loaded_weight)
zhuwenwen's avatar
zhuwenwen committed
619
            if loaded_shard_id == "v" and len(param_data.shape) == 2:
zhuwenwen's avatar
zhuwenwen committed
620
621
622
623
624
                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)
625
626


627
class RowParallelLinear(LinearBase):
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
    """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.
650
        quant_config: Quantization configure.
651
652
653
654
655
656
657
658
659
660
661
    """

    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,
662
        quant_config: Optional[QuantizationConfig] = None,
663
    ):
664
665
666
        super().__init__(input_size, output_size, skip_bias_add, params_dtype,
                         quant_config)

667
668
669
670
671
672
        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)
673
674
        # All the linear layer supports quant method.
        assert self.quant_method is not None
675
676
677
678
679
680
681
        self.quant_method.create_weights(self,
                                         self.input_size_per_partition,
                                         [self.output_size],
                                         self.input_size,
                                         self.output_size,
                                         self.params_dtype,
                                         weight_loader=self.weight_loader)
682
683
684
685
686
687
688

        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(
689
                torch.empty(self.output_size, dtype=params_dtype))
690
691
692
693
694
695
            set_weight_attrs(self.bias, {
                "output_dim": 0,
                "weight_loader": self.weight_loader,
            })
        else:
            self.register_parameter("bias", None)
zhuwenwen's avatar
zhuwenwen committed
696
        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
697
698

    def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor):
699
700
701
702
        # Special case for Fp8 scales.
        fp8_scales_shard_indexer = getattr(param, "fp8_scales_shard_indexer",
                                           None)

703
704
705
706
707
708
709
710
        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)
711
712
713
714
715
716
        # 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)

717
        assert param_data.shape == loaded_weight.shape
zhuwenwen's avatar
zhuwenwen committed
718
        if self.use_llama_nn:
zhuwenwen's avatar
zhuwenwen committed
719
720
            loaded_weight = loaded_weight.transpose(0, 1)
            loaded_weight=loaded_weight.reshape(param_data.shape[0],-1)
721
722
723
724
725
726
727
728
729
730
731
732
733
        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.
734
        assert self.quant_method is not None
735
        output_parallel = self.quant_method.apply(self, input_parallel)
736
737
738
739
740
741
742
743
744
745
746
747
        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
748
749
750
751
752
753
754
755

    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