linear.py 59.4 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import itertools
5
from abc import abstractmethod
6

7
from typing import Any
8
from vllm import envs
9
10

import torch
11
from torch.nn.parameter import Parameter, UninitializedParameter
12

13
14
15
16
17
18
19
20
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,
)
21
from vllm.logger import init_logger
22
from vllm.model_executor.custom_op import CustomOp
23
from vllm.model_executor.layers.quantization.base_config import (
24
25
26
    QuantizationConfig,
    QuantizeMethodBase,
)
27
from vllm.model_executor.layers.utils import dispatch_unquantized_gemm
28
29
30
31
32
33
34
35
36
from vllm.model_executor.parameter import (
    BasevLLMParameter,
    BlockQuantScaleParameter,
    ModelWeightParameter,
    PackedColumnParameter,
    PackedvLLMParameter,
    PerTensorScaleParameter,
    RowvLLMParameter,
)
37
from vllm.model_executor.utils import set_weight_attrs
38
from vllm.platforms import current_platform
gaoqiong's avatar
gaoqiong committed
39

zhuwenwen's avatar
zhuwenwen committed
40
import os
41
from vllm.model_executor.utils import gemm_bank_conf
42
43
44

logger = init_logger(__name__)

45
WEIGHT_LOADER_V2_SUPPORTED = [
46
    "UnquantizedLinearMethod",
47
    "CompressedTensorsLinearMethod",
48
    "CompressedTensorsLinearTransformMethod",
49
50
    "BitBLASLinearMethod",
    "GPTQBitBLASLinearMethod",
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
    "AWQMarlinLinearMethod",
    "AWQLinearMethod",
    "GPTQMarlinLinearMethod",
    "Fp8LinearMethod",
    "MarlinLinearMethod",
    "GPTQMarlin24LinearMethod",
    "TPUInt8LinearMethod",
    "GPTQLinearMethod",
    "FBGEMMFp8LinearMethod",
    "ModelOptFp8LinearMethod",
    "IPEXAWQLinearMethod",
    "IPEXGPTQLinearMethod",
    "HQQMarlinMethod",
    "QuarkLinearMethod",
    "ModelOptNvFp4LinearMethod",
66
    "PetitNvFp4LinearMethod",
zhuwenwen's avatar
zhuwenwen committed
67
    "BlockInt8LinearMethod",
68
]
69

70

71
72
73
def adjust_bitblas_shard(param, shard_size, shard_offset):
    bitblas_tile_size = getattr(param, "bitblas_tile_size", None)
    if bitblas_tile_size is not None:
74
        return (shard_size // bitblas_tile_size, shard_offset // bitblas_tile_size)
75
76
77
78

    return shard_size, shard_offset


79
80
81
82
83
84
85
86
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


87
88
89
def adjust_bitsandbytes_4bit_shard(
    param: Parameter, shard_offsets: dict[str, tuple[int, int]], loaded_shard_id: str
) -> tuple[int, int]:
90
91
    """Adjust the quantization offsets and sizes for BitsAndBytes sharding."""

92
93
    total, _ = shard_offsets["total"]
    orig_offset, orig_size = shard_offsets[loaded_shard_id]
94
95
96
97
98
99
100
101

    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


102
103
104
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
105
106
    is an array of length N. The loaded_weight corresponds to
    one of the shards on disk. Here, we slice the param based on
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
    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]

122
123
124
125
    if envs.VLLM_USE_NN:
        return param[shard_id], loaded_weight.t()
    else:
        return param[shard_id], loaded_weight
126
127


128
129
130
131
132
133
134
135
# TODO(Isotr0py): We might need a more flexible structure to handle
# bitsandbytes shard offsets.
def left_shift_bitsandbytes_4bit_shard(bnb_weight_attrs: dict[str, Any]):
    """
    Separate the BitsAndBytes 4-bit shard.

    For example, given bnb weight attributes as below:
    {
136
        'bnb_shard_offsets': array([0, 4, 8, 16]),
137
138
139
140
141
        'bnb_quant_state': {0: ..., 1: ..., 2: ...},
    }

    The function will return:
    {
142
        'bnb_shard_offsets': array([0, 4]),
143
144
145
146
147
148
149
150
151
152
153
154
155
156
        'bnb_quant_state': {0: ...},
    }
    and
    {
        'bnb_shard_offsets': array([0, 4, 12]),
        'bnb_quant_state': {0: ..., 1: ...},
    }
    """
    shard_offsets = bnb_weight_attrs["bnb_shard_offsets"]
    offset_l = shard_offsets[:2]
    offset_r = shard_offsets[1:] - shard_offsets[1]
    quant_state_l = {0: bnb_weight_attrs["bnb_quant_state"][0]}
    quant_state_r = {
        i - 1: bnb_weight_attrs["bnb_quant_state"][i]
157
        for i in range(1, len(shard_offsets) - 1)
158
159
160
161
162
163
    }
    left = dict(bnb_shard_offsets=offset_l, bnb_quant_state=quant_state_l)
    right = dict(bnb_shard_offsets=offset_r, bnb_quant_state=quant_state_r)
    return left, right


164
class LinearMethodBase(QuantizeMethodBase):
165
166
167
    """Base class for different (maybe quantized) linear methods."""

    @abstractmethod
168
169
170
171
172
173
174
175
176
177
178
    def create_weights(
        self,
        layer: torch.nn.Module,
        input_size_per_partition: int,
        output_partition_sizes: list[int],
        input_size: int,
        output_size: int,
        params_dtype: torch.dtype,
        **extra_weight_attrs,
    ):
        """Create weights for a linear layer.
179
           The weights will be set as attributes of the layer.
180

181
182
183
        Args:
            layer: The layer that is using the LinearMethodBase factory.
            input_size_per_partition: Size of the weight input dim on rank X.
184
            output_partition_sizes: Sizes of the output dim of each logical
185
186
187
188
189
190
                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.
        """
191
192
193
        raise NotImplementedError

    @abstractmethod
194
195
196
197
    def apply(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
198
        bias: torch.Tensor | None = None,
199
    ) -> torch.Tensor:
200
201
        """Apply the weights in layer to the input tensor.
        Expects create_weights to have been called before on the layer."""
202
203
204
205
        raise NotImplementedError


class UnquantizedLinearMethod(LinearMethodBase):
206
    """Linear method without quantization."""
207
208
    
    def __init__(self):
zhuwenwen's avatar
zhuwenwen committed
209
        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
210
        self.use_gemm_pad = os.environ.get('GEMM_PAD') == '1'
gaoqiong's avatar
gaoqiong committed
211
        
212

213

214
215
216
217
218
219
220
221
222
223
    def create_weights(
        self,
        layer: torch.nn.Module,
        input_size_per_partition: int,
        output_partition_sizes: list[int],
        input_size: int,
        output_size: int,
        params_dtype: torch.dtype,
        **extra_weight_attrs,
    ):
224
225
226
227
        # This method creates unquantized linear weights.
        # The weights are not quantized, and they are not sharded.
        # The amount of memory allocated for the weights is
        # sum(output_partition_sizes) * input_size_per_partition.
228
        weight_loader = extra_weight_attrs.pop("weight_loader")
229
230
231
        if envs.VLLM_USE_NN:
            weight = ModelWeightParameter(
                data=torch.empty(
232
233
                    input_size_per_partition,
                    sum(output_partition_sizes),
234
235
236
237
238
239
240
241
242
                    dtype=params_dtype,
                ),
                input_dim=1,
                output_dim=0,
                weight_loader=weight_loader,
            )
        else:
            weight = ModelWeightParameter(
                data=torch.empty(
243
244
                    sum(output_partition_sizes),
                    input_size_per_partition,
245
246
247
248
249
250
                    dtype=params_dtype,
                ),
                input_dim=1,
                output_dim=0,
                weight_loader=weight_loader,
            )
251

252
253
        layer.register_parameter("weight", weight)
        set_weight_attrs(weight, extra_weight_attrs)
254

255
    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
256
        if current_platform.is_cpu():
257
            from vllm.model_executor.layers.utils import dispatch_cpu_unquantized_gemm
258

259
            dispatch_cpu_unquantized_gemm(layer, remove_weight=True)
260

261
262
263
264
    def apply(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
265
        bias: torch.Tensor | None = None,
266
    ) -> torch.Tensor:
zhuwenwen's avatar
zhuwenwen committed
267
        if self.use_llama_nn:
zhuwenwen's avatar
zhuwenwen committed
268
269
            # if os.environ['GEMM_PAD'] == '1' and gemm_bank_conf(layer.weight.shape[1] - 32):
            #     layer.weight = layer.weight[:,:-32]
zhuwenwen's avatar
zhuwenwen committed
270
            if bias is not None:
zhuwenwen's avatar
zhuwenwen committed
271
                if len(x.shape) == 2: 
272
                    return torch.addmm(bias, x, layer.weight)
zhuwenwen's avatar
zhuwenwen committed
273
                else:
274
                    return torch.matmul(x, layer.weight) + bias
zhuwenwen's avatar
zhuwenwen committed
275
            else:
276
                return torch.matmul(x, layer.weight)
zhuwenwen's avatar
zhuwenwen committed
277
        else:
zhuwenwen's avatar
zhuwenwen committed
278
279
280
281
282
283
284
285
286
287
            # if envs.VLLM_USE_NN and x.shape[-1] == layer.weight.shape[0]:
            #     return dispatch_unquantized_gemm()(layer, x, layer.weight.t(), bias)
            if envs.VLLM_USE_NN:
                if bias is not None:
                    if len(x.shape) == 2: 
                        return torch.addmm(bias, x, layer.weight)
                    else:
                        return torch.matmul(x, layer.weight) + bias
                else:
                    return torch.matmul(x, layer.weight)
288
            else:
zhuwenwen's avatar
zhuwenwen committed
289
                return dispatch_unquantized_gemm()(layer, x, layer.weight, bias)
290

291

292
class LinearBase(CustomOp):
293
    """Base linear layer.
294
295
296
297
298
299

    Args:
        input_size: input dimension of the linear layer.
        output_size: output dimension of the linear layer.
        skip_bias_add: If true, skip adding bias but instead return it.
        params_dtype: Data type for the parameters.
300
        quant_config: Quantization configure.
301
        prefix: Prefix for parameter names.
302
        return_bias: If true, return bias together with outputs in forward pass.
303
        disable_tp: If true, tensor parallelism will be disabled for this layer.
304
305
306
307
308
309
310
    """

    def __init__(
        self,
        input_size: int,
        output_size: int,
        skip_bias_add: bool = False,
311
312
        params_dtype: torch.dtype | None = None,
        quant_config: QuantizationConfig | None = None,
313
        prefix: str = "",
314
315
        *,
        return_bias: bool = True,
316
        disable_tp: bool = False,
317
318
319
320
321
322
323
324
325
326
    ):
        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
327
328
        self.quant_config = quant_config
        self.prefix = prefix
329
        if quant_config is None:
330
            self.quant_method: QuantizeMethodBase | None = UnquantizedLinearMethod()
331
        else:
332
            self.quant_method = quant_config.get_quant_method(self, prefix=prefix)
333
        self.return_bias = return_bias
334
        self.disable_tp = disable_tp
335
336
        self.tp_rank = get_tensor_model_parallel_rank() if not disable_tp else 0
        self.tp_size = get_tensor_model_parallel_world_size() if not disable_tp else 1
337

338
    def update_param_tp_status(self):
339
340
341
342
        for param in self.parameters():
            if isinstance(param, BasevLLMParameter):
                param.tp_rank = self.tp_rank
                param.tp_size = self.tp_size
343
344


345
@CustomOp.register("replicated_linear")
346
347
348
349
350
351
352
353
354
355
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.
356
357
        prefix: The name of the layer in the state dict, including all parents
                        (e.g. model.layers.0.qkv_proj)
358
        return_bias: If true, return bias together with outputs in forward pass.
359
        disable_tp: Take no effect for replicated linear layers.
360
361
    """

362
363
364
365
366
367
    def __init__(
        self,
        input_size: int,
        output_size: int,
        bias: bool = True,
        skip_bias_add: bool = False,
368
369
        params_dtype: torch.dtype | None = None,
        quant_config: QuantizationConfig | None = None,
370
        eps: float | None = 1e-6,
371
372
373
        prefix: str = "",
        *,
        return_bias: bool = True,
374
        disable_tp: bool = False,
375
    ):
376
377
378
379
380
        # If MergedReplicatedLinear, use output size of each partition.
        if hasattr(self, "output_sizes"):
            self.output_partition_sizes = self.output_sizes
        else:
            self.output_partition_sizes = [output_size]
381
            
382
        self.eps = eps
383

384
385
386
387
388
389
390
391
392
393
        super().__init__(
            input_size,
            output_size,
            skip_bias_add,
            params_dtype,
            quant_config,
            prefix=prefix,
            return_bias=return_bias,
            disable_tp=disable_tp,
        )
394

395
396
        # All the linear layer supports quant method.
        assert self.quant_method is not None
397
398
399
400
401
402
403
404
405
        self.quant_method.create_weights(
            self,
            self.input_size,
            self.output_partition_sizes,
            self.input_size,
            self.output_size,
            self.params_dtype,
            weight_loader=self.weight_loader,
        )
406

407
408
        if bias:
            self.bias = Parameter(
409
410
411
412
413
414
415
416
417
                torch.empty(self.output_size, dtype=self.params_dtype)
            )
            set_weight_attrs(
                self.bias,
                {
                    "output_dim": 0,
                    "weight_loader": self.weight_loader,
                },
            )
418
419
        else:
            self.register_parameter("bias", None)
420
        self.is_quantization = not isinstance(self.quant_method, UnquantizedLinearMethod)
421

422
423
424
    def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor):
        # If the weight on disk does not have a shape, give it one
        # (such scales for AutoFp8).
425
426
427
428
429
430
431
432
433
434
435
        # Special case for GGUF

        is_gguf_weight = getattr(param, "is_gguf_weight", False)
        is_gguf_weight_type = getattr(param, "is_gguf_weight_type", False)
        if is_gguf_weight_type:
            param.weight_type = loaded_weight.item()

        # Materialize GGUF UninitializedParameter
        if is_gguf_weight and isinstance(param, UninitializedParameter):
            param.materialize(loaded_weight.shape, dtype=loaded_weight.dtype)

436
437
438
        if len(loaded_weight.shape) == 0:
            loaded_weight = loaded_weight.reshape(1)

439
        if envs.VLLM_USE_NN and not self.is_quantization:
440
441
            loaded_weight = loaded_weight.t()
            
442
443
        assert param.size() == loaded_weight.size(), (
            f"Tried to load weights of size {loaded_weight.size()}"
444
445
            f"to a parameter of size {param.size()}"
        )
446
447
        param.data.copy_(loaded_weight)

448
    def forward(
449
        self,
zhuwenwen's avatar
zhuwenwen committed
450
        x: torch.Tensor,
451
    ) -> torch.Tensor | tuple[torch.Tensor, Parameter | None]:
zhuwenwen's avatar
zhuwenwen committed
452
453
        bias = self.bias if not self.skip_bias_add else None
        assert self.quant_method is not None
454

zhuwenwen's avatar
zhuwenwen committed
455
456
        output = self.quant_method.apply(self, x, bias)
        output_bias = self.bias if self.skip_bias_add else None
457

zhuwenwen's avatar
zhuwenwen committed
458
459
460
        if not self.return_bias:
            return output
        return output, output_bias
461

462
463
464
465
466
467
    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

468

469
@CustomOp.register("column_parallel_linear")
470
class ColumnParallelLinear(LinearBase):
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
    """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.
487
        quant_config: Quantization configure.
James Fleming's avatar
James Fleming committed
488
489
        output_sizes: list of output sizes packed into one output, like for QKV
                       the list would be size 3.
490
        prefix: The name of the layer in the state dict, including all parents
491
                        (e.g. model.layers.0.qkv_proj)
492
493
        return_bias: If true, return bias together with outputs in forward pass.
        disable_tp: If true, weights matrix won't be sharded through tp rank.
494
495
    """

496
497
498
499
500
501
502
    def __init__(
        self,
        input_size: int,
        output_size: int,
        bias: bool = True,
        gather_output: bool = False,
        skip_bias_add: bool = False,
503
504
505
        params_dtype: torch.dtype | None = None,
        quant_config: QuantizationConfig | None = None,
        output_sizes: list[int] | None = None,
506
        eps: float | None = 1e-6,
507
508
509
        prefix: str = "",
        *,
        return_bias: bool = True,
510
        disable_tp: bool = False,
511
    ):
512
        # Divide the weight matrix along the last dimension.
513
514
        self.tp_rank = get_tensor_model_parallel_rank() if not disable_tp else 0
        self.tp_size = get_tensor_model_parallel_world_size() if not disable_tp else 1
515
516
        self.input_size_per_partition = input_size
        self.output_size_per_partition = divide(output_size, self.tp_size)
517
518
519
520
        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 = [
521
                divide(output_size, self.tp_size) for output_size in self.output_sizes
522
523
            ]

524
525
526
527
528
529
530
531
532
533
        super().__init__(
            input_size,
            output_size,
            skip_bias_add,
            params_dtype,
            quant_config,
            prefix,
            return_bias=return_bias,
            disable_tp=disable_tp,
        )
534

535
        self.eps = eps
536
537
        self.gather_output = gather_output

James Fleming's avatar
James Fleming committed
538
539
        if output_sizes is None:
            output_sizes = [output_size]
540

541
        assert self.quant_method is not None
542
543
        self.quant_method.create_weights(
            layer=self,
544
            input_size_per_partition=self.input_size_per_partition,
545
546
547
548
            output_partition_sizes=self.output_partition_sizes,
            input_size=self.input_size,
            output_size=self.output_size,
            params_dtype=self.params_dtype,
549
            weight_loader=(
550
551
552
553
554
                self.weight_loader_v2
                if self.quant_method.__class__.__name__ in WEIGHT_LOADER_V2_SUPPORTED
                else self.weight_loader
            ),
        )
555
556
        if bias:
            self.bias = Parameter(
557
558
559
560
561
562
563
564
565
                torch.empty(self.output_size_per_partition, dtype=params_dtype)
            )
            set_weight_attrs(
                self.bias,
                {
                    "output_dim": 0,
                    "weight_loader": self.weight_loader,
                },
            )
566
567
        else:
            self.register_parameter("bias", None)
568
        self.update_param_tp_status()
569
        self.is_quantization = not isinstance(self.quant_method, UnquantizedLinearMethod)
570

571
572
    def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor):
        output_dim = getattr(param, "output_dim", None)
573

574
575
576
577
578
579
        is_sharded_weight = getattr(param, "is_sharded_weight", False)
        use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit", False)
        # bitsandbytes loads the weights of the specific portion
        # no need to narrow
        is_sharded_weight = is_sharded_weight or use_bitsandbytes_4bit

580
581
582
583
584
585
586
587
        # Special case for GGUF
        is_gguf_weight = getattr(param, "is_gguf_weight", False)
        is_gguf_weight_type = getattr(param, "is_gguf_weight_type", False)
        if is_gguf_weight_type:
            param.weight_type = loaded_weight.item()

        # Materialize GGUF UninitializedParameter
        if is_gguf_weight and isinstance(param, UninitializedParameter):
588
589
            final_shape = list(loaded_weight.shape)
            if output_dim is not None:
590
                assert final_shape[output_dim] % self.tp_size == 0
591
                final_shape[output_dim] = final_shape[output_dim] // self.tp_size
592
            param.materialize(final_shape, dtype=loaded_weight.dtype)
593

594
        param_data = param.data
595
        if output_dim is not None and not is_sharded_weight:
596
            if not envs.VLLM_USE_NN or len(param_data.shape)==1 or self.is_quantization:
597
598
599
                shard_size = param_data.shape[output_dim] 
            else:
                shard_size = param_data.shape[int(not(output_dim))]
600
            start_idx = self.tp_rank * shard_size
601
            loaded_weight = loaded_weight.narrow(output_dim, start_idx, shard_size)
602

603
604
605
606
        # 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)
607

608
        if envs.VLLM_USE_NN and not self.is_quantization:
609
610
            loaded_weight = loaded_weight.t()
            
611
612
613
        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)

614
    def weight_loader_v2(self, param: BasevLLMParameter, loaded_weight: torch.Tensor):
615
616
617
618
619
        # 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:
            assert loaded_weight.numel() == 1
            loaded_weight = loaded_weight.reshape(1)
620
        param.load_column_parallel_weight(loaded_weight=loaded_weight, is_quantization=self.is_quantization)
621

622
    def forward(
623
624
        self,
        input_,
625
    ) -> torch.Tensor | tuple[torch.Tensor, Parameter | None]:
zhuwenwen's avatar
zhuwenwen committed
626
627
628
629
630
631
632
633
634
        bias = self.bias if not self.skip_bias_add else None

        # Matrix multiply.
        assert self.quant_method is not None
        output_parallel = self.quant_method.apply(self, input_, bias)

        if self.gather_output and self.tp_size > 1:
            # All-gather across the partitions.
            output = tensor_model_parallel_all_gather(output_parallel)
635
        else:
zhuwenwen's avatar
zhuwenwen committed
636
637
638
639
640
            output = output_parallel
        output_bias = self.bias if self.skip_bias_add else None
        if not self.return_bias:
            return output
        return output, output_bias
641

642
643
644
645
    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}"
646
        s += f", tp_size={self.tp_size}"
647
648
649
        s += f", gather_output={self.gather_output}"
        return s

650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668

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.
669
        quant_config: Quantization configure.
670
671
        prefix: The name of the layer in the state dict, including all parents
                        (e.g. model.layers.0.qkv_proj)
672
        return_bias: If true, return bias together with outputs in forward pass.
673
674
        disable_tp: If true, all weights matrix won't be sharded, this layer
                    will be treated as a "Replicated" MergedLinear.
675
676
    """

677
678
679
680
681
682
683
    def __init__(
        self,
        input_size: int,
        output_sizes: list[int],
        bias: bool = True,
        gather_output: bool = False,
        skip_bias_add: bool = False,
684
685
        params_dtype: torch.dtype | None = None,
        quant_config: QuantizationConfig | None = None,
zhuwenwen's avatar
zhuwenwen committed
686
        eps: float | None = 1e-6,
687
688
689
        prefix: str = "",
        *,
        return_bias: bool = True,
690
        disable_tp: bool = False,
691
    ):
692
        self.eps = eps
693
        self.output_sizes = output_sizes
694
695
        self.tp_size = get_tensor_model_parallel_world_size() if not disable_tp else 1
        self.tp_rank = get_tensor_model_parallel_rank() if not disable_tp else 0
696

697
698
699
700
701
702
703
704
705
706
707
708
709
        assert all(output_size % self.tp_size == 0 for output_size in output_sizes)
        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,
            prefix=prefix,
            return_bias=return_bias,
            disable_tp=disable_tp,
        )
710
        self.is_quantization = not isinstance(self.quant_method, UnquantizedLinearMethod)
James Fleming's avatar
James Fleming committed
711

712
713
714
715
    def weight_loader(
        self,
        param: Parameter,
        loaded_weight: torch.Tensor,
716
        loaded_shard_id: int | None = None,
717
    ):
718
719
720
721
722
        # Special case for GGUF
        # initialize GGUF param after we know the quantize type
        is_gguf_weight = getattr(param, "is_gguf_weight", False)
        is_gguf_weight_type = getattr(param, "is_gguf_weight_type", False)
        if is_gguf_weight_type:
723
724
725
726
727
            if loaded_shard_id is not None:
                param.data[loaded_shard_id].copy_(loaded_weight)
                param.shard_weight_type[loaded_shard_id] = loaded_weight.item()
            else:
                param.shard_weight_type = {
728
                    i: loaded_weight.item() for i, _ in enumerate(self.output_sizes)
729
                }
730
731
            return

732
733
        if is_gguf_weight:
            output_dim = getattr(param, "output_dim", None)
734
735
            shard_size = loaded_weight.size(output_dim) // self.tp_size
            start_idx = self.tp_rank * shard_size
736

737
            if loaded_shard_id is not None:
738
                loaded_weight = loaded_weight.narrow(output_dim, start_idx, shard_size)
739
740
741
742
                param.shard_id.append(loaded_shard_id)
                param.shard_id_map[loaded_shard_id] = len(param.data_container)
                param.data_container.append(loaded_weight)
                return
743

744
745
        param_data = param.data
        output_dim = getattr(param, "output_dim", None)
746
747
        # Special case for per-tensor scale to load scalar into fused array.
        needs_scalar_to_array = getattr(param, "needs_scalar_to_array", False)
748

749
        if loaded_shard_id is None:
750
751
            # Loaded weight is already fused on disk (mlp).
            # (e.g., Phi-3's gate_up_proj).
752
            if output_dim is None:
753
                if needs_scalar_to_array:
754
                    param_data, loaded_weight = adjust_scalar_to_fused_array(
755
756
                        param_data, loaded_weight, 0
                    )
757

758
759
760
761
                assert param_data.shape == loaded_weight.shape
                param_data.copy_(loaded_weight)
                return
            current_shard_offset = 0
762
            use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit", False)
763
            shard_offsets: list[tuple[int, int, int]] = []
764
765
766
767
768
            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:
769
                # Special case for Quantization.
770
771
772
                # If quantized, we need to adjust the offset and size to account
                # for the packing.
                if packed_dim == output_dim:
773
774
                    shard_size = shard_size // param.packed_factor
                    shard_offset = shard_offset // param.packed_factor
775
                    # Special case for Marlin.
776
                    shard_size, shard_offset = adjust_marlin_shard(
777
778
                        param, shard_size, shard_offset
                    )
779

780
                shard_size, shard_offset = adjust_bitblas_shard(
781
782
                    param, shard_size, shard_offset
                )
783

784
                if use_bitsandbytes_4bit:
785
786
787
788
789
790
791
                    index = list(itertools.accumulate([0] + self.output_sizes))
                    orig_offsets = {
                        str(i): (index[i], size)
                        for i, size in enumerate(self.output_sizes)
                    }
                    orig_offsets["total"] = (self.output_size, 0)
                    shard_size, shard_offset = adjust_bitsandbytes_4bit_shard(
792
793
                        param, orig_offsets, str(shard_id)
                    )
794

795
                loaded_weight_shard = loaded_weight.narrow(
796
797
                    output_dim, shard_offset, shard_size
                )
798
799
800
801
802
                self.weight_loader(param, loaded_weight_shard, shard_id)
            return

        assert loaded_shard_id < len(self.output_sizes)
        if output_dim is not None:
803
            shard_offset = sum(self.output_sizes[:loaded_shard_id]) // self.tp_size
804
            shard_size = self.output_sizes[loaded_shard_id] // self.tp_size
805
            # Special case for quantization.
806
807
808
809
            # 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:
810
811
                shard_size = shard_size // param.packed_factor
                shard_offset = shard_offset // param.packed_factor
812
                # Special case for Marlin.
813
                shard_size, shard_offset = adjust_marlin_shard(
814
815
                    param, shard_size, shard_offset
                )
816
            shard_size, shard_offset = adjust_bitblas_shard(
817
818
                param, shard_size, shard_offset
            )
gaoqiong's avatar
gaoqiong committed
819

820
            use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit", False)
821
822
823
824
825
            is_sharded_weight = getattr(param, "is_sharded_weight", False)
            # bitsandbytes loads the weights of the specific portion
            # no need to narrow
            is_sharded_weight = is_sharded_weight or use_bitsandbytes_4bit

826
            if use_bitsandbytes_4bit:
827
                shard_size = loaded_weight.shape[output_dim]
828
                shard_offset = loaded_weight.shape[output_dim] * loaded_shard_id
829

830
            if not envs.VLLM_USE_NN or self.is_quantization or (envs.VLLM_USE_NN and param_data.dim()==1):
831
832
833
                param_data = param_data.narrow(output_dim, shard_offset, shard_size)
            else:
                param_data = param_data.narrow(int(not(output_dim)), shard_offset, shard_size)
834
            start_idx = self.tp_rank * shard_size
835
            if not is_sharded_weight:
836
                loaded_weight = loaded_weight.narrow(output_dim, start_idx, shard_size)
837
838
839
        # Special case for per-tensor scales in fused case.
        elif needs_scalar_to_array:
            param_data, loaded_weight = adjust_scalar_to_fused_array(
840
841
                param_data, loaded_weight, loaded_shard_id
            )
842

843
        else:
CHU Tianxiang's avatar
CHU Tianxiang committed
844
845
846
847
848
            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 "
849
850
                    "the same for all partitions."
                )
851

852
        if envs.VLLM_USE_NN and not self.is_quantization:
853
854
            loaded_weight = loaded_weight.t()
            
gaoqiong's avatar
gaoqiong committed
855
856
        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)
857

858
859
860
    def _load_fused_module_from_checkpoint(
        self, param: BasevLLMParameter, loaded_weight: torch.Tensor
    ):
861
862
863
        """
        Handle special case for models where MLP layers are already
        fused on disk. In this case, we have no shard id. This function
864
        determines the shard id by splitting these layers and then calls
865
866
867
868
869
870
871
        the weight loader using the shard id.

        An example of a model with these fused layers:
        https://huggingface.co/microsoft/Phi-3-mini-4k-instruct
        """

        current_shard_offset = 0
872
        shard_offsets: list[tuple[int, int, int]] = []
873
874
875
876
877
878
879
880
        for i, output_size in enumerate(self.output_sizes):
            shard_offsets.append((i, current_shard_offset, output_size))
            current_shard_offset += output_size

        for shard_id, shard_offset, shard_size in shard_offsets:
            # Special case for Quantization.
            # If quantized, we need to adjust the offset and size to account
            # for the packing.
881
882
883
884
885
886
887
888
889
890
891
            if (
                isinstance(param, (PackedColumnParameter, PackedvLLMParameter))
                and param.packed_dim == param.output_dim
            ):
                shard_size, shard_offset = param.adjust_shard_indexes_for_packing(
                    shard_size=shard_size, shard_offset=shard_offset
                )

            loaded_weight_shard = loaded_weight.narrow(
                param.output_dim, shard_offset, shard_size
            )
892
893
            self.weight_loader_v2(param, loaded_weight_shard, shard_id)

894
895
896
897
    def weight_loader_v2(
        self,
        param: BasevLLMParameter,
        loaded_weight: torch.Tensor,
898
        loaded_shard_id: int | None = None,
899
    ):
900
        if loaded_shard_id is None:
901
            if isinstance(param, PerTensorScaleParameter):
902
                param.load_merged_column_weight(loaded_weight=loaded_weight, shard_id=0)
903
                return
904
            elif type(param) in (RowvLLMParameter, BasevLLMParameter):
905
                param.load_merged_column_weight(loaded_weight=loaded_weight)
906
                return
907
            # TODO: @dsikka - move to parameter.py
908
909
910
911
912
            self._load_fused_module_from_checkpoint(param, loaded_weight)
            return

        assert loaded_shard_id < len(self.output_sizes)

913
914
        if isinstance(param, BlockQuantScaleParameter):
            assert self.quant_method is not None
915
916
917
            # Assume the weight block size has been set by quant method
            assert hasattr(self, "weight_block_size")
            weight_block_size = self.weight_block_size
918
919
920
            assert weight_block_size is not None
            block_n, _ = weight_block_size[0], weight_block_size[1]
            shard_offset = (
921
922
923
924
925
926
927
                (sum(self.output_sizes[:loaded_shard_id]) + block_n - 1) // block_n
            ) // self.tp_size
            shard_size = (
                (self.output_sizes[loaded_shard_id] + block_n - 1)
                // block_n
                // self.tp_size
            )
928
        else:
929
            shard_offset = sum(self.output_sizes[:loaded_shard_id]) // self.tp_size
930
            shard_size = self.output_sizes[loaded_shard_id] // self.tp_size
931

932
933
934
935
936
937
        param.load_merged_column_weight(
            loaded_weight=loaded_weight,
            shard_id=loaded_shard_id,
            shard_offset=shard_offset,
            shard_size=shard_size,
            tp_rank=self.tp_rank,
938
            is_quantization=self.is_quantization
939
        )
940

941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962

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.
963
        quant_config: Quantization configure.
964
965
        prefix: The name of the layer in the state dict, including all parents
                        (e.g. model.layers.0.qkv_proj)
966
        return_bias: If true, return bias together with outputs in forward pass.
967
        disable_tp: If true, weights matrix won't be sharded through tp rank.
968
969
    """

970
971
972
973
974
    def __init__(
        self,
        hidden_size: int,
        head_size: int,
        total_num_heads: int,
975
        total_num_kv_heads: int | None = None,
976
977
        bias: bool = True,
        skip_bias_add: bool = False,
978
979
        params_dtype: torch.dtype | None = None,
        quant_config: QuantizationConfig | None = None,
980
981
982
        prefix: str = "",
        *,
        return_bias: bool = True,
983
        disable_tp: bool = False,
984
    ):
985
986
987
988
989
990
991
        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.
992
        tp_size = get_tensor_model_parallel_world_size() if not disable_tp else 1
993
994
995
        self.num_heads = divide(self.total_num_heads, tp_size)
        if tp_size >= self.total_num_kv_heads:
            self.num_kv_heads = 1
996
            self.num_kv_head_replicas = divide(tp_size, self.total_num_kv_heads)
997
998
999
1000
        else:
            self.num_kv_heads = divide(self.total_num_kv_heads, tp_size)
            self.num_kv_head_replicas = 1
        input_size = self.hidden_size
1001
1002
1003
        output_size = (
            (self.num_heads + 2 * self.num_kv_heads) * tp_size * self.head_size
        )
1004
1005
1006
        self.output_sizes = [
            self.num_heads * self.head_size * tp_size,  # q_proj
            self.num_kv_heads * self.head_size * tp_size,  # k_proj
1007
            self.num_kv_heads * self.head_size * tp_size,  # v_proj
James Fleming's avatar
James Fleming committed
1008
        ]
gaoqiong's avatar
gaoqiong committed
1009

1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
        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,
            prefix=prefix,
            return_bias=return_bias,
            disable_tp=disable_tp,
        )
1022
        self.is_quantization = not isinstance(self.quant_method, UnquantizedLinearMethod)
1023

1024
1025
1026
1027
1028
    def _get_shard_offset_mapping(self, loaded_shard_id: str):
        shard_offset_mapping = {
            "q": 0,
            "k": self.num_heads * self.head_size,
            "v": (self.num_heads + self.num_kv_heads) * self.head_size,
1029
            "total": (self.num_heads + 2 * self.num_kv_heads) * self.head_size,
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
        }
        return shard_offset_mapping.get(loaded_shard_id)

    def _get_shard_size_mapping(self, loaded_shard_id: str):
        shard_size_mapping = {
            "q": self.num_heads * self.head_size,
            "k": self.num_kv_heads * self.head_size,
            "v": self.num_kv_heads * self.head_size,
        }
        return shard_size_mapping.get(loaded_shard_id)

1041
1042
1043
    def _load_fused_module_from_checkpoint(
        self, param: BasevLLMParameter, loaded_weight: torch.Tensor
    ):
1044
        """
1045
        Handle special case for models where QKV layers are already
1046
        fused on disk. In this case, we have no shard id. This function
1047
        determines the shard id by splitting these layers and then calls
1048
1049
1050
1051
1052
1053
1054
1055
        the weight loader using the shard id.

        An example of a model with these fused layers:
        https://huggingface.co/microsoft/Phi-3-mini-4k-instruct
        """
        shard_offsets = [
            # (shard_id, shard_offset, shard_size)
            ("q", 0, self.total_num_heads * self.head_size),
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
            (
                "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,
            ),
1066
1067
1068
1069
1070
1071
        ]

        for shard_id, shard_offset, shard_size in shard_offsets:
            # Special case for Quantization.
            # If quantized, we need to adjust the offset and size to account
            # for the packing.
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
            if (
                isinstance(param, (PackedColumnParameter, PackedvLLMParameter))
                and param.packed_dim == param.output_dim
            ):
                shard_size, shard_offset = param.adjust_shard_indexes_for_packing(
                    shard_size=shard_size, shard_offset=shard_offset
                )

            loaded_weight_shard = loaded_weight.narrow(
                param.output_dim, shard_offset, shard_size
            )
1083
1084
            self.weight_loader_v2(param, loaded_weight_shard, shard_id)

1085
1086
1087
1088
    def weight_loader_v2(
        self,
        param: BasevLLMParameter,
        loaded_weight: torch.Tensor,
1089
        loaded_shard_id: str | None = None,
1090
    ):
1091
        if loaded_shard_id is None:  # special case for certain models
1092
            if isinstance(param, PerTensorScaleParameter):
1093
1094
1095
                param.load_qkv_weight(
                    loaded_weight=loaded_weight, shard_id=0, tp_rank=self.tp_rank
                )
1096
                return
1097
            elif type(param) in (RowvLLMParameter, BasevLLMParameter):
1098
                param.load_qkv_weight(loaded_weight=loaded_weight, tp_rank=self.tp_rank)
1099
                return
1100
            # TODO: @dsikka - move to parameter.py
1101
1102
1103
1104
1105
1106
1107
1108
            self._load_fused_module_from_checkpoint(param, loaded_weight)
            return

        assert loaded_shard_id in ["q", "k", "v"]

        shard_offset = self._get_shard_offset_mapping(loaded_shard_id)
        shard_size = self._get_shard_size_mapping(loaded_shard_id)

1109
1110
1111
        # Note(simon): This is needed for Qwen3's fp8 quantization.
        if isinstance(param, BlockQuantScaleParameter):
            assert self.quant_method is not None
1112
1113
1114
1115
            # Assume the weight block size has been set by quant method
            assert hasattr(self, "weight_block_size")
            weight_block_size = self.weight_block_size
            assert weight_block_size is not None
1116
1117
1118
1119
            block_n, _ = weight_block_size[0], weight_block_size[1]
            shard_offset = (shard_offset + block_n - 1) // block_n
            shard_size = (shard_size + block_n - 1) // block_n

1120
1121
1122
1123
1124
1125
1126
        param.load_qkv_weight(
            loaded_weight=loaded_weight,
            num_heads=self.num_kv_head_replicas,
            shard_id=loaded_shard_id,
            shard_offset=shard_offset,
            shard_size=shard_size,
            tp_rank=self.tp_rank,
1127
            is_quantization=self.is_quantization, 
1128
        )
1129

1130
1131
1132
1133
    def weight_loader(
        self,
        param: Parameter,
        loaded_weight: torch.Tensor,
1134
        loaded_shard_id: str | None = None,
1135
    ):
1136
1137
1138
1139
        # Special case for GGUF
        # initialize GGUF param after we know the quantize type
        is_gguf_weight = getattr(param, "is_gguf_weight", False)
        is_gguf_weight_type = getattr(param, "is_gguf_weight_type", False)
1140
        if is_gguf_weight_type:
1141
            idx_map = {"q": 0, "k": 1, "v": 2}
1142
1143
1144
1145
            if loaded_shard_id is not None:
                param.data[idx_map[loaded_shard_id]].copy_(loaded_weight)
                param.shard_weight_type[loaded_shard_id] = loaded_weight.item()
            else:
1146
                param.shard_weight_type = {k: loaded_weight.item() for k in idx_map}
1147
1148
            return

1149
1150
        if is_gguf_weight:
            output_dim = getattr(param, "output_dim", None)
1151
1152
            shard_size = loaded_weight.size(output_dim) // self.tp_size
            start_idx = self.tp_rank * shard_size
1153

1154
            if loaded_shard_id is not None:
1155
                loaded_weight = loaded_weight.narrow(output_dim, start_idx, shard_size)
1156
1157
1158
1159
                param.shard_id.append(loaded_shard_id)
                param.shard_id_map[loaded_shard_id] = len(param.data_container)
                param.data_container.append(loaded_weight)
                return
1160

1161
1162
        param_data = param.data
        output_dim = getattr(param, "output_dim", None)
1163

1164
1165
        # Special case for per-tensor scales in fused case.
        needs_scalar_to_array = getattr(param, "needs_scalar_to_array", False)
1166

1167
        if loaded_shard_id is None:
1168
1169
            # Loaded weight is already fused on disk (qkv).
            # (e.g., Phi-3's qkv_proj).
1170
            if output_dim is None:
1171
                if needs_scalar_to_array:
1172
                    param_data, loaded_weight = adjust_scalar_to_fused_array(
1173
1174
                        param_data, loaded_weight, 0
                    )
1175

1176
1177
1178
1179
1180
1181
                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),
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
                (
                    "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,
                ),
1192
            ]
1193
            use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit", False)
1194

1195
1196
            packed_dim = getattr(param, "packed_dim", None)
            for shard_id, shard_offset, shard_size in shard_offsets:
1197
                # Special case for Quantized Weights.
1198
1199
1200
                # If quantized, we need to adjust the offset and size to account
                # for the packing.
                if packed_dim == output_dim:
1201
1202
                    shard_size = shard_size // param.packed_factor
                    shard_offset = shard_offset // param.packed_factor
1203

1204
                    # Special case for Marlin.
1205
                    shard_size, shard_offset = adjust_marlin_shard(
1206
1207
                        param, shard_size, shard_offset
                    )
1208

1209
1210
1211
                if use_bitsandbytes_4bit:
                    orig_qkv_offsets = {
                        "q": (0, self.total_num_heads * self.head_size),
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
                        "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,
                        ),
                        "total": (
                            (self.total_num_heads + 2 * self.total_num_kv_heads)
                            * self.head_size,
                            0,
                        ),
1226
1227
1228
                    }

                    shard_size, shard_offset = adjust_bitsandbytes_4bit_shard(
1229
1230
                        param, orig_qkv_offsets, shard_id
                    )
1231

1232
                loaded_weight_shard = loaded_weight.narrow(
1233
1234
                    output_dim, shard_offset, shard_size
                )
1235
1236
1237
1238
                self.weight_loader(param, loaded_weight_shard, shard_id)
            return

        assert loaded_shard_id in ["q", "k", "v"]
1239
1240

        # If output dim is defined, use the default loading process.
1241
1242
1243
1244
1245
1246
1247
1248
        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":
1249
                shard_offset = (self.num_heads + self.num_kv_heads) * self.head_size
1250
                shard_size = self.num_kv_heads * self.head_size
1251
            # Special case for Quantized Weights.
1252
1253
1254
1255
            # 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:
1256
1257
                shard_size = shard_size // param.packed_factor
                shard_offset = shard_offset // param.packed_factor
1258

1259
                # Special case for Marlin.
1260
                shard_size, shard_offset = adjust_marlin_shard(
1261
1262
                    param, shard_size, shard_offset
                )
1263

1264
            use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit", False)
1265
1266
1267
1268
1269
            is_sharded_weight = getattr(param, "is_sharded_weight", False)
            # bitsandbytes loads the weights of the specific portion
            # no need to narrow
            is_sharded_weight = is_sharded_weight or use_bitsandbytes_4bit

1270
            if use_bitsandbytes_4bit:
1271
1272
                orig_qkv_offsets = {
                    "q": (0, self.num_heads * self.head_size),
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
                    "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,
                    ),
1285
                }
1286
                shard_size, shard_offset = adjust_bitsandbytes_4bit_shard(
1287
1288
                    param, orig_qkv_offsets, loaded_shard_id
                )
gaoqiong's avatar
gaoqiong committed
1289

1290
            if not envs.VLLM_USE_NN or len(param_data.shape)==1 or self.is_quantization:
1291
                param_data = param_data.narrow(output_dim, shard_offset, shard_size)
1292
1293
1294
            else:
                param_data = param_data.narrow(int(not(output_dim)), shard_offset,
                                               shard_size)
zhuwenwen's avatar
zhuwenwen committed
1295
            if loaded_shard_id == "q":
1296
                shard_rank = self.tp_rank
1297
            else:
1298
1299
                shard_rank = self.tp_rank // self.num_kv_head_replicas
            start_idx = shard_rank * shard_size
1300

1301
            if not is_sharded_weight:
1302
                loaded_weight = loaded_weight.narrow(output_dim, start_idx, shard_size)
1303

1304
1305
1306
        # Special case for per-tensor scales in fused case.
        elif needs_scalar_to_array:
            param_data, loaded_weight = adjust_scalar_to_fused_array(
1307
1308
                param_data, loaded_weight, loaded_shard_id
            )
1309
        else:
CHU Tianxiang's avatar
CHU Tianxiang committed
1310
1311
1312
1313
1314
            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 "
1315
1316
                    "for all partitions."
                )
gaoqiong's avatar
gaoqiong committed
1317

1318
        if envs.VLLM_USE_NN and not self.is_quantization:
1319
1320
            loaded_weight = loaded_weight.t()
            
gaoqiong's avatar
gaoqiong committed
1321
1322
        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)
1323
1324


1325
@CustomOp.register("row_parallel_linear")
1326
class RowParallelLinear(LinearBase):
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
    """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.
1349
1350
1351
        reduce_results: If true, call all-reduce on output and make Y available
                       to all GPUs, otherwise, every GPU will have its output
                       which is Y = X_iA_i
1352
        quant_config: Quantization configure.
1353
1354
1355
        prefix: The name of the layer in the state dict, including all parents
                        (e.g. model.layers.0.down_proj)
        return_bias: If true, return bias together with outputs in forward pass.
1356
        disable_tp: If true, weights matrix won't be sharded through tp rank.
1357
1358
    """

1359
1360
1361
1362
1363
1364
1365
    def __init__(
        self,
        input_size: int,
        output_size: int,
        bias: bool = True,
        input_is_parallel: bool = True,
        skip_bias_add: bool = False,
1366
        params_dtype: torch.dtype | None = None,
1367
        reduce_results: bool = True,
1368
        quant_config: QuantizationConfig | None = None,
1369
1370
1371
        prefix: str = "",
        *,
        return_bias: bool = True,
1372
        disable_tp: bool = False,
1373
    ):
1374
        # Divide the weight matrix along the first dimension.
1375
1376
        self.tp_rank = get_tensor_model_parallel_rank() if not disable_tp else 0
        self.tp_size = get_tensor_model_parallel_world_size() if not disable_tp else 1
1377
1378
1379
1380
        self.input_size_per_partition = divide(input_size, self.tp_size)
        self.output_size_per_partition = output_size
        self.output_partition_sizes = [output_size]

1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
        super().__init__(
            input_size,
            output_size,
            skip_bias_add,
            params_dtype,
            quant_config,
            prefix,
            return_bias=return_bias,
            disable_tp=disable_tp,
        )
1391

1392
1393
1394
        self.input_is_parallel = input_is_parallel
        self.reduce_results = reduce_results

1395
        assert self.quant_method is not None
1396
1397
1398
        self.quant_method.create_weights(
            layer=self,
            input_size_per_partition=self.input_size_per_partition,
1399
            output_partition_sizes=self.output_partition_sizes,
1400
1401
1402
            input_size=self.input_size,
            output_size=self.output_size,
            params_dtype=self.params_dtype,
1403
            weight_loader=(
1404
1405
1406
1407
1408
                self.weight_loader_v2
                if self.quant_method.__class__.__name__ in WEIGHT_LOADER_V2_SUPPORTED
                else self.weight_loader
            ),
        )
1409
        if not reduce_results and (bias and not skip_bias_add):
1410
1411
1412
1413
            raise ValueError(
                "When not reduce the results, adding bias to the "
                "results can lead to incorrect results"
            )
1414
1415

        if bias:
1416
1417
1418
1419
1420
1421
1422
1423
            self.bias = Parameter(torch.empty(self.output_size, dtype=params_dtype))
            set_weight_attrs(
                self.bias,
                {
                    "output_dim": 0,
                    "weight_loader": self.weight_loader,
                },
            )
1424
1425
        else:
            self.register_parameter("bias", None)
1426

1427
        self.update_param_tp_status()
1428
        self.is_quantization = not isinstance(self.quant_method, UnquantizedLinearMethod)
1429
1430
1431

    def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor):
        input_dim = getattr(param, "input_dim", None)
1432
        use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit", False)
1433
1434
1435
1436
        is_sharded_weight = getattr(param, "is_sharded_weight", False)
        # bitsandbytes loads the weights of the specific portion
        # no need to narrow
        is_sharded_weight = is_sharded_weight or use_bitsandbytes_4bit
1437
1438
1439
1440
1441
1442
1443
1444
1445

        # Special case for GGUF
        is_gguf_weight = getattr(param, "is_gguf_weight", False)
        is_gguf_weight_type = getattr(param, "is_gguf_weight_type", False)
        if is_gguf_weight_type:
            param.weight_type = loaded_weight.item()

        # Materialize GGUF UninitializedParameter
        if is_gguf_weight and isinstance(param, UninitializedParameter):
1446
1447
            weight_shape = list(loaded_weight.shape)
            if input_dim:
1448
                weight_shape[input_dim] = weight_shape[input_dim] // self.tp_size
1449
            param.materialize(tuple(weight_shape), dtype=loaded_weight.dtype)
1450

1451
        param_data = param.data
1452
        if input_dim is not None and not is_sharded_weight:
1453
            if not envs.VLLM_USE_NN or self.is_quantization:
1454
1455
1456
                shard_size = param_data.shape[input_dim]
            else:
                shard_size = param_data.shape[int(not(input_dim))]
1457
            start_idx = self.tp_rank * shard_size
1458
            loaded_weight = loaded_weight.narrow(input_dim, start_idx, shard_size)
1459

1460
1461
1462
        # 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:
1463
1464
            loaded_weight = loaded_weight.reshape(1)

1465
        if envs.VLLM_USE_NN and not self.is_quantization:
1466
1467
            loaded_weight = loaded_weight.t()
            
1468
1469
1470
        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)

1471
    def weight_loader_v2(self, param: BasevLLMParameter, loaded_weight: torch.Tensor):
1472
1473
1474
1475
1476
1477
        # 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:
            assert loaded_weight.numel() == 1
            loaded_weight = loaded_weight.reshape(1)

1478
        param.load_row_parallel_weight(loaded_weight=loaded_weight, is_quantization=self.is_quantization)
1479

1480
    def forward(
1481
1482
        self,
        input_,
1483
    ) -> torch.Tensor | tuple[torch.Tensor, Parameter | None]:
1484
1485
1486
1487
        if self.input_is_parallel:
            input_parallel = input_
        else:
            splitted_input = split_tensor_along_last_dim(
1488
1489
                input_, num_partitions=self.tp_size
            )
1490
            input_parallel = splitted_input[self.tp_rank].contiguous()
1491
1492

        # Matrix multiply.
1493
        assert self.quant_method is not None
1494
1495
1496
        # Only fuse bias add into GEMM for rank 0 (this ensures that
        # bias will not get added more than once in TP>1 case)
        bias_ = None if (self.tp_rank > 0 or self.skip_bias_add) else self.bias
1497
        output_parallel = self.quant_method.apply(self, input_parallel, bias_)
1498

1499
        if self.reduce_results and self.tp_size > 1:
zhuwenwen's avatar
zhuwenwen committed
1500
            output = tensor_model_parallel_all_reduce(output_parallel)
1501
        else:
1502
1503
1504
            output = output_parallel

        output_bias = self.bias if self.skip_bias_add else None
1505

1506
1507
        if not self.return_bias:
            return output
1508
        return output, output_bias
1509
1510

    def extra_repr(self) -> str:
1511
        s = f"in_features={self.input_size_per_partition}"
1512
1513
1514
1515
1516
        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